A method and system for full life-cycle traceability of the catalyst industry chain

Abstract

This invention relates to the field of lifecycle management technology, specifically a method and system for full lifecycle traceability of the catalyst industry chain. The method involves acquiring state attribute data of the same batch of recycled products and parsing it to obtain an efficiency feature vector; inputting the efficiency feature vector into a pre-set quality evaluation model for matching to identify quality deviation data of the recycled products; parsing the batch identification code to obtain production lifecycle history data; constructing a multi-dimensional data association model, using quality deviation data as the target variable and history data as input features, analyzing the weight coefficients and attribution confidence indices of each production control variable on the quality deviation, and screening out key control elements; performing production regulation and optimization operations based on the key control elements, calculating the target parameter offset used to correct the quality deviation, and dynamically calculating the dynamic tolerance constraint range of the production control variables in conjunction with the attribution confidence index, and encapsulating it into an execution control protocol.

Description

Technical Field

[0001] This invention relates to the field of lifecycle management technology, specifically to a method and system for full lifecycle traceability of the catalyst industry chain. Background Technology

[0002] Catalysts are core materials in the petrochemical, automotive exhaust treatment, and new energy industries, and their industrial chain encompasses multiple key links, including raw material preparation, precision manufacturing, industrial application, and recycling. With the advancement of industrial digital transformation, the management model of the catalyst industrial chain is gradually transitioning from traditional physical management to digital full lifecycle management. In existing technological systems, Enterprise Resource Planning (ERP) systems and Manufacturing Execution Systems (MES) are widely used for material control and parameter recording in the production process, while logistics tracking systems are mainly used to monitor the in-transit status and delivery progress of products. However, existing catalyst industrial chain management generally suffers from technical pain points such as unidirectional data flow and information silos across different stages.

[0003] Current traceability technologies primarily focus on forward tracking, recording the flow of products from the factory to the user, mainly addressing anti-counterfeiting and inventory management issues. However, once a catalyst is put into use and eventually becomes obsolete and enters the recycling stage, its entire lifecycle data chain often breaks down. Used catalysts, as a reference for recording product performance under actual operating conditions, contain a wealth of feedback data regarding product quality defects, tolerance bottlenecks, and potential hidden dangers in the production process. However, in existing management processes, the testing data at the recycling end is typically only used to assess the recycling value of residual precious metals, failing to be digitally analyzed into failure characteristic vectors, and even less able to establish a connection across time and space with production data from a year ago or even earlier.

[0004] This lack of a data loop leads to delayed decision-making and inefficient resource allocation at the production management level. When unexpected failures occur in the same batch of products at the user end, the manufacturing end often lacks attribution analysis methods based on quantitative data, making it difficult to accurately pinpoint which fluctuation in production control variables caused the quality deviation.

[0005] To this end, a method and system for tracing the entire life cycle of the catalyst industry chain is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for full life cycle traceability of the catalyst industry chain. It optimizes production regulation through key control elements, calculates the offset of target parameters used to correct quality deviations, and dynamically calculates the dynamic tolerance constraint range of production control variables in combination with the attribution confidence index. The results are then encapsulated into an execution control protocol to achieve full life cycle traceability of the catalyst.

[0007] To achieve the above objectives, this invention provides a method for full life-cycle traceability of the catalyst industry chain, comprising:

[0008] Obtain status attribute data of the same batch of recycled products, parse and process the status attribute data, and construct a standardized performance feature vector; input the performance feature vector into a preset quality evaluation model for matching, and identify the quality deviation data of the recycled products.

[0009] The batch identification code of the recycled products is parsed to obtain the production life cycle history data, which includes raw material batch information and production control variables in the production process;

[0010] A multidimensional data association model is constructed, with the quality deviation data as the target variable and the history data as the input feature. The influence weight coefficients and attribution confidence indices of each production control variable on the quality deviation are analyzed. Variables with influence weight coefficients higher than a preset threshold are selected and defined as key control elements.

[0011] Based on the key control elements, production regulation and optimization operations are performed, the target parameter offset used to correct quality deviations is calculated, and the dynamic tolerance constraint range of production control variables is dynamically calculated in combination with the attribution confidence index; the target parameter offset and the dynamic tolerance constraint range are encapsulated into an execution control protocol.

[0012] The state attribute data covers performance evaluation indicators at three levels: microstructure, surface deposition, and macroscopic mechanics.

[0013] The first evaluation index is the microstructure characterization index, including the specific surface area decay index, the pore volume distribution skewness index, and the average pore size collapse index.

[0014] The second evaluation index is the surface chemical deposition index, which includes the total carbon deposition index, the carbon graphitization degree index, and the deposition density index of exogenous poison elements on the catalyst surface.

[0015] The third evaluation indicator is the macroscopic mechanical performance index, which includes the average index of particle lateral compressive strength, wear index, and the rate of change index of bulk density.

[0016] The state attribute data is parsed and processed to construct a standardized performance feature vector, including:

[0017] Perform consistency checks on various status attribute data in the same batch of recycled products, and use preset outlier detection logic to identify abnormal data points that exceed the compliance range;

[0018] For the identified abnormal data points, data validity audits are conducted in conjunction with sampling location labels. If the data is determined to be due to sampling error, it is removed. If the data is determined to be due to local operating condition abnormalities, it is retained and marked as an abnormal subsample, resulting in high-fidelity cleaned data. A dimensionless mapping protocol is used to map the cleaned data to a preset standardized evaluation range, generating an efficiency feature vector of the recycled products.

[0019] The quality deviation data includes anomaly type labels and deviation feature data; the anomaly type labels include high-temperature sintering type failure, chemical poisoning type failure, pore blockage type failure and mechanical breakage type failure; the deviation feature data is the failure severity index, which is used to quantitatively describe the severity of the failure type.

[0020] The identification logic of the quality evaluation model is designed as follows: a business feature reference library containing various standard failure modes is pre-constructed, and the library stores standard benchmark models corresponding to different failure types; during the matching process, similarity evaluation rules are applied to calculate the matching correlation between the current input performance feature vector and the standard benchmark models in the business feature reference library; the abnormal type labels and deviation feature data are output to obtain quality deviation data.

[0021] The multidimensional data association model includes:

[0022] Perform inverse mapping and alignment of spatiotemporal data; calculate the time lag from raw material input to finished product output based on the logistics flow rules of the production line; accurately map the batch identification code of the recycled products back to the operation data slices flowing through each production process; and establish an input feature matrix.

[0023] Perform weight evaluation based on multidimensional attribution analysis; import the input feature matrix and target variables into the multidimensional data association analysis model for learning; and quantify the influence weight coefficients by evaluating the information contribution value of each production control variable in the model node splitting process.

[0024] Perform confidence assessment based on repeated validation; use random sampling strategy to reconstruct the dataset multiple times and repeat the analysis, statistically analyze the frequency stability of production control variables identified as key elements in repeated evaluation, and define the frequency stability value as the attribution confidence index.

[0025] The production control and optimization operation adopts an optimization design based on a reverse compensation strategy, including:

[0026] Based on historical batch data, a correlation response model between key control elements and quality deviation data is constructed to characterize the sensitivity relationship between changes in production control variables and final product quality deviation.

[0027] In the correlation response model, a management objective function is defined, which aims to minimize the magnitude of the quality deviation data. A multi-objective optimization strategy is used to search for the optimal control strategy within a feasible process window, and the numerical change that should occur in the key control elements to offset the currently identified quality deviation is calculated. The numerical change is defined as the target parameter offset.

[0028] Calculating the dynamic tolerance constraint range of production control variables includes: identifying the attribution confidence index and distinguishing between high confidence intervals, medium confidence intervals, and low confidence intervals; and dynamically adjusting the data range of the dynamic tolerance constraint range based on the confidence interval in which the attribution confidence index is located.

[0029] A lifecycle traceability system for the catalyst industry chain includes:

[0030] The data acquisition module acquires the status attribute data of the same batch of recycled products, parses and processes the status attribute data, and constructs a standardized performance feature vector; the performance feature vector is then input into a preset quality evaluation model for matching, and the quality deviation data of the recycled products is identified.

[0031] The data parsing module parses the batch identification code of the recycled products to obtain the production life cycle history data, which includes raw material batch information and production control variables during the production process.

[0032] The element identification module constructs a multi-dimensional data association model, using the quality deviation data as the target variable and the history data as the input feature, to analyze the influence weight coefficients and attribution confidence indices of each production control variable on the quality deviation; variables with influence weight coefficients higher than a preset threshold are selected and defined as key control elements.

[0033] The production control module performs production regulation and optimization operations based on the key control elements, calculates the target parameter offset used to correct quality deviations, and dynamically calculates the dynamic tolerance constraint range of production control variables in combination with the attribution confidence index; the target parameter offset and the dynamic tolerance constraint range are encapsulated into an execution control protocol.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention constructs a standardized performance feature vector to quantitatively characterize the failure state of recycled products at the microstructure, surface chemistry, and macroscopic mechanical levels. Utilizing reverse mapping alignment technology, it can calculate lag based on logistics flow rules, accurately tracing the batch code of recycled products back to a production operation segment from a year ago or even earlier. This enables the manufacturing end to reverse-optimize processes based on the product's full lifecycle performance under real-world operating conditions. This not only eliminates information silos but also makes production decisions based on objective quantitative data, achieving closed-loop iteration of product quality management.

[0036] 2. This invention accurately identifies and eliminates logistics-damaged outliers caused by transportation bumps or dampness by monitoring the cumulative vibration energy spectral density and the integral value of environmental humidity during the logistics process, ensuring that the data input for analysis purely reflects production quality. In the attribution analysis stage, a chemical mechanism knowledge graph is used to verify the path of key elements screened by the statistical model. This dual verification of "physical + data" ensures the scientificity and uniqueness of the traceability results, avoiding the misattribution of logistics problems to the production process or the misleading process adjustments due to statistical coincidences.

[0037] 3. This invention proposes a flexible control strategy based on confidence level. It not only calculates the offset of the target parameter used to correct deviations, but also introduces the core dimension of attribution confidence index. Based on the level of this index, the control behavior is divided into different levels. The mechanism of transforming statistical probability into management strength generates a machine-readable execution control protocol, which not only gives the production system the ability to self-correct, but also retains the necessary fault tolerance flexibility through dynamic tolerance mechanism, thus achieving the best balance between quality improvement and safe production operation. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a method for tracing the entire lifecycle of a catalyst industry chain according to the present invention.

[0039] Figure 2 This is a schematic diagram illustrating the process of executing the control protocol of the present invention;

[0040] Figure 3 This is a schematic diagram of the structure of a full life-cycle traceability system for the catalyst industry chain according to the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1:

[0043] This invention proposes a method for full life-cycle traceability of the catalyst industry chain, the process of which is as follows: Figure 1 As shown, the process of obtaining the execution control protocol is as follows: Figure 2 As shown, it includes:

[0044] Obtain status attribute data of the same batch of recycled products, parse and process the status attribute data, and construct a standardized performance feature vector; input the performance feature vector into a preset quality evaluation model for matching, and identify the quality deviation data of the recycled products.

[0045] The batch identification code of the recycled products is parsed to obtain the production life cycle history data, which includes raw material batch information and production control variables in the production process;

[0046] A multidimensional data association model is constructed, with the quality deviation data as the target variable and the history data as the input feature. The influence weight coefficients and attribution confidence indices of each production control variable on the quality deviation are analyzed. Variables with influence weight coefficients higher than a preset threshold are selected and defined as key control elements.

[0047] Based on the key control elements, production regulation and optimization operations are performed, the target parameter offset used to correct quality deviations is calculated, and the dynamic tolerance constraint range of production control variables is dynamically calculated in combination with the attribution confidence index; the target parameter offset and the dynamic tolerance constraint range are encapsulated into an execution control protocol.

[0048] The state attribute data covers performance evaluation indicators at three levels: microstructure, surface deposition, and macroscopic mechanics.

[0049] The first evaluation index is the microstructure characterization index, including the specific surface area decay index, the pore volume distribution skewness index, and the average pore size collapse index.

[0050] The second evaluation index is the surface chemical deposition index, which includes the total carbon deposition index, the carbon graphitization degree index, and the deposition density index of exogenous poison elements on the catalyst surface.

[0051] The third evaluation indicator is the macroscopic mechanical performance index, which includes the average index of particle lateral compressive strength, wear index, and the rate of change index of bulk density.

[0052] The microstructure characterization index was obtained by nitrogen adsorption-desorption experiments on the recovered particles;

[0053] Specific surface area decay index (SSA): This index quantifies the thermal stability of the catalyst support framework and is calculated by comparing the specific surface area of ​​the recovered catalyst with that of the fresh catalyst. First, the recovered catalyst particle samples are degassed under vacuum, and their specific surface area is measured using the gas adsorption principle. Simultaneously, the specific surface area of ​​the fresh catalyst at the time of manufacture is retrieved from the historical database. The difference between the fresh catalyst's specific surface area and the recovered catalyst's specific surface area is calculated. This difference is then divided by the fresh catalyst's specific surface area; the quotient is the SSA. A higher SSA value indicates a more severe degree of thermal collapse or sintering of the catalyst support framework.

[0054] Pore ​​volume distribution skewness index: The pore size inside a catalyst typically follows a normal distribution; this index measures whether the pore distribution has undergone non-uniform distortion. Based on the nitrogen adsorption-desorption isotherm, a pore size distribution curve is derived using a pore size distribution calculation model. This curve shows the volume proportion of pores of different sizes. The statistical skewness value of the pore size distribution curve is calculated. Specifically, it is the cube moment of the difference between the volume distribution probability density at each pore size point and the average pore size, and then standardized to obtain the skewness value. If the index deviates significantly from zero, it indicates that the pore structure inside the catalyst has undergone non-uniform collapse or blockage of specific pore sizes; if the skewness is too large, it indicates that the channels of specific pore sizes have experienced concentrated collapse.

[0055] Average pore size collapse index: used to characterize the patency of gas transport channels; the average pore size values ​​of fresh catalyst and recycled catalyst are obtained separately. The difference between the average pore size value of fresh catalyst and the average pore size value of recycled catalyst is calculated, and this difference is divided by the average pore size value of fresh catalyst. This index is used to quantify the degree of shrinkage of gas transport channels.

[0056] The surface chemical deposition index is mainly calculated using a high-frequency infrared carbon-sulfur analyzer, a laser Raman spectrometer, and an X-ray fluorescence spectrometer to assess the coverage and poisoning of active sites.

[0057] Total carbon deposition index: Characterizes the amount of carbon deposits covering the active sites; the carbon content in the sample is determined by detecting the amount of carbon dioxide gas released through high-temperature combustion of a catalyst sample in an oxygen-rich environment using a high-frequency infrared carbon-sulfur analyzer. The carbon element mass percentage value output by the instrument is read; the higher the value, the more coke deposits covering the active sites, and the smaller the effective reaction area of ​​the catalyst.

[0058] The graphitization degree index of carbon deposits characterizes the hardness of the carbon deposits. A higher index indicates that the carbon deposits are closer to a graphite structure, making regeneration more difficult. The surface carbon deposits on the catalyst are scanned using a laser Raman spectroscopy system to obtain Raman spectra. Defect peaks representing disordered carbon structures and graphite peaks representing graphitic carbon structures are identified in the spectra. The ratio of the peak intensity of the graphite peaks to the peak intensity of the defect peaks is calculated. A higher ratio indicates greater hardness and denser structure of the deposited carbon, meaning greater difficulty in removing the carbon deposits through conventional regeneration methods and stronger irreversibility of deactivation.

[0059] Exogenous poison deposition density index: This index focuses on detecting sulfur, phosphorus, arsenic, and heavy metals, quantifying their aggregation density per unit surface area to determine the presence of chemical poisoning. X-ray fluorescence spectrometry is used to scan the surface of catalyst particles, focusing on elements not inherent to the catalyst itself, such as sulfur, phosphorus, arsenic, and heavy metals. The mass percentage values ​​of each detected poison element are summed, or the ratio of the content of a specific poison element to the specific surface area of ​​the catalyst is calculated. This index characterizes the density of poison adsorbed per unit surface area, quantifying the severity of chemical poisoning.

[0060] Calculation of the average index of particle lateral compressive strength: A certain number (e.g., 50) of intact catalyst particles are randomly selected, and radial pressure is applied to each particle individually using a particle strength tester until the particles break. The critical pressure value at which each particle breaks is recorded. The calculation logic is as follows: the arithmetic mean of all recorded critical pressure values ​​is calculated to obtain the average lateral compressive strength. To construct the index, the percentage decrease of this average value relative to the strength of fresh catalyst can be further calculated.

[0061] Calculation of the wear index: A certain weight of catalyst sample is weighed and placed in a standard rotary drum wear tester, where it undergoes rotary wear at a specified speed and time. After the test, the sample is sieved using a standard aperture sieve, and the weight of the fine powder that passes through the sieve is collected and weighed. The calculation logic is as follows: the weight of the fine powder passing through the sieve is divided by the total weight of the sample before the test. This value directly reflects the particle's resistance to pulverization under fluid erosion.

[0062] Calculation of the rate of change of bulk density index: Using a standard volume graduated cylinder, the catalyst sample is freely stacked to the graduation mark. The weight of the sample in the graduated cylinder is measured, and the weight per unit volume is calculated to obtain the bulk density. The calculation logic is as follows: calculate the bulk density of the recovered catalyst and subtract the bulk density of the fresh catalyst, then divide by the bulk density of the fresh catalyst. If this index is positive and large, it usually means that the particles have broken down, and the fine particles fill the gaps between the particles, resulting in a more compact overall packing and increased fluid resistance.

[0063] This invention overcomes the limitations of traditional methods that focus only on a single physical indicator by establishing a three-dimensional indicator system. In particular, the introduction of deeper indicators such as distribution skewness and graphitization degree for solid particle groups can more accurately capture potential quality hazards caused by minor fluctuations in the production process, providing a high-quality data foundation for subsequent accurate attribution.

[0064] The state attribute data is parsed and processed to construct a standardized performance feature vector, including:

[0065] Perform consistency checks on various status attribute data in the same batch of recycled products, and use preset outlier detection logic to identify abnormal data points that exceed the compliance range;

[0066] For the identified abnormal data points, data validity audits are conducted in conjunction with sampling location labels. If the data is determined to be due to sampling error, it is removed. If the data is determined to be due to local operating condition abnormalities, it is retained and marked as an abnormal subsample, resulting in high-fidelity cleaned data. A dimensionless mapping protocol is used to map the cleaned data to a preset standardized evaluation range, generating an efficiency feature vector of the recycled products.

[0067] Consistency Verification Process: Given the frequent presence of noise in industrial big data, a consistency verification is performed first. For multiple sample data within the same batch, their dispersion is calculated. Using pre-defined outlier detection logic, such as statistical rules based on quartiles, outlier data points whose values ​​significantly deviate from the normal range are automatically identified.

[0068] Data validity audit process: Instead of blindly rejecting outliers, a secondary audit is conducted based on the sampling location labels. Specifically, this includes: Since solid particles may segregate within the transport container, fine particles tend to settle at the bottom; if the outlier is a sample from the bottom of the container and shows a high abrasion index, it is considered a normal physical settling phenomenon and is rejected or corrected; if the outlier is from the middle or upper part of the container, it is considered a local operating condition anomaly, is retained and marked as a special subsample to prevent the loss of critical fault information.

[0069] The specific design of the outlier detection logic is illustrated below with an example of the specific surface area decay index:

[0070] Step 1: Calculate the quartile index; Based on the surface area decay index data of all samples in the same batch, sort them in ascending order, calculate the first quartile (Q1, i.e., the 25th percentile) and the third quartile (Q3, i.e., the 75th percentile), and then calculate the interquartile range IQR = Q3 - Q1.

[0071] Step 2: Determine the outlier criteria boundary; let the lower boundary for outlier criteria be Q1-k×IQR, and the upper boundary be Q3+k×IQR, where the coefficient k is 1.5. This coefficient is dynamically adjusted based on industry experience and actual surface area decay data to prevent over-removal. If the surface area decay index of a sample falls outside this boundary range, it is determined to be an outlier.

[0072] Step 3: Multi-dimensional Outlier Assessment; For cases where multiple indicators simultaneously exhibit outliers, a logical AND rule is used. A sample is classified as an outlier only if it exceeds the boundary for two or more key indicators simultaneously. Single-indicator outliers typically reflect differences in operating conditions at different sampling locations and should not be directly excluded.

[0073] Step 4: For other indicators, such as pore size, carbon deposits, and toxins, the same quartile-IQR method is applied, but the coefficient k can be adjusted according to the physical meaning of the indicator. For example, for outliers in the wear index caused by transportation vibration, k=2.0 can be set to increase the tolerance; for outliers in the severity of failure caused by abnormal production processes, k=1.2 should be set to increase the sensitivity.

[0074] Dimensionless mapping process: Using a dimensionless mapping protocol, such as the range standardization method, the cleaned data of each dimension are uniformly mapped to a standardized evaluation range of 0 to 1, eliminating differences in physical units and generating a standardized performance feature vector.

[0075] Furthermore, the outlier determination logic also includes a preprocessing elimination mechanism based on state-logistics dual-dimensional verification, aimed at eliminating physical damage interference caused by non-production factors; simultaneously acquiring the macroscopic mechanical performance index of the batch of recycled products in the state attribute data, as well as the logistics vibration monitoring data during transportation; preset vibration intensity threshold and wear anomaly threshold; when the peak acceleration in the logistics vibration monitoring data exceeds the vibration intensity threshold, and the wear index in the corresponding macroscopic mechanical performance index simultaneously exceeds the wear anomaly threshold, a correlation determination is performed; if a strong positive correlation is determined between the two, the sample is defined as a logistics-induced damage outlier and removed from the input dataset of the attribution analysis, retaining only samples without significant transportation vibration interference for subsequent analysis.

[0076] Read the macroscopic mechanical performance index detected at the recycling end, with a focus on the wear index; read the triaxial acceleration sensor data of the transport vehicle through the logistics interface to obtain the maximum vibration acceleration throughout the transportation process.

[0077] The preset vibration intensity threshold is 2.5G (an empirical value; particles are prone to breakage if this value is exceeded); the preset abnormal wear threshold is 1.5% (normal wear is usually less than 1%).

[0078] If the wear index of a certain batch is 0.8% (normal), or if the wear index is high but the maximum vibration acceleration value during transportation is only 0.5G (smooth road conditions), it is determined that the wear is due to a production quality problem and it is retained.

[0079] One batch had a wear index as high as 2.0%, and logistics data from the same period showed that it had experienced severe 3.0G jolting, such as sudden braking or falling. It was determined that this high wear was directly caused by transportation vibration and was a logistics-induced outlier.

[0080] This invention effectively eliminates noise interference in the logistics process by incorporating transportation vibration data, ensuring that the wear data fed back to the production end represents genuine internal quality defects. This guarantees the accuracy of subsequent adjustments to production control variables and avoids misoperation. In particular, it achieves calibration and removal of the unique friction and wear phenomena of solid particles, significantly improving the accuracy of data cleaning and attribution analysis.

[0081] The quality deviation data includes anomaly type labels and deviation feature data; the anomaly type labels include high-temperature sintering type failure, chemical poisoning type failure, pore blockage type failure and mechanical breakage type failure; the deviation feature data is the failure severity index, which is used to quantitatively describe the severity of the failure type.

[0082] The identification logic of the quality evaluation model is designed as follows: a business feature reference library containing various standard failure modes is pre-constructed, and the library stores standard benchmark models corresponding to different failure types; during the matching process, similarity evaluation rules are applied to calculate the matching correlation between the current input performance feature vector and the standard benchmark models in the business feature reference library; the abnormal type labels and deviation feature data are output to obtain quality deviation data.

[0083] The quality assessment model in this embodiment adopts a center point matching architecture based on feature space, which is based on highly interpretable geometric space metric logic; the model contains two core components:

[0084] Business Feature Reference Library: This is a high-dimensional database that stores standard benchmark models. Each benchmark model represents a typical failure state recognized in the industry. Specifically, it includes four types of failures: high-temperature sintering failure, chemical poisoning failure, pore blockage failure, and mechanical breakage failure.

[0085] Similarity calculation engine: A vector operation unit based on Euclidean distance or cosine similarity, used to calculate the spatial positional relationship between the input vector and the benchmark model.

[0086] The quality assessment model is constructed based on an iterative process of historical data clustering. The standard benchmark model is constructed using a feature space clustering method. Each benchmark model is represented as a D-dimensional vector, where D equals the dimension of the standardized performance feature vector. In this embodiment, D=9, corresponding to three microstructure indicators, three surface chemical indicators, and three mechanical performance indicators. The basic structure of the benchmark model consists of the spatial distribution range of the benchmark feature vector center points and its multidimensional features. The model uses the K-means clustering algorithm for iterative optimization. The initial number of clusters K is set to 4, corresponding to the four standard failure types. Iteration stops when the ratio of the sum of squared intra-cluster distances to the inter-cluster distances converges. The clustering convergence condition is defined as: the sum of the Euclidean distances of the cluster centers in two consecutive iterations is less than 1×10⁻⁶. -6 Or, it can reach the maximum number of iterations, which is set to 500.

[0087] Step 1 (Sample Accumulation): Collect the state attribute data of past accumulated batches of waste catalysts and their corresponding expert judgment results, such as the failure type determined manually.

[0088] Step 2 (Feature Space Mapping): Transform these historical data into standardized performance feature vectors and map them into a multi-dimensional feature space.

[0089] Step 3 (Cluster Center Extraction): Using the logic of the K-means clustering algorithm, dense sample points in space are grouped into clusters. For example, all samples manually determined to have failed at high temperature sintering are clustered into one class, and the arithmetic mean of all feature vectors of this class is calculated. This average vector is defined as the center point of the standard feature vector for high temperature sintering failure.

[0090] Step 4 (Baseline Model Solidification): Solidify and store the calculated center points in the business feature reference library as a benchmark for subsequent matching.

[0091] Specifically, when a new batch performance feature vector is received, the following identification logic is executed:

[0092] Distance Calculation: Calculate the Euclidean distance between the input vector and the center point of each standard feature vector in the reference library. The specific calculation method is as follows: calculate the difference between the value of each dimension of the input vector and the value of the corresponding dimension of the center point vector, find the sum of the squares of the differences, and finally take the square root of the sum.

[0093] Correlation calculation: For each failure dimension, apply similarity evaluation rules, such as Euclidean distance calculation, to calculate the spatial distance between the input vector and the benchmark model.

[0094] Severity inverse mapping: The calculated spatial distance is converted into a failure severity index. Inverse mapping logic or a normalization function is used; the smaller the distance value, the more typical the feature in that dimension, and the higher the converted severity index; the greater the distance, the lower the severity index.

[0095] The quality deviation data is constructed as a multi-dimensional feature vector, with each dimension representing the severity of a specific failure mode; if the catalyst quality is perfect and there is no deviation, then this vector should theoretically be a zero vector.

[0096] Furthermore, the history data also covers environmental logistics factors of the product during the warehousing and transportation stages; the environmental logistics factors include the cumulative vibration energy spectral density, the integral value of ambient humidity, and the number of temperature change cycles during transportation; when constructing a multidimensional data association model, the environmental logistics factors and production control variables are used together as input features to identify non-production quality deviations caused by improper logistics links.

[0097] Cumulative vibration energy spectral density: used to quantify the total amount of cumulative fatigue damage to catalyst particles caused by vibration energy within a specific frequency range, i.e., the mechanical resonance sensitive frequency band of catalyst particles, during transportation; this parameter differs from ordinary average acceleration and focuses on assessing the cumulative energy that triggers resonant breakage.

[0098] Calculation method: First, triaxial acceleration data during transportation is collected at a set high sampling rate; second, a fast Fourier transform is performed on the acceleration data to convert it into a frequency domain signal to obtain the power spectral density; then, only the power spectral density value in the catalyst's vulnerable frequency band is extracted and integrated in the frequency domain to obtain the instantaneous vibration intensity (the square of the root mean square acceleration); finally, the instantaneous vibration intensity is accumulated over the entire transportation time axis to calculate the cumulative vibration fatigue damage index for the entire process.

[0099] Ambient humidity integral value: used to characterize the cumulative destructive effect of water molecules entering the pores through capillary coagulation caused by exposure to high humidity during the storage and transportation of catalysts; this parameter not only considers the peak humidity, but also focuses on the cumulative effect of the duration of high humidity and the combined effect of temperature.

[0100] Calculation method: First, a critical moisture absorption threshold is set according to the physical properties of the catalyst material, such as a relative humidity of 65%; second, real-time relative humidity and temperature data in the environment are continuously collected; then, the time period in which the relative humidity value exceeds the critical moisture absorption threshold is selected; finally, the humidity difference exceeding the threshold is integrally calculated with respect to time. In the calculation process, the diffusion coefficient based on real-time temperature can be introduced as a correction factor to obtain the final accumulated integral value.

[0101] Temperature cycle count: used to characterize the frequency of drastic fluctuations in ambient temperature (i.e., thermal shock), reflecting the number of microscopic thermal stress shocks caused by the difference in thermal expansion coefficients between the catalyst support and the active component.

[0102] Calculation method: First, collect real-time temperature data throughout the transportation cycle and calculate the derivative of temperature with respect to time, i.e., the rate of temperature change; second, set a thermal shock threshold (e.g., a temperature change exceeding ±5 degrees Celsius per minute); finally, apply the rainflow counting method to count the number of complete hysteresis loops whose rate of temperature change exceeds the above thermal shock threshold, and this number is the number of temperature change cycles.

[0103] By reading data from IoT sensor tags attached to catalyst packaging drums, accelerometer and thermo-hygrometer records were obtained during transportation. Attribution analysis revealed that the mechanical breakage failures of a certain batch were primarily unrelated to production parameters, but rather showed a high correlation (0.9) with the cumulative vibration energy spectral density during transportation. This was because the route involved bumpy roads and insufficient packaging cushioning. Therefore, the resulting control protocol was not to adjust production pressure, but rather to issue a logistics instruction to upgrade packaging cushioning materials.

[0104] By introducing refined parameters such as cumulative vibration energy spectral density, ambient humidity integral value, and temperature change cycle number, it is possible to distinguish between channel collapse caused by the production process and channel collapse caused by moisture during transportation. Currently, physical breakage at the recycling end is often mistakenly attributed to insufficient binder strength at the production end, leading to incorrect process adjustments, such as excessive addition of binder resulting in decreased activity. This invention can effectively eliminate such non-production noise and protect the stability of the production process. It also improves the definition of the entire life cycle, preventing quality problems caused by logistics and transportation from being mistakenly attributed to the manufacturing process, avoiding incorrect adjustments to the production process, and improving the accuracy of traceability responsibility determination.

[0105] The multidimensional data association model includes: performing inverse mapping and alignment of spatiotemporal data; calculating the time lag from raw material input to finished product output based on the logistics flow rules of the production line; accurately mapping the batch identification code of the recycled products back to the operation data slices flowing through each production process; and establishing an input feature matrix.

[0106] Perform weight evaluation based on multidimensional attribution analysis; import the input feature matrix and target variables into the multidimensional data association analysis model for learning; and quantify the influence weight coefficients by evaluating the information contribution value of each production control variable in the model node splitting process.

[0107] Perform confidence assessment based on repeated validation; use random sampling strategy to reconstruct the dataset multiple times and repeat the analysis, statistically analyze the frequency stability of production control variables identified as key elements in repeated evaluation, and define the frequency stability value as the attribution confidence index.

[0108] Inverse mapping and alignment of spatiotemporal data is a key step in data preparation before model training, aiming to eliminate the time misalignment between production and recycling;

[0109] Lag time calculation: For each production process, such as mixing, extrusion, drying, and calcination, the average residence time of the material in that process is calculated based on the time distribution of the historical production process, i.e., the time lag.

[0110] Slice mapping: Parse the batch identification code of the recycled product to obtain its production time point; then, subtract the dwell time of each subsequent process in turn to deduce the specific timestamp of the batch of material at each moment such as raw material mixing, extrusion molding, drying, and roasting.

[0111] Feature matrix establishment: Extract the process parameters corresponding to the specific timestamps mentioned above, such as temperature, pressure, and flow rate, to establish a precise microscopic spatiotemporal correspondence and form the input feature matrix;

[0112] The model employs an ensemble learning-based gradient boosting decision tree (GBDT) architecture, which consists of decision trees and can handle non-linear relationships and automatically evaluate the importance of features.

[0113] Training process: The input feature matrix (production control variables) is used as input, and the quality deviation data (such as the severity values ​​of each failure mode) is used as the target variable and imported into the model for learning.

[0114] Weight Calculation: During model training, whenever a production control variable is selected as a split node in the decision tree, the reduction in squared error or the improvement in purity resulting from that split is calculated as its information contribution value. The improvements brought by this variable in all decision trees are summed and divided by the sum of the improvements of all variables to quantify the influence weight coefficient of that variable. This coefficient reflects the degree of contribution of the production control variable to the final quality deviation.

[0115] Confidence assessment based on repeated validation: To ensure the reliability of the attribution results, a confidence assessment based on repeated validation is performed, using a Bootstrap resampling strategy.

[0116] Resampling settings: Set the number of resampling times to 100. Each time, randomly draw 80% of the samples with replacement from the original dataset to construct a new training subset.

[0117] Repeated training: Train 100 independent association models on each of these 100 subsets.

[0118] Frequency statistics: For variables identified as key control elements, count how many times their ranking exceeds a preset threshold in these 100 training sessions.

[0119] Index definition: The frequency stability of a statistical production control variable as a key element in repeated evaluations. For example, if a variable ranks in the top 3 in 95 training iterations, its frequency stability is high, and it is defined as the attribution confidence index, such as 95%. This indicates that the causal relationship is extremely robust and not random noise.

[0120] This invention solves the common problems of spatiotemporal misalignment and spurious correlation in industrial big data analysis by using physical-level temporal alignment and statistical-level repeated verification. It ensures that the key control elements found are not only mathematically correlated, but also physically real causal, providing a solid scientific basis for subsequent process adjustments.

[0121] Furthermore, for products with complex production processes, since most catalyst production involves processes such as mixing, kneading, extrusion, or precipitation, these processes have clearly defined residence time distributions in chemical engineering principles. The material in the reactor or mixer is not in an ideal plug flow, but rather experiences backmixing. This means that the finished particles produced at a particular moment are actually the result of the mixing of raw materials and process parameters input over a past period (perhaps several minutes or even hours). Using only data slices may not accurately reflect the production process.

[0122] Therefore, the construction of the input feature matrix can also be designed as follows: based on the logistics flow rules of the production line and the average dwell time distribution of each process, calculate the time lag interval from raw material input to finished product output, map the batch identification code of the recycled product back to the time period weighted operation data when it flows through each production process, and establish the input feature matrix; that is, regard the relatively stable operation process as a data point; for example, cluster the production control variables of the mixed process to identify several relatively stable time periods as data points.

[0123] Furthermore, the multidimensional data association model can also integrate a pre-built chemical mechanism knowledge graph; the chemical mechanism knowledge graph includes entity nodes representing production control variables, result nodes representing quality deviation data, and relational edges representing chemical reaction principles; when calculating the influence weight coefficients, a mechanism path verification step is performed: using a graph traversal algorithm, a connected path from the high-weight production control variable node to the quality deviation data node is retrieved in the chemical mechanism knowledge graph;

[0124] Execute a weight adjustment strategy based on the search results of the connectivity path:

[0125] If no connected path exists, the statistical correlation between the production control variable and the quality deviation data is determined to be a spurious correlation, and a damping penalty factor is introduced to reduce its influence weight coefficient.

[0126] If a connected path exists, it is determined to be strongly correlated with the mechanism, and a gain enhancement factor is introduced to improve its attribution confidence index.

[0127] Among them, the chemical mechanism knowledge graph is defined as a structured graph database (such as one built on Neo4j) used to store expert knowledge and chemical reaction principles in the field of catalyst production.

[0128] Node design includes parameter entities, intermediate entities, and result entities;

[0129] Parameter entities: including specific process parameters, such as calcination temperature, nitric acid solution concentration, kneader speed, etc.; intermediate entities: including microscopic physicochemical changes, such as grain growth, pore collapse, active component migration, and carbon precursor formation; result entities: i.e. quality deviation data, such as specific surface area decay, poisoning failure, and insufficient lateral pressure strength.

[0130] The design of the relationship edges includes: defining causal or promoting relationships between nodes, such as: excessively high calcination temperature leads to grain growth, and grain growth leads to a decrease in specific surface area; each edge can be attached with attributes, such as threshold conditions for the reaction to occur, for example, this path is activated when the temperature is >500℃; when the multidimensional data association model identifies that a certain variable X has a high statistical correlation with the quality deviation Y, the graph verification is initiated, with X as the starting point and Y as the ending point, and a depth-first search algorithm is applied to search for one or more connected paths in the knowledge graph; if a path is found, it is determined that there is a supporting mechanism.

[0131] False correlation elimination, also known as damping penalty mechanism: When a strong statistical correlation is detected between a non-process-related characteristic variable (such as a batch produced on Tuesday afternoon) and a product quality deviation (such as a high product wear rate), but in the chemical mechanism knowledge graph, this timestamp node cannot be connected to the wear rate node through any known chemical reaction path or physical action mechanism; the following action is taken: the statistical correlation is determined to be a statistical coincidence without any physical or chemical mechanism support, i.e., a false correlation. A damping penalty factor is introduced to attenuate the original influence weight of the variable. Specifically, the weight value of the variable is reduced by a preset ratio, resulting in a significant reduction in the corrected weight value. After this operation, the priority of the variable is downgraded, and it is no longer included in the subsequent management and control system as a key control element.

[0132] Mechanism confirmation, i.e., gain enhancement mechanism: When a correlation is found between a certain process parameter (e.g., steam partial pressure) and a specific quality defect (e.g., skeleton dealuminization), and a clear connection path is found in the chemical mechanism knowledge graph (e.g., the path shows: increased steam partial pressure leads to accelerated hydrothermal aging, which in turn induces skeleton dealuminization); Execution: Determine that the correlation is a true physicochemical causal relationship, introduce a gain enhancement factor to amplify the original influence weight of the variable, specifically by increasing the weight value of the variable by a preset amplification factor, thereby significantly improving the confidence value of the attribution result and ensuring that the key control element is preferentially encapsulated into the execution control protocol.

[0133] In complex chemical production processes, numerous coincidental statistical events occur, making purely statistical models prone to false alarms. This invention utilizes graph-based reasoning as a filter to automatically identify and eliminate spurious correlation variables lacking physicochemical support, preventing the issuance of incorrect process adjustment instructions and avoiding misoperation of the production line. When the sample data is small or noisy, statistical models are prone to distortion. In such cases, the prior expert knowledge provided by the knowledge graph can serve as a stable reference system, strengthening the weight of the correct path through gain enhancement factors, ensuring accurate tracing judgments consistent with chemical principles even under small sample conditions.

[0134] The production control and optimization operation adopts an optimization design based on a reverse compensation strategy, including:

[0135] Based on historical batch data, a correlation response model between key control elements and quality deviation data is constructed to characterize the sensitivity relationship between changes in production control variables and final product quality deviation.

[0136] In the correlation response model, a management objective function is defined, which aims to minimize the magnitude of the quality deviation data. A multi-objective optimization strategy is used to search for the optimal control strategy within a feasible process window, and the numerical change that the key control elements should undergo to offset the currently identified quality deviation is calculated and defined as the target parameter offset.

[0137] Based on historical batch big data, a second-order response surface regression method is used to fit and construct a correlation response model between key control elements and quality deviation data. This model is a multinomial regression model, which includes not only linear terms (the variables themselves), but also quadratic terms (the squares of the variables) and interaction terms (the products between the variables), representing how small changes in production control variables will lead to changes in the final product quality deviation, i.e., the sensitivity relationship.

[0138] Definition of Management Objective Function: Define a management objective function whose mathematical meaning aims to minimize the magnitude of the quality deviation data. The quality deviation data is a multi-dimensional vector, and the magnitude represents the comprehensive distance of this vector from the perfect zero point in multi-dimensional space, i.e., the comprehensive quality loss value. The goal is to find a set of production parameters that minimizes the output value of this function.

[0139] Multi-objective optimization strategy: Use sequential quadratic programming algorithm or genetic algorithm for optimization.

[0140] Constraints: The optimization process is strictly limited to a feasible process window. For example, the upper limit of the calcination temperature search cannot exceed the equipment's maximum tolerance temperature, and the lower limit cannot be lower than the reaction initiation temperature.

[0141] Iterative search: The algorithm continuously tries to adjust the values ​​of key control elements within the constraints, inputs the prediction deviation magnitude of the response model, until it finds a parameter combination that minimizes the magnitude.

[0142] Offset Calculation: Calculate the difference between the optimal parameter combination and the parameter settings currently being used on the production line. Calculate the amount of numerical change that the key control elements should undergo to offset the currently identified quality deviation, and define it as the target parameter offset.

[0143] This invention achieves a leap from qualitative adjustment to quantitative and precise compensation, no longer relying on manual experience to estimate the adjustment range, but using mathematical models to accurately calculate the optimal solution that minimizes quality loss; making the adjustment of the production process more scientific and precise, and effectively suppressing the recurrence of similar failure modes in the next batch of production.

[0144] Calculating the dynamic tolerance constraint range of production control variables includes: identifying the attribution confidence index and distinguishing between high confidence intervals, medium confidence intervals, and low confidence intervals; and dynamically adjusting the data range of the dynamic tolerance constraint range based on the confidence interval in which the attribution confidence index is located.

[0145] The dynamic tolerance constraint range of production control variables is dynamically adjusted based on the output attribution confidence index.

[0146] High confidence interval (e.g., ≥90%): Extremely high confidence, requiring strong correction. Tolerance calculation: Dynamic tolerance constraints are narrowed to the limits of equipment capability or a small multiple of historical standard tolerances (e.g., 30% of standard tolerance). This means that production equipment must execute the new parameters with extreme precision.

[0147] Medium confidence interval (e.g., 60% to 90%): Highly suspicious of anomalies, but with some uncertainty; moderate flexibility is maintained. Tolerance calculation: Dynamic tolerance constraints maintain historical standard tolerances (e.g., 100% of standard tolerances).

[0148] Low confidence intervals (e.g., less than or equal to 60%): Insufficient evidence; avoid excessive intervention that could lead to production disruptions. Tolerance calculation: Expand the dynamic tolerance constraint range to a larger multiple of the historical standard tolerance (e.g., 1.2 to 1.5 times), or issue only suggested values ​​without locking modification permissions.

[0149] Finally, the calculated target parameter offsets and dynamic tolerance constraint ranges are merged and encoded to generate a machine-readable execution control protocol. This protocol is directly sent to the production system via a data interface, automatically updating the control logic of the production workflow.

[0150] This invention transforms statistical confidence into managerial control, establishing an adaptive risk management mechanism. It can swiftly resolve problems when the cause is clear, and steadily maintain production when the cause is uncertain, achieving the best balance between quality improvement and production safety, and minimizing the risk of misoperation caused by data noise.

[0151] Example 2:

[0152] This invention proposes a full life-cycle traceability system for the catalyst industry chain, the structure of which is as follows: Figure 3 As shown, it includes: a data acquisition module, a data parsing module, an element identification module, and a production control module; it is deployed on a cloud server and interconnected with the laboratory management system, manufacturing execution system, and logistics tracking system through data interfaces.

[0153] The data acquisition module acquires the status attribute data of the same batch of recycled products, parses and processes the status attribute data, and constructs a standardized performance feature vector; the performance feature vector is then input into a preset quality evaluation model for matching, and the quality deviation data of the recycled products is identified.

[0154] This module is responsible for interpreting the current state of waste materials and is the starting point of the entire traceability process.

[0155] Multidimensional index collection: After the recovered catalyst enters the processing center, the data acquisition module first retrieves the multidimensional state attribute data of this batch from the laboratory management system. The collected data covers three levels:

[0156] Microstructure characterization indices include specific surface area decay index, pore volume distribution skewness index, and average pore size collapse index.

[0157] Surface chemical deposition index, including total carbon deposition index, carbon graphitization degree index, and deposition density index of exogenous poison elements;

[0158] Macroscopic mechanical performance indices include the average index of particle lateral compressive strength, the wear index, and the rate of change index of bulk density.

[0159] Data cleaning and feature construction: The module performed a consistency check and found that two samples had abnormally high wear indices. The module automatically audited the sampling location tags and logistics data, finding that these two samples were taken from the bottom of the packaging barrels and no severe vibration was recorded during transportation. Therefore, they were determined to be local operating condition anomalies rather than sampling errors, and were retained and marked as anomalous subsamples. Subsequently, the range standardization method was used to map all cleaned data to the range of 0 to 1, constructing a standardized performance feature vector.

[0160] Quality Deviation Identification: The aforementioned performance feature vectors are input into a pre-set quality evaluation model. This model internally stores a business feature reference library, including standard benchmark models such as high-temperature sintering and chemical poisoning.

[0161] Matching process: By calculating the Euclidean distance between the input vector and each benchmark model, it was found that the features of this batch were closest to the benchmark model of "high temperature sintering failure" and had a very high similarity. Among them, the abnormal type labels include high temperature sintering failure, chemical poisoning failure, pore blockage failure and mechanical breakage failure.

[0162] Output: The module identifies quality deviation data, including the abnormal type label "high temperature sintering failure", with a deviation feature value of 0.85, indicating that the sintering condition is serious.

[0163] The data parsing module parses the batch identification code of the recycled products to obtain the production life cycle history data, which includes raw material batch information and production control variables during the production process.

[0164] This module is responsible for data alignment across time and space, establishing a connection between failures at the recycling end and parameters at the production end.

[0165] Full lifecycle history acquisition: The module parses the batch identification code of the recycled products and traces them back to their production records. The acquired history data includes batch information of raw materials, such as the origin of the alumina carrier, as well as production control variables during the production process, such as temperature, pressure, and feed rate in each temperature zone.

[0166] Spatiotemporal reverse mapping alignment: In order to solve the lag problem in the production process, the module performs a reverse mapping alignment operation.

[0167] Lag Calculation: Based on the logistics flow rules of the production line, the system calculates the average time lag of four hours from the granulation process to the finished product packaging.

[0168] Precise Slicing: The module extrapolates the production time of the recycled products to pinpoint the specific time window when the batch of materials flows through the three-stage roasting furnace, and captures slices of operational data within that window, such as the three-stage roasting temperature, heating rate, and ventilation volume, thus establishing an input feature matrix.

[0169] The element identification module constructs a multi-dimensional data association model, using the quality deviation data as the target variable and the history data as the input feature, to analyze the influence weight coefficients and attribution confidence indices of each production control variable on the quality deviation; variables with influence weight coefficients higher than a preset threshold are selected and defined as key control elements.

[0170] Multidimensional correlation analysis: The module constructs a multidimensional data correlation model, using quality deviation data as the target variable and the above-mentioned input feature matrix as the input features;

[0171] An exemplary construction of the input feature matrix and feature engineering includes:

[0172] The set of production control variable features includes, but is not limited to, the following, totaling 18 numerical features;

[0173] Mixing process: raw material mixing time, mixing speed, mixing temperature;

[0174] Extrusion process: extrusion pressure, extrusion temperature, and extrusion output;

[0175] Drying process: drying temperature, drying time, drying humidity;

[0176] Firing process: first-stage firing temperature, second-stage firing temperature, third-stage firing temperature, heating rate, ventilation volume;

[0177] Cooling process: cooling rate, final cooling temperature.

[0178] For raw material batch information, such as the origin of alumina, a unique thermal coding method is used to convert it into classification features. If there are 5 original suppliers, 5 binary features are generated, where the feature value corresponding to the product of a certain supplier is 1, and the rest are 0.

[0179] The set of environmental logistics factor characteristics includes: cumulative vibration energy spectral density, environmental humidity integral value, and temperature change cycle number, totaling three characteristics.

[0180] After spatiotemporal mapping, the final input feature matrix dimension is (number of samples × 26 feature dimensions).

[0181] Multi-output design for target variable:

[0182] The quality deviation data is a 4-dimensional vector, corresponding to the severity index of four failure modes [high temperature sintering, chemical poisoning, pore blockage, mechanical breakage].

[0183] A multi-output GBDT model architecture is adopted, which contains four independent but shared ensemble learning hyperparameters of decision trees, each learning the relationship between the severity of a failure and production control variables. This design preserves the indirect correlations that may exist between different failure modes while allowing for refined learning based on the specific characteristics of each failure mode.

[0184] The hyperparameter settings for the GBDT model are as follows: Maximum decision tree depth: max_depth=5, to prevent overfitting while retaining sufficient nonlinear fitting ability; Learning rate: learning_rate=0.1; Subsampling ratio: subsample=0.8, randomly sampling 80% of the samples in each iteration; Feature subsampling ratio: colsample_bytree=0.8, randomly sampling 80% of the features in each split; Minimum number of leaf samples: min_child_weight=5, to prevent overfitting; L1 regularization parameter: reg_alpha=0.1; L2 regularization parameter: reg_lambda=1.0; Number of iterations: an early stopping strategy is adopted to monitor the root mean square error on the validation set; training stops when there is no improvement in the validation set RMSE in 20 consecutive iterations, with a maximum of 500 iterations.

[0185] The training data was partitioned and multicollinearity was handled by dividing the historical sample set into three parts in chronological order.

[0186] Training set: The first 70% of samples used for model training;

[0187] Validation set: The middle 15% of samples, used for parameter tuning and early stopping determination;

[0188] Test set: The latest 15% of samples, used to independently evaluate the model's generalization performance.

[0189] During the feature engineering phase, the Pearson correlation coefficient matrix is ​​calculated among all input features. For feature pairs with an absolute correlation coefficient exceeding 0.9, they are identified as having strong multicollinearity, and one of these features is selected for deletion, retaining the one with a stronger correlation to the target variable.

[0190] After GBDT training, the contribution value of each feature in the model is quantified by the amount of loss reduction brought about by the feature being used for splitting during iteration. The contribution values ​​of all features are normalized to obtain the influence weight coefficient.

[0191] Training and Weight Calculation: The gradient boosting decision tree algorithm was used for training, and the information contribution value of each variable in node splitting was evaluated. The results showed that the average temperature of the three-stage calcining furnace had the highest influence weight coefficient.

[0192] Mechanism verification and confidence assessment: To prevent statistical misjudgment, the module performs double verification;

[0193] Mechanism path verification: By calling the chemical mechanism knowledge graph, a clear connection path was found between excessively high calcination temperature and specific surface area decay (sintering), which was confirmed as a strong mechanism correlation. A gain enhancement factor was given to increase its weight.

[0194] Repeated validation: The data were resampled 100 times using a random sampling strategy. The results showed that the average temperature of the three-stage calcining furnace was identified as a key factor in 95 analyses, therefore its attribution confidence index was determined to be 95%.

[0195] Key element identification: Finally, the module screened out variables whose influence weight coefficients were higher than the preset threshold, and formally defined the average temperature of the three-stage calcining furnace as the key control element that caused the failure of this high-temperature sintering.

[0196] The production control module performs production regulation and optimization operations based on the key control elements, calculates the target parameter offset used to correct quality deviations, and dynamically calculates the dynamic tolerance constraint range of production control variables in combination with the attribution confidence index; the target parameter offset and the dynamic tolerance constraint range are encapsulated into an execution control protocol.

[0197] This module is responsible for generating closed-loop control strategies and transforming analysis results into machine-executable instructions.

[0198] Reverse compensation optimization: Based on historical data, the module fits and constructs a correlation response model between the average temperature of the three-stage calcining furnace and the sintering severity; a management objective function is defined in the model to minimize the quality deviation modulus;

[0199] The correlation response model includes a constant term, a first-order linear term, a quadratic term, and an interaction term. To prevent the model from becoming too complex and causing overfitting, a stepwise regression method is used to select significant interaction terms. The specific process is as follows:

[0200] The initial model contains only first-order linear and quadratic terms, totaling 26 + 26 = 52 coefficients. Based on the significance determination of the F-test (p-value < 0.05), interaction terms are gradually added. After each interaction term is added, the adjusted R² is recalculated. If the increase in R² is less than 2%, the addition is stopped. R² represents the proportion of the variance (variation) of the response variable (i.e., failure mode) that the model can explain to the total variance. In this way, the number of interaction terms can be compressed from the theoretical 325 to the usual 15-20, which not only preserves important nonlinearities and coupling relationships, but also avoids parameter explosion.

[0201] For the multi-output scenario (four failure modes), four independent response surface models are established, allowing different failure modes to have different sensitivities to the same production parameter;

[0202] The least squares method is used for model fitting, and the number of samples should be no less than 10 times the number of coefficients. If the number of available samples is insufficient, ridge regression or Lasso regression is used to introduce a regularization term to improve model stability.

[0203] Offset calculation: Using a multi-objective optimization strategy to search for the optimal solution, it is calculated that in order to offset the current sintering deviation, the set temperature should be reduced by 8 degrees Celsius, that is, the target parameter offset is -8 degrees Celsius.

[0204] Dynamic tolerance calculation: The module reads that the attribution confidence index output by the feature recognition module is 95%, and determines that it is in the high confidence interval;

[0205] Tolerance narrowing: The system executes a strong correction strategy, dynamically adjusting the dynamic tolerance constraint range. This involves narrowing the historical standard tolerance of the temperature variable (…). The tolerance (in degrees Celsius) is narrowed to 30% of the equipment's limit capability, meaning the new dynamic tolerance is... Celsius, to ensure absolute precision in temperature control;

[0206] Protocol encapsulation and delivery: The module will adjust the target parameter offset (down by 5 degrees Celsius) and the dynamic tolerance constraint range ( The temperature (in degrees Celsius) is encapsulated into a standardized execution control protocol. This protocol is automatically pushed to the manufacturing execution system via an interface. In the next batch of production, the temperature control program of the calcining furnace will automatically apply this more stringent process standard, thereby achieving closed-loop quality management throughout the entire life cycle.

[0207] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A catalyst industry chain-oriented full life cycle traceability method, characterized in that, The method comprises the following steps: Obtain state attribute data of the same batch of recycled products, analyze the state attribute data, and construct a standardized performance feature vector; input the performance feature vector into a pre-set quality evaluation model for matching to identify the quality deviation data of the recycled products; The state attribute data covers performance evaluation indicators at three levels of microstructure, surface deposition, and macro mechanics: The first evaluation indicator is a microstructure characterization index, including a specific surface area attenuation index, a pore volume distribution skewness index, and an average pore size collapse index; The second evaluation indicator is a surface chemical deposition index, including a total amount of carbon deposition on the catalyst surface, a carbon deposition graphitization degree index, and a deposition density index of exogenous toxic elements; The third evaluation indicator is a macro mechanical property index, including an average value index of particle side pressure strength, a wear index, and a change rate index of bulk density; The quality deviation data includes an abnormal type label and deviation feature data; the abnormal type label includes high-temperature sintering failure, chemical poisoning failure, pore blockage failure, and mechanical crushing failure; the deviation feature data, i.e., a failure severity index, is used to quantitatively describe the severity value of the failure type; Analyze the batch identity code of the recycled products to obtain production life cycle history data, which covers raw material batch information and production control variables during the production process; Construct a multi-dimensional data correlation model, taking the quality deviation data as the target variable and the history data as the input feature, analyze the influence weight coefficient and the attribution confidence index of each production control variable on the quality deviation, and select the variable with an influence weight coefficient higher than a pre-set threshold as a key control element; The multi-dimensional data correlation model comprises: Perform reverse mapping alignment of space-time data; according to the logistics flow rule of the production line, calculate the time lag of the product from raw material input to finished product output, accurately map the batch identity code of the recycled products back to the operation data slice when flowing through each production process, and establish an input feature matrix; Perform weight evaluation based on multi-dimensional attribution analysis; import the input feature matrix and the target variable into the multi-dimensional data correlation analysis model for learning, and quantitatively obtain the influence weight coefficient by evaluating the information contribution value of each production control variable in the model node splitting process; Perform confidence evaluation based on repeated verification; use a random sampling strategy to reconstruct and repeatedly analyze the data set multiple times, statistically analyze the frequency stability of the production control variable being identified as a key element in repeated evaluation, and define the frequency stability value as the attribution confidence index; The multi-dimensional data correlation model can also integrate a pre-constructed chemical mechanism knowledge graph; the chemical mechanism knowledge graph contains entity nodes representing production control variables, result nodes representing quality deviation data, and relationship edges representing chemical reaction principles; when calculating the influence weight coefficient, perform a mechanism path verification step: use a graph traversal algorithm to search for a connected path from the high-weight production control variable node to the quality deviation data node in the chemical mechanism knowledge graph; According to the search results of the connected path, perform a weight correction strategy: If there is no connected path, it is determined that the statistical correlation of the production control variable and the quality deviation data is a pseudo correlation, a damping penalty factor is introduced to attenuate the influence weight coefficient; If there is a connected path, it is determined to be a mechanism strong correlation, a gain enhancement factor is introduced to improve the attribution confidence index; Based on the key control element, the production control optimization operation is performed, the target parameter offset for correcting the quality deviation is calculated, and the dynamic tolerance constraint range of the production control variable is dynamically calculated combined with the attribution confidence index; The target parameter offset and the dynamic tolerance constraint range are packaged as an execution control protocol.

2. The full life cycle tracing method for catalyst industry chain according to claim 1, characterized in that: The state attribute data is analyzed and processed to construct a standardized performance feature vector, including: Performing consistency verification of each state attribute data in the same batch of recovered products, identifying abnormal data points outside the compliance range using a preset outlier judgment logic; For the identified abnormal data points, combined with the sampling position label, data validity audit is performed, if it is determined as sampling error, it is excluded, if it is determined as local working condition abnormality, it is reserved and marked as an abnormal sub-sample, to obtain high-fidelity cleaned data; Adopting dimensionless mapping protocol, the cleaned data is mapped into the preset standardized evaluation interval, generating the performance feature vector of the recovered product.

3. The full life cycle tracing method for catalyst industry chain according to claim 1, characterized in that: The identification logic of the quality evaluation model is designed as follows: a business feature reference library containing various standard failure modes is constructed in advance, which stores standard benchmark models corresponding to different failure types; In the matching process, the similarity evaluation rule is applied to calculate the matching correlation degree of the current input performance feature vector and the standard benchmark model in the business feature reference library; The abnormal type label and the deviation feature data are correspondingly output, and the quality deviation data is obtained.

4. The full life cycle tracing method for catalyst industry chain according to claim 1, characterized in that: The production control optimization operation adopts an optimization design based on the reverse compensation strategy, including: Based on the data of historical batches, an associated response model between the key control elements and the quality deviation data is fitted and constructed, representing the sensitivity relationship between the change of the production control variable and the final product quality deviation; In the associated response model, a management objective function is defined, which aims to minimize the modulus of the quality deviation data; Using multi-objective optimization strategy, the optimal control strategy is searched in the feasible process window, the numerical change amount of the key control element that should occur to offset the currently identified quality deviation is calculated, and the numerical change amount is defined as the target parameter offset.

5. The full life cycle tracing method for catalyst industry chain according to claim 1, characterized in that: The calculation of the dynamic tolerance constraint range of the production control variable includes: identifying the attribution confidence index to distinguish high confidence interval, medium confidence interval and low confidence interval; Based on the confidence interval of the attribution confidence index, the data interval of the dynamic tolerance constraint range is dynamically adjusted.

6. A catalyst industry chain-oriented whole life cycle traceability system, characterized in that, Including: The data acquisition module is configured to acquire state attribute data of the same batch of recycled products, analyze and process the state attribute data, and construct a standardized performance feature vector; and input the performance feature vector into a preset quality evaluation model for matching to identify quality deviation data of the recycled products. The state attribute data covers performance evaluation indexes in three aspects of microscopic structure, surface deposition, and macro mechanics: The first evaluation index is a microscopic structure representation index, including a specific surface area attenuation index, a pore volume distribution skewness index, and an average pore size collapse index; The second evaluation index is a surface chemical deposition index, including a total amount of deposited carbon on the surface of the catalyst, a degree of graphitization of the deposited carbon, and a deposition density of exogenous toxic elements; The third evaluation index is a macro mechanical performance index, including an average value index of particle side pressure strength, a wear index, and a change rate index of bulk density; The quality deviation data includes an abnormal type label and deviation feature data; the abnormal type label includes high-temperature sintering failure, chemical poisoning failure, pore blockage failure, and mechanical crushing failure; and the deviation feature data is a failure severity index for quantitatively describing a severity value of the failure type. The data analysis module is configured to analyze batch identity codes of the recycled products, and acquire history data of the whole production life cycle, the history data covering raw material batch information and production control variables in the production process. The factor identification module is configured to construct a multi-dimensional data correlation model, take the quality deviation data as a target variable, take the history data as input features, analyze influence weight coefficients and attribution confidence indexes of each production control variable on the quality deviation, and screen out variables with an influence weight coefficient higher than a preset threshold value as key control factors. The multi-dimensional data correlation model includes: Performing reverse mapping alignment of space-time data; calculating a time lag amount of products from raw material input to finished product output according to a logistics flow rule of a production line, accurately mapping the batch identity codes of the recycled products back to operation data slices when the products flow through each production process, and establishing an input feature matrix; Performing weight evaluation based on multi-dimensional attribution analysis; inputting the input feature matrix and the target variable into a multi-dimensional data correlation analysis model for learning, quantitatively obtaining the influence weight coefficients by evaluating information contribution values of each production control variable in a model node splitting process; Performing confidence evaluation based on repeated verification; repeatedly analyzing a data set by using a random sampling strategy to reconstruct the data set multiple times, and statistically obtaining frequency stability of production control variables identified as key factors in repeated evaluation, and defining the frequency stability value as an attribution confidence index; The multi-dimensional data correlation model can also fuse a pre-constructed chemical mechanism knowledge graph; the chemical mechanism knowledge graph includes entity nodes representing production control variables, result nodes representing quality deviation data, and relationship edges representing chemical reaction principles; and when calculating the influence weight coefficients, a mechanism path verification step is performed: a graph traversal algorithm is used to search a connected path from a high-weight production control variable node to a quality deviation data node in the chemical mechanism knowledge graph. According to the search result of the connected path, a weight correction strategy is performed: If there is no connected path, it is determined that the statistical correlation between the production control variable and the quality deviation data is a pseudo correlation, and a damping penalty factor is introduced to attenuate the influence weight coefficient; If there is a connected path, it is determined to be a strong correlation mechanism, and a gain enhancement factor is introduced to improve the attribution confidence index; The production control module is used for production control optimization operation based on the key control elements, calculating the target parameter offset for correcting the quality deviation, and dynamically calculating the dynamic tolerance constraint range of the production control variable combined with the attribution confidence index; the target parameter offset and the dynamic tolerance constraint range are packaged as an execution control protocol.

Images