Sludge treatment composite compensation method and system for precise reagent dosing

Abstract

The present application provides a sludge treatment composite compensation method and system for accurate reagent dosing, comprising: arranging an in-situ multi-parameter high-frequency sensing array downstream of the reagent dosing point, collecting time series signals such as Zeta potential, floc particle size distribution and interfacial tension, applying phase space reconstruction and chaotic dynamics topology index extraction technology to generate reagent reaction time series fingerprint vector. Through the establishment of fingerprint-working condition mapping library and improved probability distance algorithm, accurate matching of current working condition and historical working condition is realized, and the reagent dosing compensation amount is dynamically calculated in combination with the time series weight template of each influence factor, and the pump flow is adjusted by the real-time actuator; in addition, there are drift detection and small sample incremental learning modules, which can automatically update the template parameters when the working condition deviates, continuously optimize the compensation accuracy, and the present application improves the real-time accuracy of reagent dosing and the adaptive ability of the system.

Description

Technical Field

[0001] This invention relates to the field of sludge treatment process control and intelligent chemical dosing technology, and particularly to a composite compensation method and system for sludge treatment oriented towards precise chemical dosing. Background Technology

[0002] Precise dosing control of chemicals in sludge treatment has become a core focus of the industry. The key objective is to achieve the optimal match between chemical dosage and the requirements of the sludge mixing system, significantly improving flocculation efficiency, reducing chemical waste, and optimizing effluent quality. Mainstream control schemes largely rely on traditional proportional-integral-derivative (PID) regulation, static or semi-dynamic multi-factor weighted models, adaptive dosing methods based on fuzzy logic or expert rule bases, and the increasingly popular data-driven predictive control technologies. For example, some sludge treatment plants input multiple parameters such as pH, temperature, mixed liquor suspended solids concentration (MLSS), and sludge volume index (SVI) into a weighting function, setting weights based on historical statistical patterns to achieve dynamic correction of chemical dosage. Other methods utilize parameter identification techniques such as particle swarm optimization and genetic algorithms, combined with fluid dynamics modeling, to compensate for and optimize reaction time delay characteristics, thereby reducing control lag. With the development of intelligent sensing and big data analytics technologies, some enterprises and research institutions are attempting to apply artificial intelligence technologies such as LSTM, fuzzy inference networks, and autoregressive ensemble methods to predict pesticide dosage, using large-sample training models to improve their adaptive adjustment capabilities under complex operating conditions. Meanwhile, some high-end systems integrate in-situ detection modules for online Zeta potential and floc particle size distribution, aiming to achieve quantitative perception of microscopic reaction processes. However, current mainstream technologies still primarily rely on parametric modeling, fixed weight allocation, or online closed-loop correction based solely on macroscopic factors, paying insufficient attention to the microscopic mechanisms of mixing between pesticides and sludge after dosage and the time-delayed nature of the entire reaction kinetic process. Even when some systems introduce time-delay compensation, they still mainly rely on the identification or empirical estimation of delay time parameters, lacking detailed analysis of the dynamic response of multidimensional signal coupling under actual operating conditions. Existing technologies are suitable for small to medium-sized sludge treatment plants with relatively stable operating conditions and minimal changes in dosage. However, when dealing with process disturbances (such as rapid changes in influent water quality or sudden temperature changes) or adjustments to reagent types, the dosing strategy is prone to mismatch, resulting in frequent overshoots, response lags, or decreased reagent utilization, making it difficult to meet the requirements of second-level response and long-term precise control. Especially in real-world scenarios with large-scale processes, highly coupled influencing factors, and significant chaotic characteristics in the reaction process, traditional solutions relying on static models or lag correction parameters struggle to capture key kinetic index changes, often leading to reagent overdosing and decreased system stability. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a composite compensation method and system for sludge treatment with precise dosing of reagents.

[0004] The technical solution of this invention is implemented as follows: a composite compensation method for sludge treatment oriented towards precise dosing of reagents, comprising: S1: Deploy a high time resolution in-situ sensing array 5 to 15 cm downstream of the sludge treatment agent dosing point to simultaneously collect three types of time-series signals: Zeta potential, floc particle size distribution dynamic spectrum, and interfacial tension change rate within 0 to 120 seconds after agent injection, with a sampling frequency of not less than 10 Hz. S2: Based on the three types of time-series signals collected by S1, a three-dimensional attractor trajectory is generated through phase space reconstruction technology. The embedding dimension is set to 5 and the delay time is determined by the mutual information method. The box dimension, correlation dimension and maximum Lyapunov exponent of the trajectory are calculated to form the drug reaction time-series fingerprint vector. S3: Using the reagent reaction time-series fingerprint vector generated in S2, combined with 128 sets of typical operating condition data covering pH values ​​of 5.5 to 8.5, temperatures of 10 to 35 degrees Celsius, sludge volume index of 50 to 250 ml / g, and mixed liquor suspended solids concentration of 2 to 12 g / L, an offline fingerprint-operating condition mapping library was established. Sixteen fingerprint clusters were formed through clustering algorithms and bound to the time-series weight templates of each influencing factor. S4: During the sludge treatment operation, the reagent reaction time sequence fingerprint vector of the current working condition is generated in real time, and the improved Wasserstein distance algorithm is used to compare it with the offline fingerprint-working condition mapping library established in S3. After identifying the fingerprint cluster with the highest matching degree, its bound time sequence weight template is retrieved. S5: Based on the time-series weight template retrieved by S4, the weights of pH at 60 seconds, temperature at 30 seconds, and sludge volume index at 90 seconds are dynamically allocated, and the reagent dosage compensation amount is generated by combining the nonlinear response function of each factor deviation. S6: Input the reagent dosage compensation amount generated in S5 into the actuator to control the output flow rate of the reagent dosing pump, so that the actual dosage dynamically matches the real-time requirements of the sludge treatment process. S7: Every 24 hours, detect the deviation between the drug reaction timing fingerprint vector generated by S2 and the center of its fingerprint cluster, and determine whether the deviation exceeds the preset threshold for 5 consecutive times. If the condition is met, trigger the fingerprint drift signal. S8: In response to the fingerprint drift signal triggered by S7, it uses the newly acquired time-series signal to perform small-sample incremental learning on the drifting fingerprint clusters, and only updates the time-series weight template parameters in the interval of 40 seconds to 100 seconds to complete the dynamic correction of the drug reaction time-series fingerprint database.

[0005] The present invention also provides a composite compensation system for sludge treatment with precise dosing of chemicals, which uses the above-mentioned composite compensation method for sludge treatment with precise dosing of chemicals to compensate for the dosing of chemicals during the sludge treatment process.

[0006] The sludge treatment composite compensation method and system for precise dosing of reagents provided by this invention have the following beneficial effects: (1) This invention effectively overcomes the technical bottlenecks of high system complexity, weak generalization ability, and slow response caused by relying on explicit process modeling and empirical rule bases in traditional water treatment dosing control by constructing a "reaction time sequence fingerprint" characterization system based on chaotic dynamics characteristics. By utilizing the phase space reconstruction technology of high-frequency in-situ sensing signals, the box dimension, correlation dimension, and maximum Lyapunov exponent are extracted to form a low-dimensional but information-dense fingerprint vector, which directly captures the inherent chaotic characteristics of micro-particle aggregation behavior during coagulation, and realizes second-level resolution perception of the reaction process. This method does not require the establishment of an explicit time delay model or coupling strength matrix, avoids the computational burden and hyperparameter sensitivity caused by complex algorithms, and significantly improves the real-time performance and deployment convenience of the system while ensuring control accuracy. (2) This invention combines offline fingerprint clustering with dynamic weight template matching to achieve temporal decoupling and adaptive compensation of the contributions of multiple influencing factors. Unlike static weight allocation or fixed rule judgment, this invention decomposes the temporal contribution weights of key factors such as pH, temperature, and SVI to the turbidity improvement rate in historical data through partial least squares regression, and clusters them according to typical reaction stages to form a reusable "temporal weight template", so that the optimal response mode can be automatically matched according to the real-time fingerprint during operation. In particular, by setting different factors to trigger the nonlinear compensation function ΔQ at specific time points (such as pH at t=60s and temperature at t=30s), the kinetic differences of each physicochemical action path are accurately responded to, which greatly improves the scientificity and pertinence of the control strategy. At the same time, the improved Wasserstein distance is introduced for fingerprint matching, which enhances the sensitivity to the identification of small operating condition drifts, ensures the accuracy and stability of weight calling, thereby significantly reducing the over-dosing rate of reagents and achieving efficient resource utilization while ensuring the quality of effluent. (3) This invention constructs a closed-loop optimization system with long-term robustness and adaptive evolution capability by designing a lightweight incremental learning and local template update mechanism, overcoming the technical limitations of traditional fixed models that are susceptible to environmental drift and performance degradation. Conventional control strategies, once deployed, are difficult to self-adjust with the slow changes in influent characteristics, eventually leading to control failure; while full-cycle retraining brings high computational costs. This invention proposes to trigger fingerprint drift detection once every 24 hours, and only start small-sample incremental learning when the fingerprint deviates from the center of its cluster by more than a set threshold for five consecutive observations, and limit the weight update range to the critical reaction period, which effectively captures the long-term evolution trend of the system and avoids the disturbance risk and resource consumption caused by global retraining. This mechanism can complete model self-correction without manual intervention, significantly enhancing the continuous applicability and reliability of the system in complex real-world scenarios, especially suitable for stable operation under seasonal water quality fluctuations or sudden pollution events. Attached Figure Description

[0007] Figure 1 The flowchart is a composite compensation method for sludge treatment based on precise dosing of reagents according to the present invention. Figure 2 This is a sub-flowchart of the sludge treatment composite compensation method for precise dosing of reagents according to the present invention; Figure 3 This is another sub-flowchart of the sludge treatment composite compensation method for precise dosing of reagents according to the present invention. Detailed Implementation

[0008] 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.

[0009] 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.

[0010] like Figure 1 As shown, this invention provides a composite compensation method for sludge treatment aimed at precise dosing of reagents, specifically including: S1: Deploy a high time resolution in-situ sensing array 5 to 15 cm downstream of the sludge treatment agent dosing point to simultaneously collect three types of time-series signals: Zeta potential, floc particle size distribution dynamic spectrum, and interfacial tension change rate within 0 to 120 seconds after agent injection, with a sampling frequency of not less than 10 Hz. S2: Based on the three types of time-series signals collected by S1, a three-dimensional attractor trajectory is generated through phase space reconstruction technology. The embedding dimension is set to 5 and the delay time is determined by the mutual information method. The box dimension, correlation dimension and maximum Lyapunov exponent of the trajectory are calculated to form the drug reaction time-series fingerprint vector. S3: Using the reagent reaction time-series fingerprint vector generated in S2, combined with 128 sets of typical operating condition data covering pH values ​​of 5.5 to 8.5, temperatures of 10 to 35 degrees Celsius, sludge volume index of 50 to 250 ml / g, and mixed liquor suspended solids concentration of 2 to 12 g / L, an offline fingerprint-operating condition mapping library was established. Sixteen fingerprint clusters were formed through clustering algorithms and bound to the time-series weight templates of each influencing factor. S4: During the sludge treatment operation, the reagent reaction time sequence fingerprint vector of the current working condition is generated in real time, and the improved Wasserstein distance algorithm is used to compare it with the offline fingerprint-working condition mapping library established in S3. After identifying the fingerprint cluster with the highest matching degree, its bound time sequence weight template is retrieved. S5: Based on the time-series weight template retrieved by S4, the weights of pH at 60 seconds, temperature at 30 seconds, and sludge volume index at 90 seconds are dynamically allocated, and the reagent dosage compensation amount is generated by combining the nonlinear response function of each factor deviation. S6: Input the reagent dosage compensation amount generated in S5 into the actuator to control the output flow rate of the reagent dosing pump, so that the actual dosage dynamically matches the real-time requirements of the sludge treatment process. S7: Every 24 hours, detect the deviation between the drug reaction timing fingerprint vector generated by S2 and the center of its fingerprint cluster, and determine whether the deviation exceeds the preset threshold for 5 consecutive times. If the condition is met, trigger the fingerprint drift signal. S8: In response to the fingerprint drift signal triggered by S7, it uses the newly acquired time-series signal to perform small-sample incremental learning on the drifting fingerprint clusters, and only updates the time-series weight template parameters in the interval of 40 seconds to 100 seconds to complete the dynamic correction of the drug reaction time-series fingerprint database.

[0011] Step S1: Deploy a high-temporal-resolution in-situ sensing array 5 to 15 centimeters downstream of the sludge treatment agent dosing point to simultaneously collect three types of time-series signals within 0 to 120 seconds after agent injection: Zeta potential, floc particle size distribution dynamic spectrum, and interfacial tension change rate. The sampling frequency is not less than 10 Hz. Specifically, this includes: S1.1: Spatial positioning analysis of the fluid region 5 to 15 cm downstream of the sludge treatment agent dosing point is performed. Based on the fluid dynamics mixing characteristics, the optimal observation section is determined to generate an in-situ sensing array spatial layout scheme that includes the installation sites of the Zeta potential probe, the laser backscatter particle size dynamic analysis module, and the microfluidic interface tension sensor. For the fluid region located 5 to 15 centimeters downstream of the sludge treatment agent dosing point, a three-dimensional laser scanning flow field mapping method (parameter settings: scanning resolution 0.5 mm, sampling rate 2000 points / second) was used to achieve spatial analysis of the flow velocity distribution, vortex structure, and mixing layer thickness. Furthermore, by calculating the local turbulence intensity index of the fluid velocity vector field (parameter: calculated based on the standard deviation of instantaneous velocity fluctuation), a quantitative assessment of the mixing uniformity is achieved, and flow uniformity judgment data at different cross-sectional locations are obtained; Numerical simulation method was adopted (parameters: three-dimensional incompressible Navier-Stokes equations, The turbulence model is used to perform CFD inversion calculations on the flow field data obtained from the scan, so as to realize multi-scale prediction of fluid mixing characteristics and generate mixing efficiency index matrix for each candidate observation section. Furthermore, by using a flow characteristic matching algorithm (parameter: correlation coefficient threshold 0.85), the mixing efficiency index of candidate sections is matched and screened with the distinguishability requirements of the reagent reaction signal, and the best candidate section set that meets the requirements of high-resolution acquisition is output. A multi-objective optimization algorithm (parameter: the objective function is a weighted sum of maximizing signal strength and minimizing flow velocity gradient) is used to calculate the comprehensive score of each location within the candidate cross-section set and obtain the coordinates of the optimal observation cross-section. Furthermore, through physical accessibility verification methods (parameters: probe installation space ≥ 30 mm, no obstructions from structures), the feasibility of installing the optimal observation section is verified, and the range of installable sections is determined. By using a spatial layout generation algorithm (parameters: layout constraints include probe spacing ≥ 15 mm and signal interference coefficient ≤ 0.05), the installable cross-sectional area is transformed into an in-situ sensing array spatial layout scheme that includes Zeta potential probe mounting sites, laser backscatter particle size dynamic analysis module mounting sites, and microfluidic interface tension sensor mounting sites, thereby achieving precise allocation of effective observation positions. For example, in the sludge return pipeline at the outlet of the secondary sedimentation tank of a municipal wastewater treatment plant, within a downstream range of 8 to 12 centimeters from the reagent dosing point, flow field scanning revealed a maximum longitudinal velocity distribution of 1.2 m / s, an average turbulence intensity of 0.18, and a mixing layer thickness of 22 mm. Based on CFD inversion, the mixing efficiency index matrix for this area showed a comprehensive score of 0.92 at the 9 cm section, 0.94 at the 10 cm section, and 0.89 at the 11 cm section. The multi-objective optimization algorithm output the optimal observation section coordinates as the 10 cm position. Installation feasibility verification showed that this position has a 45 mm space margin and no structural obstruction. The spatial layout generation algorithm allocates three sensor mounting sites within the cross-sectional area: the Zeta potential probe is located 5 mm to the left of the center of the cross-section, the laser backscatter particle size module is located 7 mm to the right of the center of the cross-section, and the microfluidic interfacial tension sensor is located 3 mm above the center of the cross-section. The measured signal interference coefficient is 0.04. Finally, an in-situ sensor array spatial layout scheme that meets the requirements of high time resolution synchronous acquisition is formed, ensuring the accuracy and stability of subsequent multi-dimensional time-series signal acquisition. S1.2: Based on the in-situ sensor array spatial layout scheme, the micro Zeta potential probe, the laser backscatter particle size dynamic analysis module and the microfluidic interface tension sensor are hardware integrated and triggered synchronously configured. A high-precision clock source is used to unify the sampling start time of each sensor in order to generate a multi-channel synchronous acquisition and execution command with time alignment capability. The coordinate data of each sensor installation point in the in-situ sensor array spatial layout scheme output by S1.1 are obtained. The hardware integration optimization method (parameters: layout scheme coordinate accuracy ≤1mm, module interface unified standardization protocol) is adopted to realize the physical connection and signal channel unified processing of the micro Zeta potential probe, laser backscatter particle size dynamic analysis module and microfluidic interface tension sensor. Furthermore, by using a bus synchronous trigger configuration method (parameters: control bus type is SPI high-speed mode, maximum transmission delay ≤ 0.5ms), the clock signals of all sensor modules are synchronously distributed, and a hardware handshake signal for a unified trigger input port is obtained; Furthermore, a high-precision clock source is used to perform time base calibration operation on the unified trigger signal (parameters: clock accuracy ±1μs, temperature compensation frequency correction coefficient 0.0001Hz / °C) to achieve absolute time alignment of each sampling start time and generate a multi-channel acquisition time index table; Furthermore, by using a multi-channel acquisition command generation algorithm (parameters: sampling frequency setting value ≥ 10Hz, acquisition duration setting value 120s), the time index table is bound to the sampling frequency parameter to generate synchronous acquisition execution commands with time alignment capability that can be executed in parallel by various sensor hardware modules; Through the above hardware integration and synchronous triggering processing, the spatial layout scheme of the previous step is transformed into hardware configuration data that can stably achieve synchronous acquisition of multiple signals during operation, thereby achieving time base consistency and accuracy of multi-dimensional time-series signal acquisition. For example, in the operation of a municipal wastewater treatment plant with a daily processing capacity of 5000 cubic meters, the selected Zeta potential probe installation point in the in-situ sensor array layout is 9cm downstream of the injection point, the laser backscattering module installation point is 12cm downstream of the injection point, and the interface tension sensor installation point is 14cm downstream of the injection point. Each module uses a unified M12 waterproof interface connected to the SPI bus. The clock source is a temperature-compensated quartz oscillator with a nominal frequency of 10MHz and an accuracy of ±0.5μs. After time base calibration, the deviation of the sampling start time is less than 2μs. The acquisition command sets the sampling frequency to 15Hz and the acquisition window length to 120s. The timestamp difference of each channel signal in the acquisition execution result remains constant within ±1μs, ensuring the time synchronization of multi-dimensional signals in subsequent phase space reconstruction and improving the stability and accuracy of dynamic weight calculation. S1.3: Response to multi-channel synchronous acquisition execution command, high-frequency scanning of sludge mixture is performed within a time window of 0 to 120 seconds after the agent injection action. The potential fluctuation sequence, light intensity scattering spectrum sequence and interface mechanical change sequence are continuously captured with a sampling frequency of not less than 10 Hz to generate an original multidimensional time series dataset containing Zeta potential time series signal, floc particle size distribution dynamic spectrum time series signal and interface tension change rate time series signal. S1.4: The original multidimensional time series dataset is subjected to noise suppression and baseline drift correction. The sliding window filtering algorithm is used to remove high-frequency electrical noise and compensate for sensor zero-point drift, so as to generate high-quality pre-processed Zeta potential cleaning time series signal, floc particle size distribution dynamic spectrum cleaning time series signal and interfacial tension change rate cleaning time series signal. S1.5: Based on the high-quality Zeta potential cleaning time series signal, floc particle size distribution dynamic spectrum cleaning time series signal, and interfacial tension change rate cleaning time series signal, the data is structured and encapsulated. The three types of signals are mapped into a unified three-dimensional matrix format according to the timestamp index to generate a standardized reagent reaction multidimensional time series input tensor for subsequent phase space reconstruction calculation.

[0012] Step S2: Based on the three types of time-series signals acquired in S1, a three-dimensional attractor trajectory is generated using phase space reconstruction technology. The embedding dimension is set to 5, and the delay time is determined by mutual information. The box dimension, correlation dimension, and maximum Lyapunov exponent of this trajectory are calculated to form a drug response time-series fingerprint vector. Specifically, this includes: S2.1: Obtain the pre-processed high-quality Zeta potential cleaning time-series signal, floc particle size distribution dynamic spectrum cleaning time-series signal, and interfacial tension change rate cleaning time-series signal. Perform mutual information function calculation processing on the three types of cleaning time-series signals respectively to determine the optimal delay time parameter that can minimize redundant information and retain the maximum dynamic correlation. The preprocessed high-quality Zeta potential cleaning time-series signal, floc particle size distribution dynamic spectrum cleaning time-series signal, and interfacial tension change rate cleaning time-series signal are used as input objects for mutual information calculation. The mutual information function calculation method is used (parameter: time delay). The enumeration range is from 1 to N, where N is the sampling window length, to achieve a quantitative measurement of the autocorrelation and information redundancy of each time-series signal under different time delays; Furthermore, by using the minimum search algorithm of the mutual information curve (constraint: the mutual information value drops to below 1 / e of the global maximum value), the time delay parameter is located to minimize redundant information and maximize dynamic correlation, and a candidate set of time delays is obtained. Furthermore, by employing smoothing filtering and parameter stability testing methods (test criterion: the variance of the delay parameter within the sliding window is less than a preset threshold), the candidate set of time delays is accurately selected, and the optimal delay time parameter is generated. ; The mutual information value is calculated using the following formula. :

[0013] in, This represents the i-th sampling point of the original time-series signal. Indicates delay The j-th sampling point after that, Represents the probability distribution function. for and The joint probability, and These are the sampling point indices; By using the mutual information function calculation method and parameter stability test, the cleaning signal from the previous step is transformed into the optimal delay time parameter data, thereby realizing the time embedding reference setting for subsequent phase space reconstruction operations. For example, after the sludge treatment agent is added, mutual information calculation is performed on a high-quality Zeta potential cleaning timing signal with a sampling frequency of 12Hz. The step size τ is set to 1 to 60, and a curve showing the mutual information value changing with τ is plotted. In the curve... When the mutual information value drops below 1 / e of the global maximum value at a value of 8, and the stability test variance is 0.002, it is determined that... =8 is the optimal delay time parameter. =8 is applied to phase space reconstruction, which can fully preserve the dynamic correlation of the reaction stage when generating three-dimensional attractor trajectories, thus improving the accuracy of chaotic feature extraction. Under the same working conditions, the floc particle size distribution dynamic spectrum cleaning signal is obtained through the same calculation. =6, the cleaning signal of the interfacial tension change rate is obtained. =10, used for phase space embedding of the three data streams respectively, to achieve precise matching of delay times for different observations. Finally, the output of this step is... The parameters significantly improve the ability to perceive the time delay characteristics of the reaction process in the subsequent fingerprint vector generation, ensuring that the compensation analysis module can respond quickly to changes in operating conditions. S2.2: Based on the determined optimal delay time parameter and the preset five-dimensional embedding dimension, perform phase space reconstruction operation on the high-quality Zeta potential cleaning time series signal, floc particle size distribution dynamic spectrum cleaning time series signal and interfacial tension change rate cleaning time series signal, and map the one-dimensional time series into a three-dimensional attractor trajectory dataset containing system state evolution information. S2.3: Using the generated three-dimensional attractor trajectory dataset, the box counting algorithm is used to calculate its geometric fractal dimension to obtain the box dimension. At the same time, the correlation integral method is used to analyze the spatial distribution density of trajectory points to obtain the correlation dimension. The small data volume method is used to estimate the divergence rate of adjacent orbits to obtain the maximum Lyapunov exponent, thereby extracting three topological invariants that characterize the chaotic characteristics of the reaction process. Based on the 3D attractor trajectory dataset generated by S2.2, the box counting algorithm was used (parameter: segmentation scale ε ranges from 10). -3 Up to 10 -1 (with a step size of 0.5 times the scaling factor), to achieve multi-scale segmentation of the trajectory space and count the number of non-empty boxes at each scale; Furthermore, the box dimension is obtained by calculating the geometric fractal dimension through proportional relationships. The calculation formula is as follows:

[0014] in, This refers to the number of non-empty boxes. The segmentation scale; The correlation integral method is used (parameter: minimum distance threshold). Value range 10 -3 Up to 10 -1 Using 20 discrete nodes, the distance statistical distribution between trajectory point pairs is realized, and the correlation integral function is calculated. The correlation dimension is obtained by fitting the slope using the following formula. :

[0015] in, For the distance between two points is less than The probability distribution function; The divergence rate of adjacent orbits is estimated using a small data approach (parameters: taking adjacent trajectory points within the embedding dimension, iteration step size n=10), and the maximum Lyapunov exponent is calculated according to the following formula. :

[0016] in, The initial track spacing, For the process Track spacing after time; Using the three algorithms mentioned above, the three-dimensional attractor trajectory dataset is transformed into three topological invariants: box dimension, correlation dimension, and maximum Lyapunov exponent, thereby enabling the extraction of chaotic characteristics in the reagent-sludge mixing reaction process. For example, in a sludge reagent dosing experiment at a wastewater treatment plant, the collected three-dimensional attractor trajectories included 1200 signal points across three categories: zeta potential changes, particle size distribution changes, and interfacial tension changes. The segmentation scale of the box counting algorithm... The values ​​range from 0.001 to 0.1, with a step size of 0.5 times the scaling factor; the number of non-empty boxes is statistically obtained. When the value is 0.005, the value is 350. Substituting this into the formula, the box dimension is calculated as follows: Distance threshold of correlation integral method The values ​​range from 0.001 to 0.1, and 20 nodes are discretely analyzed. The fitted slope yields the correlation dimension. The embedding dimension in the small data method is 5, the iteration step size of adjacent trajectory points is n=10, and the initial track spacing is... , It takes 1 second to pass The track spacing after time is The maximum Lyapunov exponent is calculated as follows: The three topological invariants output in this step are encapsulated and passed to S2.4 to generate the reaction time-series fingerprint vector under the current operating conditions. This ensures that the compensation analysis module can accurately perceive the chaotic dynamic characteristics of the reaction and significantly improve the response capability to time-delay changes in this operating scenario. S2.4: Perform vector encapsulation processing on the calculated box dimension, correlation dimension and maximum Lyapunov exponent, and arrange the three topological invariants in a fixed order to generate a reagent reaction time sequence fingerprint vector containing the unique characteristics of the reagent and sludge mixing reaction kinetics under the current operating conditions.

[0017] like Figure 2 As shown, step S3 involves using the reagent reaction time-series fingerprint vector generated in S2, combined with 128 sets of typical operating condition data covering pH values ​​of 5.5 to 8.5, temperatures of 10 to 35 degrees Celsius, sludge volume index of 50 to 250 ml / g, and mixed liquor suspended solids concentration of 2 to 12 g / L, to establish an offline fingerprint-operating condition mapping library. This is then used to form 16 fingerprint clusters through a clustering algorithm, and each cluster is bound to a time-series weight template for its influencing factors. Specifically, this includes: S3.1: Acquire multiple sets of typical working condition combination data covering the preset pH range, temperature range, sludge volume index gradient, and mixed liquor suspended solids concentration gradient. Perform batch reagent dosing experiments on each set of typical working condition combination data and simultaneously collect the corresponding reagent reaction time series fingerprint vector. Associate and store the collected reagent reaction time series fingerprint vector with the time series profile of the contribution of each influencing factor under the corresponding working condition to generate an original fingerprint working condition dataset containing the mapping relationship between multidimensional state parameters and kinetic features. Multiple sets of typical operating condition combinations covering preset pH ranges, temperature ranges, sludge volume index gradients, and mixed liquor suspended solids concentration gradients were obtained. Using the operating condition matrix generation method (parameters: pH∈[5.5,8.5], temperature∈[10,35]℃, SVI∈[50,250]mL / g, MLSS∈[2,12]g / L), all factors were arranged to form 128 experimental configurations. Furthermore, through a batch dosing experiment method (parameters: number of repetitions per group = 10, dosing concentration of the agent is fixed at the baseline value ± 5%), the mixing reaction of the agent and sludge under various working conditions was realized, and multi-dimensional time-series signals within 0–120 seconds after the agent was added were collected, and the turbidity change curve of the effluent was recorded simultaneously. Furthermore, the reaction time fingerprint extraction algorithm generated in the previous steps is invoked (input: preprocessed Zeta potential, floc particle size distribution dynamic spectrum, and interfacial tension change rate signal) to generate the reaction time fingerprint vector corresponding to each set of experiments, and obtain a three-dimensional set of topological invariants. ; Furthermore, by using the contribution curve construction method (parameters: measured values ​​of deviation of each influencing factor and corresponding turbidity improvement rate records), the time-series contribution profiles of pH, temperature, SVI and MLSS under operating conditions are generated, and a contribution sequence matrix with multiple time nodes is obtained. By using an associative storage processing method, an index mapping is established between the drug reaction time-series fingerprint vector and the multi-dimensional state parameters and contribution time-series profiles of the corresponding working conditions. This generates an original fingerprint working condition dataset containing the correlation between working condition parameters, chaotic dynamic characteristics and factor contribution, enabling multi-working condition data to support the subsequent construction of weight templates. For example, in the sludge treatment of the secondary sedimentation tank of a sewage treatment plant, 128 sets of operating conditions were selected: pH values ​​of 5.5, 6.0, 6.5, 7.0, 7.5, 8.0, and 8.5; temperature points of 10, 15, 20, 25, 30, and 35℃; SVI values ​​of 50, 100, 150, 200, and 250 mL / g; and MLSS values ​​of 2, 4, 6, 8, 10, and 12 g / L. The operating condition matrix was generated by combining all factors. In each set of operating conditions, a batch dosing test was conducted with a reagent concentration 5% higher than the baseline dosage. A high-temporal-resolution sensor array installed 10 cm downstream of the dosing point was used to collect Zeta potential, particle size distribution spectrum, and interfacial tension change rate signals from 0 to 120 seconds after reagent dosing. The box dimension, correlation dimension, and maximum Lyapunov exponent were calculated through phase space reconstruction (embedding dimension m=5, delay time τ determined by mutual information). For example, the fingerprint vector corresponding to a certain set of operating conditions was... The corresponding turbidity improvement rates were measured at t=10s, 30s, 60s, and 90s. Partial least squares regression was used to calculate the contribution of each factor at different time points. For example, at t=30s, the temperature contribution was... Finally, the fingerprint vector, along with the contribution sequence matrices of pH, temperature, SVI, and MLSS, and the operating condition parameters, are stored in the original fingerprint operating condition dataset, realizing the mapping output of multi-operating condition features and contributions in a real-world scenario for this sub-step. S3.2: Based on the reagent reaction time-series fingerprint vector contained in the original fingerprint working condition dataset, the partial least squares regression decomposition algorithm is used to decouple the local weights of each influencing factor at multiple discrete time nodes, extract the time-series weight sequence that characterizes the contribution of pH value, temperature, sludge volume index and mixed liquor suspended solids concentration to the effluent turbidity improvement rate at different reaction stages, and generate influencing factor time-series weight profile data with specific numerical distributions; S3.3: Using the drug reaction time-series fingerprint vectors in the original fingerprint working condition dataset as clustering input objects, an improved clustering analysis algorithm is executed to identify sample groups with similar chaotic dynamic topological invariant characteristics. Drug reaction time-series fingerprint vectors with similar box dimension, correlation dimension and maximum Lyapunov exponent spectrum are merged into the same category to generate a preliminary working condition classification structure composed of multiple independent fingerprint clusters. The drug reaction time-series fingerprint vectors are obtained from the original fingerprint condition dataset. An improved clustering analysis algorithm (with parameters including distance metric selection, initial cluster center number, and convergence iteration threshold) is used as the input processing object to achieve sample group identification with chaotic dynamic topological invariants as clustering features. Furthermore, a feature space is constructed by using the joint eigenvectors of box dimension, correlation dimension, and maximum Lyapunov exponent. The dimensional differences of various indicators are eliminated by using the feature standardization method (zero mean standard deviation normalization), and a feature matrix that can be used for distance calculation is obtained. Furthermore, by using a composite distance formula that weights and fuses improved Euclidean distance and topological spectral similarity, the similarity between fingerprint vectors is quantified, where the composite distance... The calculation formula is:

[0018] in, These are the weighting coefficients for each indicator. For fingerprint vector elements, As the cluster center element, The number of elements; Furthermore, based on the composite distance calculation results, a preliminary cluster center set is generated using a cluster center initialization strategy based on density peaks, and an iterative update algorithm is applied (number of iterations ≤ 200, convergence threshold ≤ 100). The cluster center positions are adjusted to generate clustering results with stable similarity segmentation; Furthermore, the clustering structure is validated by intra-cluster consistency evaluation indicators (such as average silhouette coefficient), and a category division structure with high consistency and significant inter-cluster differences is selected to form a preliminary working condition classification structure composed of multiple independent fingerprint clusters. By combining clustering analysis algorithms with topological invariant feature fusion processing, the time-series fingerprint vectors from the previous step are classified and transformed into clustered data with the ability to distinguish chaotic response patterns. This enables the initial grouping of reagent reaction patterns under different working conditions, providing a data foundation for subsequent binding and mapping of weight templates. For example, in a sampling experiment of sludge treatment process operation, the input is a raw dataset containing 128 fingerprint vectors. Each vector contains three elements: box dimension, correlation dimension, and maximum Lyapunov exponent, which are then normalized to zero mean and standard deviation. Weighting coefficients are set. The sample size is [0.4, 0.35, 0.25]. The distance is calculated by substituting the feature vector and the candidate cluster center vector into the composite distance formula, yielding the distance matrix from the sample to the cluster center. During the initialization of the density peak cluster centers, the distance threshold is set to 0.15, and the similarity weight adjustment coefficient is 1.2. The cluster center is iterated and updated 120 times, ultimately reaching the convergence threshold. Through verification using the average profile coefficient, 16 clusters were selected as preliminary classification structures for operating conditions. Under these experimental conditions, the resulting clusters can stably distinguish reagent reaction modes under different pH ranges, temperature intervals, and SVI gradients, significantly improving the accuracy of subsequent matching and the reliability of the dynamic weight template. S3.4: Based on the fingerprint clusters formed in the preliminary working condition classification structure, statistically match and average the impact factor time-series weight profile data generated in the previous steps with all drug reaction time-series fingerprint vectors in the fingerprint cluster, calculate the central weight distribution curve representing the typical response mode of the cluster, and generate target fingerprint cluster units bound with standardized time-series weight templates. Based on the fingerprint clusters in the preliminary working condition classification structure, the index matching method is used to establish a correspondence between the drug reaction time sequence fingerprint vectors contained in the cluster and the influence factor time sequence weight profile data generated in the previous step, so as to achieve accurate mapping between samples in the cluster and weight profiles. Furthermore, a statistical matching algorithm (parameter: matching tolerance is set to absolute difference of topological invariants ≤ 0.02) is used to calculate the coupling consistency coefficient between the fingerprint vector and the corresponding weight profile at each time point, and outlier samples with consistency coefficients lower than 0.85 are removed to obtain a highly reliable matching set. Furthermore, a mean-averaging method (parameter: weighting coefficients are allocated according to the consistency coefficient) is used to perform a weighted average operation on the weight values ​​of each time node in the high-reliability matching set, resulting in the center weight distribution curve of the fingerprint cluster. The weighted average formula is as follows:

[0019] in, The center weight value at time t. Let be the weight value of the i-th sample at time t. Let be the consistency coefficient of the i-th sample; Furthermore, the central weight distribution curve is smoothed and interpolated using a curve smoothing method (parameters: smoothing window width = 5 seconds, interpolation method = cubic spline) to eliminate noise fluctuations and fill in missing time nodes, forming a continuously defined weight time series. By using a binding mapping process, the smoothed center weight time series is encapsulated into a standardized time series weight template, and a bidirectional index relationship is established with the corresponding fingerprint cluster unit to realize the weight calling function of the cluster in future matching. For example, in a sludge treatment experimental scenario, the pH range was 6.8 to 7.2, the temperature range was 22 to 24 degrees Celsius, the SVI was 120 to 135 ml / g, and the MLSS was 4.5 to 5.0 g / L. The fingerprint cluster corresponding to this scenario contained 12 sets of reagent reaction time sequence fingerprint vectors and corresponding influencing factor weight profile data. After consistency coefficient calculation, two samples with a coefficient below 0.85 were removed, and the remaining 10 samples underwent weighted averaging. The temperature weight at a specific 30-second time point was calculated as follows: Substituting the temperature weight values ​​and consistency coefficients of each sample, the center weight value is obtained as 0.412. After three spline smoothing processes, a continuous temperature weight curve covering 0 to 120 seconds is formed, and a bidirectional index mapping is established with the fingerprint cluster unit. When applied to runtime matching, it can significantly improve the timing allocation accuracy of the compensation amount and ensure the stability and response consistency of the drug dosing control. S3.5: Integrate all generated target fingerprint cluster units and their bound standardized temporal weight templates to construct an offline fingerprint condition mapping library containing index key values ​​and mapping relationships. Deploy this offline fingerprint condition mapping library to the storage medium of the compensation analysis module to form a benchmark knowledge base that supports fast retrieval of matches using the improved Wasserstein distance algorithm. Obtain the target fingerprint cluster unit and its bound standardized temporal weight template generated in the preceding step S3.4, and use them as input objects for mapping construction. Employ a key-value index generation method (the index key is composed of cluster identifier, box dimension, association dimension, and maximum Lyapunov exponent) to achieve a unique and searchable marker for the fingerprint cluster unit. Furthermore, a mapping table is generated between the target fingerprint cluster unit and its corresponding standardized time-series weight template through a mapping relationship construction algorithm (parameters: set of index key values, set of time-series weight templates, and mapping mode set to one-to-one binding), and the basic mapping relationship matrix data is obtained. Furthermore, an association structure optimization method (parameters: index reference order, storage address allocation strategy, retrieval priority) is adopted to optimize the physical location of the mapping relationship matrix in storage and generate an index linked list structure that can be quickly retrieved; Furthermore, by using a storage format encapsulation method (parameters: data format specification, check code generation rules), a unified encapsulation of the mapping relationship matrix and index linked list is achieved, and an offline fingerprint condition mapping library file that can be directly deployed on the storage medium of the compensation analysis module is generated. Furthermore, by utilizing the improved Wasserstein distance algorithm interface registration method (parameters: call path, retrieval timeout, matching accuracy threshold), the offline fingerprint condition mapping library and the real-time condition matching function are quickly connected, and a benchmark knowledge base with second-level response capability is obtained. By constructing and optimizing the mapping relationship, the target fingerprint cluster unit and the standardized temporal weight template in the previous step are transformed into an indexed knowledge structure that supports fast matching of the improved Wasserstein distance algorithm, thereby achieving real-time technical effects for working condition identification and compensation weight invocation. For example, in the offline database construction stage of the sludge treatment process, the index key value is set to consist of cluster ID, , , The cluster ID is generated by combining four elements, with the cluster ID ranging from 1 to 16. and Round to three decimal places. Two decimal places are retained. A mapping algorithm is used to bind each index key value with a corresponding time-series weight template, which contains contribution curve data of pH value, temperature, and sludge volume index at multiple discrete time points. The mapping matrix is ​​stored in solid-state storage medium using an association structure optimization method, and its retrieval time is controlled within 5 milliseconds using a linked list structure. In the improved Wasserstein distance retrieval, the matching accuracy threshold is set to 0.002, and the retrieval timeout is set to 10 milliseconds. Validation results show that when the matching call is triggered by real-time operating conditions during operation, fingerprint cluster identification and corresponding weight template invocation can be completed within 8 milliseconds, realizing a rapid response of the compensation analysis module and significantly improving the real-time performance and accuracy of reagent dosing control.

[0020] like Figure 3 As shown, step S4 involves generating a real-time reagent reaction time-series fingerprint vector for the current operating condition during sludge treatment operation. This vector is then compared with the offline fingerprint-operating condition mapping library established in S3 using an improved Wasserstein distance algorithm. After identifying the fingerprint cluster with the highest matching degree, its associated time-series weight template is retrieved. Specifically, this includes: S4.1: Acquire the real-time collected Zeta potential time series signal, floc particle size distribution dynamic spectrum time series signal, and interfacial tension change rate time series signal. Based on phase space reconstruction technology, perform embedding dimension setting and delay time calculation processing on the three types of time series signals to generate a current working condition reagent reaction time series fingerprint vector containing box dimension, correlation dimension and maximum Lyapunov exponent. The real-time Zeta potential time-series signal, floc particle size distribution dynamic spectrum time-series signal, and interfacial tension change rate time-series signal output by the in-situ sensing array were acquired. A multi-channel data synchronization calibration method (parameters: unified time reference, sampling accuracy 10 Hz) was adopted to achieve accurate time alignment of the three types of time-series signals at the sampling start time. Furthermore, by using a mutual information function calculation method (parameters: scanning delay interval 1–50 sample points, step size 1 sample point), the redundancy information and dynamic correlation of each type of time-series signal are measured, and the optimal delay time is obtained. Values ​​are used for phase space reconstruction; Furthermore, by adopting the embedding dimension setting method (parameter: m=5, dimension saturation is verified according to the Cao rule), the three types of one-dimensional time series signals with time alignment and determined delay time are mapped into a multi-dimensional phase space trajectory matrix containing system state evolution information. Furthermore, using the box counting algorithm (parameter: segmentation scale range 2) -1 Up to 2 -6 This allows for the geometric fractal analysis of three-dimensional attractor trajectories and the determination of the box dimension. ; Furthermore, the correlation integral method is adopted (parameter: relative distance threshold). (Range 0.01–0.2) to analyze the distribution density of trajectory point pairs in phase space and obtain the correlation dimension. ; Furthermore, the small data divergence rate method (parameters: sample size ≤ 200, orbital separation distance threshold) was used. = 0.1), to calculate the exponential divergence rate of adjacent orbits of the attractor trajectory, and to obtain the maximum Lyapunov exponent. ; By using vector assembly processing, the box dimension, correlation dimension, and maximum Lyapunov exponent are encapsulated in a fixed order into a reagent reaction time sequence fingerprint vector, thereby realizing a unique identifier of the kinetic characteristics of the reagent and sludge mixing reaction under the current operating conditions; For example, under certain sludge treatment process operating conditions, the real-time sensor array collects Zeta potential signals within the range of... 18 mV to At 8 mV, the main peak of the flocculent particle size distribution dynamic spectrum is located at 120 μm, and the amplitude range of the interfacial tension change rate is 0.05–0.12 mN / m / s. The mutual information method was used to scan the delay interval 1–50 sample points, obtaining optimal delay times of 7, 10, and 8 sample points for the three types of signals, respectively. The embedding dimension was set to 5 after verification using the Cao rule. Phase space reconstruction was performed to generate a three-dimensional attractor trajectory matrix, and the box-counting algorithm was used to calculate the box dimension. The correlation dimension is obtained by the correlation integral method. The maximum Lyapunov exponent is calculated using the small data method. The fingerprint vector [2.14, 1.87, 0.042] was encapsulated. This vector had the highest matching degree with the ID#07 cluster in the historical fingerprint database in subsequent comparisons and was used to call the time-series weight template bound to it, thereby realizing the dynamic adjustment of the drug dosage by the compensation analysis module. S4.2: Obtain the current working condition drug reaction time sequence fingerprint vector and the historical fingerprint cluster center vector stored in the offline fingerprint-working condition mapping library. Use the improved Wasserstein distance algorithm to perform probability distribution difference measurement processing on the current working condition drug reaction time sequence fingerprint vector and the center vector of each historical fingerprint cluster, so as to output a set of distance measurement values ​​that characterize the similarity between the current working condition and each historical fingerprint cluster. The current working condition reagent reaction time sequence fingerprint vector generated by S4.1 and the center vectors of each historical fingerprint cluster stored in the offline fingerprint working condition mapping library constructed by S3.5 are obtained. An improved Wasserstein distance calculation method (parameter settings include the number of sample points n≥3, the number of distribution segments k=50, and the smoothing coefficient δ=0.01) is used to realize the quantitative measurement of the probability distribution difference between each vector. Furthermore, a cumulative distribution curve is constructed through a probability distribution matching function, and the area of ​​the difference region between the current fingerprint vector and the target center vector is solved using the piecewise integration method, and the distance value based on the comprehensive correction of the first moment and the second moment is obtained. Furthermore, the single-dimensional Wasserstein distance component is calculated using the following MathML formula. :

[0021] in and Let these represent the cumulative distribution functions of the current fingerprint vector and the centers of historical fingerprint clusters, respectively. This represents the total number of sampling points; Furthermore, the total distance value is calculated using the Euclidean norm of the multidimensional component distance vectors. :

[0022] in Let be the Wasserstein component of the j-th dimension; Furthermore, weighted normalization is used to scale each distance value to ensure that the distance metrics under different feature dimensions are comparable, and a set of distance metric values ​​containing the corresponding distance values ​​of all target clusters is generated. By using the improved Wasserstein distance algorithm, the fingerprint vector distribution differences from the previous step are transformed into a multi-dimensional distance index, thereby achieving a quantitative expression of the similarity between the current working condition and historical fingerprint clusters. For example, in a sludge treatment system, the current operating condition reagent reaction time-series fingerprint vector is [1.52, 2.11, 0.336], one of the historical fingerprint cluster center vectors is [1.48, 2.09, 0.340], the number of sampling points n is 100, the number of segments k is 50, and the smoothing coefficient δ is 0.01. The cumulative distribution functions F and G are calculated, and the W components obtained by applying the above formula on each dimension are 0.009, 0.012, and 0.005, respectively. These components are then applied to the total distance calculation formula, resulting in D of 0.0161. After normalization adjustment, this value in the distance value set corresponding to this cluster is 0.82, indicating a high degree of similarity. During system operation, this distance value will participate in the minimum distance optimization principle of S4.3 to achieve accurate fingerprint cluster matching, improve the accuracy of weight template calling, and significantly improve the time-series response performance of reagent dosing control in subsequent compensation calculations. S4.3: Obtain a set of distance metrics, and perform extreme value retrieval processing on the set of distance metrics based on the minimum distance selection principle to identify the target fingerprint cluster identifier that has the highest matching degree with the current working condition drug reaction time sequence fingerprint vector; S4.4: Obtain the target fingerprint cluster identifier, and perform an index query on the target fingerprint cluster identifier according to the pre-established binding mapping relationship between the fingerprint cluster and the time-series weight template, so as to retrieve the time-series weight template bound to the target fingerprint cluster, which includes the contribution of pH value, temperature and sludge volume index at different times.

[0023] Step S5: Based on the time-series weight template retrieved in S4, the weights of pH at 60 seconds, temperature at 30 seconds, and sludge volume index at 90 seconds are dynamically allocated, and the reagent dosage compensation amount is generated by combining the nonlinear response function of each factor deviation. Specifically, this includes: S5.1: Obtain the time-series weight template bound to the fingerprint cluster with the highest matching degree output by S4, and perform time axis parsing processing on the time-series weight template to extract the first weight coefficient of pH value at 60 seconds, the second weight coefficient of temperature at 30 seconds, and the third weight coefficient of sludge volume index at 90 seconds. Data loading processing is performed on the time-series weight template bound to the target fingerprint cluster with the highest matching degree output by S4 to obtain a structured weight curve dataset containing the contribution of multiple factors at different time nodes. A time axis parsing algorithm (parameters: sampling time resolution 1 second, time axis range 0–120 seconds) is used to perform time series expansion processing on the above structured weight curve dataset and generate a multi-factor weight index table arranged in timestamp order. Furthermore, by using a time node positioning algorithm (parameters: target time nodes are 60 seconds, 30 seconds, and 90 seconds respectively), the function of retrieving specific nodes in the multi-factor weight index table is realized, and the initial set of weight coefficient values ​​for the corresponding time nodes is obtained. Furthermore, a factor label mapping method is adopted (rule: the factor labels of the weight curve correspond one-to-one with the operating condition influencing factors) to realize the factor attribution analysis of the initial value set of weight coefficients at time nodes, and generate accurate values ​​of pH value as the first weight coefficient, temperature as the second weight coefficient, and sludge volume index as the third weight coefficient. By using a coefficient mapping process, the three types of weight coefficients retrieved in the previous step are transformed into a standardized weight parameter vector, achieving a numerical format consistent with the input interface of the subsequent nonlinear response function. For example, under a certain sludge treatment condition, the pH value weighting coefficient of the target fingerprint cluster binding weight template with the highest matching degree at time 60 seconds is: The temperature weighting coefficient at 30 seconds is The sludge volume index weighting coefficient at 90 seconds is The timeline parsing algorithm uses a sampling time resolution. The weight curve is unfolded in seconds, and the corresponding coefficients are retrieved from the three time points mentioned above. The factor label mapping method will... Identified as the first weighting factor for pH value, Identified as the second weighting factor for temperature, These are identified as the third weighting coefficients of the sludge volume index. These coefficients are encapsulated as a weight parameter vector. This is directly used for the nonlinear response mapping of the deviation in the subsequent step S5.2, ensuring that the input weights of the compensation calculation are accurately matched with the time characteristics of the working condition, and improving the time sensitivity and factor contribution accuracy of the compensation calculation. S5.2: Based on the real-time collected sludge state data and the preset standard operating condition threshold, calculate the current pH value deviation, the current temperature deviation, and the current sludge volume index deviation respectively, and perform normalization processing on the three deviations to obtain a standardized pH value deviation vector, a standardized temperature deviation vector, and a standardized sludge volume index deviation vector. S5.3: Using the S-shaped curve obtained by fitting historical operating data as a nonlinear response function, nonlinear mapping transformation is performed on the standardized pH value deviation vector, standardized temperature deviation vector, and standardized sludge volume index deviation vector to generate pH value nonlinear response component, temperature nonlinear response component, and sludge volume index nonlinear response component. S5.4: Multiply the first weighting coefficient with the pH nonlinear response component to obtain a pH weighted compensation term, multiply the second weighting coefficient with the temperature nonlinear response component to obtain a temperature weighted compensation term, and multiply the third weighting coefficient with the sludge volume index nonlinear response component to obtain a sludge volume index weighted compensation term. S5.5: Perform a cumulative summation operation on the pH-weighted compensation item, temperature-weighted compensation item, and sludge volume index-weighted compensation item to synthesize the final reagent dosage compensation amount, and output the reagent dosage compensation amount to the actuator to correct the flow rate setpoint of the reagent dosing pump.

[0024] Step S6: Input the reagent dosage compensation amount generated in S5 into the actuator to control the output flow rate of the reagent dosing pump, so that the actual dosage dynamically matches the real-time requirements of the sludge treatment process. Specifically, this includes: S6.1: Perform safety threshold verification and amplitude limiting on the agent addition compensation amount generated in S5 to eliminate abnormal compensation values ​​that exceed the rated operating range of the agent addition pump and obtain the target flow increment signal corrected by safety constraints; The reagent dosage compensation amount generated by S5 is used as the input signal, and an anomaly detection algorithm is employed (parameter: minimum rated flow rate of the dosing pump). and rated maximum flow rate This enables the parsing of the upper and lower limits of the compensation amount, and the construction of a safety threshold interval dataset; Furthermore, through a limiting algorithm (parameter: threshold range) This allows for the elimination and replacement of compensation values ​​that exceed the limit, and the acquisition of a set of candidate compensation values ​​after safety limits. Furthermore, a numerical constraint optimization method is adopted (parameter: the anomaly replacement strategy is boundary value replacement) to replace the outliers in the compensation quantity candidate values ​​with the boundary flow values ​​of the corresponding safe interval, and generate a safe candidate sequence of compensation quantities without anomalies; Furthermore, a signal smoothing algorithm (parameter: moving average window width w=3) is used to achieve dynamic smoothing of the compensation amount safety candidate sequence, so as to eliminate the impact on the actuator caused by sudden adjustment and generate a smoothed target flow increment signal. By using the above-mentioned limiting and smoothing methods, the compensation data from the previous step is transformed into a target flow increment signal that meets the safety operating range constraints of the drug dosing pump and has stable transition characteristics, thereby achieving the expected technical effect of avoiding mechanical shock to the pump body and instability of the control response. For example, in the sludge treatment section of a wastewater treatment plant, the rated minimum flow rate of the dosing pump is configured as 4.5 L / min, and the rated maximum flow rate is configured as 8.2 L / min. The system calculates a reagent dosing compensation amount of 9.0 L / min. An anomaly detection algorithm identifies this value as exceeding the safe threshold range of the rated maximum flow rate of 8.2 L / min. Through threshold limiting, the excess value of 9.0 L / min is replaced with 8.2 L / min, generating a candidate compensation value sequence [7.6, 8.2, 7.8] after safety threshold limiting. Further smoothing filtering with a moving average window width w=3 is applied to perform smoothing calculations on the above sequence. The smoothing process formula is as follows:

[0025] in, To compensate for candidate values, , , These represent the previous sample value, the current sample value, and the next sample value at the current time point, respectively. The target flow rate increment signal is used. In this embodiment, the smoothed calculation result outputs a target flow rate increment signal of 8.0 L / min. After being applied to the superposition calculation in the subsequent S6.2 step, the stability of the added flow rate response is significantly improved, avoiding pump overload and flow oscillation caused by over-limit compensation. S6.2: The target flow increment signal corrected by safety constraints is superimposed with the basic set flow of the current sludge treatment process to synthesize an instantaneous target total flow command that reflects the real-time operating conditions. Based on the target flow increment signal corrected by safety constraints and the basic set flow parameters of the current sludge treatment process, a numerical superposition operation method (operation mode: floating-point double precision accumulation) is used to achieve accurate synthesis of the two types of flow data. Furthermore, by using a vectorized flow synthesis algorithm (parameters: target flow increment signal vector, basic set flow vector), the elements of the two types of vectors are summed one by one, and a flow synthesis result set bound to the time index is obtained; Furthermore, a traffic merging processing method based on time series consistency constraints is adopted (constraint condition: timestamp matching accuracy ≤ 0.001 seconds) to achieve synchronous fusion of traffic data from different sources under a unified time series benchmark, and to generate a preliminary numerical matrix of the instantaneous target total traffic command; Furthermore, the flow synthesis process is expressed using mathematical formulas, and the instantaneous target total flow command is calculated using the following formula:

[0026] in, The command is for the instantaneous target total flow rate. Set the traffic based on the base. The target flow increment signal is corrected for safety constraints; Furthermore, by using a low-latency cache write algorithm (cache time setting: ≤5 ms), the instantaneous target total flow instruction value matrix is ​​written to the execution control bus to ensure that the downstream execution module can receive and respond to the instruction within milliseconds. Through the above superposition calculation and buffer transmission processing method, the target flow increment signal of the previous step is transformed into complete instantaneous target total flow command data, realizing the real-time generation and operating condition matching of the agent dosing pump flow command; For example, in the sludge treatment process of the secondary sedimentation tank in a wastewater treatment plant, the basic set flow rate is 5.000 L / min, and the target flow rate increment signal after safety constraint correction is 0.275 L / min. Using double-precision cumulative calculation, the instantaneous target total flow rate command is... = L / min. During the flow vectorization process, the time index precision was set to 0.001 seconds to ensure the timing consistency of the synthesized results. During the low-latency cache writing process, the instantaneous target total flow command of 5.275 L / min was written to the PLC control bus. The execution module responded within 3 ms, and the flow feedback value of the reagent dosing pump stabilized within the range of 5.275 L / min ± 0.002 L / min, verifying that this step can significantly improve the accuracy of flow command generation and execution response speed in real-world scenarios. S6.3: The deviation between the instantaneous target total flow command and the actual operating flow fed back by the built-in flow meter of the reagent dosing pump is calculated to generate a flow error vector characterizing the difference between the current dosing state and the target state; S6.4: Perform proportional-integral-derivative (PID) adjustment algorithm on the flow error vector to convert the flow deviation into a pulse width modulation duty cycle control parameter with dynamic response characteristics; S6.5: The variable frequency motor of the reagent dosing pump is driven by the pulse width modulation duty cycle control parameter to perform speed adjustment action, so as to change the pump impeller rotation speed and finally output the precise reagent dosing flow rate that matches the real-time requirements of the sludge treatment process.

[0027] Step S7: Every 24 hours, detect the deviation between the drug reaction time sequence fingerprint vector generated in S2 and the center of its corresponding fingerprint cluster, and determine whether the deviation exceeds a preset threshold for 5 consecutive times. If the condition is met, trigger a fingerprint drift signal. Specifically, this includes: S7.1: Trigger detection at a fixed 24-hour cycle, calculate the Euclidean distance between the reagent reaction time sequence fingerprint vector collected in the current operating cycle and the reference coordinates of the fingerprint cluster center in the offline fingerprint condition mapping library, so as to obtain a real-time deviation value sequence that characterizes the degree of deviation of the current operating condition. S7.2: Execute sliding window counting logic based on real-time deviation value sequence, count the cumulative number of five consecutive real-time deviation values ​​exceeding the preset drift threshold, and generate a drift judgment flag bit to determine whether the system has experienced significant operating condition drift; The input conditions include the real-time deviation value sequence output by S7.1, which is obtained by calculating the Euclidean distance between the drug reaction timing fingerprint vector and the reference coordinates of the center of the fingerprint cluster in the current running cycle; A sliding window counting method (parameter: window length = 5 sampling points) is adopted to perform window sliding operation on the real-time deviation value sequence, extract data groups of five consecutive deviation values ​​within each window, and realize time series alignment of group boundaries; Furthermore, a threshold comparison algorithm (parameter: the drift threshold is determined by the 95th percentile of the historical stable operating condition error distribution) is used to compare the deviation values ​​in each window point by point. Deviation values ​​exceeding the threshold are marked as 1, and those below the threshold are marked as 0, thus generating a binary decision sequence within the window. Furthermore, the cumulative count is calculated by performing a summation operation on the binary decision sequence within the window using a cumulative counting function, and then using the following MathML formula:

[0028] in, To accumulate the number of times, This is a binary determination function for the deviation value within the window. For the position index within the window; Furthermore, through a logical decision-making algorithm (condition: Determine whether the current window meets the drift detection condition of exceeding the threshold five times consecutively, and generate a drift determination flag for the time period corresponding to the window that meets the condition; By using sliding window counting and threshold comparison, the real-time deviation value sequence from the previous step is transformed into a drift judgment flag, thereby enabling automatic detection and triggering of significant operating condition drift. For example, in a sludge treatment system, a drift threshold of 0.85 (unit: Euclidean distance) is set, and the sliding window length is 5 sampling points. A sliding window is performed on the real-time deviation value sequence collected within a day. Assuming the deviation values ​​within the window are [0.91, 0.88, 0.90, 0.86, 0.87], a binary decision sequence [1,1,1,1,1] is generated using a threshold comparison algorithm. The cumulative counting formula is then used to calculate... The logic judgment algorithm outputs a drift judgment flag of 1. When multiple windows output the drift judgment flag, the system will trigger a single operating condition drift signal and enter the S7.3 small sample incremental learning dataset construction process. In this scenario, the accuracy of drift detection is significantly improved, enabling rapid capture of significant changes in drug reaction kinetics characteristics and timely correction of the time-series weight template. S7.3: Responding to the trigger signal of the drift determination flag, extract local drug reaction time fingerprint feature data within the time interval of 40 to 100 seconds from the newly acquired time series signal to construct a small sample incremental learning dataset for correcting model parameters; S7.4: Use the small-sample incremental learning dataset to perform gradient descent optimization iterations on the temporal weight template parameters bound to the drifting fingerprint clusters to output updated temporal weight template parameters that only cover the time interval from forty seconds to one hundred seconds. S7.5: Write the updated timing weight template parameters into the storage unit of the corresponding fingerprint cluster in the offline fingerprint condition mapping library to complete the dynamic correction of the drug reaction timing fingerprint library and restore the system's adaptive compensation capability.

[0029] Step S8: In response to the fingerprint drift signal triggered by S7, small-sample incremental learning is performed on the drifting fingerprint clusters using the newly acquired time-series signal. Only the time-series weight template parameters within the 40-second to 100-second interval are updated to complete the dynamic correction of the drug reaction time-series fingerprint database. Specifically, this includes: S8.1: Obtain the drift signal triggered by the fingerprint drift detection module and the corresponding fingerprint cluster identifier, extract the center vector of the fingerprint cluster and the drug reaction time sequence fingerprint vector sequence of five consecutive deviations, and construct a small sample incremental learning dataset containing the latest working condition features; S8.2: Based on the small sample incremental learning dataset, three types of time-series signals are extracted within a time window of 40 to 100 seconds after drug injection: Zeta potential, dynamic spectrum of floc particle size distribution, and rate of change of interfacial tension. Sliding window resampling processing is performed to generate local dynamic response signal segments with high temporal resolution. S8.3: Utilize local dynamic response signal segments, reuse phase space reconstruction technology to set the embedding dimension and delay time, recalculate the box dimension, correlation dimension and maximum Lyapunov exponent, and generate a corrected local drug response time-series fingerprint vector to characterize the dynamic features of the current drift state; S8.4: Based on the corrected local agent reaction time fingerprint vector, the partial least squares regression decomposition algorithm is called to analyze the time profile of the contribution of pH value, temperature and sludge volume index to the effluent turbidity improvement rate in the interval of 40 seconds to 100 seconds, and calculate the updated local factor contribution matrix. S8.5: Based on the updated local factor contribution matrix, replace the weight parameters in the 40-100 second interval of the original time-series weight template, generate the corrected time-series weight template, and write it into the corresponding fingerprint cluster in the offline fingerprint condition mapping library to complete the dynamic correction of the drug reaction time-series fingerprint library.

[0030] The present invention also provides a composite compensation system for sludge treatment with precise dosing of chemicals, which uses the above-mentioned composite compensation method for sludge treatment with precise dosing of chemicals to compensate for the dosing of chemicals during the sludge treatment process.

[0031] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0032] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A composite compensation method for sludge treatment aimed at precise dosing of chemicals, characterized in that, Includes the following steps: S1: Deploy a high time resolution in-situ sensing array at the sludge treatment agent dosing point to synchronously collect time-series signals of Zeta potential, floc particle size distribution dynamic spectrum, and interfacial tension change rate after agent injection. S2: Based on the time-series signal, attractor traject ... S3: Using the drug reaction time sequence fingerprint vector and typical working condition data, establish an offline fingerprint-working condition mapping library; S4: During the sludge treatment operation, the reagent reaction time sequence fingerprint vector of the current working condition is generated in real time, and the reagent reaction time sequence fingerprint vector of the current working condition is compared with the offline fingerprint-working condition mapping library. After identifying the fingerprint cluster with the highest matching degree, its bound time sequence weight template is retrieved. S5: Based on the time-series weight template, dynamically allocate the pH weight, temperature weight, and sludge volume index weight, and generate the reagent addition compensation amount by combining the nonlinear response function of each factor deviation. S6: Input the reagent dosage compensation amount into the actuator to control the output flow rate of the reagent dosing pump, so that the actual dosage dynamically matches the real-time requirements of the sludge treatment process; S7: Detect the real-time deviation between the drug reaction timing fingerprint vector collected in the current running cycle and the center of the fingerprint cluster to which it belongs, and determine whether the deviation exceeds the preset drift threshold for 5 consecutive times. If the condition is met, trigger the fingerprint drift signal.

2. The sludge treatment composite compensation method for precise dosing of reagents according to claim 1, characterized in that, The process following step S7 also includes: S8: In response to the fingerprint drift signal, perform small-sample incremental learning on the drifting fingerprint cluster using the newly acquired time-series signal, and only update the time-series weight template parameters in the interval of 40 seconds to 100 seconds to complete the dynamic correction of the drug reaction time-series fingerprint database.

3. The sludge treatment composite compensation method for precise dosing of reagents according to claim 1, characterized in that, The topological invariants include the box dimension, the correlation dimension, and the maximum Lyapunov exponent.

4. The sludge treatment composite compensation method for precise dosing of reagents according to claim 1, characterized in that, In the generation of attractor trajectories through phase space reconstruction technology, the embedding dimension is set to 5, and the delay time parameter is calculated by the mutual information function.

5. The sludge treatment composite compensation method for precise dosing of reagents according to claim 1, characterized in that, Step S3 specifically includes: Acquire typical working condition combination data, perform batch drug addition experiments on each typical working condition combination data and simultaneously collect the corresponding drug reaction time sequence fingerprint vector, associate and store the drug reaction time sequence fingerprint vector with the time sequence profile of the contribution of each influencing factor under the corresponding working condition, and generate the original fingerprint working condition dataset. Based on the drug reaction time-series fingerprint vector contained in the original fingerprint working condition dataset, the local weights of each influencing factor at multiple discrete time nodes are decoupled and calculated to generate influencing factor time-series weight profile data. Using the drug reaction time-series fingerprint vectors in the original fingerprint working condition dataset, an improved clustering analysis algorithm is executed to identify sample groups with similar chaotic dynamic topological invariant characteristics. Drug reaction time-series fingerprint vectors with similar box dimension, correlation dimension, and maximum Lyapunov exponent spectrum are merged into the same category to generate a preliminary working condition classification structure. Based on the fingerprint clusters formed in the preliminary working condition classification structure, the time-series weight profile data of the influencing factors are statistically matched and averaged with all drug reaction time-series fingerprint vectors in the current specific fingerprint cluster in each fingerprint cluster. The center weight distribution curve representing the typical response mode of the specific fingerprint cluster is calculated, and a target fingerprint cluster unit bound with a standardized time-series weight template is generated. Integrate all generated target fingerprint cluster units and their bound standardized temporal weight templates to construct an offline fingerprint condition mapping library.

6. The sludge treatment composite compensation method for precise dosing of reagents according to claim 5, characterized in that, The typical operating condition combination data covers preset pH range, temperature range, sludge volume index gradient, and mixed liquor suspended solids concentration gradient.

7. The sludge treatment composite compensation method for precise dosing of reagents according to claim 1, characterized in that, Step S4 specifically includes: The real-time acquired Zeta potential time-series signal, floc particle size distribution dynamic spectrum time-series signal, and interfacial tension change rate time-series signal are processed by setting the embedding dimension and calculating the delay time based on phase space reconstruction technology to generate the current working condition reagent reaction time-series fingerprint vector. Obtain the current working condition drug reaction time sequence fingerprint vector and the historical fingerprint cluster center vector stored in the offline fingerprint-working condition mapping library. Use the improved Wasserstein distance algorithm to perform probability distribution difference measurement processing on the current working condition drug reaction time sequence fingerprint vector and each historical fingerprint cluster center vector, and output a set of distance measurement values. Based on the minimum distance optimization principle, extreme value retrieval processing is performed on the distance metric value set to identify the target fingerprint cluster identifier with the highest matching degree with the current working condition drug reaction time sequence fingerprint vector. Based on the pre-established binding mapping relationship between fingerprint clusters and time-series weight templates, the target fingerprint cluster identifier is indexed and queried to retrieve the time-series weight templates that are bound to the target fingerprint clusters and contain the contributions of pH value, temperature and sludge volume index at different times.

8. The sludge treatment composite compensation method for precise dosing of reagents according to claim 1, characterized in that, Step S5 specifically includes: Obtain the temporal weight template of the fingerprint cluster with the highest matching degree output in step S4, perform time axis parsing processing on the temporal weight template, and extract the first weight coefficient of pH value at 60 seconds, the second weight coefficient of temperature at 30 seconds, and the third weight coefficient of sludge volume index at 90 seconds. Based on real-time collected sludge state data and preset standard operating condition thresholds, the current pH value deviation, current temperature deviation, and current sludge volume index deviation are calculated and normalized to obtain standardized pH value deviation vector, standardized temperature deviation vector, and standardized sludge volume index deviation vector. Using the S-shaped curve obtained by fitting historical operating data as a nonlinear response function, nonlinear mapping transformations are performed on the standardized pH value deviation vector, the standardized temperature deviation vector, and the standardized sludge volume index deviation vector to generate pH value nonlinear response components, temperature nonlinear response components, and sludge volume index nonlinear response components. The first weighting coefficient is multiplied by the pH nonlinear response component to obtain the pH weighted compensation term; the second weighting coefficient is multiplied by the temperature nonlinear response component to obtain the temperature weighted compensation term; and the third weighting coefficient is multiplied by the sludge volume index nonlinear response component to obtain the sludge volume index weighted compensation term. The pH-weighted compensation term, the temperature-weighted compensation term, and the sludge volume index-weighted compensation term are summed to synthesize the final reagent dosage compensation amount.

9. The sludge treatment composite compensation method for precise dosing of reagents according to claim 8, characterized in that, Step S5 further includes extracting pH weight at 60 seconds, temperature weight at 30 seconds, and sludge volume index weight at 90 seconds from the retrieved time-series weight template. The real-time deviations of each factor are normalized and nonlinearly mapped from the historical S-curve, and then multiplied by the weight coefficients. The results are then summed to form the final reagent dosage compensation amount.

10. A composite compensation system for sludge treatment aimed at precise dosing of chemicals, characterized in that: The sludge treatment compound compensation method for precise dosing of chemicals described in any one of claims 1-9 is used to compensate for the dosing of chemicals during the sludge treatment process.

Images