Dispersant suspension system anti-layering stability prediction method and system

Abstract

The invention relates to the technical field of dispersant suspension system monitoring, and discloses a dispersant suspension system anti-layering stability prediction method and system.The method comprises the steps that an original time sequence turbidity data flow of a dispersant suspension system is collected, a microscopic agglomeration index is calculated after sliding window processing, and probability density distribution is generated; performing asymmetric Gaussian fitting on the probability density distribution to separate characteristic components, and calculating an attenuation amount of an interface coupling factor; building a healthy phase space of the suspension system, mapping the indexes into real-time state points, and calculating a Mahalanobis distance and direction angle sequence; finally, judging a dominant risk type and predicting a macroscopic layering time node; according to the method, the microcosmic features are deeply excavated by constructing the multi-dimensional phase space, small dynamic changes and interface failure signals can be captured in advance when the macroscopic turbidity value is pseudo-stable, and the sensitivity and foresight of stability prediction are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of dispersant suspension system monitoring technology, and more specifically, to a method and system for predicting the anti-stratification stability of dispersant suspension systems. Background Technology

[0002] Dispersant suspension systems, such as pesticide suspensions and paint pastes, are heterogeneous, thermodynamically unstable systems formed by the uniform dispersion of solid particles in a liquid medium with the aid of dispersants. Maintaining the physical stability of the suspended particles and preventing irreversible aggregation, sedimentation, and stratification during long-term storage and transportation is crucial to ensuring the release of pesticide efficacy and the performance of the product. Currently, the monitoring of the stability of suspension systems has shifted from simple end-product quality testing to dynamic prediction throughout the entire lifecycle. This aims to predict the deterioration trend of the system in advance through real-time monitoring data, enabling timely formulation adjustments or inventory intervention.

[0003] To assess the stability of suspended systems, various detection methods have emerged in the prior art. For example, Chinese patent CN120197407B discloses a method and system for predicting and optimizing the dispersion stability of herbicide suspensions. This method accelerates the deterioration process of the suspension by simulating various environmental stress conditions (such as temperature, humidity, and water hardness) and measures turbidity data in real time during the deterioration process. The rate of turbidity change is used to assess dispersion stability, focusing on addressing the problem of assessing the impact of environmental factors on stability. Furthermore, Chinese patent application CN121275571A discloses a method for testing the performance of macromolecular dispersants used in the production of polymer polyols. This method focuses on the manufacturing process, constructing multidimensional indicators by collecting dispersion time, process energy consumption, particle size distribution (D50, D90), and rheological parameters, thereby achieving intelligent diagnosis and optimization of the dispersant production process performance.

[0004] However, the aforementioned technologies primarily rely on significant changes in macroscopic physical quantities as the basis for judgment, exhibiting lag and blind spots when facing the microscopic dynamic evolution that occurs in the initial stage of the dispersant suspension system. Specifically, when the small-molecule encapsulation ability of the dispersant begins to fail, the "adhesion" between suspended particles often begins at the microscopic level. At this time, although the particles approach each other through van der Waals forces, forming tiny agglomeration precursors, they remain suspended in the liquid medium and have not yet undergone macroscopic sedimentation. From the perspective of light scattering principles, this microscopic agglomeration leads to a slight increase in the equivalent particle size, thereby increasing the light scattering cross-section. This largely offsets the turbidity reduction effect caused by the slight decrease in local particle number density. This antagonistic effect of the physical mechanism causes the overall turbidity value to remain in a "pseudo-stable" state for a considerable period of time, or even show an abnormal slight increase. Existing monitoring methods are easily deceived by this numerical appearance, misjudging the system as being in a stable period, thus ignoring microscopic instability signals such as a decrease in Brownian motion frequency. Once the micro-agglomeration accumulates and exceeds the critical point, the system will quickly undergo demulsification and macro-stratification. If the alarm is triggered at this time, the best opportunity to carry out redispersion or formula remediation has often been missed, resulting in economic losses and product waste. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, this invention provides a method and system for predicting the anti-stratification stability of dispersant suspension systems, aiming to solve the problem of traditional monitoring methods being misled by the "pseudo-stability" phenomenon of macroscopic turbidity. This invention constructs a two-dimensional healthy phase space including the micro-agglomeration index and the decay of the interfacial coupling factor. By utilizing Mahalanobis distance and orientation angle depth to deconstruct the masked Brownian motion changes and interfacial wetting failure characteristics in time-series data, it achieves accurate identification and early warning of microscopic instability risks in suspension systems at an early stage when they are invisible to the naked eye and conventional turbidity values ​​remain unchanged, upgrading passive post-event alarms to proactive pre-event predictions.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for predicting the anti-stratification stability of a dispersant suspension system includes:

[0008] The original time-series turbidity data stream of the dispersant suspension system is collected, and the original time-series turbidity data stream is divided into local data segment sequences by sliding window. The micro-agglomeration index sequence is calculated based on the local data segment sequence and a probability density distribution sequence is formed.

[0009] Asymmetric Gaussian fitting is performed on the probability density distribution sequence to separate the main peak component and the tail component, and the interface coupling factor attenuation sequence is calculated based on the main peak component and the tail component.

[0010] A healthy phase space for the suspended system is constructed. The micro-aggregation index sequence and the interface coupling factor decay sequence are mapped to the healthy phase space of the suspended system to form a real-time state point sequence. The Mahalanobis distance sequence and orientation angle sequence are calculated based on the real-time state point sequence. The current dominant risk type is determined according to the orientation angle sequence. The time node of macro-stratification is predicted according to the Mahalanobis distance sequence, and a full-cycle stability prediction report is output.

[0011] The method for calculating the micro-agglomeration index sequence based on local data fragment sequences includes:

[0012] A time-frequency transformation is performed on a local data segment sequence to generate a time-frequency acoustic pattern sequence. The high-frequency Brownian energy sequence is extracted from the time-frequency acoustic pattern sequence, and the micro-aggregation index sequence is calculated.

[0013] The method for generating the time-frequency acoustic pattern sequence includes:

[0014] A short-time Fourier transform is performed on each local data segment in the local data segment sequence to generate a corresponding time-frequency acoustic pattern, which is then summarized to form a time-frequency acoustic pattern sequence; each pixel of the time-frequency acoustic pattern represents the energy density of the corresponding time and frequency.

[0015] The method for extracting the high-frequency Brownian energy sequence includes:

[0016] In each time-frequency acoustic pattern sequence, a high-frequency characteristic interval corresponding to the Brownian motion of suspended particles is defined. The energy density within the high-frequency characteristic interval is integrated to obtain the high-frequency Brownian energy value, and the results are summarized to form a high-frequency Brownian energy sequence.

[0017] The method for generating the micro-agglomeration index sequence includes:

[0018] The first local data segment in the local data segment sequence is marked as the initial reference data segment. The high-frequency Brown energy value corresponding to the initial reference data segment is used as the reference value. The decay ratio of each high-frequency Brown energy value in the high-frequency Brown energy sequence relative to the reference value is calculated to obtain the micro-aggregation index sequence.

[0019] The method for generating the probability density distribution sequence includes:

[0020] Statistical analysis is performed on all turbidity values ​​within each local data segment in the local data segment sequence to calculate the probability density distribution corresponding to the local data segment. The probability density distributions corresponding to all local data segments are then summarized in chronological order to form a probability density distribution sequence.

[0021] The method for generating the interface coupling factor attenuation sequence includes:

[0022] For each probability density distribution in the probability density distribution sequence, calculate the offset distance between the centroid position of the corresponding tail component and the centroid position of the main peak component to obtain the centroid offset, and summarize them to form a centroid offset sequence.

[0023] Using the centroid offset corresponding to the initial reference data segment as the offset reference value, the attenuation ratio of each centroid offset in the centroid offset sequence relative to the offset reference value is calculated to obtain the interface coupling factor attenuation sequence.

[0024] The healthy phase space of the suspension system uses the micro-aggregation index as the horizontal axis variable and the attenuation of the interface coupling factor as the vertical axis variable.

[0025] The methods for generating the Mahalanobis distance sequence and orientation angle sequence include:

[0026] In the healthy phase space of the suspended system, the origin of the coordinate system is defined as the ideal homogeneous origin. The Mahalanobis distance from each real-time state point in the real-time state point sequence to the ideal homogeneous origin is calculated and summarized to form a Mahalanobis distance sequence. The deviation angle of each real-time state point in the real-time state point sequence from the ideal homogeneous origin is calculated and summarized to form a direction angle sequence.

[0027] The method for determining the current dominant risk type based on the direction angle sequence includes:

[0028] Obtain the deviation direction angle corresponding to the current moment in the direction angle sequence. When the deviation direction angle at the current moment is less than a preset angle threshold, determine that the current dominant risk type is small molecule failure risk. When the deviation direction angle at the current moment is greater than or equal to the preset angle threshold, determine that the current dominant risk type is spreadability failure risk.

[0029] A system for predicting the anti-stratification stability of a dispersant suspension system, used to implement the above-mentioned method for predicting the anti-stratification stability of a dispersant suspension system, the system comprising:

[0030] Microagglomeration calculation module: Collects the original time-series turbidity data stream of the dispersant suspension system, performs sliding window segmentation on the original time-series turbidity data stream to obtain local data segment sequences, calculates the microagglomeration index sequence based on the local data segment sequences, and forms a probability density distribution sequence;

[0031] Interface coupling calculation module: Performs asymmetric Gaussian fitting on the probability density distribution sequence to separate the main peak component and the tail component, and calculates the interface coupling factor attenuation sequence based on the main peak component and the tail component;

[0032] Stability prediction module: Constructs a healthy phase space for the suspended system, maps the micro-agglomeration index sequence and the interface coupling factor decay sequence to the healthy phase space of the suspended system to form a real-time state point sequence, calculates the Mahalanobis distance sequence and orientation angle sequence based on the real-time state point sequence, determines the current dominant risk type based on the orientation angle sequence, predicts the time node of macro-stratification based on the Mahalanobis distance sequence, and outputs a full-cycle stability prediction report.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] This invention achieves early quantitative tracking of the microscopic instability process of dispersant suspension systems by deconstructing the original time-series turbidity data stream into multidimensional features and constructing a healthy phase space for the suspension system. The method utilizes asymmetric Gaussian fitting separation technology of probability density distribution to decouple the signal characterizing the bulk concentration from the edge signal characterizing the interface wetting state, and maps the micro-agglomeration index and the attenuation of the interface coupling factor to an orthogonal phase space. By calculating the Mahalanobis distance of the real-time state point relative to the ideal homogeneous origin, and leveraging the Mahalanobis distance's ability to standardize multivariate covariance, it effectively overcomes the "pseudo-stable" phenomenon of macroscopic turbidity values ​​caused by the increase in light scattering cross-section due to the slight increase in particle size during the initial stage of micro-agglomeration. Thus, even in the signal masking stage before the turbidity reading has significantly decreased, it keenly captures the directional evolution trajectory of the state point deviating from the ideal state within the phase space, achieving accurate identification and early warning of dispersant encapsulation failure and the initiation of micro-adhesion. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A flowchart illustrating a method for predicting the anti-stratification stability of a dispersant suspension system provided in this embodiment of the invention;

[0037] Figure 2 This is a schematic diagram of the micro-agglomeration and Brownian motion state of the dispersant suspension system provided in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of the time-frequency acoustic pattern and high-frequency feature range provided in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the Mahalanobis distance and deviation direction angle of real-time state points in phase space provided in an embodiment of the present invention;

[0040] Figure 5 A functional block diagram of a dispersant suspension system anti-stratification stability prediction system provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1:

[0043] Please see Figure 1 As shown, this embodiment provides a method for predicting the anti-stratification stability of a dispersant suspension system, including:

[0044] Step S10: Collect the original time-series turbidity data stream of the dispersant suspension system, perform sliding window segmentation on the original time-series turbidity data stream to obtain local data segment sequences, perform time-frequency transformation on the local data segment sequences to generate time-frequency acoustic pattern sequences, extract the high-frequency Brownian energy sequence from the time-frequency acoustic pattern sequence and calculate the micro-agglomeration index sequence;

[0045] A dispersant suspension system refers to a heterogeneous mixture formed by a dispersant and various pesticide active ingredients in a liquid medium. In this system, the dispersant encapsulates pesticide particles and suspends them uniformly in the medium. The small-molecule encapsulation ability of the dispersant determines the particle size distribution and stability of the suspended particles. Turbidity is a physical quantity characterizing the degree of scattering of incident light by particles in a suspension system. Its value is directly related to the concentration, particle size, and distribution of the suspended particles. Traditional turbidity monitoring typically focuses on the absolute trend of turbidity values, determining particle sedimentation or stratification when the turbidity value decreases significantly. A raw time-series turbidity data stream refers to a sequence of turbidity values ​​continuously collected at fixed sampling intervals, with each value accompanied by a corresponding collection time identifier. This data stream carries information about the dynamic evolution of the suspension system over time.

[0046] In actual monitoring scenarios of dispersant suspension systems, the adhesion and aggregation of pesticide macromolecules often occur at the microscopic level, before forming visible precipitation or stratification. (See [link to relevant documentation]). Figure 2 This is a schematic diagram of the micro-agglomeration and Brownian motion state of the dispersant suspension system provided in the embodiments of this application. Figure 2 The diagram illustrates the microscopic process of particles evolving from a monodisperse state to an aggregated state in a liquid medium environment. This microscopic aggregation manifests as... Figure 2 The originally independent suspended particles move closer together and aggregate due to the failure of the dispersant coating. Figure 2The diagram shows larger aggregates. Because the early stages of aggregation involve only a slight increase in particle size, the aggregates remain suspended in the liquid medium. The overall turbidity value may remain unchanged or even slightly increase due to the increased light scattering cross-section. This pseudo-stability at the numerical level prevents traditional turbidity monitoring methods from identifying the microscopic adhesion process hidden behind the stable values. The system may misjudge that the small molecule encapsulation ability of the dispersant is still effective until the micro-aggregates accumulate to a critical level, at which point catastrophic demulsification and stratification suddenly occur, missing the opportunity for remediation and formulation adjustment. Step S10 extracts microscopic dynamic information neglected by traditional methods from the original time-series turbidity data stream. By quantifying and analyzing the minute fluctuations in the turbidity signal, the initiation signal of micro-aggregation is captured in advance, before the macroscopic turbidity value changes.

[0047] Brownian motion is the random thermal motion of suspended particles in a liquid medium caused by irregular collisions with medium molecules. According to the physical laws described by the Stokes-Einstein equations, the diffusion coefficient of a particle is inversely proportional to its hydrodynamic radius; the smaller the particle size, the larger the diffusion coefficient, and the higher the frequency of positional changes per unit time. This high-frequency positional fluctuation causes rapid fluctuations in the intensity of incident light scattering, thus manifesting as a high-frequency component in the turbidity signal. When the dispersant's ability to encapsulate small molecules is effective, pesticide particles are encapsulated by the dispersant into smaller particle size units, corresponding to... Figure 2 Suspended particles with relatively long motion trajectories exhibit vigorous Brownian motion, resulting in high-frequency jitter components with abundant energy in the turbidity signal. When micro-aggregation occurs, multiple small particles adhere to form larger aggregates, such as... Figure 2 As shown, the increase in aggregate mass leads to a shortening of its trajectory, meaning that the increase in aggregate mass leads to a decrease in Brownian motion frequency, and the energy of the high-frequency components in the turbidity signal decreases accordingly. Step S10 utilizes this physical mechanism to use the change in the Brownian motion characteristics of suspended particles as a characterization index of micro-aggregation, and achieves a quantitative description of the micro-aggregation process by monitoring the attenuation law of the high-frequency component energy.

[0048] Further, step S10 includes:

[0049] Step S11: Deploy a high-frequency turbidity sensor to continuously monitor the dispersant suspension system and collect raw time-series turbidity data streams with timestamps; set the sliding time window width, divide the raw time-series turbidity data streams into sliding window segments, generate a sequence of local data segments with serial numbers, and mark the first local data segment in the local data segment sequence as the initial reference data segment.

[0050] The deployment of high-frequency turbidity sensors must meet two technical requirements: firstly, the sampling frequency must be at least twice the frequency of turbidity fluctuations caused by Brownian motion of suspended particles to ensure compliance with the Nyquist sampling theorem and thus fully preserve high-frequency information; secondly, the sensor's signal-to-noise ratio must be high enough to distinguish between the weak fluctuations caused by Brownian motion and the inherent noise of electronic components. The specific value of the sampling frequency is determined based on the typical particle size range in the suspension system. The smaller the particle size, the higher the Brownian motion frequency, and the larger the required sampling frequency. For example, for pesticide suspension systems with particle sizes ranging from submicron to micron, the sampling frequency can be set in the range of hundreds of hertz to kilohertz to balance high-frequency information capture with data storage and computing resources. The generation process of the original time-series turbidity data stream is as follows: the sensor emits incident light into the suspension system according to the set sampling interval and receives the scattered light intensity signal. The light intensity signal is converted into a voltage signal and then converted into discrete values ​​through analog-to-digital conversion. Each value is accompanied by a timestamp accurate to the sampling interval level. The sequence of timestamped values ​​formed by continuous acquisition is the original time-series turbidity data stream. Continuous monitoring by a high-frequency turbidity sensor allows for the complete recording of the dynamic evolution of the suspended system throughout the entire storage period. Compared to traditional timed sampling methods, continuous monitoring avoids the risk of missing transient events due to sampling gaps, providing a complete data foundation for subsequent extraction of micro-aggregation features.

[0051] If the sliding time window is too narrow, the number of data points within a single window is insufficient for reliable spectral analysis, reducing frequency resolution and decreasing the accuracy of energy estimation for high-frequency components. Conversely, if the window is too wide, the temporal resolution decreases, as the window may contain multiple stages of significant state changes in the suspended system, obscuring the temporal boundaries of these changes. The method for determining the width of the sliding time window is as follows: calculate the period of Brownian motion of typical particles in the suspended system based on their characteristic frequencies, and set the window width to an integer multiple of this period to ensure that the window contains complete Brownian motion period information. Simultaneously, the window width should be smaller than the expected time constant of the micro-agglomeration process to ensure the capture of the phased changes in the agglomeration process. For example, if the characteristic frequency of Brownian motion of a typical particle is 100 Hz, its period is 10 ms. The window width can be set to 100 ms to contain approximately 10 complete Brownian motion periods. This setting satisfies the requirement for a large number of data points in spectral analysis while maintaining time sensitivity to minute-level micro-agglomeration processes. The sliding window segmentation process involves sequentially extracting data segments along the time axis of the original time-series turbidity data stream, with the window width as the step size. Each data segment contains all turbidity values ​​and their timestamps within the time range of the window width. Sequence numbers are assigned to each data segment according to the extraction order, generating a local data fragment sequence. The sequence numbers start from 0 and increment to facilitate subsequent indexing. Sliding window segmentation discretizes the continuous data stream into independent, analyzable data units, allowing the spectral characteristics within each time window to be calculated and tracked individually. This segmentation process lays a structured data organization foundation for the time-series evolution analysis of micro-aggregation processes.

[0052] The first local data segment in the local data segment sequence is marked as the initial reference data segment. The initial reference data segment corresponds to the moment when the dispersant suspension system has just been prepared or just started storage and monitoring. At this time, the small molecule encapsulation ability of the dispersant is in its initial state, and the particle size distribution and Brownian motion characteristics of the suspended particles represent the ideal homogeneity of the system. Using the spectral characteristics of the initial reference data segment as a reference benchmark, the degree of deviation of the spectral characteristics of subsequent time windows relative to this benchmark can quantify the degradation process of the suspension system state. This relative comparison method based on the initial state eliminates the absolute numerical differences caused by factors such as different batch formulations, different sensor devices, and different environmental conditions, making the micro-agglomeration index comparable across batches and devices, and facilitating the establishment of unified early warning thresholds and judgment criteria.

[0053] Step S12: Perform a short-time Fourier transform on each local data segment in the local data segment sequence to generate the corresponding time-frequency acoustic pattern, and summarize them to form a time-frequency acoustic pattern sequence; each pixel of the time-frequency acoustic pattern represents the energy density of the corresponding time and frequency;

[0054] Step S12 uses Short-Time Fourier Transform (SFT) to transform the local data segment from the time domain to the time-frequency domain. SFT is a time-frequency analysis method. Its principle is to apply a window function to the signal within a sliding time window and then perform a Fourier transform to obtain the local spectral characteristics of the signal at different times. The time-frequency acoustic pattern is a visual representation of the output of the SFT. Its horizontal axis represents the relative time within the window, and the vertical axis represents the frequency. The brightness or color intensity of each pixel in the image represents the energy intensity of the frequency component at that moment. The implementation process of SFT is as follows: a smoothing window function such as a Hanning window or Hamming window is applied to the turbidity numerical sequence in the local data segment to reduce spectral leakage. The Fast Fourier Transform (FFT) algorithm is executed on the windowed sequence to calculate its discrete spectrum. The square of the magnitude of the spectrum is taken to obtain the energy density of each frequency component. The energy density is arranged by frequency to form the local spectral vector of the time window. The local spectral vectors are stacked along the time axis to form a two-dimensional time-frequency acoustic pattern. The short-time Fourier transform is chosen because the frequency components after the Fourier transform directly correspond to the periodic oscillation characteristics of the signal, making it easy to establish a correspondence with the physical frequency of Brownian motion; short-time analysis preserves the time-varying characteristics of the signal, making the evolution of the spectral characteristics traceable.

[0055] Time-frequency acoustic signature mapping transforms one-dimensional turbidity time-series signals into two-dimensional time-frequency feature representations. This dimensionality increase brings information gain. In one-dimensional time-series signals, high-frequency jitter caused by Brownian motion is superimposed with fluctuations caused by sensor noise, environmental disturbances, and other factors, making direct separation difficult. In two-dimensional time-frequency acoustic signature mapping, different frequency components are mapped to different vertical axis positions. The characteristic frequency range of Brownian motion and the frequency range of noise usually differ, allowing the target signal and interference signal to be naturally separated in the frequency dimension. For example, turbidity fluctuations caused by Brownian motion are mainly concentrated in the mid-to-high frequency band corresponding to the particle diffusion time constant, while sensor electronic noise typically exhibits broadband white noise characteristics and is uniformly distributed across the entire frequency band. Environmental mechanical vibrations mainly manifest as low-frequency components. By defining specific high-frequency feature intervals in the time-frequency acoustic signature mapping, Brownian motion-related spectral features can be selectively extracted, eliminating interference from irrelevant frequency bands. After performing a short-time Fourier transform on each local data segment in step S12, the generated time-frequency acoustic pattern is arranged according to the sequence number of the local data segment to form a time-frequency acoustic pattern sequence. This sequence completely records the evolution of the spectral characteristics of the suspension system along the time axis, providing a structured time-frequency data carrier for subsequent feature extraction and trend analysis.

[0056] Step S13: In each time-frequency acoustic pattern of the time-frequency acoustic pattern sequence, define the high-frequency characteristic interval corresponding to the Brownian motion of the suspended particles, perform integral calculation on the energy density within the high-frequency characteristic interval to obtain the high-frequency Brownian energy value, and summarize to form a high-frequency Brownian energy sequence.

[0057] See Figure 3 This is a schematic diagram of the time-frequency acoustic pattern and high-frequency feature range provided in the embodiments of this application. Figure 3 The diagram illustrates a time-frequency acoustic signature coordinate system with relative time on the horizontal axis and frequency on the vertical axis, as well as a coordinate system defined by the upper limit frequency F. max Lower limit frequency F min The defined high-frequency characteristic interval is represented by each point in the time-frequency acoustic waveform, which indicates the energy density at the corresponding relative time and frequency. The boundary of the high-frequency characteristic interval is determined by the physical properties of the Brownian motion of the suspended particles. According to the Stokes-Einstein equation, the diffusion coefficient of the particles is related to the Boltzmann constant, absolute temperature, dynamic viscosity of the medium, and the hydrodynamic radius of the particles. Dividing the diffusion coefficient by the square of the particle radius yields the characteristic frequency of particle position fluctuations. The method for determining the lower limit of the high-frequency characteristic interval is as follows: calculate the corresponding Brownian motion characteristic frequency based on the maximum expected particle size of the target particles in the dispersant suspension system, and use this frequency as the lower limit F of the high-frequency characteristic interval. min Components below this frequency mainly originate from the slow motion of large particles or aggregates and low-frequency environmental disturbances, and are not within the scope of concern. The method for determining the upper limit of the high-frequency characteristic range is as follows: calculate the corresponding Brownian motion characteristic frequency based on the minimum expected particle size of the target particle, and use this frequency as the upper limit F of the high-frequency characteristic range. max At the same time, it is necessary to ensure that the upper limit does not exceed half of the sensor sampling frequency to avoid aliasing distortion. For example, for pesticide suspension particles with a particle size distribution in the range of 1 micrometer to 10 micrometers, in a room-temperature aqueous medium, the Brownian motion characteristic frequency distribution is in the range of tens of hertz to hundreds of hertz. Figure 3 As shown in the example, the high-frequency characteristic range can be set to cover the frequency range accordingly.

[0058] The calculation process for the high-frequency Brownian energy value is as follows: see [link to relevant documentation] Figure 3 In the time-frequency acoustic signature map, the frequency coordinates corresponding to the upper and lower limits of the high-frequency characteristic interval are located along the vertical axis. Then, all relative times within the window are traversed along the horizontal axis. The energy density within the high-frequency characteristic interval at each time point is integrated and summed. Finally, the average high-frequency energy density of the time window is obtained by averaging the integral values ​​over all times. This average value is the high-frequency Brownian energy value corresponding to the local data segment. The formula for calculating the high-frequency Brownian energy value can be expressed as: ,in Let F represent the high-frequency Brownian energy value corresponding to the i-th local data segment, Δt represent the width of the sliding time window, and F min With F max These represent the lower and upper frequency limits of the high-frequency characteristic range, respectively. This represents the energy density value at frequency f and relative time τ in the i-th time-frequency acoustic pattern. This represents the integration over the relative times from 0 to Δt within the time window. This indicates the high-frequency characteristic interval F min To F max The frequency integral. The physical meaning of this formula is: for the target's high-frequency characteristic range (F... min ~F max The total energy of the frequency band is obtained by double integration of the energy density within the time window interval (0~Δt). This total energy is then divided by the time window width Δt to obtain the average energy per unit time. This average energy reflects the activity level of Brownian motion of suspended particles within the window period. The formula uses integral form rather than discrete summation, reflecting a mathematical abstraction of the continuous physical process. In actual calculations, numerical integration is used based on the discrete resolution of the time-frequency acoustic pattern. After performing the above calculation on each time-frequency acoustic pattern in step S13, the high-frequency Brownian energy values ​​are summarized in chronological order to form a high-frequency Brownian energy sequence. This sequence records the dynamic change trajectory of Brownian motion energy characteristics using time as an index.

[0059] Step S14: Using the high-frequency Brown energy value corresponding to the initial reference data segment as the reference value, calculate the attenuation ratio of each high-frequency Brown energy value in the high-frequency Brown energy sequence relative to the reference value to obtain the micro-agglomeration index sequence.

[0060] Step S14 converts the high-frequency Brownian energy sequence into a microagglomeration index sequence. The microagglomeration index is a dimensionless measure of the degree of micro-agglomeration of suspended particles. The calculation of the microagglomeration index is based on the following: when the dispersant's ability to encapsulate small molecules is effective, the suspended particles maintain a small particle size, Brownian motion is vigorous, and the high-frequency Brownian energy value remains at a high level; when microagglomeration occurs, particles adhere to form larger aggregates, Brownian motion slows down, and the high-frequency Brownian energy value decreases. The high-frequency Brownian energy value E corresponding to the initial baseline data segment is used as the starting point. brown Using (0) as the baseline, calculate the high-frequency Brownian energy value E for each subsequent time window. brown (i) Retention rate relative to the benchmark value R brown (i)=E brown (i) / E brown (0), the retention rate reflects the degree to which the Brownian motion energy is retained relative to the initial state at the current moment. The micro-agglomeration index is defined as the degree of decay of the retention rate, and the calculation formula is: , where Kagg(i) represents the microagglomeration index corresponding to the i-th time window. The design logic of this formula is as follows: when the high-frequency Brownian energy is fully maintained, the retention rate is 1, and the microagglomeration index is 0, indicating that no agglomeration has occurred and the suspension system is in an initial homogeneous state; when the high-frequency Brownian energy decays, the retention rate is less than 1, and the microagglomeration index is greater than 0, with a larger value indicating a more severe degree of agglomeration; when the high-frequency Brownian energy completely disappears, the retention rate approaches 0, and the microagglomeration index approaches 1, indicating that the suspended particles have completely agglomerated into large particles, and Brownian motion has basically stopped. Normalization using a ratio relative to the initial baseline limits the value range of the microagglomeration index to between 0 and 1, facilitating the setting of a unified warning threshold and grading standard, while eliminating the influence of differences in absolute energy values ​​caused by differences in suspension concentration between different batches of formulations on the judgment results.

[0061] The generation process of the micro-agglomeration index sequence is as follows: The micro-agglomeration index corresponding to each time window is calculated sequentially according to the time window number and arranged into a sequence. This sequence uses a unified normalized index to characterize the trajectory of the evolution of micro-agglomeration degree over time. The decision logic is implemented by setting a Brownian threshold K. threshold As the criterion for initiating micro-agglomeration, the Brown threshold is determined based on historical data statistical analysis of the dispersant suspension system. The upper bound of the micro-agglomeration index fluctuation range of normal samples from historical batches that have not experienced stratification failure is used as the Brown threshold. For example, if the micro-agglomeration index fluctuation range of normal samples is 0 to 0.1, the Brown threshold can be set to 0.15 to allow for a safety margin. When the real-time calculated micro-agglomeration index Kagg(i) exceeds the Brown threshold, even if the macroscopic mean of the original time-series turbidity data stream remains stable without a downward trend, the system determines that hidden aggregation has occurred within the suspension system, and the small molecule encapsulation function of the dispersant is failing. This necessitates triggering an early warning and initiating subsequent risk assessment and intervention procedures.

[0062] Step S10 forms a progressive and synergistic relationship with subsequent steps S20 and S30. The micro-agglomeration index sequence output from step S10 serves as the horizontal axis input for constructing the healthy phase space of the suspended system in step S30. Together with the interface coupling factor decay sequence output from step S20, they form a two-dimensional orthogonal coordinate system, enabling the state of the suspended system to be accurately located in a multi-dimensional space. Without the micro-agglomeration index output from step S10, step S30 would degenerate into a one-dimensional judgment based solely on the interface coupling factor, unable to distinguish between the two different failure modes: "agglomeration-type degradation caused by small molecule failure" and "wetting-type degradation caused by surfactant failure." The accuracy of the comprehensive prediction and the precision of risk type identification would be significantly reduced. The time window segmentation method of step S10 is consistent with that of step S20. The two steps share the same set of local data segment sequences as input, ensuring strict alignment of the micro-agglomeration index and the interface coupling factor decay in the time dimension. This avoids state mapping distortion caused by time misalignment and provides data synchronization assurance for the pairing mapping of the two indicators at the same time in step S30. Step S10 employs a technical approach of mining high-frequency Brownian energy characteristics from the original time-series turbidity data stream. This redefines the minute fluctuations, which are considered noise and discarded in traditional turbidity monitoring, as effective signals carrying microscopic dynamic characteristics. This information reuse extracts previously unavailable information on suspended particle size changes from existing data without adding additional sensor hardware or changing the data acquisition method. This overcomes the limitations of traditional turbidity monitoring, which can only reflect macroscopic sedimentation and stratification, and enables real-time quantitative tracking of microscopic aggregation processes. It identifies the initiation signals of microscopic adhesion in advance during the pseudo-stationary phase where turbidity values ​​remain stable, shifting the warning time forward to before macroscopic stratification occurs, thus providing a buffer window for formula adjustments or product disposal. A unified judgment standard is established based on a normalized microscopic aggregation index, making the warning logic universal across batches and devices, reducing the complexity of system deployment and maintenance. Step S10 addresses the long-standing technical contradiction in the field of dispersant suspension system stability prediction: the pseudo-steady state of macroscopic turbidity values ​​masks the true microscopic unstable state. This is achieved by transforming the time-domain signal to the time-frequency domain to expose the masked high-frequency characteristics, associating it with Brownian motion physical mechanisms to give the high-frequency characteristics clear physical meaning, and constructing comparable quantitative indicators through normalization. This process extracts dynamic information on microscopic particle size from the apparent stability of turbid liquids, enabling continuous monitoring of the small-molecule state of the suspension system, quantitative tracking of the degradation process of dispersant encapsulation ability, early prediction of the risk of pesticide macromolecule adhesion, and timely detection of potential changes in chemical properties. Ultimately, this ensures that the dispersant suspension system maintains a uniform mixture of multiple pesticides without adhesion or aggregation during the storage period.

[0063] Step S20: Form a probability density distribution sequence based on the local data fragment sequence, perform asymmetric Gaussian fitting on the probability density distribution sequence to separate the main peak component and the tail component, and calculate the interface coupling factor attenuation sequence based on the main peak component and the tail component.

[0064] The goal of step S20 is to extract interfacial wetting characteristics that characterize the functional state of the surfactant in the dispersant from time-series turbidity data. Application scenarios for dispersant suspension systems require the mixture to spread and remain wet after being sprayed onto target surfaces such as leaves. This performance is determined by the interfacial tension regulation capability provided by the surfactant component in the dispersant. Traditional turbidity monitoring only reflects the distribution of suspended particles in the liquid medium and cannot directly measure interfacial physicochemical parameters such as surface tension or wetting angle, resulting in a disconnect between storage stability monitoring and application dispersibility assessment. Step S20 utilizes the optical edge effect generated at the interface between the turbidity sensor and the liquid during measurement, treating the contact area between the sensor probe and the liquid surface as a miniature wetting experiment model. By analyzing the statistical distribution changes in turbidity sampling data, the interfacial wetting capability of the liquid is indirectly inferred, enabling prediction of spreading performance without opening the container or conducting physical spreading experiments. The interfacial coupling factor is a quantitative indicator characterizing the strength of the wetting interaction between the liquid and solid surfaces. When the surfactant function of the dispersant is normal, the liquid exhibits good wettability on the sensor probe surface, forming an upward-curving concave meniscus along the probe surface. The curvature of the meniscus is determined by both the liquid surface tension and the contact angle. The liquid layer thickness in the meniscus region exhibits a gradient distribution along the probe surface. When incident light passes through this region, it undergoes refraction and shift, resulting in a systematic difference in the intensity of scattered light received by the sensor between the meniscus region and the bulk liquid region. This difference manifests as a specific asymmetric morphological feature in the statistical distribution of time-series turbidity data. Step S20 captures and quantifies the evolution of this asymmetric feature, mapping it to the attenuation of the interfacial coupling factor, thereby achieving dynamic tracking of dispersion performance.

[0065] Further, step S20 includes;

[0066] Step S21: Perform statistical analysis on all turbidity values ​​in each local data segment of the local data segment sequence, calculate the probability density distribution corresponding to the local data segment, and summarize the probability density distributions corresponding to all local data segments in chronological order to form a probability density distribution sequence.

[0067] Each local data segment contains all turbidity sample values ​​collected continuously within a time window. The probability density distribution, a function describing the probability of a random variable's values, is obtained by frequency statistics and normalization of the turbidity sample values ​​within the local data segment. The calculation process is as follows: the range of turbidity sample values ​​is evenly divided into multiple non-overlapping sub-intervals. The number of sampling points in each sub-interval is counted. The probability density estimate for each sub-interval is obtained by dividing the number of sampling points in each sub-interval by the product of the total number of sampling points and the width of the sub-interval. A distribution curve is plotted with the midpoint of the sub-interval as the x-axis and the probability density estimate as the y-axis; this curve represents the probability density distribution corresponding to the local data segment. The number of sub-intervals is selected according to Sturgess's rule or Scott's rule in statistics, adaptively determined based on the number of sampling points to balance the smoothness and detail resolution of the distribution curve. In an ideal homogeneous medium, the measurement noise of a turbidity sensor mainly originates from the thermal noise and quantization error of the photoelectric conversion circuit. These noises have zero mean and are independent of each sampling point. According to the central limit theorem, the superposition of a large number of independent and identically distributed random variables tends towards a Gaussian normal distribution. Therefore, ideally, the probability density distribution of turbidity sampling values ​​should present a symmetrical bell-shaped curve centered on the true turbidity value. When the surfactant function of the dispersant is normal, the liquid has good wetting ability on the sensor probe surface. A significant concave meniscus is formed at the interface between the probe and the liquid surface. The liquid layer thickness in the meniscus region gradually increases from the probe surface towards the main liquid mass. When the incident light beam passes through the meniscus, refraction and scattering path shift occur, causing the light intensity signal received by the sensor in the meniscus region to systematically deviate from the light intensity signal in the main liquid mass region. This optical path shift effect causes the turbidity readings at some sampling points to deviate from the main turbidity value, forming a tail component extending to one side in the probability density distribution, resulting in an asymmetrical distribution curve. When dispersants fail and surfactants degrade, the wetting ability of the liquid decreases, the meniscus curvature decreases or even becomes straight, the edge optical path offset effect weakens, the tail component gradually shrinks, and the probability density distribution tends to a symmetrical normal distribution. Step S21 calculates the probability density distribution to transform the interface wetting information hidden in the discrete sampling points of the time-series data into an analyzable statistical distribution shape, laying a data foundation for subsequent morphological feature extraction.

[0068] After performing the above probability density distribution calculation on each local data segment in the local data segment sequence, the probability density distributions are arranged and summarized according to the time sequence number of the local data segments to form a probability density distribution sequence. The probability density distribution sequence records the evolution of the statistical distribution pattern of turbidity data along the time axis with time as the index, so that the dynamic change trend of wetting ability can be tracked in the sequence dimension. Step S21 and step S10 share the same set of local data segment sequences as input sources. This data sharing mechanism ensures strict alignment of the micro-agglomeration index and the attenuation of the interface coupling factor in the time dimension, providing time synchronization guarantee for the pairing mapping of the two indicators in step S30. Step S21 reorganizes the sampling data originally used for single turbidity mean calculation into a probability density distribution form, so that the distribution pattern information ignored in the traditional method can be revealed. This change in data perspective expands the information utilization dimension of time-series turbidity data without increasing hardware costs.

[0069] Step S22: Perform asymmetric Gaussian fitting on each probability density distribution in the probability density distribution sequence to separate the main peak component representing the bulk turbidity inside the liquid and the tail component representing the edge optical path anomaly.

[0070] Step S22 performs parameterized decomposition on each probability density distribution in the probability density distribution sequence, splitting the composite distribution into two independent components with clear physical meaning. Asymmetric Gaussian fitting represents the probability density distribution to be analyzed as a weighted superposition of two Gaussian distribution components. One component is a symmetric Gaussian distribution as the dominant peak component, and the other component is a skewed Gaussian distribution or an exponentially corrected Gaussian distribution as the tail component. The dominant peak component corresponds to the turbidity signal distribution in the bulk liquid region; its distribution center position reflects the intrinsic turbidity level of the suspension system, and its distribution width reflects the random noise level of the measurement system. The tail component corresponds to the edge optical path anomaly signal distribution in the meniscus region; its distribution center is offset relative to the dominant peak component, and its distribution shape exhibits an asymmetric tailing characteristic extending to one side. The implementation process of asymmetric Gaussian fitting is as follows: Using the probability density distribution to be analyzed as the objective function, the center position, standard deviation, and weighting coefficient of the main peak component, and the center position, standard deviation, skewness parameter, and weighting coefficient of the tail component are used as optimization parameters. The nonlinear least squares method or expectation-maximization algorithm is used to iteratively adjust the values ​​of each parameter until the sum of squared residuals between the weighted superposition curve of the two components and the target probability density distribution reaches its minimum. After the fitting converges, the parameters of the main peak component and the tail component are extracted and stored separately for subsequent centroid position calculation. The reason for choosing asymmetric Gaussian fitting is that: the Gaussian distribution has a clear physical background; the normal distribution characteristics of measurement noise make the main peak component naturally conform to the Gaussian shape; the introduction of the asymmetric tail can flexibly characterize the unidirectional characteristics of the meniscus optical path offset effect, that is, the direction of the edge signal deviating from the main signal is definite rather than random; parameterized expression allows for the quantitative extraction of component features, facilitating subsequent time-series comparison and trend analysis.

[0071] The separation of the main peak component and the tail component decouples the two types of physical signals mixed in a single probability density distribution. The main peak component carries the main signal scattered by suspended particles, and its morphological changes mainly reflect the overall state of suspension concentration and particle distribution; the tail component carries the edge-additional signal generated by the interface wetting effect, and its morphological changes directly reflect the evolution of the meniscus geometry. This signal decoupling ensures that wettability changes are no longer masked by the main turbidity signal, and even if the main turbidity of the suspended system remains stable, the morphological degradation of the tail component can still be independently identified. After performing the above fitting separation on each probability density distribution in the probability density distribution sequence in step S22, a main peak component sequence and a tail component sequence corresponding one-to-one with the probability density distribution sequence are formed. These two sequences respectively characterize the temporal evolution trajectory of the main turbidity signal and the edge wetting signal.

[0072] Step S23: For each probability density distribution in the probability density distribution sequence, calculate the offset distance between the centroid position of the corresponding tail component and the centroid position of the main peak component to obtain the centroid offset, and summarize them to form a centroid offset sequence.

[0073] Step S23 extracts the centroid position parameters of the main peak component and the tail component obtained in step S22, and calculates the offset distance between them as a quantitative measure of the interface wetting effect. The centroid position is the first moment of the probability density distribution, and is the abscissa of the centroid of the area under the distribution curve. It is calculated by multiplying the abscissa value by the corresponding probability density value, integrating over the entire range, and then dividing by the total area under the probability density curve. For a normalized probability density distribution, the total area is unit 1, and the centroid position is the integral of the product of the abscissa and the probability density. The centroid position of the main peak component is denoted as μ. main The centroid position of the trailing component is denoted as μ. tail Both are expressed in units of turbidity sampling values. The centroid offset ΔL is defined as the absolute distance between the centroid positions of the tail component and the main peak component, calculated using the following formula: , where i is the time sequence number of the local data segment.

[0074] Centroid offset characterizes the degree of deviation of the edge optical path anomaly signal from the main turbidity signal. When the dispersant's wetting function is normal, the drug solution forms a concave meniscus with a large curvature on the sensor probe surface. The incident light undergoes significant refraction and shift in the meniscus region, and the turbidity readings at the edge sampling points systematically deviate from the readings in the main region. This manifests as the centroid of the tail component moving away from the centroid of the main peak component, resulting in a large centroid offset value. When the wetting ability decreases, the meniscus tends to be flatter, the edge optical path offset effect weakens, the tail component moves closer to the main peak component, and the centroid offset value decreases. This variation in centroid offset corresponds to the change in liquid surface tension: the lower the surface tension, the better the wettability, the larger the meniscus curvature, and the larger the centroid offset; the higher the surface tension, the worse the wettability, the flatter the meniscus, and the smaller the centroid offset. This correspondence allows the centroid offset to indirectly characterize the dispersant's interfacial tension regulation capability. The centroid offsets corresponding to each time window are calculated sequentially according to the time sequence number of the local data segments and arranged to form a centroid offset sequence. The centroid offset sequence characterizes the trajectory of the interface wetting effect intensity over time with a unified physical dimension. Step S23 further condenses the component parameters output from step S22 into a single scalar index, thus condensing the complex distribution morphology characteristics into a directly comparable numerical sequence, reducing the computational complexity of subsequent trend analysis and threshold determination. The centroid offset is calculated based on the first-order moment difference between the main peak component and the tail component. This index is sensitive to translational changes in the distribution morphology but relatively robust to scale changes, effectively reflecting the directional offset characteristics of the meniscus effect.

[0075] Step S24: Using the centroid offset corresponding to the initial reference data segment as the offset reference value, calculate the attenuation ratio of each centroid offset in the centroid offset sequence relative to the offset reference value, and obtain the interface coupling factor attenuation sequence.

[0076] Step S24 transforms the centroid offset sequence into a normalized interface coupling factor decay sequence, making the degree of wettability variation comparable across different batches and sensor conditions. The initial reference data segment corresponds to the first element of the local data segment sequence marked in step S11. Its acquisition time is the initial state of the dispersant suspension system after preparation or at the beginning of monitoring. At this time, the surfactant function of the dispersant is in its designed state, and the wetting ability represents the ideal level of the system. The centroid offset corresponding to the initial reference data segment is denoted as ΔL(0), serving as the offset reference value for subsequent relative comparisons. Wetting retention rate W ret (i) is defined as the ratio of the current centroid offset to the offset reference value, calculated using the formula W. ret (i) = ΔL(i) / ΔL(0). The wetting retention rate reflects the degree to which the interfacial wetting effect is retained relative to the initial state at the current moment, and its value ranges from 0 to 1. When the wetting ability is fully maintained, the centroid offset remains at the initial level, and the wetting retention rate is equal to 1; when the wetting ability decreases, the centroid offset decreases, and the wetting retention rate is less than 1; when the wetting ability is completely lost and the meniscus effect disappears, the centroid offset approaches 0, and the wetting retention rate approaches 0.

[0077] The interface coupling factor attenuation σdec(i) is defined as the degree of attenuation of the wetting retention rate, and is calculated using the following formula: The design logic of this formula is symmetrical to the definition of the microagglomeration index in step S14: when the wetting ability is fully maintained, the attenuation of the interfacial coupling factor is equal to 0, indicating that no wetting degradation has occurred and the spreading function of the dispersant is in normal condition; when the wetting ability decreases, the attenuation of the interfacial coupling factor is greater than 0, and the larger the value, the more severe the wetting degradation; when the wetting ability is completely lost, the attenuation of the interfacial coupling factor approaches 1, indicating that the surfactant function of the dispersant has completely failed. Normalization is performed using a ratio relative to the initial baseline, limiting the value range of the interfacial coupling factor attenuation to between 0 and 1, consistent with the value range of the microagglomeration index, facilitating equal-weighted multidimensional mapping in the healthy phase space in step S30. The normalization process also eliminates the influence of differences in the absolute centroid offset caused by differences in the initial surfactant concentration between different batches of formulations on the judgment result, making the early warning logic universal across formulations.

[0078] The generation process of the interface coupling factor decay sequence is as follows: The decay of the interface coupling factor corresponding to each time window is calculated sequentially according to the time window number, and then arranged into a sequence. This sequence uses a unified normalized index to characterize the trajectory of the wettability degradation degree over time. The implementation of the judgment logic is as follows: a wettability threshold σ is set. threshold As the criterion for spreadability failure, the wetting threshold is determined based on historical data statistical analysis and actual spreadability experiments of the dispersant suspension system. The upper bound of the fluctuation range of the interfacial coupling factor attenuation of qualified samples in historical batches is used as the wetting threshold. For example, if the fluctuation range of the interfacial coupling factor attenuation of qualified samples is 0 to 0.08, the wetting threshold can be set to 0.12 to leave a safety margin. When the real-time calculated interfacial coupling factor attenuation σdec(i) exceeds the wetting threshold, even if the macroscopic mean of the original time-series turbidity data stream remains stable and no precipitation occurs, the system determines that the surfactant component of the dispersant is failing. It predicts that the mixture will not be able to spread sufficiently and maintain a wet state after being sprayed onto the target plane, requiring the triggering of a spreadability warning and the initiation of subsequent risk assessment procedures.

[0079] Step S20 solves the technical challenge of traditional turbidity monitoring failing to detect the wetting performance of liquid interfaces by constructing a virtual meniscus tension inversion model. Utilizing the edge optical path effect at the sensor probe, and through probability density distribution statistics and asymmetric Gaussian fitting, the time-series turbidity data is decoupled into a main peak component representing the bulk concentration and a tail component representing the wetting effect. By quantifying the relative shift of their centroids, the edge signal, originally considered noise, is transformed into an interface coupling factor characterizing the surfactant function. This technology not only achieves non-invasive inversion of the "outside-bottle spreading performance" of the drug solution without disrupting the sealed environment or increasing hardware costs, filling the blind spot in applications of dispersed monitoring, but also achieves strict data synchronization by sharing a time window with step S10. This provides a crucial vertical axis input for step S30 to construct a two-dimensional healthy phase space, ensuring that the system can simultaneously identify two distinct risk modes: "micro-agglomeration" and "spreading failure," significantly improving the completeness and accuracy of the prediction results.

[0080] Step S30: Construct a healthy phase space for the suspended system, map the micro-aggregation index sequence and the interface coupling factor decay sequence to the healthy phase space of the suspended system to form a real-time state point sequence, calculate the Mahalanobis distance sequence and orientation angle sequence based on the real-time state point sequence, determine the current dominant risk type based on the orientation angle sequence, predict the time node of macro-stratification based on the Mahalanobis distance sequence, and output a full-cycle stability prediction report.

[0081] Step S30 integrates the two independent feature indices extracted in steps S10 and S20 to construct a comprehensive evaluation framework that simultaneously reflects the microscopic particle size state and macroscopic interface wetting state of the dispersant suspension system. The stability failure of the dispersant suspension system is not a single-dimensional degradation process, but a complex evolutionary process in which two failure modes—microscopic aggregation and decreased interface wetting ability—may occur independently, alternately, or deteriorate simultaneously. Traditional single-indicator monitoring methods can only identify one failure mode. When the other mode dominates the degradation process, it will result in missed detections. When both modes occur simultaneously, it is impossible to distinguish their respective contributions, leading to a lack of targeted warning signals and failing to provide a clear direction for formulation adjustments. Step S30 constructs a two-dimensional orthogonal healthy phase space for the suspension system, using the microscopic aggregation index and the attenuation of the interface coupling factor as two independent coordinate axes. This allows the state of the suspension system at any given time to be uniquely mapped to a state point in the phase space. The position of the state point simultaneously encodes the degree information of both failure modes, and the trajectory of the state point completely depicts the dynamic evolution path of system degradation. Phase space is a mathematical space describing the state of a dynamic system. Each point in the space corresponds to a complete state description of the system at a certain moment, and the temporal evolution of the system is manifested as trajectory motion in the phase space. The healthy phase space of the suspension system is a dedicated phase space constructed for the stability monitoring scenario of dispersant suspension systems. Its dimensions are determined by the number of independent failure modes that need to be monitored simultaneously. Since the micro-agglomeration index extracted in step S10 reflects the ability to encapsulate small molecules, and the attenuation of the interfacial coupling factor extracted in step S20 reflects the wetting function of the surfactant, these two parameters correspond to two core functions of the dispersant, and their variation patterns are independent rather than subordinate to each other. Therefore, a two-dimensional orthogonal coordinate system is chosen as the geometric structure of the healthy phase space of the suspension system. Orthogonality ensures that the information dimensions represented by the two coordinate axes are perpendicular to each other and do not interfere with each other. Changes in one dimension will not produce projection components in the other dimension, allowing the two failure modes to be quantified and tracked independently.

[0082] Further, step S30 includes:

[0083] Step S31: Construct a two-dimensional orthogonal healthy phase space of the suspension system with the micro-aggregation index as the horizontal axis variable and the interface coupling factor attenuation as the vertical axis variable, and define the origin of the coordinate system in the healthy phase space of the suspension system as the ideal homogeneous origin.

[0084] The healthy phase space of the suspended system adopts a Cartesian coordinate system, with the horizontal axis X and the vertical axis Y perpendicular to each other, and their intersection point defined as the origin O. The coordinate variable of the horizontal axis X is the micro-agglomeration index Kagg, whose value range is limited to a closed interval of 0 to 1 by the normalization process in step S14. The positive direction of the horizontal axis corresponds to the direction in which the micro-agglomeration index increases, i.e., the direction in which the adhesion of suspended particles intensifies and the Brownian motion activity decreases. The coordinate variable of the vertical axis Y is the interfacial coupling factor attenuation σdec, whose value range is also limited to a closed interval of 0 to 1 by the normalization process in step S24. The positive direction of the vertical axis corresponds to the direction in which the interfacial coupling factor attenuation increases, i.e., the direction in which the liquid wetting ability decreases and the meniscus effect weakens. The consistency of the value ranges of the two coordinate axes ensures that the phase space is geometrically presented as a square region with a side length of 1, avoiding the imbalance of coordinate axis scale caused by differences in dimensions or numerical ranges.

[0085] The origin O of the coordinate system is set at the intersection of the x-axis coordinate Kagg equal to 0 and the y-axis coordinate σdec equal to 0, defined as the ideal homogeneous origin. The ideal homogeneous state refers to the initial state of the dispersant suspension system after formulation. At this state, the dispersant's ability to encapsulate small molecules is intact and effective, the suspended particles maintain the designed particle size distribution without any aggregation, corresponding to a microagglomeration index of 0; simultaneously, the surfactant component of the dispersant functions normally, the liquid's wetting ability on the solid surface is at its initial level without any degradation, corresponding to an interfacial coupling factor attenuation of 0. The ideal homogeneous origin serves as a reference point for the health of the suspension system. The deviation of subsequent state points from the origin directly reflects the degree of degradation of the system relative to the initial ideal state. Setting the origin at the zero point of the coordinate system, rather than at other locations, gives the distance from the origin a clear physical meaning; a larger distance indicates more severe degradation. This design simplifies subsequent distance calculations and threshold determination logic.

[0086] Step S32: Pair the values ​​at the same time in the micro-aggregation index sequence and the interface coupling factor decay sequence to map them as real-time state points in the healthy phase space of the suspension system, and arrange them in chronological order to form a real-time state point sequence.

[0087] Step S32 merges the two one-dimensional time-series index sequences output from steps S10 and S20 into a two-dimensional state point sequence in the healthy phase space of the suspended system. The pairing mapping process is as follows: traverse the micro-agglomeration index sequence and the interface coupling factor decay sequence. For each time window number i, extract the element Kagg(i) with index i in the micro-agglomeration index sequence as the horizontal axis coordinate value, and extract the element σdec(i) with index i in the interface coupling factor decay sequence as the vertical axis coordinate value. The coordinate pair (Kagg(i), σdec(i)) is taken as the i-th real-time state point Pnow(i) in the healthy phase space of the suspended system. The real-time state point reflects the instantaneous state of the suspended system within the corresponding time window. As the monitoring time progresses, new time window data is collected and processed, and new real-time state points are continuously generated and appended to the end of the sequence. The generation process of the real-time state point sequence is as follows: according to the time window number increasing from 0, the above pairing mapping operation is executed sequentially, and the real-time state points corresponding to each time window are arranged by number to form an ordered sequence. The real-time state point sequence, indexed by time, fully records the motion trajectory of the suspended system's state in phase space. The first element of the sequence corresponds to the state of the initial reference data segment, with its coordinates at the origin O=(0,0). As the monitoring time increases, if the suspended system deteriorates, the state points will gradually deviate from the origin and move towards the positive quadrant of phase space. The shape and direction of the movement trajectory encode the dynamic characteristics of the deterioration process.

[0088] The pairing mapping operation in step S32 relies on the consistency of the time window division between steps S10 and S20. Both steps share the same set of local data fragment sequences generated in step S11 as input sources, employing the same sliding time window width and the same window number encoding rules. This ensures that the micro-agglomeration index with sequence number i and the interface coupling factor decay with sequence number i correspond to the exact same physical time period. This time synchronization mechanism eliminates the risk of state mapping distortion caused by sampling time misalignment, enabling the paired real-time state points to accurately reflect the joint state of the suspension system in the two failure dimensions at the same time. The generation of the real-time state point sequence allows the multidimensional state evolution of the dispersant suspension system to be visualized in the form of geometric trajectories. Compared to observing two independent one-dimensional time series curves separately, the two-dimensional phase space trajectory can simultaneously display the changing trends and interrelationships of the two dimensions, revealing joint change patterns that may be missed when observing each dimension individually. For example, if the phase space trajectory of a batch of samples shows a broken line shape, moving first along the positive horizontal axis and then along the positive vertical axis, it indicates that the batch first experienced micro-agglomeration and degradation, followed by wettability degradation, and that the two failure modes have a temporal relationship. If the trajectory of another batch of samples shows a straight line shape moving along the diagonal direction, it indicates that the two failure modes occurred simultaneously and to an equal degree. Identifying such joint change modes is difficult to achieve when analyzing time-series curves in each dimension separately. Phase space representation provides an intuitive geometric perspective for the mechanism analysis of failure modes and formulation improvement.

[0089] Step S33: Calculate the Mahalanobis distance from each real-time state point in the real-time state point sequence to the ideal uniform origin, and summarize them to form a Mahalanobis distance sequence; calculate the deviation angle of each real-time state point in the real-time state point sequence relative to the ideal uniform origin, and summarize them to form a direction angle sequence;

[0090] Step S33 extracts two complementary geometric features from the real-time state point sequence: the Mahalanobis distance sequence characterizes the overall degree of deviation of state points from the origin, and the orientation angle sequence characterizes the dominant direction of deviation of state points from the origin. These two features are used for subsequent stratified time prediction and risk type determination, respectively. Mahalanobis distance is a standardized distance metric that considers the correlation and scale differences between variables. Its calculation requires the introduction of the covariance matrix of the variables. The covariance matrix is ​​a 2×2 matrix describing the statistical correlation between the micro-clustering index and the decay of the interface coupling factor. The diagonal elements are the variances of the two variables, reflecting the fluctuation amplitude of each variable; the off-diagonal elements are the covariances of the two variables, reflecting the tendency of the two variables to change in the same or opposite directions. The covariance matrix is ​​obtained by collecting the micro-clustering index sequence and the interface coupling factor decay sequence of each batch of samples from historical monitoring data, forming a two-dimensional vector sample from the values ​​of the two indicators at each time point, and calculating the sample covariance matrix for all samples. This matrix serves as the statistical parameter for Mahalanobis distance calculation. The covariance matrix needs to be estimated based on a sufficient number of historical samples to ensure statistical stability. For example, complete monitoring data from no fewer than 30 batches can be collected for covariance matrix estimation.

[0091] See Figure 4 This is a schematic diagram of the Mahalanobis distance and deviation direction angle of the real-time state point in phase space provided in the embodiments of this application. Figure 4 The diagram illustrates a two-dimensional coordinate system with the microagglomeration index as the horizontal axis and the attenuation of the interfacial coupling factor as the vertical axis. The origin represents the ideal homogeneous state, and the position of the real-time state point in the coordinate system reflects the current degradation state. The Mahalanobis distance is calculated as follows: a quadratic operation is performed on the coordinate vector of the real-time state point and the inverse of the covariance matrix to obtain a scalar value, and then the square root of this scalar value is taken to obtain the Mahalanobis distance. The reason for choosing Mahalanobis distance instead of Euclidean distance is that there may be a correlation at the chemical mechanism level between the microagglomeration index and the attenuation of the interfacial coupling factor. When the surfactant component in the dispersant simultaneously performs the functions of encapsulation and wetting, its failure may cause both indicators to rise simultaneously. Euclidean distance cannot distinguish between this correlated change and the case where the two independent dimensions change separately. Mahalanobis distance performs decorrelation processing on the variables through the inverse transformation of the covariance matrix, so that the distance calculation reflects the degree of deviation of the variables in the standardized space after decorrelation, making it more sensitive to the identification of abnormal states. The numerical characteristics of Mahalanobis distance are as follows: when the state point is located at the origin, the Mahalanobis distance is 0; when the state point deviates from the origin, the Mahalanobis distance is greater than 0, and the greater the deviation, the greater the Mahalanobis distance; the contour lines of Mahalanobis distance are elliptical in phase space, such as... Figure 4As shown by the dashed arc in the figure, this dashed line represents the elliptical contour lines of the Mahalanobis distance. The shape and orientation of the ellipse are determined by the covariance matrix. The generation process of the Mahalanobis distance sequence is as follows: traverse the real-time state point sequence, perform Mahalanobis distance calculation for each real-time state point, and arrange the Mahalanobis distance values ​​in order according to the time window number to form an ordered sequence. The Mahalanobis distance sequence uses a unified scalar index to characterize the trajectory of the overall degradation degree of the suspension system over time. A monotonically increasing trend in the sequence values ​​indicates that the system is continuously deteriorating, a stable fluctuation in the sequence values ​​indicates that the system remains stable, and a sudden jump in the sequence values ​​indicates that a rapid deterioration event has occurred. The Mahalanobis distance sequence compresses the trajectory information in two-dimensional phase space into a one-dimensional distance time series curve, which facilitates subsequent extrapolation of the degradation trend using time series prediction models. To adapt to the training and prediction efficiency of the Long Short-Term Memory network, the Mahalanobis distance sequence is downsampled at the minute level: the time interval is divided into 1-minute intervals, and the mean of the Mahalanobis distance in each interval is calculated as the representative value of that interval, forming a downsampled Mahalanobis distance sequence. The time resolution after downsampling is 1 time / minute, and the subsequent training and prediction of the LSTM network are all based on this downsampled sequence.

[0092] The deviation direction angle θ(i) is defined as the angle between the direction vector from the ideal uniform origin O to the real-time state point Pnow(i) and the positive direction of the horizontal axis X. Its calculation formula is θ(i) = arctan(σdec(i) / Kagg(i)), where arctan(·) is the arctangent function. (See also...) Figure 4 In the diagram, the solid line with arrows connects the origin and the real-time state point, representing the current direction vector. The angle between this direction vector and the horizontal axis representing the micro-agglomeration index is the deviation direction angle marked in the diagram. The direction angle ranges from 0 degrees to 90 degrees in an open interval because both the micro-agglomeration index and the interface coupling factor decay are non-negative, and the state point always lies in the first quadrant of the phase space. Figure 4Taking the example shown, a smaller azimuth angle indicates that the direction of the state point's deviation from the origin is closer to the positive direction of the horizontal axis, meaning that the micro-agglomeration index dominates the overall degradation. A larger azimuth angle indicates that the direction of the state point's deviation from the origin is closer to the positive direction of the vertical axis representing the attenuation of the interface coupling factor, meaning that the attenuation of the interface coupling factor dominates the overall degradation. An azimuth angle of 45 degrees indicates that the contributions of the two indices are equal. When the micro-agglomeration index Kagg(i) equals 0, if the attenuation of the interface coupling factor σdec(i) is greater than 0, the azimuth angle is defined as 90 degrees, indicating that the degradation is entirely dominated by wettability degradation. If both indices are equal to 0, the state point is located at the origin, and the azimuth angle is meaningless, so it can be defined as a null value or no azimuth angle calculation can be performed. The process of generating the azimuth angle sequence is as follows: traverse the real-time state point sequence, perform azimuth angle calculation for each real-time state point, and arrange the azimuth angle values ​​into an ordered sequence according to the time window number. The orientation angle sequence depicts the trajectory of the dominant degradation direction over time. The stability of the sequence values ​​reflects the persistence of the failure mode, while abrupt changes in the sequence values ​​reflect the transformation of the failure mode.

[0093] The Mahalanobis distance sequence and orientation angle sequence generated in step S33 deconstruct the geometric features of the phase space trajectory from a complementary perspective. The Mahalanobis distance sequence quantitatively characterizes the overall degradation degree of the system deviating from an ideal homogeneous state, reflecting the amplitude characteristics of state evolution; the orientation angle sequence identifies the dominant failure mode in the system degradation process, reflecting the tendency characteristics of state evolution. The synergistic effect of both constitutes a complete description of the system's stability state. If only the Mahalanobis distance is calculated without calculating the orientation angle, it is impossible to distinguish between degradation dominated by micro-agglomeration and degradation dominated by wettability degradation, and the warning signal will lack directional guidance. Conversely, if only the orientation angle is calculated without calculating the Mahalanobis distance, the severity of degradation cannot be quantified, and it is impossible to determine whether the safety boundary has been exceeded. Step S33, by simultaneously extracting both types of features, provides the necessary input data for the risk type determination and stratification time prediction in step S34.

[0094] Step S34: Determine the current dominant risk type based on the direction angle sequence, and predict the time node for macro-level stratification based on the Mahalanobis distance sequence; generate a full-cycle stability prediction report based on the determination result of the current dominant risk type and the time node for macro-level stratification.

[0095] The method for determining the current dominant risk type based on the orientation angle sequence includes: obtaining the deviation orientation angle corresponding to the current moment in the orientation angle sequence; when the deviation orientation angle at the current moment is less than a preset angle threshold, the current dominant risk type is determined to be a small molecule failure risk; when the deviation orientation angle at the current moment is greater than or equal to the preset angle threshold, the current dominant risk type is determined to be a spreading failure risk.

[0096] Methods for predicting the time point of macroscopic stratification based on Mahalanobis distance sequences include: using long short-term memory networks to predict the trend of Mahalanobis distance sequences, and determining the expected time when the Mahalanobis distance sequence exceeds the critical distance threshold as the time point of macroscopic stratification.

[0097] The determination of the dominant risk type is based on a comparison between the current value of the direction angle sequence and a preset angle threshold. The preset angle threshold is θ. threshold The determination method is based on a balanced consideration of the impact of the two failure modes on application performance. When the impact of the two modes is comparable, the preset angle threshold can be set to 45 degrees to achieve a symmetrical judgment boundary. When one mode has a more severe impact on application performance, the preset angle threshold can be shifted to enhance the sensitivity detection of that mode. For example, if it is believed that the damage of spreading failure to the pesticide application effect is greater than the damage of micro-aggregation to storage stability, the preset angle threshold can be set to a value less than 45 degrees, such as 40 degrees, making the judgment logic more sensitive to the risk of spreading failure. The implementation process of the judgment logic is as follows: obtain the deviation direction angle θ(i) corresponding to the current time in the direction angle sequence, and compare θ(i) with the preset angle threshold θ. threshold Comparison, when θ(i) is less than θ threshold When the current dominant risk type is determined to be small molecule failure risk, it indicates that the system degradation is mainly driven by micro-agglomeration, and the small molecule encapsulation ability of the dispersant is failing, which corresponds to the user's functional requirements of preventing pesticide macromolecule adhesion and maintaining unchanged chemical properties; when θ(i) is greater than or equal to θ threshold When the dominant risk type is determined to be spreading failure, it indicates that the system degradation is mainly driven by wetting degradation, and the surfactant function of the dispersant is failing, corresponding to the user's functional requirements for full spreading and maintaining wetting. The determination of the dominant risk type provides directional guidance for formulation adjustments. When the determination is for small molecule failure, formulation adjustments should focus on enhancing the encapsulation stability of the dispersant, such as increasing the amount of high molecular weight dispersant or selecting dispersant varieties with stronger steric hindrance effects. When the determination is for spreading failure, formulation adjustments should focus on maintaining the wetting activity of the surfactant, such as supplementing with nonionic surfactants or optimizing the surfactant blending ratio. This directional guidance cannot be achieved in single-index monitoring systems. Traditional methods can only identify system degradation but cannot distinguish the type of degradation. Formulation adjustments can only adopt a blind strategy of comprehensive strengthening, which increases costs and may introduce formulation compatibility issues.

[0098] The prediction of macroscopic stratification time points is based on the trend extrapolation of the Mahalanobis distance sequence. Macroscopic stratification refers to the failure state in a suspension system where suspended particles accumulate and settle to the bottom of the container due to gravity, resulting in a clear upper layer and a concentrated lower layer. The occurrence of macroscopic stratification signifies the complete failure of the dispersant function and the loss of product usability. Macroscopic stratification corresponds to the Mahalanobis distance exceeding the critical distance threshold D. limit The method for determining the critical distance threshold is as follows: collect samples from historical batches that have experienced macroscopic stratification failure, trace back the value levels of their Mahalanobis distance sequences before stratification occurred, and take the statistical lower bound of the Mahalanobis distance at the critical moment of stratification for each sample as the critical distance threshold, ensuring that the threshold setting can cover the early warning needs of the vast majority of stratification events. For example, if the Mahalanobis distance distribution range of historical stratified samples at the critical moment of stratification is 0.8 to 1.2, then the critical distance threshold can be set to 0.75 to reserve a safety margin.

[0099] The trend extrapolation of Mahalanobis distance sequences is implemented using a Long Short-Term Memory (LSTM) network. LSM is a recurrent neural network architecture specifically designed for processing time-series data. It solves the gradient vanishing problem in long-series training of traditional recurrent neural networks through a gating mechanism, effectively capturing long-term dependencies in time-series data. The input of the LSM network is a historical segment of the Mahalanobis distance sequence, and the output is the predicted Mahalanobis distance at future time steps. The network structure consists of an input layer, an LSM layer, and a fully connected output layer: the input layer receives a historical Mahalanobis distance sequence of length Lhist, where Lhist is the historical window length parameter; the LSM layer contains N hidden units, each of which updates its cell state and hidden state through a combination of forget gate, input gate, and output gate. The forget gate determines the proportion of information to be discarded from the previous cell state, the input gate determines the proportion of information to be written into the cell state from the current input, and the output gate determines the proportion of information to be output to the hidden state from the cell state; the fully connected output layer maps the final hidden state of the LSM layer to the predicted Mahalanobis distance at future time steps. The training process of a Long Short-Term Memory (LSTM) network is as follows: Mahalanobis distance sequences are extracted from historical monitoring data as training samples. Each sequence is divided into historical segments and paired with corresponding future true values ​​using a sliding window method. The historical segments serve as network input, and the future true values ​​serve as supervision labels. Mean squared error is used as the loss function to measure the deviation between the predicted and true values. The Adam optimizer is used to iteratively update the network weight parameters to minimize the loss function. During training, the validation set loss is monitored to prevent overfitting. If the validation set loss does not decrease for several consecutive rounds, an early stopping mechanism is triggered to terminate training. For example, the historical window length Lhist can be set to the number of downsampled time steps covering the most recent 24 hours, downsampling at 1 minute / step, corresponding to 1440 time steps. The number of hidden units Nhidden can be set to 64, the initial learning rate can be set to 0.001, the training batch size can be set to 32, and the number of early stopping rounds can be set to 10. After training, the network parameters are fixed for online prediction.

[0100] The prediction process for the stratification time node is as follows: The historical Mahalanobis distance sequence before the current time is input into a trained Long Short-Term Memory (LSTM) network. The network outputs a sequence of predicted Mahalanobis distance values ​​for multiple future time points. The first time exceeding the critical distance threshold Dlimit is retrieved from the prediction sequence; this time is the predicted macroscopic stratification time node T. future If the predicted Mahalanobis distance values ​​at all times in the predicted sequence do not exceed the critical distance threshold, it indicates that macroscopic stratification will not occur in the system within the current prediction range, and a prompt message "Remaining stable period exceeds the prediction range" can be output. The prediction results of the stratification time nodes can be converted into the expression of remaining shelf life, which is equal to the time difference between the predicted stratification time node and the current time.

[0101] The full-cycle stability prediction report integrates the dominant risk type determination results and stratification timeline prediction results into a structured output document. The report includes the following components: a micro-agglomeration status assessment module outputs the current micro-agglomeration index value and its temporal trend, assessing the current state and remaining shelf life of the small molecule function; a macro-spreading performance assessment module outputs the current interface coupling factor decay value and its temporal trend, assessing the current state and remaining shelf life of the wetting and spreading ability; a comprehensive stability prediction module outputs the current Mahalanobis distance value and its trend prediction curve, providing the expected timeline for macro-stratification and the remaining shelf life; a risk type identification module outputs the determination result of the current dominant risk type, labeling it as a small molecule failure risk, a spreadability failure risk, or a combined risk of equal severity; and an intervention suggestion module provides corresponding formulation adjustment directions or product usage suggestions based on the risk type. The report can be transmitted to the production control system via a data interface for automated quality monitoring and early warning response.

[0102] Step S30 achieves the joint expression of multidimensional degradation characteristics of the dispersant suspension system by constructing a healthy phase space, realizes the decorrelation quantification of the comprehensive degradation degree through Mahalanobis distance calculation, realizes the directional identification of the dominant failure mode through orientation angle calculation, realizes the temporal extrapolation of degradation trends through long short-term memory networks, and realizes the systematic output of prediction results through structured reports. These methods upgrade the stability prediction of the dispersant suspension system from single-dimensional turbidity monitoring to multi-dimensional functional state assessment, from passive post-event detection to proactive trend prediction, and from vague anomaly alarms to precise risk type identification.

[0103] The construction of the healthy phase space of the suspended system integrates two independent one-dimensional index sequences into a two-dimensional phase space trajectory, enabling the joint change relationship of the two failure modes to be visualized within a unified geometric framework. This provides an intuitive geometric perspective for the mechanism analysis of failure modes and formulation improvement, revealing joint change modes that are difficult to detect when observing time-series curves in individual dimensions. The introduction of Mahalanobis distance compresses two-dimensional coordinate information into a single scalar distance. Through decorrelation transformation of the covariance matrix, the interference of statistical correlation between the two indices on distance calculation is eliminated. The quantification of the overall degradation degree is no longer affected by the potential chemical correlation between the two failure modes, improving the sensitivity of abnormal state identification. The calculation of orientation angles transforms the position information of state points into directional indicators, quantifying the contribution ratio of each failure mode to the overall degradation. This gives the early warning signal directional guidance capabilities, allowing formulation adjustments to precisely target the dominant failure mode rather than blindly and comprehensively strengthen it. The application of Long Short-Term Memory (LSTM) networks extrapolates historical trend information to future moments, enabling the prediction of the timing of stratified final states based on the temporal patterns of degradation trajectories. This allows for the quantitative prediction of the remaining shelf life and safe usage window of products, providing a forward-looking basis for production scheduling and inventory management.

[0104] Step S30 simplifies the complex scheme that originally required the independent deployment of two monitoring systems to assess the microscopic particle size state and interface wetting state separately into a unified prediction process based on a single turbidity sensor. This replaces the multi-sensor configuration at the hardware level with multi-dimensional feature decoupling and fusion at the data level, significantly reducing system complexity and deployment costs. It upgrades the original coarse-grained early warning mode, which could only output a binary "stable / unstable" judgment, to a fine-grained prediction mode that can distinguish failure types and quantify the remaining validity period, qualitatively improving the granularity and operability of the early warning information. Furthermore, it transforms the post-event detection mode, which required waiting for macroscopic stratification to occur before confirming failure, into a pre-event prediction mode that can trigger early warnings and predict the final state time at the microscopic degradation initiation stage, expanding the intervention window from almost zero to a considerable period before stratification occurs. These beneficial effects are not achieved through additional physical measurement methods or expensive analytical instruments, but through in-depth mining and intelligent analysis of existing time-series turbidity data, demonstrating the full release of data value and the application of algorithmic capabilities.

[0105] Example 2:

[0106] This embodiment, based on Embodiment 1, provides a system for predicting the anti-stratification stability of dispersant suspension systems, such as... Figure 5 As shown, it includes:

[0107] Microagglomeration calculation module: Collects the original time-series turbidity data stream of the dispersant suspension system, performs sliding window segmentation on the original time-series turbidity data stream to obtain local data segment sequences, calculates the microagglomeration index sequence based on the local data segment sequences, and forms a probability density distribution sequence;

[0108] Interface coupling calculation module: Performs asymmetric Gaussian fitting on the probability density distribution sequence to separate the main peak component and the tail component, and calculates the interface coupling factor attenuation sequence based on the main peak component and the tail component;

[0109] Stability prediction module: Constructs a healthy phase space for the suspended system, maps the micro-agglomeration index sequence and the interface coupling factor decay sequence to the healthy phase space of the suspended system to form a real-time state point sequence, calculates the Mahalanobis distance sequence and orientation angle sequence based on the real-time state point sequence, determines the current dominant risk type based on the orientation angle sequence, predicts the time node of macro-stratification based on the Mahalanobis distance sequence, and outputs a full-cycle stability prediction report.

[0110] Furthermore, in the micro-agglomeration calculation module, the method for calculating the micro-agglomeration index sequence based on the local data fragment sequence includes:

[0111] A time-frequency transformation is performed on a local data segment sequence to generate a time-frequency acoustic pattern sequence. The high-frequency Brownian energy sequence is extracted from the time-frequency acoustic pattern sequence, and the micro-aggregation index sequence is calculated.

[0112] The method for generating the time-frequency acoustic pattern sequence includes:

[0113] A short-time Fourier transform is performed on each local data segment in the local data segment sequence to generate a corresponding time-frequency acoustic pattern, which is then summarized to form a time-frequency acoustic pattern sequence; each pixel of the time-frequency acoustic pattern represents the energy density of the corresponding time and frequency.

[0114] The method for extracting the high-frequency Brownian energy sequence includes:

[0115] In each time-frequency acoustic pattern sequence, a high-frequency characteristic interval corresponding to the Brownian motion of suspended particles is defined. The energy density within the high-frequency characteristic interval is integrated to obtain the high-frequency Brownian energy value, and the results are summarized to form a high-frequency Brownian energy sequence.

[0116] Furthermore, in the interface coupling calculation module, the method for generating the interface coupling factor attenuation sequence includes:

[0117] For each probability density distribution in the probability density distribution sequence, calculate the offset distance between the centroid position of the corresponding tail component and the centroid position of the main peak component to obtain the centroid offset, and summarize them to form a centroid offset sequence.

[0118] Using the centroid offset corresponding to the initial reference data segment as the offset reference value, the attenuation ratio of each centroid offset in the centroid offset sequence relative to the offset reference value is calculated to obtain the interface coupling factor attenuation sequence.

[0119] Furthermore, in the stability prediction module, the methods for generating the Mahalanobis distance sequence and orientation angle sequence include:

[0120] In the healthy phase space of the suspended system, the origin of the coordinate system is defined as the ideal homogeneous origin. The Mahalanobis distance from each real-time state point in the real-time state point sequence to the ideal homogeneous origin is calculated and summarized to form a Mahalanobis distance sequence. The deviation angle of each real-time state point in the real-time state point sequence from the ideal homogeneous origin is calculated and summarized to form a direction angle sequence.

[0121] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0122] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0123] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the anti-stratification stability of a dispersant suspension system, characterized in that, The method includes: The original time-series turbidity data stream of the dispersant suspension system is collected, and the original time-series turbidity data stream is divided into local data segment sequences by sliding window. The micro-agglomeration index sequence is calculated based on the local data segment sequence and a probability density distribution sequence is formed. Asymmetric Gaussian fitting is performed on the probability density distribution sequence to separate the main peak component and the tail component, and the interface coupling factor attenuation sequence is calculated based on the main peak component and the tail component. A healthy phase space for the suspended system is constructed. The micro-aggregation index sequence and the interface coupling factor decay sequence are mapped to the healthy phase space of the suspended system to form a real-time state point sequence. The Mahalanobis distance sequence and orientation angle sequence are calculated based on the real-time state point sequence. The current dominant risk type is determined according to the orientation angle sequence. The time node of macro-stratification is predicted according to the Mahalanobis distance sequence, and a full-cycle stability prediction report is output.

2. The method for predicting the anti-stratification stability of a dispersant suspension system according to claim 1, characterized in that, The method for calculating the micro-agglomeration index sequence based on local data fragment sequences includes: A time-frequency transformation is performed on a local data segment sequence to generate a time-frequency acoustic pattern sequence. The high-frequency Brownian energy sequence is extracted from the time-frequency acoustic pattern sequence, and the micro-aggregation index sequence is calculated.

3. The method for predicting the anti-stratification stability of a dispersant suspension system according to claim 2, characterized in that, The method for generating the time-frequency acoustic pattern sequence includes: A short-time Fourier transform is performed on each local data segment in the local data segment sequence to generate a corresponding time-frequency acoustic pattern, which is then summarized to form a time-frequency acoustic pattern sequence; each pixel of the time-frequency acoustic pattern represents the energy density of the corresponding time and frequency.

4. The method for predicting the anti-stratification stability of a dispersant suspension system according to claim 3, characterized in that, The method for extracting the high-frequency Brownian energy sequence includes: In each time-frequency acoustic pattern sequence, a high-frequency characteristic interval corresponding to the Brownian motion of suspended particles is defined. The energy density within the high-frequency characteristic interval is integrated to obtain the high-frequency Brownian energy value, and the results are summarized to form a high-frequency Brownian energy sequence.

5. The method for predicting the anti-stratification stability of a dispersant suspension system according to claim 4, characterized in that, The method for generating the micro-agglomeration index sequence includes: The first local data segment in the local data segment sequence is marked as the initial reference data segment. The high-frequency Brown energy value corresponding to the initial reference data segment is used as the reference value. The decay ratio of each high-frequency Brown energy value in the high-frequency Brown energy sequence relative to the reference value is calculated to obtain the micro-aggregation index sequence.

6. The method for predicting the anti-stratification stability of a dispersant suspension system according to claim 5, characterized in that, The method for generating the probability density distribution sequence includes: Statistical analysis is performed on all turbidity values ​​within each local data segment in the local data segment sequence to calculate the probability density distribution corresponding to the local data segment. The probability density distributions corresponding to all local data segments are then summarized in chronological order to form a probability density distribution sequence.

7. The method for predicting the anti-stratification stability of a dispersant suspension system according to claim 6, characterized in that, The method for generating the interface coupling factor attenuation sequence includes: For each probability density distribution in the probability density distribution sequence, calculate the offset distance between the centroid position of the corresponding tail component and the centroid position of the main peak component to obtain the centroid offset, and summarize them to form a centroid offset sequence. Using the centroid offset corresponding to the initial reference data segment as the offset reference value, the attenuation ratio of each centroid offset in the centroid offset sequence relative to the offset reference value is calculated to obtain the interface coupling factor attenuation sequence.

8. The method for predicting the anti-stratification stability of a dispersant suspension system according to claim 7, characterized in that, The healthy phase space of the suspension system uses the micro-aggregation index as the horizontal axis variable and the attenuation of the interface coupling factor as the vertical axis variable.

9. The method for predicting the anti-stratification stability of a dispersant suspension system according to claim 8, characterized in that, The methods for generating the Mahalanobis distance sequence and orientation angle sequence include: In the healthy phase space of the suspended system, the origin of the coordinate system is defined as the ideal homogeneous origin. The Mahalanobis distance from each real-time state point in the real-time state point sequence to the ideal homogeneous origin is calculated and summarized to form a Mahalanobis distance sequence. The deviation angle of each real-time state point in the real-time state point sequence from the ideal homogeneous origin is calculated and summarized to form a direction angle sequence.

10. The method for predicting the anti-stratification stability of a dispersant suspension system according to claim 9, characterized in that, The method for determining the current dominant risk type based on the direction angle sequence includes: Obtain the deviation direction angle corresponding to the current moment in the direction angle sequence. When the deviation direction angle at the current moment is less than a preset angle threshold, determine that the current dominant risk type is small molecule failure risk. When the deviation direction angle at the current moment is greater than or equal to the preset angle threshold, determine that the current dominant risk type is spreadability failure risk.

11. A system for predicting the anti-stratification stability of a dispersant suspension system, used to implement the method for predicting the anti-stratification stability of a dispersant suspension system according to any one of claims 1-10, characterized in that, The system includes: Microagglomeration calculation module: Collects the original time-series turbidity data stream of the dispersant suspension system, performs sliding window segmentation on the original time-series turbidity data stream to obtain local data segment sequences, calculates the microagglomeration index sequence based on the local data segment sequences, and forms a probability density distribution sequence; Interface coupling calculation module: Performs asymmetric Gaussian fitting on the probability density distribution sequence to separate the main peak component and the tail component, and calculates the interface coupling factor attenuation sequence based on the main peak component and the tail component; Stability prediction module: Constructs a healthy phase space for the suspended system, maps the micro-agglomeration index sequence and the interface coupling factor decay sequence to the healthy phase space of the suspended system to form a real-time state point sequence, calculates the Mahalanobis distance sequence and orientation angle sequence based on the real-time state point sequence, determines the current dominant risk type based on the orientation angle sequence, predicts the time node of macro-stratification based on the Mahalanobis distance sequence, and outputs a full-cycle stability prediction report.

Images