A method for recycling and processing NMP refined three-tower overhead materials

Abstract

The present application relates to a kind of NMP refining three-tower overhead material's resource recycling processing method, to solve the problem of behavior inertia difficult to identify and accurate coupling control under the disturbance of complex process feed.The core technical scheme includes, based on distributed sensor array real-time acquisition multidimensional time series data, by feature extraction and dimensionality reduction clustering process semantic fingerprint knowledge base is built;Using hash fast matching algorithm realizes the semantic inference of historical behavior inertia and current dynamic working condition;With process coupling relationship topological remapping and weight distribution mechanism, drive key control path parameter issue and multi-process temperature compensation signal generation, and complete closed-loop control in single cycle.The method can significantly improve the anti-disturbance ability of complex process system, realize the collaborative control of low disturbance and high adaptability between multiple processes, help to improve product purity and system running stability.

Description

Technical Field

[0001] This invention relates to the field of automated collaborative control technology for chemical processes, and in particular to a method for the resource recovery and treatment of top materials from three NMP refining towers. Background Technology

[0002] For the automated control of a multi-stage resource recovery and processing system for top materials from a three-tower NMP refining plant, current mainstream process control modes mainly employ model predictive control, adaptive control, and multivariate collaborative regulation methods based on expert rules. Traditional solutions often rely on high-precision mechanistic modeling or deep learning time-series networks. By collecting system operating parameters, a state-space model of the system is constructed, and the control loop is dynamically corrected through parameter identification and controller retraining when process load disturbances or feed changes occur.

[0003] In recent years, to improve the accuracy and robustness of collaborative control of complex chemical processes across multiple stages, some studies have introduced reinforcement learning, attention-weighted strategies, and end-to-end neural network prediction models. These models map load fluctuations in nonlinear dynamic conditions to adaptive redefinitions of controller weights. Traditional control methods generally focus on parameter identification and strategy reconstruction. However, if the mechanistic model or training dataset does not match the current operating conditions, model mismatch may occur, leading to misjudgments, improper execution, and even system instability. While some intelligent methods can achieve partial adaptive adjustment, their reliance on the black-box nature of end-to-end neural networks results in poor transparency and interpretability of the control process, hindering process engineers from tracing faults or extending operating conditions. Furthermore, large-scale time-series data acquisition and complex model inference heavily rely on high-performance computing resources. Deploying these at edge devices can easily create computational bottlenecks and communication delays, further amplifying the overall system response lag.

[0004] Therefore, there is an urgent need to propose a new multi-process collaborative control strategy that can adaptively switch process collaboration logic within milliseconds, automatically match adaptive process behavior characteristics, and greatly improve control real-time performance, so as to provide a solid technical foundation for large-scale industrial applications. Summary of the Invention

[0005] This application provides a method for the resource recovery and treatment of the top material of the NMP refining three towers, aiming to solve the problems mentioned in the background art, such as poor transparency and interpretability of the control process, which are not conducive to process personnel's fault tracing or external expansion of operating conditions.

[0006] This application provides a method for the resource recovery and treatment of the top material of an NMP refining three-tower system, specifically including: S1: For the four processes of NMP refining three-tower aeration, neutralization, aging crystallization and solid-liquid separation, multi-dimensional time-series process data are collected in the continuous operation cycle to generate the original operating condition response sequence. S2: Based on the original working condition response sequence, perform fractal dimension analysis and morphological similarity measurement, and extract the curvature integral ratio, phase lag index and response hysteresis entropy value as morphological invariants to construct the process semantic fingerprint. S3: Classify and store the semantic fingerprints of the process steps to establish a fingerprint knowledge base; S4: Real-time acquisition of current feed spectrum preprocessing data and initial response features of the first process, and retrieval of historical fingerprint combinations with the smallest semantic distance in the fingerprint knowledge base to generate current dynamic working condition matching results; S5: Based on the current dynamic working condition matching result, start the dynamic topology remapping engine to identify key process pairs where the coupling dominance has shifted, so as to generate inter-process coupling relationship topology remapping instructions; S6: According to the inter-process coupling relationship topology remapping instruction, freeze non-critical control paths and decouple the parameter transfer coefficients on the reweighted new activated paths to generate a critical control path weight allocation scheme. S7: Generate multi-process collaborative control execution instructions based on the critical control path weight allocation scheme; S8: Execute the multi-process collaborative control execution instructions to adjust the operating parameters of each process, and complete the control update within a single sampling period to realize the resource recycling and processing of materials.

[0007] The resource recovery and treatment method for the top material of the NMP refining three-tower column provided in this application has the following beneficial effects: To address the issues of response lag and model mismatch-induced performance degradation in traditional industrial process control under dynamic conditions, this solution constructs a lightweight knowledge base based on process semantic fingerprints. This enables efficient identification and interpretable representation of behavioral patterns in complex multi-process systems. Compared to existing model predictive control or deep time-series modeling methods, this solution does not rely on online parameter identification or high-computing-power neural network inference. Instead, it performs unsupervised clustering and fractal dimension analysis on the multidimensional time-series characteristics of key processes such as aeration, neutralization, aging and crystallization, and solid-liquid separation under different loads and impurity spectra. This extracts morphological invariants with process discrimination significance as "process semantic fingerprints," such as curvature integral ratio, phase lag exponent, and response hysteresis entropy, forming a physically interpretable operating condition identification system. This fingerprint retains only the normalized semantic labels and mapping relationships, significantly reducing data storage and transmission overhead. The entire identification process can be completed in milliseconds at the edge gateway, effectively overcoming the control latency problem caused by long model training cycles and poor adaptability in traditional methods, and greatly improving the system's real-time response capability under sudden disturbances. Attached Figure Description

[0008] Figure 1This is the main flow chart of a method for the resource recovery and treatment of the top material of a three-tower NMP refining system.

[0009] Figure 2 This is a sub-flowchart of a method for the resource recovery and treatment of the top material of a three-tower NMP refining system.

[0010] Figure 3 This is another sub-flowchart of a method for the resource recovery and treatment of the top material of a three-tower NMP refining process. Detailed Implementation

[0011] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0012] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0013] like Figure 1 As shown, this application provides a method for the resource recovery and treatment of the top material of an NMP refining tower, specifically including: S1: For the four processes of NMP refining three-tower aeration, neutralization, aging crystallization and solid-liquid separation, multi-dimensional time-series process data are collected in the continuous operation cycle to generate the original operating condition response sequence. S2: Based on the original working condition response sequence, perform fractal dimension analysis and morphological similarity measurement, and extract the curvature integral ratio, phase lag index and response hysteresis entropy value as morphological invariants to construct the process semantic fingerprint. S3: Classify and store the semantic fingerprints of the process steps to establish a fingerprint knowledge base; S4: Real-time acquisition of current feed spectrum preprocessing data and initial response features of the first process, and retrieval of historical fingerprint combinations with the smallest semantic distance in the fingerprint knowledge base to generate current dynamic working condition matching results; S5: Based on the current dynamic working condition matching result, start the dynamic topology remapping engine to identify key process pairs where the coupling dominance has shifted, so as to generate inter-process coupling relationship topology remapping instructions; S6: According to the inter-process coupling relationship topology remapping instruction, freeze non-critical control paths and decouple the parameter transfer coefficients on the reweighted new activated paths to generate a critical control path weight allocation scheme. S7: Generate multi-process collaborative control execution instructions based on the critical control path weight allocation scheme; S8: Execute the multi-process collaborative control execution instructions to adjust the operating parameters of each process, and complete the control update within a single sampling period to realize the resource recycling and processing of materials.

[0014] Step S1: For the four processes of NMP refining—aeration, neutralization, aging and crystallization, and solid-liquid separation—multi-dimensional time-series process data are collected within a continuous operating cycle to generate the original operating condition response sequence. Specifically, this includes: S1.1: Acquire the raw analog voltage signal from the distributed sensor array of the NMP refining three-tower process. Use a high-speed analog-to-digital converter to synchronously sample and process the pH electrode potential, redox electrode millivolt value, turbidity light scattering intensity, and centrifuge motor current waveform to generate a multi-channel raw digital signal stream with a unified timestamp.

[0015] A distributed sensor array deployed throughout the entire NMP refining process of three towers acquires raw analog signals in real time, including pH electrode potential in the aeration section, millivolt values ​​of the redox electrode in the neutralization section, turbidity light scattering intensity in the crystallization section, and motor current waveform in the centrifugation section. A high-speed analog-to-digital converter (ADC) is used to synchronously sample these multi-channel analog signals, setting the sampling frequency to 1 kHz to ensure the capture of transient fluctuations, discretizing the continuous voltage signals into a high-precision digital sequence. A unified timestamp is applied to the discretized digital signals, and a global clock synchronization protocol is used to eliminate data transmission delay differences between sensors, constructing a time-aligned multi-channel raw digital signal stream. Hardware filtering circuits suppress high-frequency electromagnetic interference noise, retaining effective frequency band signals reflecting process state changes, ensuring that the signal-to-noise ratio for subsequent feature extraction meets analytical requirements. Through synchronous sampling and clock alignment, the dispersed analog signals from the previous step are transformed into a multi-channel raw digital signal stream with a unified time base, achieving spatiotemporal consistency of multi-source heterogeneous data and providing a high-quality input foundation for subsequent sliding window differential methods.

[0016] S1.2: Based on the multi-channel raw digital signal stream, the sliding window differentiation method is used to perform real-time slope calculation and transition amplitude extraction on the pH value sequence and redox potential value sequence to generate pH change slope sequence and redox potential transition amplitude sequence.

[0017] Receive the multi-channel raw digital signal stream with a unified timestamp generated by S1.1, and extract the pH electrode potential sequence and the redox electrode millivolt value sequence as the objects to be processed.

[0018] The sliding window differential method was applied to the extracted pH numerical sequence. The window length was set to N sampling points, and the difference between adjacent sampling points within the window was calculated to obtain the instantaneous rate of change.

[0019] The central difference formula is used to eliminate the phase lag caused by unidirectional difference. The calculated slope sequence is then subjected to low-pass filtering to remove high-frequency noise interference, generating a smooth pH change slope sequence to characterize the real-time rate of acid-base neutralization reaction.

[0020] Simultaneously, the transition amplitude of the redox potential numerical sequence is extracted, local extreme points on the potential curve are identified, and the absolute value of the potential difference between adjacent extreme points is calculated.

[0021] The absolute value of the potential difference is compared with a preset dynamic threshold to filter out significant transition events that exceed the threshold, and the time and magnitude of their occurrence are recorded.

[0022] The selected transition events are aggregated within a time window, and the cumulative total number of transitions per unit time is calculated to generate a redox potential transition amplitude sequence, which reflects the degree of abrupt change in the redox state of the system.

[0023] By using the above-mentioned sliding window differentiation and transition feature extraction processing method, the original voltage signal from the previous step is transformed into a pH change slope sequence and a redox potential transition amplitude sequence that characterize the reaction kinetics. This achieves a quantitative mapping from physical signals to process state features, providing high signal-to-noise ratio input data for the construction of semantic fingerprints for subsequent processes.

[0024] S1.3: Using the turbidity light scattering intensity sequence, an adaptive threshold inflection point detection method is used to identify the curvature abrupt change point in the crystallization nucleation stage. Combined with the centrifuge motor current waveform, transient peak capture technology is applied to extract the load impact characteristics in the solid-liquid separation process, so as to generate turbidity rise inflection point time data and centrifugal current transient peak data.

[0025] A sliding window smoothing filter is applied to the turbidity light scattering intensity sequence to eliminate high-frequency sensor noise interference and generate a denoised turbidity time series curve.

[0026] The second derivative sequence was calculated for the denoised turbidity time series curve, and the rate of curvature change was quantified to locate the morphological abrupt change region in the crystallization nucleation stage.

[0027] An adaptive threshold method is used to dynamically set the curvature judgment benchmark. This benchmark is automatically adjusted based on the statistical variance of the pH stable zone in the preceding neutralization process to eliminate misjudgments caused by background fluctuations.

[0028] Traverse the second derivative sequence, identify the first positive extreme point that exceeds the adaptive threshold, and mark its corresponding timestamp as the inflection point time data of turbidity increase, characterizing the moment of crystal nucleus explosion generation.

[0029] The current waveform of the centrifuge motor is acquired synchronously, and the instantaneous envelope sequence of the current amplitude is obtained by applying the envelope extraction method, which reflects the dynamic mechanical load during the solid-liquid separation process.

[0030] Peak search was performed on the instantaneous current envelope sequence to lock the maximum impact current value during the separation start-up phase, which was used as the transient peak data of the centrifugal current to characterize the resistance characteristics in the early stage of filter cake formation.

[0031] Through the above processing method, the original signal from the previous step is transformed into turbidity rise inflection point time data marking the start of crystallization and centrifugal current transient peak data characterizing separation load fluctuation, thereby achieving accurate quantification of key phase transition and load characteristics.

[0032] S1.4: Based on the pH change slope sequence, redox potential transition amplitude sequence, turbidity rise inflection point time data, and centrifugal current transient peak data, a multidimensional feature vector space is constructed. Data fusion processing is performed through time alignment and missing value interpolation mechanism to generate multidimensional time-series process data covering the entire cycle of aeration, neutralization, aging crystallization, and solid-liquid separation.

[0033] The system receives pH change slope sequences, redox potential transition amplitude sequences, turbidity rise inflection point times, and transient peak values ​​of centrifugal current to construct a four-dimensional feature vector space. Timestamp alignment is performed on each sequence, and linear interpolation is used to fill in data gaps caused by sensor response delays, ensuring strict synchronization of data across all dimensions on the time axis, based on a unified sampling frequency. Z-score standardization is applied to the aligned multidimensional data to eliminate numerical differences between different physical dimensions, generating a standard feature matrix with zero mean and one variance. Based on the standard feature matrix, the time lag cross-correlation coefficient between adjacent processes is calculated to quantify the dynamic coupling strength between processes such as aeration to neutralization and neutralization to crystallization. Principal component analysis is used to reduce the dimensionality of the high-dimensional feature matrix, extracting principal component vectors with a cumulative contribution rate exceeding a preset threshold and removing redundant noise information. The principal component vectors are then fused with the original time-series data to generate a standardized multidimensional time-series process dataset covering the entire process cycle. The above processing method transforms the discrete features from the previous step into a structured multidimensional time-series dataset, achieving spatiotemporal consistency fusion of multi-source heterogeneous data and providing high-quality input for semantic fingerprint extraction in subsequent processes.

[0034] S1.5: Based on the multi-dimensional time-series process data, segment and encapsulate the data according to the process type label and label the metadata. Integrate the feature parameters within the continuous operation cycle into structured data blocks to generate the original operating condition response sequence.

[0035] Step S2: Based on the original operating condition response sequence, fractal dimension analysis and morphological similarity measurement are performed to extract the curvature integral ratio, phase hysteresis exponent, and response hysteresis entropy value as morphological invariants, in order to construct a process semantic fingerprint. Specifically, this includes: S2.1: Perform sliding window segmentation and detrending processing on the original operating condition response sequence to generate a standardized time series fragment set, eliminate the interference of different operating cycle lengths on subsequent feature extraction, and ensure the time alignment and amplitude consistency of the input data.

[0036] The system receives the raw operating condition response sequence generated by S1.5, which contains multi-dimensional time-series data such as pH change slope and redox potential transition amplitude. A sliding window segmentation process is performed on the raw operating condition response sequence, setting a fixed time width to cut the continuous long sequence into several short time-series segments with overlapping regions. The overlap ratio between segments is adjusted through a window step size control mechanism to ensure the continuity of dynamic features at the boundaries, generating a preliminary set of time-series segments.

[0037] For each time segment, detrending processing is performed, using a polynomial fitting method to extract low-frequency trend components from the data. The residual sequence between the original signal and the fitted trend line is calculated to eliminate baseline offsets caused by equipment drift or long-term load variations. The residual sequence is then zero-mean processed by subtracting the arithmetic mean within the segment to eliminate DC component interference.

[0038] Amplitude normalization is performed on the zero-mean residual sequence, and the standard deviation of the data within each segment is calculated. Each data point is divided by the standard deviation to eliminate amplitude inconsistencies caused by dimensional differences between different processes and fluctuations in sensor sensitivity. This generates a standardized time series segment set with uniform time alignment and amplitude consistency.

[0039] By using sliding window segmentation, detrending, and normalization, the original operating condition response sequence from the previous step is transformed into a standardized time-series fragment set of data, achieving the expected technical effect of eliminating differences in operating cycle length and baseline drift interference.

[0040] For example, a sliding window segmentation is performed on the pH change slope sequence, with a window width of 60 seconds, a step size of 30 seconds, and an overlap rate of 50%. A second-order polynomial is used to extract the trend term, and the residuals are calculated and subtracted from the mean. The residual sequence is then normalized by dividing by the standard deviation using Z-score. If the standard deviation of a segment is less than the threshold of 0.01, it is identified as steady-state noise and marked for removal. Finally, a standardized set of time series segments with uniform length, a mean of 0, and a variance of 1 is generated, significantly improving the convergence speed and accuracy of subsequent fractal dimension calculations.

[0041] S2.2: Based on the standardized time-series fragment set, the box counting method is used to calculate the fractal dimension sequence to quantify the complexity and self-similarity characteristics of the response curves of each process, and generate fractal dimension feature vectors as the first-level discrimination criterion for distinguishing process behavior patterns under different impurity spectrum compositions.

[0042] Spatial gridding is performed on the standardized time series fragment set, and a two-dimensional phase plane coordinate system is constructed based on the amplitude range and time span of the response curve. The initial grid side length is set to one-tenth of the maximum amplitude of the response data, and the phase plane is uniformly divided into several square sub-regions to form a basic grid matrix covering the entire operating condition response trajectory.

[0043] The number of data points falling into each grid sub-region is counted, and non-empty grids are marked as valid covering units. The total number of valid covering units is recorded as the box count value at the current scale. By iteratively reducing the grid side length to half of the original side length, the grid division and counting operations are repeated to generate a series of box count value sequences at different scales, establishing a mapping relationship between scale and coverage number.

[0044] Linear regression fitting is performed on the logarithm of the scale and the logarithm of the box count using the least squares method. The absolute value of the slope of the regression line is calculated, and this slope is the fractal dimension. The fractal dimension is calculated using the following formula. : in, For grid coverage, This represents the grid side length.

[0045] The standardized time segments of all processes are traversed, and the fractal dimension value corresponding to each segment is calculated. The fractal dimension values ​​are then combined into an eigenvector. This eigenvector quantifies the filling density and self-similarity complexity of the response curve in phase space, reflecting the differences in kinetic behavior caused by changes in the impurity spectrum.

[0046] By calculating the fractal dimension sequence using the box counting method, the standardized time series fragments from the previous step are transformed into fractal dimension feature vectors that characterize the complexity of process behavior. This enables the first-level quantitative discrimination of process behavior patterns under different impurity spectrum compositions, providing a highly discriminative geometric feature basis for subsequent clustering.

[0047] S2.3: The fractal dimension feature vector is used to guide the unsupervised temporal clustering algorithm to perform cluster center iterative optimization, so as to group the working condition samples with similar dynamic characteristics into the same category and generate a working condition cluster label set, which lays the grouping foundation for the subsequent extraction of representative morphological invariants of each category.

[0048] The system receives a set of standardized time series segments that have undergone detrending processing and maps them to a high-dimensional feature space to construct an initial cluster center matrix. Based on the Euclidean distance metric of the fractal dimension feature vectors, it calculates the similarity weight between each time series segment and the current cluster center, generating a preliminary probability distribution of the working condition affiliation.

[0049] Based on the probability distribution of the operating conditions, the K-Means++ initialization strategy is used to optimize the cluster center position. The cluster center coordinates are updated by iteratively minimizing the intra-cluster sum of squares error function to ensure that operating condition samples with similar dynamic characteristics are closely clustered in the feature space.

[0050] The silhouette coefficient is introduced as a clustering effectiveness evaluation index to quantitatively evaluate the intra-cluster compactness and inter-cluster separation after each iteration. When the silhouette coefficient converges to a preset threshold or reaches the maximum number of iterations, the cluster center iterative optimization process is terminated.

[0051] The set of operating condition samples corresponding to the converged cluster centers are labeled as the same category and assigned a unique operating condition cluster label, forming a set of operating condition cluster labels covering different impurity spectrum compositions and load fluctuation patterns.

[0052] By performing cluster center iterative optimization through unsupervised temporal clustering algorithm, the standardized temporal segments of the previous step are transformed into a set of working condition clustering labels with clear process semantics, realizing automatic grouping of behavioral patterns under complex working conditions, and laying the grouping foundation for subsequent extraction of representative morphological invariants of each category.

[0053] S2.4: For each cluster sample in the working condition clustering label set, calculate the curve curvature integral ratio, phase lag exponent and response hysteresis entropy value respectively, so as to extract the set of morphological invariants that are not affected by time axis scaling and amplitude translation, and form candidate fingerprint data that can uniquely identify the behavioral inertia under the working condition.

[0054] For each cluster of samples in the working condition clustering label set, standardized time-series segments are extracted as computational objects. Second-order differentiation is performed on these time-series segments to obtain a curvature sequence characterizing the degree of curve bending. The trapezoidal integral rule is used to sum the absolute curvature values ​​to calculate the total curvature integral. Dividing the total curvature integral by the time span of the time-series segment generates a normalized curvature integral ratio index, eliminating the influence of time axis scaling on morphological description.

[0055] A phase lag exponent calculation model was constructed, selecting the redox potential transition sequence as the reference signal and the pH change slope sequence as the response signal. Cross-correlation function calculations were performed on the two signals, and the time delay offset corresponding to the maximum cross-correlation coefficient was searched. The ratio of this time delay offset to the main period length of the signal was defined as the phase lag exponent, quantifying the synchronicity differences in reaction kinetics between different processes.

[0056] Information entropy theory is introduced to evaluate the hysteresis characteristics of the system response, and probability density estimation is performed on the detrended residual sequence. The residual interval is divided into several equal-width bins, and the sample frequencies within each interval are statistically analyzed and normalized to a probability distribution.

[0057] The curvature integral ratio, phase hysteresis exponent, and response hysteresis entropy are combined into a three-dimensional morphological invariant vector. Euclidean distance is applied to this vector to ensure its numerical value remains constant under time translation and amplitude scaling transformations. Through this processing, the clustered samples from the previous step are transformed into candidate fingerprint data with unique identifiers, achieving a dimensionality reduction mapping from dynamic time-series waveforms to static morphological features, providing a robust feature basis for subsequently building a fingerprint knowledge base.

[0058] For example, a standardized pH response segment of 60 seconds was selected for the "high organic acid entrainment" cluster in the neutralization process. The calculated curvature integral was 12.5, the time span was 60 seconds, and the curvature integral ratio was 0.208. The maximum delay of the cross-correlation between the redox potential and the pH signal was 2.5 seconds, the main period was 20 seconds, and the phase lag exponent was 0.125. The response hysteresis entropy value calculated from the residual probability distribution was 1.85 bits. The final generated candidate fingerprint data was [0.208, 0.125, 1.85]. This triplet maintained a numerical deviation of less than 0.01 when the feed fluctuation range changed by ±30%, significantly improving the stability of feature extraction.

[0059] S2.5: Normalize and encode the candidate fingerprint data and perform semantic association mapping to generate process semantic fingerprints containing curvature integral ratio, phase lag exponent and response hysteresis entropy value, thus completing the final transformation from raw multidimensional time series data to interpretable and searchable behavioral inertia representation.

[0060] like Figure 2 As shown, step S3 involves classifying and storing the semantic fingerprints of the work processes to establish a fingerprint knowledge base. Specifically, this includes: S3.1: Perform standardized encoding processing on the morphological invariant data in the semantic fingerprint of the process to generate fingerprint feature codes with unique identification attributes, ensuring that the curvature integral ratio, phase hysteresis exponent and response hysteresis entropy value under different working conditions are converted into digital signatures in a unified format.

[0061] S3.2: Perform an association mapping operation based on the fingerprint feature code and the predefined typical disturbance source label to generate a composite semantic label containing process type identifier and disturbance scenario description, thereby realizing the logical transformation of abstract mathematical features into interpretable process states.

[0062] The system receives a fingerprint feature code with unique identification attributes generated by S3.1. This feature code is formed by normalizing and encoding the curvature integral ratio, phase hysteresis exponent, and response hysteresis entropy value. A predefined typical disturbance source tag library is constructed, covering common abnormal operating conditions in the NMP refining three-tower system, such as high chloride ion feed, low-boiling-point solvent entrainment, and sudden changes in organic acid concentration. The fingerprint feature code is logically associated with the typical disturbance source tags, establishing a semantic bridge between mathematical feature vectors and physical process states. A multi-dimensional spatial distance metric is used to calculate the Euclidean distance between the current fingerprint feature code and historical standard disturbance templates in the feature space, determining the most matching disturbance source type. Based on the matching result, the corresponding process type identifier is extracted, clarifying the specific process step (aeration, neutralization, crystallization, or separation) to which the fingerprint belongs. The process type identifier and the determined disturbance source descriptor are concatenated and structurally encapsulated to generate a composite semantic tag. This composite semantic tag includes a process ID, disturbance category code, and confidence score, realizing the transformation from abstract numerical values ​​to interpretable process states. Through the above association mapping operation, the fingerprint feature code of the previous step is transformed into a composite semantic label containing process type identifier and disturbance scenario description, realizing the logical transformation of abstract mathematical features into interpretable process state, and providing standardized input for subsequent hash index construction.

[0063] S3.3: Utilize a hash algorithm to perform index key-value calculation on the composite semantic tag to generate a lightweight retrieval index table, completely remove the original time-series waveform data and retain only the pointer links pointing to the fingerprint feature code, thereby completing the dimensionality reduction optimization of the knowledge base storage structure.

[0064] Receive the composite semantic tag generated by S3.2, which contains process type identifier and disturbance scenario description, as the input data source for building a lightweight retrieval index.

[0065] Perform UTF-8 encoding conversion on the text description field in the composite semantic tag to convert the unstructured process state description into a fixed-length byte sequence, eliminating the risk of hash collisions caused by character set differences.

[0066] The SHA-256 hash algorithm is used to perform a one-way hash operation on the encoded byte sequence to generate a 256-bit binary digest information, ensuring that different working condition labels are mapped to unique and evenly distributed hash values.

[0067] The first 16 bits of the hash digest are used as the high-order index key to establish a fast location bucket in memory, and the remaining bits are used as the low-order check code to resolve secondary confirmation when hash collisions occur.

[0068] A mapping table between hash key values ​​and fingerprint feature code pointers is established, and the original high-dimensional time-series waveform data is replaced with pointer addresses that occupy only a few bytes, thereby achieving exponential compression of storage space.

[0069] The mapping table is optimized with a B+ tree index, and the data is sorted and stored in lexicographical order according to the hash key values. Range queries and prefix matching are supported, improving retrieval efficiency in concurrent scenarios with multiple disturbance sources.

[0070] Through the hash index construction and pointer mapping mechanism described above, the composite semantic tags from the previous step are transformed into a lightweight index structure with millisecond-level retrieval capabilities, thereby achieving dimensionality reduction optimization of the knowledge base storage structure and query acceleration.

[0071] S3.4: Perform classification storage operations based on the lightweight retrieval index table to construct a fingerprint knowledge base organized by process type and typical disturbance source, and establish a static data foundation that supports fast hash matching.

[0072] It receives a lightweight retrieval index table and the corresponding process semantic fingerprint feature code as the input data source for building the knowledge base.

[0073] Based on the process flow logic of the NMP refining three towers, the fingerprint data is divided into four independent data partitions: aeration, neutralization, aging and crystallization, and solid-liquid separation. Each partition corresponds to a unique process type identifier.

[0074] Within each process zone, a secondary sub-directory structure is further established based on typical disturbance source labels. These typical disturbance sources include specific operating condition categories such as high chloride ion feed, low boiling point solvent entrainment, and organic acid concentration fluctuations.

[0075] A key-value pair storage architecture is adopted, using the hash key value generated in the lightweight retrieval index table as the primary key, and the process semantic fingerprint feature code containing the curvature integral ratio, phase lag exponent and response hysteresis entropy value as the value object for persistent writing.

[0076] During the writing process, data integrity checks are performed to ensure that each hash key is uniquely mapped to a set of definite form invariants, thus avoiding data redundancy and conflicts.

[0077] A dual-dimensional index matrix is ​​constructed, with the first dimension being the process type vector and the second dimension being the disturbance source category vector. Fingerprint data can be quickly located by using matrix coordinate positioning.

[0078] The storage structure is statically fixed to lock the topological organization of the knowledge base and prevent dynamic modification of the directory hierarchy at runtime, so as to ensure the determinism of the retrieval path and the stability of latency.

[0079] By using the above-mentioned classification storage and dual-dimensional index construction method, the discrete fingerprint data generated in the previous step is transformed into a structured fingerprint knowledge base that can be quickly retrieved, realizing a static data foundation that supports millisecond-level hash matching, and significantly reducing the retrieval computation load under dynamic operating conditions.

[0080] S3.5: Perform integrity verification and version solidification processing based on the fingerprint knowledge base to generate a final usable process behavior inertial mapping relationship table, ensuring that the subsequent dynamic topology remapping engine can directly call the standardized semantic matching results.

[0081] Obtain the index table and corresponding fingerprint feature code data block stored in the fingerprint knowledge base according to process type and typical disturbance source classification, and use them as input objects for integrity verification.

[0082] Traverse each entry in the index table, extract its associated fingerprint feature code, composite semantic label, and the pointer to the storage address, and construct the dataset to be verified.

[0083] Perform Cyclic Redundancy Check (CRC) calculation on the extracted fingerprint feature code to generate the checksum value of the current data block, and compare and verify it with the standard checksum preset during storage.

[0084] If the checksum does not match, the entry is marked as corrupted data and an isolation mechanism is triggered to prevent erroneous fingerprints from being used by subsequent search engines, thus ensuring the reliability of the data source.

[0085] For data entries that pass the verification, check whether the process type identifier and disturbance scenario description in their composite semantic tags conform to the predefined dictionary specifications, and remove illegal or incomplete tag records.

[0086] Based on the verified and labeled fingerprint data, a memory mirror structure of the process behavior inertia mapping relationship table is constructed, and a bidirectional index link is established between the hash key value and the topology remapping parameter.

[0087] Version solidification is performed on the memory image structure to generate a metadata header file containing timestamps, version numbers, and data integrity information, and the current knowledge base state is locked to prevent concurrent write conflicts.

[0088] The solidified mapping table is persistently written to non-volatile memory, and the system global configuration pointer is updated to point to the latest available standardized semantic matching result set.

[0089] By using integrity verification and version solidification processing, the discrete fingerprint data from the previous step is transformed into a highly reliable and instantly callable process behavior inertial mapping table, achieving the expected technical effect of millisecond-level error-free access to historical working condition data by the dynamic topology remapping engine.

[0090] For example, a lightweight knowledge base containing 5000 fingerprint records is validated by calculating the checksum of each fingerprint feature using the CRC-32 method. Three records with mismatched checksums are automatically marked as invalid and isolated. For the remaining 4997 valid records, their tags are verified to conform to preset specifications such as 'aeration-high chloride ions'. A memory mapping table is constructed, linking the SHA-256 hash index with the corresponding neutralization-crystallization coupling weight coefficients. A metadata header with version number V20231027_01 is generated and stored on an industrial-grade SSD. After a system restart, this mapping table is directly loaded, and the retrieval latency remains stable within 2ms, ensuring the real-time performance and accuracy of control command generation.

[0091] like Figure 3 As shown, step S4 involves: real-time acquisition of the current feed spectrum preprocessing data and the initial response features of the first process, and retrieving the historical fingerprint combination with the smallest semantic distance from the fingerprint knowledge base to generate the current dynamic operating condition matching result. Specifically, this includes: S4.1: Baseline correction and scattering elimination processing are performed on the raw near-infrared spectrum signal of the NMP tower top material feed acquired in real time to generate a standardized feed spectrum preprocessing data sequence.

[0092] S4.2: Based on the feed spectral preprocessing data sequence, the instantaneous values ​​of the pH change slope and the instantaneous values ​​of the redox potential transition amplitude at the outlet of the aeration process are extracted simultaneously to construct the initial response feature vector of the first process that characterizes the current disturbance source.

[0093] Receive the standardized feed spectral preprocessing data sequence generated by S4.1 as the initial input condition to characterize the fluctuation of the current feed composition.

[0094] The online pH sensor and redox potential (ORP) sensor deployed at the outlet of the aeration process are invoked simultaneously to obtain real-time electrochemical signal streams that are strictly aligned with the timestamps of the spectral data.

[0095] A first-order forward difference operation is performed on the real-time acquired pH value sequence to calculate the ratio of the potential difference between adjacent sampling points to the time interval, so as to quantify the instantaneous rate change of the acid-base neutralization reaction.

[0096] The calculated pH change rate is noise-suppressed by using a sliding window filtering method. The absolute value of the maximum slope within the window is extracted to generate the instantaneous value of the pH change slope that characterizes the current acid-base impact intensity of the feed.

[0097] The second derivative detection is performed on the synchronously acquired ORP millivolt value sequence to identify the inflection point position on the potential curve and calculate the potential transition amplitude within a preset time window before and after the inflection point.

[0098] The extracted instantaneous values ​​of pH change slope and ORP transition amplitude are vector-concatenated to construct a two-dimensional initial response feature vector for the first process.

[0099] Through the above processing method, the spectral data and real-time electrochemical response from the previous step are transformed into structured feature vectors that characterize the current disturbance source properties, realizing rapid digital characterization of feed fluctuation properties and providing a highly discriminative input basis for subsequent hash matching.

[0100] For example, the sampling period is set to 100ms, and the sliding window length is 50 points. When a sudden increase in the organic acid content in the feed is detected, the pH electrode potential drops by 0.5 units within 2 seconds, and the instantaneous pH change slope is calculated to be -0.25pH / s. Simultaneously, the ORP potential exhibits a sharp fluctuation from +150mV to +80mV, with a calculated transition amplitude of 70mV. The system combines [-0.25, 70] into an initial response feature vector. This vector accurately reflects the strong acid impact and enhanced reducing properties under high organic acid entrainment conditions, significantly improving the sensitivity and response speed for identifying complex impurity profiles compared to traditional single-parameter monitoring.

[0101] S4.3: Use the locality-sensitive hashing algorithm to perform high-dimensional space mapping encoding on the initial response feature vector of the first process to generate a current working condition binary hash fingerprint code with fast retrieval capability.

[0102] Receive the initial response feature vector of the first process generated by S4.2. This vector contains the instantaneous value of the pH change slope and the instantaneous value of the redox potential transition amplitude, which serves as the input data source for locality-sensitive hashing.

[0103] The initial response feature vector is subjected to random hyperplane projection mapping to construct a set of random projection matrices that follow a standard normal distribution, thus mapping the high-dimensional continuous feature space to the low-dimensional binary Hamming space.

[0104] Calculate the dot product of the eigenvector and the random projection matrix, and determine the binary state of each bit of the hash value using the following formula: in, For the i-th bit hash value, This is the normalized initial response feature vector. Let i be the i-th random projection vector. This indicates the transpose operation.

[0105] Traverse all projection dimensions and concatenate the calculated binary bit sequences in order from high bit to low bit to generate a fixed-length binary hash fingerprint.

[0106] The generated binary hash fingerprint code is subjected to a check bit appending process. The check code is calculated using the cyclic redundancy check method and appended to the end of the fingerprint code to ensure the integrity of data transmission.

[0107] By using Local Sensitive Hash Mapping (LSH) to process the high-dimensional continuous feature vector from the previous step, the binary hash fingerprint of the current working condition is transformed into a binary hash fingerprint with fast retrieval capability, achieving the expected technical effect of millisecond-level working condition similarity matching.

[0108] S4.4: Based on the current working condition binary hash fingerprint code, perform Hamming distance calculation and similarity sorting operations in the fingerprint knowledge base to filter out the three candidate historical fingerprint combination indexes with the smallest semantic distance.

[0109] The system receives the current working condition binary hash fingerprint code generated by S4.3 and loads it as the query key into the fingerprint knowledge base's index retrieval engine. It then calls the Hamming distance calculation module to perform a bitwise XOR operation between the current binary hash fingerprint code and the feature codes of all historical process semantic fingerprints pre-stored in the knowledge base. The number of bits with a value of 1 in the XOR result is counted; this value represents the Hamming distance between the current working condition and each historical working condition, characterizing the degree of semantic difference in behavioral inertia. Finally, a similarity scoring function is constructed, generating a matching confidence score for each historical fingerprint based on the inverse Hamming distance mapping.

[0110] The above calculations are performed on all entries in the knowledge base to generate a full matching list containing historical fingerprint IDs and their corresponding confidence scores. The full matching list is then sorted in descending order based on the confidence scores to establish a priority sequence with the highest semantic similarity. The first three entries of the sorted list are extracted, and their corresponding historical fingerprint combination indices are obtained. Through a fast Hamming distance measurement and sorting mechanism, the similarity search in the high-dimensional feature space is transformed into low-computational-consumption binary bit operations, achieving millisecond-level accuracy in historical condition localization, providing a highly reliable reference benchmark for subsequent topology remapping.

[0111] For example, the current operating condition binary hash fingerprint is 10110011. The knowledge base contains three historical fingerprint records with feature codes 10110010, 11010011, and 10010011, respectively. The first record's XOR result is 00000001, with a Hamming distance of 1 and a confidence score of 0.5; the second record's XOR result is 01100000, with a Hamming distance of 2 and a confidence score of 0.33; and the third record's XOR result is 00100000, with a Hamming distance of 1 and a confidence score of 0.5. After sorting in descending order, the first two records with a confidence score of 0.5 and one record with a confidence score of 0.33 are selected as candidate historical fingerprint combination indexes. This process takes less than 1 millisecond, significantly improving dynamic response speed and ensuring the real-time performance and accuracy of control strategy switching.

[0112] S4.5: Based on the candidate historical fingerprint combination index, call the corresponding process type label and typical disturbance source mapping relationship table for fusion parsing to generate the current dynamic working condition matching result containing the exact working condition category and coupling transfer trend.

[0113] The system receives candidate historical fingerprint combination indexes and extracts corresponding process type identifiers and typical disturbance source semantic labels from the fingerprint knowledge base. A multi-dimensional mapping parsing matrix is ​​constructed, and the extracted process type identifiers are key-value matched with a predefined process coupling rule base. The baseline coupling strength of each process pair under a specific disturbance source is retrieved to form an initial coupling state vector. The inverse of the semantic distance of each candidate fingerprint in the current dynamic operating condition matching result is calculated and used as the confidence weight coefficient. A weighted fusion operation is performed on the initial coupling state vector to quantify the contribution of different candidate operating conditions to the current system behavior. Key process pairs where coupling dominance shifts are identified, and the logical threshold for control path switching is determined. A structured data object containing precise operating condition category labels and coupling transfer trend direction vectors is generated. Through the above fusion and parsing processing, the discrete fingerprint index is transformed into a current dynamic operating condition matching result with clear physical meaning, achieving a precise transformation from data retrieval to control decision-making basis.

[0114] Step S5: Based on the current dynamic operating condition matching result, activate the dynamic topology remapping engine to identify key process pairs where coupling dominance has shifted, and generate inter-process coupling relationship topology remapping instructions. Specifically, this includes: S5.1: Obtain the typical disturbance source labels and process type identifiers from the current dynamic working condition matching results, and perform association retrieval processing using the predefined process coupling rule base to generate an initial coupling relationship matrix containing all potential process pairs and their basic coupling strength values.

[0115] The current dynamic operating condition matching results are analyzed, and typical disturbance source labels and process type identifiers are extracted as index keys for retrieving the predefined process coupling rule base. Based on the process type identifier, the corresponding topology template in the rule base is located. This template defines the default connection relationships between the four nodes: aeration, neutralization, crystallization, and separation. All potential process pair combinations in the template are traversed, including aeration-neutralization, neutralization-crystallization, crystallization-separation, and cross-level coupled aeration-crystallization paths. For each potential process pair, its basic coupling strength coefficient is read. This coefficient represents the weight of the impact of upstream process parameter changes on downstream processes under standard steady-state operating conditions. An initial coupling relationship matrix is ​​constructed, where the rows and columns of the matrix correspond to the four process nodes, and the matrix element values ​​are the basic coupling strength coefficients of the corresponding process pairs. For process pairs without direct physical connections or logical associations, their coupling strength coefficients are set to zero to sparse the matrix structure. Through the above processing method, the operating condition matching results of the previous step are transformed into an initial coupling relationship matrix containing all potential process pairs and their basic coupling strength values, achieving the technical effect of providing a standardized data foundation for subsequent calculation of coupling sensitivity differences.

[0116] For example, when the matching result is high organic acid entrainment perturbation, the system extracts the tag "organic acid" and the process identifier "full process". It retrieves a topology template containing six potential process pairs from the rule base. The coupling coefficients are read as follows: aeration-neutralization 0.85, neutralization-crystallization 0.92, crystallization-separation 0.78, aeration-crystallization 0.15, neutralization-separation 0.05, and aeration-separation 0.02. A 4x4 initial coupling matrix is ​​constructed, with the main diagonal elements set to 1 and non-directly connected path elements set to 0. The element in the first row and second column of the matrix is ​​assigned a value of 0.85, the element in the second row and third column is assigned a value of 0.92, and so on, completing the matrix initialization. This matrix clearly quantifies the degree of inter-process influence under standard operating conditions, providing a benchmark for identifying the shift of coupling dominance towards the aeration-crystallization path under high organic acid conditions.

[0117] The dynamic topology remapping engine is a core control module for adaptive reconfiguration of process-coupled topology to adapt to dynamic operating condition disturbances, and its functions are as follows: Initial coupling benchmark construction function: It can receive typical disturbance source labels and process type identifiers of the current dynamic working condition matching results, call the predefined process coupling rule library to complete the association retrieval, quantify the basic coupling strength of each process pair under the standard steady-state working condition, generate a structured initial coupling relationship matrix, and provide a standardized benchmark data base for dynamic coupling evaluation. Coupling dominance quantification assessment function: It can combine the initial coupling relationship matrix with the semantic distance value of the working condition matching, and quantify the response correlation of each process pair under a specific disturbance scenario through the coupling sensitivity difference calculation method, generate a coupling sensitivity deviation vector, and accurately characterize the changing trend of the coupling dominance of each process pair relative to the standard working condition. Key control path filtering function: It can perform extreme value search and threshold determination on the coupling sensitivity deviation vector, accurately filter key process pairs where the coupling dominance has shifted significantly and exceeds the preset switching threshold, lock the core control path, and filter out false triggering signals in the perturbation scenario to ensure the necessity and accuracy of topology adjustment. Topology reconfiguration action generation function: It can compare the target key process with the system's default standard topology structure diagram, perform connection state reversal operation, automatically weaken the original intermediate control path with lag response, activate the backup direct connection control path adapted to the current disturbance, and generate a structured topology adjustment action list. Standardized control command output function: It can integrate the topology adjustment action list and operating condition matching timestamp information, complete the structured coding according to the industrial control command protocol, and output the topology remapping command with clear path switching indication and effective sequence to drive the control loop to switch without disturbance.

[0118] This engine can realize the adaptive reconstruction of the process coupling topology from a static fixed structure to a dynamic one that adapts to feed disturbances. It effectively solves the problems of response lag and model mismatch of the original intermediate control path under specific impurity disturbances, and significantly improves the response speed and disturbance resistance of the whole process control.

[0119] S5.2: Based on the semantic distance values ​​in the matching results of the initial coupling relationship matrix and the current dynamic working condition, the coupling sensitivity difference calculation method is used to quantitatively evaluate the response correlation of each process under a specific disturbance scenario, so as to generate a coupling sensitivity deviation vector that characterizes the coupling-dominant change trend of each process.

[0120] It receives the semantic distance value from the initial coupling relationship matrix and the current dynamic working condition matching result as the input data source for calculating the coupling sensitivity difference.

[0121] The original coupling strength values ​​of each process pair in the initial coupling relationship matrix are normalized to eliminate the magnitude interference of different physical dimensions on the correlation evaluation and generate a standardized basic coupling coefficient matrix.

[0122] Extract the inverse semantic distance of candidate historical fingerprint combinations from the current dynamic working condition matching results, construct a weight factor vector representing the similarity of working conditions, and use it to correct the response correlation under specific disturbance scenarios.

[0123] The coupling sensitivity difference calculation method is adopted to perform Hadamard product operation on the standardized basic coupling coefficient matrix and the weight factor vector to quantify the correlation strength of each process to the dynamic response under specific disturbances.

[0124] Calculate the deviation of the coupling sensitivity of each process under the current disturbance scenario from the standard operating condition baseline value, and generate a coupling sensitivity deviation vector that reflects the changing trend of coupling dominance.

[0125] By using the coupling sensitivity difference calculation method, the static coupling matrix of the previous step is transformed into a dynamically changing coupling sensitivity deviation vector, thereby achieving a precise quantitative assessment of the correlation between each process and the response.

[0126] S5.3: Based on the coupling sensitivity deviation vector, perform extreme value search and threshold determination logic to identify specific process pairs that have undergone a significant shift in coupling dominance and exceed a preset switching threshold, so as to generate a set of target process pairs marked as critical control paths.

[0127] The system receives a coupling sensitivity deviation vector, which contains quantified values ​​of the response correlation of each process pair under a specific disturbance scenario. A traversal scan is performed on the deviation vector to extract the extreme value element with the largest absolute value and its corresponding process pair index, identifying the coupling link most significantly affected by the current feed fluctuation. The extracted extreme value absolute value is compared with a preset dynamic switching threshold, which is set based on the standard deviation of coupling strength under historical steady-state conditions. If the extreme value absolute value is greater than the dynamic switching threshold, it is determined that the coupling dominance of the process pair has significantly shifted, and it is marked as a critical control path to be activated. If the extreme value absolute value is less than or equal to the dynamic switching threshold, the system is determined to be in a perturbation state, maintaining the original topology and not generating a remapping instruction. Deduplication processing is performed on the marked critical control paths to ensure that the same process node appears only once in the target set, avoiding control loop conflicts. The selected critical process pairs are combined and encapsulated into a target process pair set, including the source process identifier, the target process identifier, and a coupling direction flag. By using extreme value search and threshold determination logic, the coupling sensitivity deviation vector is transformed into a set of target process pairs marked as critical control paths, thereby achieving accurate identification and rapid locking of the dominant coupling link under dynamic feed fluctuations.

[0128] For example, the coupling sensitivity deviation vector is [0.12, -0.85, 0.03, 0.41], corresponding to the aeration-neutralization, neutralization-crystallization, crystallization-separation, and aeration-separation process pairs, respectively. The preset dynamic switching threshold is 0.5. Traversing the vector, the element with the largest absolute value is found to be -0.85, corresponding to the neutralization-crystallization process pair. Since |-0.85|>0.5, it is determined that the dominance of the neutralization-crystallization coupling has significantly shifted. This process pair is marked as a critical control path. The absolute values ​​of the remaining elements are all less than 0.5 and are not marked. The generated set of target process pairs is {(neutralization, crystallization, negative coupling)}. Based on this, the system freezes other non-critical pathways and only reweights the control parameters from neutralization to crystallization, significantly improving the response speed to disturbances entrained by high organic acids.

[0129] S5.4: Using the target process, perform topology difference analysis and connection state reversal operation on the default standard topology diagram of the set comparison system to determine the newly added direct connection paths that need to be activated and the original intermediate paths that need to be weakened, so as to generate a specific list of topology adjustment actions.

[0130] Obtain the target process set and the system default standard topology diagram data. The standard topology diagram defines the baseline connection relationship and signal flow direction between the four processes of aeration, neutralization, crystallization and separation.

[0131] Traverse each key process pair in the target process pair set, retrieve the corresponding original connection path status in the standard topology diagram, and mark it as an intermediate path or direct path to be evaluated.

[0132] By comparing the coupling transfer trend indicated in the current dynamic operating condition matching results with the preset path in the standard topology diagram, the weakened path caused by the increase in response hysteresis due to changes in the impurity spectrum is identified.

[0133] For the identified weakened path, a connection state inversion operation is performed to set its control gain coefficient to zero or to an extremely low threshold, thereby logically disconnecting the signal transmission of the path.

[0134] At the same time, based on the combination of process pairs with high correlation in the coupling sensitivity deviation vector, the backup direct connection path that was originally dormant is activated in the topology, such as establishing a direct mapping channel from the oxidation-reduction potential of the aeration section to the temperature of the crystallization section.

[0135] Calculate the topology difference between the newly activated path and the weakened path, and generate a list of topology adjustment actions that include path ID, switching action type, and priority sequence.

[0136] By performing topology difference analysis and connection state reversal operations, the abstract coupling dominance transfer is transformed into a specific list of path switching actions, enabling millisecond-level dynamic reconstruction of the control topology.

[0137] S5.5: Integrate the timestamp information of the topology adjustment action list with the matching result of the current dynamic working condition, and perform structured encoding processing according to the industrial control instruction encapsulation protocol to generate a topology remapping instruction for inter-process coupling relationship containing clear path switching indications and effective timing.

[0138] Step S6: Based on the inter-process coupling relationship topology remapping instruction, freeze non-critical control paths and decouple the parameter transfer coefficients on the newly activated reweighted paths to generate a critical control path weight allocation scheme. Specifically, this includes: S6.1: Obtain the list of non-critical path identifiers in the topology remapping instruction for inter-process coupling relationship, and use the Boolean mask generation algorithm to perform logical masking on the full control path matrix to generate non-critical control path freeze status flags, thereby blocking parameter transmission channels between non-dominant processes and eliminating background noise interference.

[0139] The system receives a topology remapping instruction for inter-process coupling relationships and parses the list of non-critical path identifiers contained therein. This list explicitly indicates the logical connections between processes that need to be masked under the current dynamic operating conditions. A Boolean mask matrix with the same dimensions as the system's full control path matrix is ​​constructed, initializing all elements to logical true values, representing that all control paths are active by default. The list of non-critical path identifiers is traversed, updating the elements at the corresponding index positions in the Boolean mask matrix to logical false values, forming a path masking template for specific disturbance scenarios. An element-wise logical AND operation is performed between the Boolean mask matrix and the full control path matrix to generate a non-critical control path freeze state flag matrix. In the freeze state flag matrix, the path parameter transmission coefficients corresponding to logical false values ​​are forcibly set to zero, thus mathematically cutting off the signal interaction channels between non-dominant processes. Through logical masking, the topology adjustment strategy of the previous step is transformed into specific path switching state data, achieving the expected technical effect of eliminating background noise interference and focusing on critical control paths.

[0140] For example, in cases with high organic acid entrainment, the system identifies the neutralization-crystallization link as a non-critical pathway. The full control pathway matrix is ​​a 4x4 matrix, and the non-critical pathway identifier list points to matrix element (2,3). A 4x4 Boolean mask matrix is ​​initialized, with all elements set to 1. Position (2,3) in the mask matrix is ​​updated to 0. A logical AND operation is performed to generate a frozen state flag matrix, where position (2,3) is 0 and the rest are 1. This operation reduces the direct influence coefficient of the neutralization section pH setpoint on the crystallization section temperature to zero, significantly reducing cross-interference in multivariate coupled systems and ensuring a substantial improvement in the control accuracy of the aeration-crystallization direct path.

[0141] S6.2: Based on the newly activated path node pairs selected by the non-critical control path freeze status flag, the dynamic gain decoupling calculator is called to perform Jacobian determinant decomposition on the original parameter transfer coefficients on the new activated path to generate an independent variable decoupling gain matrix, thereby eliminating the cross-interference effect in the multivariable coupled system and establishing a single variable control benchmark.

[0142] The system receives newly activated path node pairs selected from the frozen status flags of non-critical control paths and extracts real-time monitoring data sequences of upstream process output variables and downstream process input variables for these paths. A local dynamic coupling model of the newly activated path is constructed, defining the nonlinear mapping relationship between changes in upstream variables and changes in downstream variable responses. The Jacobian matrix elements at the current operating point are calculated using numerical differentiation to quantify the instantaneous sensitivity of each independent control variable to the controlled variable. The Jacobian matrix is ​​calculated using the following formula. Element values ​​in: in, Let a be the a-th controlled variable. For the b-th control variable, This represents partial derivative operations. Singular value decomposition is performed on the calculated Jacobian matrix to identify strong and weak coupling directions. A decoupling compensator matrix is ​​constructed based on the decomposition results to counteract the influence of cross-coupling terms on the main control loop. The original parameter transfer coefficients are multiplied by the decoupling compensator matrix to generate an independent variable decoupling gain matrix after eliminating cross-interference. Through Jacobian determinant decomposition and decoupling compensation, the path switching command from the previous step is transformed into a decoupling gain matrix with independent control characteristics, realizing decoupling control of the multivariable coupled system and establishing a single-variable baseline.

[0143] S6.3: Based on the semantic distance reciprocal between the independent variable decoupling gain matrix and the current dynamic operating condition matching result, the adaptive weight reconstruction method is executed to normalize and scale the decoupled gain value to generate a dynamic weight coefficient vector for the key control path. This allows for real-time allocation of the adjustment priority of each control loop and quantification of coupling strength based on the similarity of the operating conditions.

[0144] It receives the semantic distance value from the independent variable decoupling gain matrix and the matching result of the current dynamic operating condition, and uses it as the input data source for adaptive weight reconstruction.

[0145] Perform a reciprocal operation on the semantic distance values ​​to construct a confidence weight factor that represents the similarity between historical operating conditions and current real-time operating conditions. The higher the similarity, the larger the weight factor.

[0146] The element-wise multiplication operation between each element value in the decoupling gain matrix and the corresponding confidence weight factor is performed to achieve preliminary gain scaling based on operating condition similarity.

[0147] The L2 norm normalization method is used to standardize the magnitude of the scaled gain vector, eliminating the interference of dimensional differences on control priority and ensuring that the sum of weight coefficients converges to a preset benchmark value. The dynamic weight coefficients of the critical control path are then calculated. Based on the calculated dynamic weight coefficient vector, the adjustment priority of each control loop is quantified and sorted, and the high-weight path obtains a higher actuator response bandwidth.

[0148] By using an adaptive weight reconstruction method, the decoupling gain from the previous step is transformed into a dynamic weight coefficient vector of the key control path with working condition awareness, thereby achieving the expected technical effect of real-time allocation and adjustment of priorities and quantification of coupling strength based on working condition similarity.

[0149] S6.4: The signal flow of the upstream process setpoint adjustment command is linearly transformed using the dynamic weight coefficient vector of the critical control path to generate a parameter transfer correction signal with weight factors. This transforms the logical switching of the topology into an executable numerical compensation quantity and ensures a smooth transition of control commands.

[0150] The system receives the dynamic weight coefficient vector of the critical control path and the upstream process setpoint adjustment command, and extracts the pH setpoint deviation of the neutralization section as the input reference variable for linear transformation. A linear mapping matrix based on the weight coefficients is constructed, mapping each element in the dynamic weight coefficient vector to a different orthogonal component of the pH deviation, thus establishing the gain allocation logic for the signal flow.

[0151] The vector dot product operation is performed, multiplying the standardized pH setpoint adjustment vector element-wise with the dynamic weight coefficient vector of the critical control path and summing the results to calculate the preliminary parameter transfer correction amplitude. This process generates the parameter correction signal using the following formula: in, To pass correction signals to parameters with weighting factors, The dynamic weight coefficients for the c-th control path are... Adjust the upstream process setpoint component for the c-th channel. The total number of activated critical control paths.

[0152] A smooth transition factor is introduced to perform first-order low-pass filtering on the calculated parameter transfer correction signal, eliminating the step jitter in control commands caused by sudden weight changes during topology switching. The filtering time constant is set to be less than half of the system sampling period to ensure that the correction signal achieves steady-state convergence within milliseconds.

[0153] The filtered and corrected signal is converted into a standard industrial analog signal format to generate a digital control word containing amplitude and phase information, which serves as the direct input command for the actuators in downstream processes. Through linear transformation, the logical weight allocation from the previous step is converted into executable numerical compensation quantities, achieving smooth transitions and disturbance-free switching of control commands.

[0154] S6.5: Integrate the parameter transfer correction signal with weighting factors and the non-critical control path freeze status flag, execute the control strategy encapsulation protocol to generate the final critical control path weight allocation scheme, and output an execution configuration file containing complete path switch status and gain values, thereby providing a definite mathematical basis for the generation of downstream process temperature control curve offset compensation signals.

[0155] Step S7: Generate multi-process collaborative control execution instructions based on the critical control path weight allocation scheme. Specifically, this includes: S7.1: Obtain the pH setpoint adjustment command and path coupling coefficient of the neutralization section in the key control path weight allocation scheme, and perform amplitude normalization processing on the pH setpoint adjustment command of the neutralization section based on the linear interpolation method to generate a standardized upstream process setpoint adjustment vector.

[0156] S7.2: Receive the standardized upstream process setpoint adjustment vector and path coupling coefficient, and use the cross-domain transfer function mapping model to perform variable translation calculation from the chemical acidity domain to the thermodynamic temperature domain to generate a reference signal for the offset of the basic downstream process temperature control curve.

[0157] The standardized upstream process setpoint adjustment vector and path coupling coefficient are received as the input benchmark for cross-domain variable translation. A chemical-thermodynamic coupling mapping model based on the Arrhenius equation is constructed to establish the influence mechanism of pH changes on crystallization kinetic constants. The baseline signal of the downstream process temperature control curve offset is calculated using the following formula: in, This serves as the reference signal for the temperature control curve offset. This is the process experience coupling constant. The path coupling coefficient is... The target pH value for the neutralization phase. The formula uses a logarithmic function to characterize the nonlinear effect of exponential changes in hydrogen ion concentration on the reaction rate, transforming the pH deviation in the chemical domain into temperature compensation requirements in the thermodynamic domain. The calculated initial offset is saturated and limited to prevent extreme pH fluctuations from causing the temperature setpoint to exceed the equipment's safe operating range. Combining the thermal inertia time constant of the aging and crystallization process, a first-order hysteresis filter is applied to the offset signal to eliminate high-frequency noise interference and generate a smooth temperature regulation trajectory. Through the aforementioned cross-domain transfer function mapping and dynamic correction processing, the standardized upstream process setpoint adjustment vector from the previous step is transformed into a physically meaningful baseline signal for the downstream process temperature control curve offset. This achieves precise variable translation from chemical pH regulation to thermodynamic crystallization control, ensuring stable operation of the crystallization process within the optimal supersaturation window.

[0158] S7.3: Extract the impurity spectrum from the current dynamic operating condition matching results to form feature labels and response hysteresis entropy values. Construct an operating condition adaptive elastic adjustment factor calculation model based on fuzzy logic reasoning mechanism to generate an elastic adjustment factor value that characterizes the fault tolerance range of crystallization kinetics.

[0159] S7.4: The baseline signal of the temperature control curve offset of the basic downstream process is fused with the value of the elastic adjustment factor. A dynamic weighted superposition method is used to perform nonlinear correction processing on the baseline signal of the temperature control curve offset of the basic downstream process to generate a final downstream process temperature control curve offset compensation signal that includes dynamic fault tolerance compensation.

[0160] The system receives the reference signal of the temperature control curve offset of the downstream process and the value of the elastic adjustment factor as input variables for nonlinear correction processing.

[0161] A dynamic weighted superposition method model is constructed, the coupling function relationship between the reference signal weight coefficient and the elastic factor compensation coefficient is defined, and the mathematical basis of nonlinear mapping is established.

[0162] The elastic adjustment factor is smoothed by using the sigmoid activation function to eliminate the step impact of control commands under abrupt operating conditions and generate a continuously differentiable compensation gain curve.

[0163] By using the processed compensation gain curve to perform point-by-point multiplication on the basic offset reference signal, adaptive amplitude scaling based on the fault tolerance range of crystallization kinetics is achieved.

[0164] A time decay factor is introduced to perform transient filtering on the scaled signal, suppressing high-frequency noise interference and ensuring the smooth transition characteristics of temperature control commands in the time domain.

[0165] Then, the offset compensation signal of the final downstream process temperature control curve is calculated, and the amplitude limiting protection logic is executed on the calculated compensation signal. Based on the upper limit of the heating power and the cooling rate constraint of the aging crystallization kettle, abnormal values ​​that exceed the physical execution boundary are truncated.

[0166] The limited signal is encapsulated into a standard floating-point number format and a timestamp is added to form the final downstream process temperature control curve offset compensation signal containing dynamic fault tolerance compensation.

[0167] By using dynamic weighted superposition and nonlinear correction processing, the results of the previous step are transformed into temperature compensation data indicators with anti-disturbance capabilities, thereby achieving the expected technical effect of robust control of feed fluctuations in the crystallization process.

[0168] S7.5: Encapsulate the final downstream process temperature control curve offset compensation signal and the actuator address code of each process, and perform data frame assembly and check code generation processing according to the industrial fieldbus communication protocol format to generate multi-process collaborative control execution instructions that can be directly sent to the edge control gateway.

[0169] Step S8: Execute the multi-process collaborative control execution instructions to adjust the operating parameters of each process, and complete the control update within a single sampling period to achieve resource-based recycling of materials. Specifically, this includes: S8.1: Obtain multi-process collaborative control execution instruction data frames, and use the industrial fieldbus protocol parsing module to perform check code verification and address code extraction on the data frames to generate standardized control instruction packages containing target actuator address identifiers and payload data.

[0170] S8.2: Based on the payload data in the standardized control instruction package, the digital form of the temperature control curve offset compensation signal and the elastic adjustment factor value are converted into an analog voltage control signal using a digital-to-analog conversion driving method to generate a continuous physical quantity control waveform that drives the heating actuator and the flow regulating valve.

[0171] S8.3: Utilize continuous physical quantity control waveforms to perform real-time parameter adjustment processing on the nitrogen flow regulating valve in the aeration process, the acid and alkali metering pump in the neutralization process, the electric heating rod in the aging and crystallization process, and the variable frequency motor in the solid-liquid separation process, so as to generate a dynamic operating state sequence of each process that reflects the latest control strategy.

[0172] S8.4: Based on the dynamic operating status sequence of each process, a single-cycle sampling synchronization mechanism is executed to perform timestamp alignment and delay compensation calculation on the sensor feedback data of the entire system, so as to generate a real-time operating condition closed-loop feedback dataset that eliminates the impact of transmission lag.

[0173] S8.5: Based on the real-time operating condition closed-loop feedback dataset and the preset dynamic feed fluctuation tolerance threshold, a deviation comparison analysis is performed to confirm that the control update has converged within a single sampling period, so as to generate a stable operation confirmation flag that characterizes the system to reach a low-disturbance coordinated control state.

[0174] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0175] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0176] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for the resource recovery and treatment of the top material of a three-tower NMP refining process, characterized in that, Specifically, it includes: S1: For the four processes of NMP refining three-tower aeration, neutralization, aging crystallization and solid-liquid separation, multi-dimensional time-series process data are collected in the continuous operation cycle to generate the original operating condition response sequence. S2: Based on the original working condition response sequence, perform fractal dimension analysis and morphological similarity measurement, and extract the curvature integral ratio, phase lag index and response hysteresis entropy value as morphological invariants to construct the process semantic fingerprint. S3: Classify and store the semantic fingerprints of the process steps to establish a fingerprint knowledge base; S4: Real-time acquisition of current feed spectrum preprocessing data and initial response features of the first process, and retrieval of historical fingerprint combinations with the smallest semantic distance in the fingerprint knowledge base to generate current dynamic working condition matching results; S5: Based on the current dynamic working condition matching result, start the dynamic topology remapping engine to identify key process pairs where the coupling dominance has shifted, so as to generate inter-process coupling relationship topology remapping instructions; S6: According to the inter-process coupling relationship topology remapping instruction, freeze non-critical control paths and decouple the parameter transfer coefficients on the reweighted new activated paths to generate a critical control path weight allocation scheme. S7: Generate multi-process collaborative control execution instructions based on the critical control path weight allocation scheme; S8: Execute the multi-process collaborative control execution instructions to adjust the operating parameters of each process, and complete the control update within a single sampling period to realize the resource recycling and processing of materials.

2. The method for resource recovery and treatment of the top material of the NMP refining three-tower system according to claim 1, characterized in that, The multidimensional time-series process data includes pH change slope, redox potential transition amplitude, turbidity rise inflection point time, and transient peak value of centrifugal current.

3. The method for resource recovery and treatment of the top material of the NMP refining three-tower as described in claim 1, characterized in that, The fractal dimension analysis and morphological similarity measurement are performed using an unsupervised temporal clustering algorithm.

4. The method for resource recovery and treatment of the top material of the NMP refining three-tower system according to claim 1, characterized in that, The step of classifying and storing process semantic fingerprints to establish a fingerprint knowledge base includes: The semantic fingerprints of the process are classified and stored according to process type and typical disturbance sources. Only interpretable semantic labels and mapping relationship tables are retained, and the original waveforms are not stored, in order to establish a fingerprint knowledge base.

5. The method for resource recovery and treatment of the top material of the NMP refining three-tower as described in claim 1, characterized in that, The process of retrieving the historical fingerprint combination with the smallest semantic distance in the fingerprint knowledge base is performed using a fast hash matching method.

6. The method for resource recovery and treatment of the top material of the NMP refining three-tower as described in claim 1, characterized in that, The generation of multi-process collaborative control execution instructions based on the critical control path weight allocation scheme includes: Using the aforementioned critical control path weight allocation scheme, the upstream process setpoint adjustment command is translated into the downstream process temperature control curve offset compensation signal and an elastic adjustment factor is injected to generate multi-process collaborative control execution commands.

7. The method for resource recovery and treatment of the top material of the NMP refining three-tower system according to claim 2, characterized in that, S1 specifically includes: The raw analog voltage signal of the distributed sensor array of the three-tower NMP refining process is obtained. The pH electrode potential, redox electrode millivolt value, turbidity light scattering intensity and centrifuge motor current waveform are synchronously sampled and processed by a high-speed analog-to-digital converter to generate a multi-channel raw digital signal stream with a unified timestamp. The pH change slope sequence and redox potential transition amplitude sequence are obtained by processing the pH numerical sequence and redox potential numerical sequence based on the multi-channel raw digital signal stream. The turbidity light scattering intensity and centrifuge motor current waveform were processed to obtain the turbidity rise inflection point time data and the transient peak value data of the centrifugal current. Based on the pH change slope sequence, redox potential transition amplitude sequence, turbidity rise inflection point time data, and centrifugal current transient peak data, a multidimensional feature vector space is constructed. Data fusion processing is performed through time alignment and missing value interpolation mechanism to generate multidimensional time-series process data covering the entire cycle of aeration, neutralization, aging crystallization, and solid-liquid separation. The original operating condition response sequence is obtained by segmenting, encapsulating, and annotating metadata based on multidimensional time-series process data.

8. The method for resource recovery and treatment of the top material of the NMP refining three-tower as described in claim 7, characterized in that, The process of processing the pH numerical sequence and redox potential numerical sequence based on the multi-channel raw digital signal stream to obtain the pH change slope sequence and redox potential transition amplitude sequence includes: Based on the multi-channel raw digital signal stream, the sliding window differentiation method is used to perform real-time slope calculation and transition amplitude extraction on the pH value sequence and redox potential value sequence to generate pH change slope sequence and redox potential transition amplitude sequence.

9. The method for resource recovery and treatment of the top material of the NMP refining three-tower column according to claim 7, characterized in that, The processing of turbidity light scattering intensity and centrifuge motor current waveforms yields turbidity rise inflection point time data and centrifuge current transient peak data, including: By utilizing the turbidity light scattering intensity sequence, an adaptive threshold inflection point detection method is used to identify curvature abrupt change points in the crystallization nucleation stage. Combined with the centrifuge motor current waveform, transient peak capture technology is applied to extract the load impact characteristics in the solid-liquid separation process, so as to generate turbidity rise inflection point time data and centrifugal current transient peak data.

10. The method for resource recovery and treatment of the top material of the NMP refining three-tower system according to claim 7, characterized in that, The original operating condition response sequence obtained by segmenting, encapsulating, and annotating metadata based on multi-dimensional time-series process data includes: Based on the multidimensional time-series process data, the data is segmented and encapsulated according to process type labels and metadata is labeled. The feature parameters within the continuous operation cycle are integrated into structured data blocks to generate the original operating condition response sequence.

Images