A PVC stabilizer production process data intelligent processing method and system

Abstract

This application provides an intelligent processing method and system for PVC stabilizer production process data, applied in the field of PVC stabilizer production technology. It acquires first-type process data reflecting the macroscopic state of the production process and second-type process data reflecting the microscopic structure of the production materials. Feature extraction is performed on the second-type process data to obtain feature signals characterizing the microscopic properties of the materials. These feature signals are then fused with the first-type process data to construct a comprehensive feature vector for the current production scenario. The comprehensive feature vector is compared with a preset set of reference scenarios to determine the deviation metric of the current production scenario relative to known scenarios. Based on the deviation metric, the uncertainty parameter of the quality prediction result is determined. Finally, a target prediction logic is matched among multiple preset prediction logics to output quality prediction information. This method effectively integrates macroscopic process data and microscopic material data, improving the accuracy and robustness of quality prediction.

Description

Technical Field

[0001] This application relates to the field of PVC stabilizer production technology, and in particular to a method and system for intelligent processing of PVC stabilizer production process data. Background Technology

[0002] In PVC stabilizer production, existing intelligent processing systems rely on conventional process information such as temperature and pressure to build quality prediction logic. Although this is effective in the early stages, when raw materials are changed (even if conventional indicators are qualified), the system cannot detect the micro-differences in raw materials that cause fluctuations in product quality, and therefore cannot provide early warnings.

[0003] To address the aforementioned issues, existing technologies have proposed a theoretical approach: installing an online spectrometer on the reactor to obtain high-dimensional spectral data at the molecular level of the materials. This theoretical approach was published by O. Kolter et al. in the journal Springer, titled "A process analyzer assembly".

[0004] However, in practical applications, the data generated by the spectrometer is high-dimensional, massive continuous curve data, which differs significantly from the original low-dimensional, sparse conventional process data in terms of structure and information density. This makes it difficult for the system to directly fuse the two types of data. Secondly, the original spectral data contains a large amount of background noise and redundant information, requiring complex feature extraction to obtain key signals that effectively characterize the material properties. Furthermore, due to the lack of ability to distinguish between different production scenarios, when the spectral features are inconsistent with conventional data, the system cannot determine whether the current production state falls within the scope of historical experience, nor can it determine the confidence level of the prediction results. Ultimately, it is still difficult to achieve effective early warning of quality anomalies caused by minor fluctuations in raw materials.

[0005] Therefore, the current technical challenge is to design a new data processing method that can extract key features representing minute differences in raw materials from high-dimensional spectral curves containing background noise, and deeply integrate these features with traditional low-dimensional process information to construct a unified quality prediction logic compatible with both types of heterogeneous data. This will solve the problem that relying solely on conventional information cannot accurately predict product quality instability caused by micro-fluctuations in raw materials. Summary of the Invention

[0006] In view of the shortcomings of the prior art, this application provides a method and system for intelligent processing of PVC stabilizer production process data, which has the advantages of effectively integrating macroscopic process data and microscopic material data, and improving the accuracy and robustness of quality prediction.

[0007] In a first aspect, a method for intelligent processing of PVC stabilizer production process data, the method comprising the following steps:

[0008] S1: Acquire the first type of process data reflecting the macroscopic state of the production process, and the second type of process data reflecting the microscopic structure of the production materials;

[0009] S2: Extract features from the second type of process data to obtain feature signals characterizing the microscopic properties of the material;

[0010] S3: The feature signal is fused with the first type of process data to construct a comprehensive feature vector of the current production situation;

[0011] S4: Compare the comprehensive feature vector with the preset reference scenario set to determine the deviation metric value of the current production scenario relative to the known scenario;

[0012] S5: Determine the uncertainty parameter of the quality prediction result based on the deviation metric value, and match the target prediction logic among a plurality of preset prediction logics based on the deviation metric value to output quality prediction information.

[0013] Furthermore, step S1 includes:

[0014] S11: Collect the first type of process data through sensors and use an industrial controller to check the range of the first type of process data. When the first type of process data exceeds the preset safe operating range, trigger an alarm.

[0015] S12: Perform spectral scanning on the production materials using an online near-infrared spectrometer to obtain the second type of process data.

[0016] Furthermore, step S12 includes:

[0017] S121: Before scanning the production materials, obtain the reference spectrum of the built-in standard reference material of the online near-infrared spectrometer;

[0018] S122: Compare the background spectrum under the current environment with the reference spectrum and calculate the spectral shift;

[0019] S123: Determine whether the spectral drift is within a preset self-compensation range;

[0020] S124: If the spectral drift is within the self-compensation range, a compensation operator is generated based on the spectral drift, and the original spectrum obtained by scanning is energy-corrected using the compensation operator to obtain the corrected second type of process data;

[0021] S125: If the spectral drift exceeds the self-compensation range, the sensor window is determined to be contaminated, the current prediction logic is stopped, and a cleaning and maintenance command is output.

[0022] Furthermore, step S2 includes:

[0023] S21: Perform smoothing and background interference elimination processing on the second type of process data;

[0024] S22: Extract absorbance feature values ​​of a specific band from the processed second type of process data, and use the absorbance feature values ​​as the feature signal.

[0025] Furthermore, step S21 includes:

[0026] S211: Apply the Savitzky-Golay smoothing algorithm to perform local polynomial fitting on the second type of process data to remove high-frequency random noise;

[0027] S212: The non-feature region of the second type of process data is fitted using a polynomial fitting method to estimate the background signal, and the background signal is subtracted from the original spectrum to highlight the absorption characteristics of the spectral curve.

[0028] Furthermore, step S22 includes:

[0029] S221: Calculate the integral area of ​​a specific wavelength band as the absorbance characteristic value; or, calculate the zero point position and intensity of the second derivative of the spectral curve at a specific wavelength as the absorbance characteristic value; or, calculate the relative height of a specific absorption peak relative to the peak height at a reference wavelength as the absorbance characteristic value, and use the absorbance characteristic value as the characteristic signal.

[0030] Furthermore, step S3 includes:

[0031] S31: Align the first type of process data with the feature signal according to the timestamp;

[0032] S32: Integrate the aligned first type of process data with the feature signal to form the comprehensive feature vector, so as to achieve synchronous description of the production status.

[0033] Furthermore, step S4 includes:

[0034] S41: Calculate the Euclidean distance between the comprehensive feature vector and each reference vector in the reference context set;

[0035] S42: Calculate the situation novelty score based on the Euclidean distance, and use the situation novelty score as the deviation metric;

[0036] Step S42 includes:

[0037] S421: Establish a context reference database, which stores the comprehensive feature vectors of historical batches and the corresponding quality results;

[0038] S422: Calculate the Euclidean distance between the integrated feature vector and each reference vector in the context reference database, and select the K closest known contexts;

[0039] S423: Calculate the average Euclidean distance of the K known scenarios, and determine the average Euclidean distance as the scenario novelty score.

[0040] Furthermore, in step S5, matching the target prediction logic among a preset set of prediction logics based on the deviation metric value includes the following steps:

[0041] S51: Determine whether the deviation measurement value exceeds a preset deviation threshold;

[0042] S52: If the deviation metric does not exceed the preset deviation threshold, then the fine association rule is determined as the target prediction logic;

[0043] S53: If the deviation metric exceeds the preset deviation threshold, the weight of the fine association rule is reduced, and the robust threshold rule based on physicochemical principles is determined as the target prediction logic.

[0044] Secondly, a smart data processing system for PVC stabilizer production process, the system being used to implement the steps of any of the methods described above, the system comprising:

[0045] The data acquisition module is used to acquire the first type of process data that reflects the macroscopic state of the production process, and the second type of process data that reflects the microscopic structure of the production materials.

[0046] The feature extraction module is used to extract features from the second type of process data to obtain feature signals that characterize the microscopic properties of the material.

[0047] The fusion processing module is used to fuse the feature signals with the first type of process data to construct a comprehensive feature vector of the current production scenario.

[0048] The similarity comparison module is used to compare the comprehensive feature vector with a preset set of reference scenarios to determine the deviation metric of the current production scenario relative to the known scenarios.

[0049] The quality prediction module is used to determine the uncertainty parameter of the quality prediction result based on the deviation metric value, and to match the target prediction logic among a number of preset prediction logics based on the deviation metric value, so as to output quality prediction information.

[0050] Beneficial Effects: The intelligent processing method and system for PVC stabilizer production process data proposed in this application acquires first-type process data reflecting the macroscopic state of the production process and second-type process data reflecting the microscopic structure of the production materials. Feature extraction is performed on the second-type process data to obtain feature signals characterizing the microscopic properties of the materials. These feature signals are then fused with the first-type process data to construct a comprehensive feature vector for the current production scenario. The comprehensive feature vector is compared with a preset set of reference scenarios to determine the deviation metric of the current production scenario relative to known scenarios. Based on the deviation metric, the uncertainty parameter of the quality prediction result is determined. Finally, a target prediction logic is matched among multiple preset prediction logics to output quality prediction information. This method effectively integrates macroscopic process data and microscopic material data, improving the accuracy and robustness of quality prediction. Attached Figure Description

[0051] Figure 1 This is a flowchart of a method for intelligent data processing in the production process of PVC stabilizers proposed in this application.

[0052] Figure 2 This is a structural diagram of an intelligent data processing system for PVC stabilizer production process proposed in this application.

[0053] Figure 3 This is a schematic diagram of an intelligent data processing system for PVC stabilizer production process proposed in this application.

[0054] Labeling Explanation: 201, Data Acquisition Module; 202, Feature Extraction Module; 203, Fusion Processing Module; 204, Similarity Comparison Module; 205, Quality Prediction Module. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and marked in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0056] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0057] In the production of polyvinyl chloride (PVC) stabilizers, ensuring the uniformity of product quality across each batch is crucial. Traditional production control heavily relies on monitoring the macroscopic conditions of the production process, such as temperature, pressure, material levels, and agitator speed within the reactor. When the production process and raw material sources remain stable, establishing correlations between these macroscopic process parameters and final product quality indicators, such as thermal stability time, can indeed achieve effective quality control and prediction. However, modern industrial production faces dynamic changes in the supply chain, with raw material changes becoming commonplace.

[0058] When raw materials are changed, subtle differences at the microscopic level can affect the macroscopic performance of the product. Existing theoretical spectral analysis algorithms are too complex. In practical applications, existing systems struggle to extract key features characterizing minute differences in raw materials from high-dimensional spectral curves. Furthermore, these features need to be deeply integrated with traditional low-dimensional process information to construct a unified quality prediction logic compatible with both types of heterogeneous data. This addresses the problem that relying solely on conventional information cannot accurately predict product quality instability caused by microscopic fluctuations in raw materials.

[0059] To solve this problem, please refer to Figure 1 This application proposes an intelligent data processing method for PVC stabilizer production process, the method comprising the following steps:

[0060] S1: Acquire the first type of process data reflecting the macroscopic state of the production process, and the second type of process data reflecting the microscopic structure of the production materials;

[0061] S2: Extract features from the second type of process data to obtain feature signals that characterize the microscopic properties of the material;

[0062] S3: The feature signals are fused with the first type of process data to construct a comprehensive feature vector for the current production situation;

[0063] S4: Compare the comprehensive feature vector with the preset set of reference scenarios to determine the deviation measure of the current production scenario relative to the known scenarios;

[0064] S5: Determine the uncertainty parameters of the quality prediction result based on the deviation metric value, and match the target prediction logic among multiple preset prediction logics based on the deviation metric value to output quality prediction information.

[0065] The method works as follows: First, comprehensive information is collected, focusing not only on macroscopic phenomena such as temperature and pressure but also exploring the internal microstructure of the material. Then, the current situation is compared with historical experience to determine familiarity with the current condition. Finally, the decision-making strategy is adjusted based on this familiarity: for familiar situations, precise judgments are made according to refined empirical rules; while for unfamiliar situations, conservative judgments are made by reverting to basic physicochemical principles, clearly identifying the inherent uncertainties.

[0066] Specifically, this method first introduces online spectral analysis technology to acquire second-type process data that directly reflects changes in the molecular structure and composition of materials, overcoming the fundamental limitation of traditional methods in failing to perceive microscopic differences in raw materials. Next, through data processing, a few key feature signals are extracted from the high-dimensional, complex spectral data. These microscopic features are then synchronized and deeply integrated with traditional macroscopic process data to construct a comprehensive feature vector that provides a panoramic description of the production status. This step solves the technical challenge of effectively integrating heterogeneous data from two completely different sources with entirely different structures and dimensions.

[0067] Subsequently, by comparing the current integrated feature vector with a reference scenario database storing historical production experience in real time, a deviation metric, or scenario novelty score, is calculated. This score directly answers the extent to which the current production scenario is unfamiliar. A high score means that the current production process has encountered an unprecedented situation, directly responding to the unknown risks caused by the introduction of new raw materials. In this unfamiliar scenario, this method does not blindly provide a seemingly accurate but potentially completely wrong prediction result like traditional logic. Instead, it first quantifies and outputs the uncertainty of the prediction result based on the magnitude of the deviation metric, for example, by providing a wider range of quality fluctuations or a clear risk level. At the same time, it dynamically switches its internal prediction logic, shifting from fine-grained rules relying on statistical correlations of historical data to robust threshold rules relying on basic physicochemical principles. This approach enables the entire decision-making process to exhibit a high degree of prudence and robustness when facing the unknown.

[0068] Finally, this method also incorporates a closed-loop mechanism for continuous learning. After each production batch is completed, its complete production context data and final quality results are added to the reference context database. This means that every unfamiliar situation encountered, regardless of the outcome, becomes a learning experience for future judgments. Through this continuous learning and accumulation, the method's cognitive boundaries for various production scenarios will continuously expand, and the accuracy and confidence of its judgments will improve accordingly.

[0069] Through the above series of steps, the method proposed in this application can effectively overcome the limitations of relying solely on macroscopic process data for quality prediction. In particular, when faced with unexpected disturbances such as minor fluctuations in raw materials, it can provide more robust and reliable quality prediction and risk warning, thereby significantly improving the stability of the PVC stabilizer production process and the qualification rate of the final product.

[0070] In a specific implementation scenario, the steps of acquiring the first type of process data reflecting the macroscopic state of the production process and the second type of process data reflecting the microscopic structure of the production materials can be divided into two parallel information acquisition processes. Specifically, step S1 includes:

[0071] S11: Collect the first type of process data through sensors and use the industrial controller to check the range of the first type of process data. When the first type of process data exceeds the preset safe operating range, an alarm is triggered.

[0072] S12: Perform spectral scanning on the production materials using an online near-infrared spectrometer to obtain second-type process data.

[0073] The first type of process data refers to a set of parameters that can describe the state of a reaction system at a macroscopic physical level. These include, but are not limited to, instantaneous temperature, static pressure, stirring shaft torque, and cumulative flow rate of the feed pump inside the reactor. These parameters determine the kinetic conditions of the reaction. The second type of process data refers to high-dimensional absorbance sequences obtained through near-infrared spectroscopy that reflect the state of chemical bonds within the molecules of the material. These sequences contain information on transitions in molecular vibrational energy levels and can identify the introduction of trace impurities or changes in the proportion of organic acid isomers. By observing these two types of data at different physical scales, a three-dimensional monitoring system from energy state to material structure can be constructed.

[0074] The core equipment in PVC stabilizer production, a large reactor, is equipped with various types of sensors. For example, armored thermocouples are used to measure temperature at different locations within the reactor, pressure transmitters monitor pressure at the top, stirring power is obtained indirectly or directly by monitoring the operating current of the stirring motor or by directly installing torque sensors, and level gauges monitor material height. These sensors convert the collected physical quantities into standard electrical signals, such as 4 to 20 mA current signals. These signals are transmitted to an industrial controller, such as a programmable logic controller (PLC). The PLC scans and digitizes these signals at millisecond intervals and uses its internal logic to perform a preliminary range check on the data. This range check is based on preset process safety procedures; for example, the normal operating temperature range is set between 50 and 250 degrees Celsius, and the safe upper limit for pressure is 0.5 MPa. When the PLC detects that any data point exceeds the preset safe operating range, it will immediately trigger an alarm. The alarm information will be sent to the human-machine interface or SCADA system in the central control room via industrial Ethernet, such as Profinet or Modbus TCP / IP protocol, to prominently prompt the operator to pay attention or intervene.

[0075] To obtain microstructural information about the materials, an online near-infrared spectrometer was installed on the external circulation pipeline of the reactor. Installation on the circulation pipeline was chosen to allow for continuous measurement of representative materials without interfering with the main reaction process. The analyzer performs a spectral scan of the material flowing through the measuring probe at preset time intervals, such as every 60 seconds. With each scan, the analyzer emits a broadband beam of near-infrared light that penetrates the material, and the detector receives and records the light absorption intensity at different wavelengths. Each scan generates a complete spectral curve consisting of thousands of data points; for example, sampling at a resolution of 2 nanometers within the wavelength range of 1000 nm to 2500 nm yields an absorbance curve containing 751 data points. This curve represents the second type of process data reflecting the molecular bond vibrational absorption characteristics of the material at the current moment. This high-dimensional spectral data is transmitted in real-time via modern industrial communication protocols such as OPC UA to a high-performance industrial computer dedicated to data processing, awaiting further in-depth analysis.

[0076] When spectrometers operate in industrial settings, their performance is affected by factors such as changes in ambient temperature and humidity, and more critically, contamination of the probe window, resulting in measurement drift. Without correction, this drift may be misinterpreted as a change in the material itself, causing the entire judgment logic to fail.

[0077] To address this problem, step S12 further includes:

[0078] S121: Before scanning production materials, obtain the reference spectrum of the built-in standard reference material of the online near-infrared spectrometer;

[0079] S122: Compare the background spectrum under the current environment with the reference spectrum and calculate the spectral shift;

[0080] S123: Determine whether the spectral drift is within the preset self-compensation range;

[0081] S124: If the spectral drift is within the self-compensation range, a compensation operator is generated based on the spectral drift, and the energy of the original spectrum obtained by scanning is corrected using the compensation operator to obtain the corrected second-type process data.

[0082] S125: If the spectral drift exceeds the self-compensation range, the sensor window is determined to be contaminated, the current prediction logic is stopped, and a cleaning and maintenance command is output.

[0083] The standard reference material is typically a material with extremely stable physical and chemical properties, such as high-purity ceramics or polystyrene sheets, whose spectral characteristics are known and do not change over time. By scanning this standard reference material, the resulting spectral curve serves as a baseline snapshot of the instrument under ideal conditions.

[0084] During production, the system periodically performs background spectrum measurements, for example, every hour. During this time, the pipeline valves are briefly switched to allow pure solvent (without reactants) to flow through the probe, or the measurement is performed with the pipeline emptied, obtaining the background spectrum for the current environment. This real-time acquired background spectrum is then mathematically compared with a previously stored reference spectrum to calculate the difference, which is quantified as a spectral drift. The comparison can be performed by calculating the root mean square error of the two spectral curves across the entire wavelength range, or by comparing the absorbance differences at several key wavelength points.

[0085] Next, it is determined whether the calculated spectral drift falls within a preset self-compensation range. This range is determined based on the instrument's historical performance data and the allowable measurement error range of the process. For example, if the root mean square error is less than 0.005 absorbance units, the drift is considered minor and can be compensated for by software algorithms.

[0086] If the spectral drift is determined to be within the self-compensation range, a compensation operator is generated based on this drift. This compensation operator can be a simple addition or subtraction of an offset. Assume that, through comparison, the calculated spectral drift represents a constant absorbance offset value ΔA across the entire wavelength range, for example, 0.002 absorbance units, and this ΔA is within the preset self-compensation range. In this case, a compensation operator is generated, which is -ΔA. When new original spectral data is acquired, this compensation operator is added point-by-point to each absorbance value of the original spectrum, i.e., the corrected absorbance value = original absorbance value + (-ΔA). For example, if the absorbance of the original spectrum at 1400 nm is 0.500, then after correction it becomes 0.500 - 0.002 = 0.498. Through this simple addition or subtraction of an offset, the overall drift of the spectral baseline is eliminated, making the corrected spectral data consistent with the spectrum under the reference state. The calibrated spectral data eliminates the systematic errors caused by slight instrument drift, becoming a second type of process data that more accurately reflects the microscopic properties of materials.

[0087] However, if the spectral drift exceeds the upper limit of the self-compensation range, it usually indicates a more serious physical problem. The most common cause is the buildup of a layer of difficult-to-remove contaminants on the sapphire window where the probe contacts the material, due to high temperatures or chemical reactions. This contamination severely distorts the spectral signal and cannot be corrected by simple mathematical compensation. In this case, it is immediately identified as sensor window contamination. At this point, all prediction logic based on spectral data is suspended to prevent misjudgments based on erroneous data. Simultaneously, a clear cleaning and maintenance instruction is output to the central control room's operating interface, such as "Online spectrometer probe is severely contaminated; please immediately execute a chemical cleaning procedure," and the time and drift data of this event are recorded for subsequent equipment maintenance and fault analysis.

[0088] Furthermore, step S2 includes:

[0089] S21: Perform smoothing and background interference removal processing on the second type of process data;

[0090] S22: Extract the absorbance feature value of a specific band from the processed second-class process data, and use the absorbance feature value as the feature signal.

[0091] Among them, the raw spectral data inevitably contains two main types of interference: one is high-frequency random noise generated by detector electronic components or environmental electromagnetic interference, which manifests as small, irregular spikes on the spectral curve; the other is the background signal caused by the scattering of light by solid particles in the material or the slow change of the instrument baseline due to temperature drift, which manifests as the baseline of the entire spectral curve tilting or bending.

[0092] To eliminate these interferences, step S21 includes:

[0093] S211: The Savitzky-Golay smoothing algorithm is applied to perform local polynomial fitting on the second type of process data to remove high-frequency random noise;

[0094] S212: Use a polynomial fitting method to fit the non-feature region of the second type of process data to estimate the background signal, and subtract the background signal from the original spectrum to make the absorption characteristics of the spectral curve stand out.

[0095] In step S211, the Savitzky-Golay smoothing algorithm works as follows: A calculation window is formed by taking several adjacent data points as the center of each data point on the spectrum. Then, a low-order polynomial, such as a quadratic or cubic polynomial, is used to fit these data points within this window. Finally, the calculated value of the fitted polynomial at the center point replaces the original measured value. By sliding this calculation window across the entire spectrum, random noise can be effectively filtered out while preserving the shape and width of the true absorption peaks representing chemical information in the spectrum to the maximum extent, avoiding peak passivation problems that may occur with methods such as simple moving averages.

[0096] Based on prior knowledge of the PVC stabilizer chemical system, it is then determined which regions in the spectrum will not exhibit characteristic absorption peaks; these regions are referred to as non-characteristic regions. Next, using only data points from these non-characteristic regions, a low-order polynomial curve, such as a third or fourth-order polynomial, is fitted. This fitted curve is considered to represent the background signal of the entire spectrum. Finally, the fitted background signal curve is subtracted point-by-point from the smoothed entire spectrum. After this operation, the spectral baseline is flattened to near zero, and the true absorption peaks directly related to the material composition and molecular structure become exceptionally clear and prominent, laying a solid foundation for subsequent quantitative analysis.

[0097] After completing the preprocessing of the spectral data to obtain clean and baseline-corrected spectral curves, step S22 further includes:

[0098] S221: Calculate the integral area of ​​a specific wavelength band as the absorbance characteristic value; or, calculate the zero point position and intensity of the second derivative of the spectral curve at a specific wavelength as the absorbance characteristic value; or, calculate the relative height of a specific absorption peak relative to the peak height at a reference wavelength as the absorbance characteristic value, and use the absorbance characteristic value as the characteristic signal.

[0099] Depending on the different microscopic properties of the material to be monitored, various characteristic value calculation methods can be flexibly selected.

[0100] In one implementation, the integral area over a specific wavelength band can be calculated as a characteristic value of absorbance. This method is particularly suitable for monitoring the content of a specific impurity or functional group. For example, experimental analysis may reveal that a trace impurity affecting the thermal stability of a product has a characteristic absorption peak in the wavelength range of 1410 nm to 1430 nm. In this case, the area enclosed by the spectral curve and the baseline in this wavelength band can be calculated. This integral area value is proportional to the concentration of the impurity and is less sensitive to peak shape changes and slight peak shifts compared to a single peak height value.

[0101] In another implementation, the zero-point position and intensity of the second derivative of the spectral curve at a specific wavelength can be calculated as absorbance characteristic values. This method has a unique advantage when dealing with overlapping absorption peaks. For example, the CH bond stretching vibrations of two key organic acids in the feedstock produce two closely spaced absorption peaks near 2300 nm that almost merge into a single broad peak in the original spectrum. In this case, calculating the second derivative of the spectral curve transforms the absorption peak in the original spectrum into a sharp negative valley in the second derivative spectrum. The two originally overlapping peaks are decomposed into two independent valleys, the wavelength corresponding to the valley bottom being the precise position of the original peak, while the valley depth is related to the intensity of the original peak. By identifying and quantifying the wavelength position and depth of these two valleys, two independent characteristic signals can be obtained, thereby enabling separate monitoring of the two components.

[0102] In another implementation, the relative height of a specific absorption peak relative to its peak height at a reference wavelength can be calculated as an absorbance characteristic value. This method aims to eliminate the influence of fluctuations in the overall sample concentration or changes in optical path length. For example, a functional group to be monitored may have a characteristic absorption peak at 1680 nm, but the overall concentration of the material in the reactor may vary between different batches or at different stages of the same batch, causing the heights of all absorption peaks to increase or decrease proportionally. To eliminate this effect, a reference peak with stable intensity in the spectrum, unaffected by the reaction process, can be selected, such as an absorption peak from solvent molecules located at 2100 nm. Then, the peak height at 1680 nm is divided by the peak height at 2100 nm to obtain a relative height value. This ratio is related to the relative content of the target functional group, but not to the overall concentration, thus providing better batch-to-batch comparability and serving as a very stable characteristic signal.

[0103] Once the characteristic signals representing the microscopic properties of the material have been extracted, and the first type of process data reflecting the macroscopic state of the production process has also been collected, it is necessary to fuse these two information streams with different sources and properties to construct a comprehensive feature vector that can provide a complete and synchronous description of the current production situation.

[0104] Furthermore, step S3 includes:

[0105] S31: Align the first type of process data and feature signals according to timestamps;

[0106] S32: Integrate the aligned first-class process data with the feature signals to form a comprehensive feature vector, so as to achieve synchronous description of the production status.

[0107] To ensure a precise correspondence between macroscopic states and microscopic structures during analysis, synchronization based on time is essential. Specifically, each generated data point, whether a single temperature value from the PLC or a characteristic signal calculated from the spectrum, can be assigned a high-precision timestamp. Then, data frames are constructed using a uniform time granularity, such as minutes. Within a data frame, the spectral characteristic signal generated within that minute is matched with the average or last value of all macroscopic data such as temperature and pressure within that minute. This ensures that at every point in time, the macroscopic and microscopic information describe the same production state.

[0108] After time alignment is completed, the aligned first-class process data and feature signals can be integrated to form a comprehensive feature vector. Operationally, this integration process involves arranging all aligned values ​​in a predetermined order to form a single, higher-dimensional vector. For example, at a specific point in time, the three macroscopic data points of temperature, pressure, and stirring speed can be concatenated with five microscopic feature signals extracted from the spectrum, including the integral area, second derivative valley depth, and relative peak height, to form a one-dimensional array containing eight elements. This array is the comprehensive feature vector of the current production scenario.

[0109] As a specific implementation, the fusion processing logic is deployed in an edge computing terminal installed on the side of the reactor. The gateway receives macroscopic parameters from the underlying controller in real time via a high-speed industrial bus, and simultaneously receives the raw data stream generated by the spectrometer via fiber optic communication. Due to the need for efficient allocation of computing resources in the edge gateway, the fusion processing employs a dynamic dimensionality reduction strategy. First, principal component dimensionality reduction is performed on the high-dimensional spectral feature signal, extracting the three principal feature components with the highest contribution. Then, these three microscopic feature components are spatially weighted and concatenated with standardized temperature, pressure, and stirring speed data to construct a comprehensive feature vector that is dimensionally simplified but highly information-rich. A lightweight reference scenario set is cached in the gateway's local memory, and similarity comparison is performed using a fast lookup table algorithm. This edge-based implementation ensures extremely low time latency from sensing microscopic differences in raw materials to outputting deviation measurements. Even under network fluctuations in the central server, the edge terminal can still independently complete scenario construction and similarity assessment, guiding the on-site actuators to perform emergency actions according to preset logic.

[0110] After constructing a comprehensive feature vector that can fully describe the current production situation, the core step of situation awareness is to compare the similarity of the comprehensive feature vector with a preset set of reference situations to determine the deviation measure of the current production situation relative to all known historical situations.

[0111] Furthermore, step S4 includes:

[0112] S41: Calculate the Euclidean distance between the integrated feature vector and each reference vector in the reference context set;

[0113] S42: Calculate the situation novelty score based on Euclidean distance and use the situation novelty score as a deviation measure.

[0114] Euclidean distance is a classic method for measuring the straight-line distance between two points in multidimensional space. If each composite feature vector is considered as a point in a high-dimensional space, then the Euclidean distance between two vectors intuitively reflects the degree of similarity between the two production scenarios. The smaller the distance, the more similar the two scenarios are; the larger the distance, the greater the difference.

[0115] Then, a situational novelty score is calculated based on the Euclidean distance, and this score is used as the final bias measure. The calculation process for this score is further refined.

[0116] Specifically, step S42 includes:

[0117] S421: Establish a context reference database, which stores the comprehensive feature vectors of historical batches and the corresponding quality results;

[0118] S422: Calculate the Euclidean distance between the integrated feature vector and each reference vector in the context reference database, and select the K closest known contexts;

[0119] S423: Calculate the average Euclidean distance of K known scenarios and determine the average Euclidean distance as the scenario novelty score.

[0120] The context reference database stores all historical comprehensive feature vectors collected at various key process stages in a large number of past production batches. Each historical vector is associated with the final quality inspection result of its corresponding batch, such as whether it is qualified or unqualified, as well as the specific thermal stability time value.

[0121] When a new production batch is underway and a new composite feature vector is generated at the current moment, a similarity calculation is performed. The Euclidean distance between this new composite feature vector and each historical reference vector stored in the context reference database is calculated. After calculating all distances, these distance values ​​are sorted, and the K closest known contexts are selected. The value of K is an adjustable parameter, typically a small odd number, such as 5 or 7. These K known contexts can be considered as the few production experiences in history that are most similar to the current situation.

[0122] Finally, the average Euclidean distance between these K known scenarios is calculated. Specifically, the K smallest Euclidean distance values ​​are summed and then divided by K. This calculated average distance is ultimately determined as the novelty score of the current scenario. Using the average distance instead of the minimum distance increases the robustness of the score, preventing drastic fluctuations due to a single, isolated outlier in the database.

[0123] This situation novelty score is a precise quantification of how unfamiliar the current production situation is: the lower the score, the more the current situation matches historical experience and everything is under control; the higher the score, the more the current situation deviates from all known historical experience and is a completely new and uncertain unfamiliar situation.

[0124] In the production process of PVC stabilizers, after the deviation metric is calculated, it is input into a preset nonlinear mapping function, such as a sigmoid function, to map the deviation metric value from 0 to 100 to an uncertainty parameter from 0 to 1. When the deviation metric value is 0, the uncertainty parameter is 0, indicating that the prediction result is highly certain; when the deviation metric value is 100, the uncertainty parameter is 1, indicating that the prediction result is completely uncertain. The determination of the uncertainty parameter ensures that the prediction information includes not only the predicted value but also the confidence level of the prediction, providing decision-makers with a more comprehensive basis for risk assessment. Simultaneously, two logics are preset. Specifically, in step S5, matching the target prediction logic among multiple preset prediction logics based on the deviation metric value includes the following steps:

[0125] S51: Determine whether the deviation measurement value exceeds the preset deviation threshold;

[0126] S52: If the deviation metric does not exceed the preset deviation threshold, then the fine association rule will be determined as the target prediction logic;

[0127] S53: If the deviation metric exceeds the preset deviation threshold, the weight of the fine association rule is reduced, and the robust threshold rule based on physicochemical principles is determined as the target prediction logic.

[0128] This deviation threshold is the key dividing line between familiar and unfamiliar situations, and its value is determined based on statistical analysis of all historical data in the situation reference database. For example, the average distance between situation vectors within all historical qualified batches can be calculated and multiplied by a safety factor to obtain the deviation threshold.

[0129] If the deviation metric does not exceed a preset deviation threshold, it indicates that the current production situation is very similar to one or more known successful situations in the past. In this case, a set of refined association rules will be identified as the target prediction logic. This set of refined association rules is usually a prediction model trained on massive amounts of historical data using machine learning algorithms, such as gradient boosting decision trees.

[0130] As a specific implementation method, the gradient boosting decision tree model first collects a complete dataset of all PVC stabilizer production batches over the past five years during the training phase. The dataset includes macroscopic process data for each batch at key reaction stages, such as reactor temperature profiles, stirring power, the amount and timing of each material added, as well as synchronously acquired online near-infrared spectral characteristic signals, such as the integral area of ​​a specific band, the zero point of the second derivative, and relative peak height. Simultaneously, each batch is associated with the final product's quality inspection results, such as thermal stabilization time.

[0131] During model training, macroscopic process data and microscopic feature signals are used as input features, and thermal stabilization time is used as the target output. Gradient boosting decision trees iteratively construct a series of weak predictors, i.e., decision trees, where each tree attempts to correct the prediction error of the previous tree. For example, the first tree might predict an initial thermal stabilization time based on temperature and pressure, while the second tree corrects the prediction of the first tree based on spectral feature signals.

[0132] Through hundreds of iterations, the model learns the complex nonlinear mapping relationship between macroscopic and microscopic data. In actual prediction, when a new production batch enters the familiar environment, its real-time collected macroscopic process data and microscopic feature signals are input into the trained gradient boosting decision tree model. Based on the complex rules it has learned, the model outputs a precise predicted thermal stabilization time, such as 125 minutes. Simultaneously, the model provides a narrow confidence interval based on its internal prediction variance or the prediction consistency of the trees in the ensemble learning, for example, predicting a thermal stabilization time between 124 and 126 minutes, indicating a high degree of certainty in the prediction result.

[0133] Conversely, if the deviation metric exceeds a preset deviation threshold, a strong warning signal is issued: the current production situation is unprecedented, and any sophisticated model based on historical experience may make unreliable inferences because it exceeds its learning range. In this case, a more cautious strategy is adopted. First, the weight of sophisticated association rules is reduced, or even abandoned entirely. Then, a set of robust threshold rules based on physicochemical principles is established as the target prediction logic. This set of robust rules does not rely on complex statistical associations but is based on a deep understanding of the reaction mechanism of PVC stabilizers and basic engineering safety knowledge. For example, the rules might include clauses such as: if the intensity of a spectral absorption peak representing a certain type of harmful impurity exceeds a certain absolute safety threshold calculated based on stoichiometry, then regardless of other macroscopic parameters, the product is directly judged to be high-risk; or, if the reaction temperature exceeds the theoretical decomposition temperature of the material, the process is directly judged as a failure. While these first-principles-based rules cannot provide precise quantitative predictions, they can provide a reliable bottom-line judgment in unknown situations, preventing catastrophic quality accidents.

[0134] Furthermore, in this unfamiliar context, the output quality prediction information will include an uncertainty parameter, the magnitude of which is positively correlated with the deviation metric. For example, the output information might be: high quality prediction uncertainty, the current situation deviates from historical experience by 75%, key spectral features are abnormal, and the product has a high risk of non-compliance.

[0135] Please refer to Figure 2 , Figure 3A smart data processing system for PVC stabilizer production process, the system being used to implement the steps of any of the above methods, the system comprising:

[0136] The data acquisition module 201 is used to acquire first-class process data reflecting the macroscopic state of the production process and second-class process data reflecting the microstructure of the production materials.

[0137] Feature extraction module 202 is used to extract features from the second type of process data to obtain feature signals characterizing the microscopic properties of the material;

[0138] The fusion processing module 203 is used to fuse the feature signals with the first type of process data to construct a comprehensive feature vector of the current production situation.

[0139] The similarity comparison module 204 is used to compare the comprehensive feature vector with the preset reference scenario set to determine the deviation measure value of the current production scenario relative to the known scenario.

[0140] The quality prediction module 205 is used to determine the uncertainty parameters of the quality prediction result based on the deviation metric value, and to match the target prediction logic among multiple preset prediction logics based on the deviation metric value to output quality prediction information.

[0141] The data acquisition module 201 can be configured to receive the first type of process data manually entered by operators through a manual input interface. For example, operators periodically input parameters such as temperature, pressure, and flow rate from the production site into the system. Simultaneously, the second type of process data can be imported through the interface of offline analysis equipment, such as uploading near-infrared spectral data files obtained from laboratory analysis to the system.

[0142] As one implementation, the data acquisition module 201 can be configured to connect directly to sensors and instruments at the production site via standard industrial communication protocols to periodically collect the first type of process data. For the second type of process data, the data acquisition module can be configured to read raw spectral data from the control software of the online analyzer via file sharing or database connection.

[0143] The feature extraction module 202 can be configured to directly output the absorbance values ​​of all wavelength points of the second type of process data, such as the original spectral data, as a high-dimensional vector as a feature signal.

[0144] As one implementation, the feature extraction module 202 can be configured to retain only wavelength point data with absorbance values ​​higher than a specific threshold in the second type of process data as feature signals through simple threshold filtering.

[0145] The fusion processing module 203 can be configured to perform a simple linear concatenation of the first type of process data and feature signals to form a comprehensive feature vector. For example, all the parameter values ​​of the first type of process data can be concatenated with all the feature signal values ​​in a predetermined order.

[0146] As one implementation method, the fusion processing module 203 can be configured to normalize the first type of process data and feature signals, and then perform a simple arithmetic average to form a comprehensive feature vector.

[0147] The similarity comparison module 204 can be configured to calculate the Euclidean distance between the current integrated feature vector and all reference vectors in the reference context set, and directly use the smallest Euclidean distance as the deviation metric.

[0148] As one implementation, the similarity comparison module 204 can be configured to calculate the cosine similarity between the current integrated feature vector and all reference vectors in the reference context set, and use the difference between 1 and the cosine similarity as the deviation metric.

[0149] Finally, the quality prediction module 205 can be configured to directly map the deviation metric value to a preset linear function to calculate the uncertainty parameter. As one implementation, the quality prediction module 205 can be configured to select from two preset prediction logics based on the magnitude of the deviation metric value. For example, when the deviation metric value is less than a certain fixed threshold, a machine learning-based prediction model is selected; when the deviation metric value is greater than or equal to the threshold, a prediction logic based on simple statistical averaging is selected.

[0150] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent processing of PVC stabilizer production process data, characterized in that, The method includes the following steps: S1: Acquire the first type of process data reflecting the macroscopic state of the production process, and the second type of process data reflecting the microscopic structure of the production materials; S2: Extract features from the second type of process data to obtain feature signals characterizing the microscopic properties of the material; S3: The feature signal is fused with the first type of process data to construct a comprehensive feature vector of the current production situation; S4: Compare the comprehensive feature vector with the preset reference scenario set to determine the deviation metric value of the current production scenario relative to the known scenario; S5: Determine the uncertainty parameter of the quality prediction result based on the deviation metric value, and match the target prediction logic among a plurality of preset prediction logics based on the deviation metric value to output quality prediction information.

2. The intelligent data processing method for PVC stabilizer production process according to claim 1, characterized in that, Step S1 includes: S11: Collect the first type of process data through sensors and use an industrial controller to check the range of the first type of process data. When the first type of process data exceeds the preset safe operating range, trigger an alarm. S12: Perform spectral scanning on the production materials using an online near-infrared spectrometer to obtain the second type of process data.

3. The intelligent data processing method for PVC stabilizer production process according to claim 2, characterized in that, Step S12 includes: S121: Before scanning the production materials, obtain the reference spectrum of the built-in standard reference material of the online near-infrared spectrometer; S122: Compare the background spectrum under the current environment with the reference spectrum and calculate the spectral shift; S123: Determine whether the spectral drift is within a preset self-compensation range; S124: If the spectral drift is within the self-compensation range, a compensation operator is generated based on the spectral drift, and the original spectrum obtained by scanning is energy-corrected using the compensation operator to obtain the corrected second type of process data; S125: If the spectral drift exceeds the self-compensation range, the sensor window is determined to be contaminated, the current prediction logic is stopped, and a cleaning and maintenance command is output.

4. The intelligent data processing method for PVC stabilizer production process according to claim 1, characterized in that, Step S2 includes: S21: Perform smoothing and background interference elimination processing on the second type of process data; S22: Extract absorbance feature values ​​of a specific band from the processed second type of process data, and use the absorbance feature values ​​as the feature signal.

5. The intelligent data processing method for PVC stabilizer production process according to claim 4, characterized in that, Step S21 includes: S211: Apply the Savitzky-Golay smoothing algorithm to perform local polynomial fitting on the second type of process data to remove high-frequency random noise; S212: The non-feature region of the second type of process data is fitted using a polynomial fitting method to estimate the background signal, and the background signal is subtracted from the original spectrum to highlight the absorption characteristics of the spectral curve.

6. The intelligent data processing method for PVC stabilizer production process according to claim 5, characterized in that, Step S22 includes: S221: Calculate the integral area of ​​a specific wavelength band as the absorbance characteristic value; or, calculate the zero point position and intensity of the second derivative of the spectral curve at a specific wavelength as the absorbance characteristic value; or, calculate the relative height of a specific absorption peak relative to the peak height at a reference wavelength as the absorbance characteristic value, and use the absorbance characteristic value as the characteristic signal.

7. The intelligent data processing method for PVC stabilizer production process according to claim 1, characterized in that, Step S3 includes: S31: Align the first type of process data with the feature signal according to the timestamp; S32: Integrate the aligned first type of process data with the feature signal to form the comprehensive feature vector, so as to achieve synchronous description of the production status.

8. The intelligent data processing method for PVC stabilizer production process according to claim 1, characterized in that, Step S4 includes: S41: Calculate the Euclidean distance between the comprehensive feature vector and each reference vector in the reference context set; S42: Calculate the situation novelty score based on the Euclidean distance, and use the situation novelty score as the deviation metric; Step S42 includes: S421: Establish a context reference database, which stores the comprehensive feature vectors of historical batches and the corresponding quality results; S422: Calculate the Euclidean distance between the integrated feature vector and each reference vector in the context reference database, and select the K closest known contexts; S423: Calculate the average Euclidean distance of the K known scenarios, and determine the average Euclidean distance as the scenario novelty score.

9. The intelligent data processing method for PVC stabilizer production process according to claim 1, characterized in that, In step S5, matching the target prediction logic among a plurality of preset prediction logics based on the deviation metric value includes the following steps: S51: Determine whether the deviation measurement value exceeds a preset deviation threshold; S52: If the deviation metric does not exceed the preset deviation threshold, then the fine association rule is determined as the target prediction logic; S53: If the deviation metric exceeds the preset deviation threshold, the weight of the fine association rule is reduced, and the robust threshold rule based on physicochemical principles is determined as the target prediction logic.

10. A smart data processing system for PVC stabilizer production process, characterized in that, The system is used to implement the steps of the method according to any one of claims 1-9, the system comprising: The data acquisition module is used to acquire the first type of process data that reflects the macroscopic state of the production process, and the second type of process data that reflects the microscopic structure of the production materials. The feature extraction module is used to extract features from the second type of process data to obtain feature signals that characterize the microscopic properties of the material. The fusion processing module is used to fuse the feature signals with the first type of process data to construct a comprehensive feature vector of the current production scenario. The similarity comparison module is used to compare the comprehensive feature vector with a preset set of reference scenarios to determine the deviation metric of the current production scenario relative to the known scenarios. The quality prediction module is used to determine the uncertainty parameter of the quality prediction result based on the deviation metric value, and to match the target prediction logic among a number of preset prediction logics based on the deviation metric value, so as to output quality prediction information.

Images