A method and system for monitoring the milk fat separation process

Abstract

This invention relates to the field of milk fat separation monitoring technology, and discloses a method and system for monitoring the milk fat separation process. The method captures optical property data of the milk fat fluid using a multispectral sensor installed at the inlet of the separation equipment; identifies the compositional change patterns of the milk fat fluid based on the optical property data, and plots a dynamic distribution cloud map of the milk fat within the separation equipment according to these patterns; extracts high-concentration milk fat aggregation regions and low-concentration milk fat dispersion regions from the dynamic distribution cloud map, and defines monitoring priority regions based on these regions; adjusts the scanning frequency of the sensor array according to the monitoring priority regions, and calculates the data correlation degree between each region; constructs a monitoring network topology based on the data correlation degree, and generates a monitoring node activation sequence using a graph traversal algorithm; and controls the data acquisition operation of the sensor array according to the monitoring node activation sequence.

Description

Technical Field

[0001] This invention relates to the field of milk fat separation monitoring technology, specifically to a method and system for monitoring the milk fat separation process. Background Technology

[0002] Milk fat separation is a crucial step in dairy processing, directly impacting product quality and yield. Current technologies primarily rely on offline monitoring of key physicochemical indicators (such as fat and protein content) in imported raw materials and finished products. These methods suffer from significant lag, failing to reflect the dynamic fluid states within the separation equipment in real time. By the time abnormal indicators are detected, the separation process may have already deviated from optimal operating conditions for some time, resulting in wasted materials and energy.

[0003] Another common approach is to install single-parameter sensors, such as online refractometers or near-infrared sensors, at key locations in the equipment to monitor the concentration or composition of localized areas. However, the internal flow field of separation equipment is complex, and the distribution of milk fat components exhibits significant spatial inhomogeneity and temporal dynamics. Measurement data from a single point or a few fixed points cannot comprehensively and accurately characterize the component distribution across the entire cross-section of the equipment or even in three-dimensional space. This prevents the monitoring system from promptly detecting localized concentration anomalies or uneven dispersion, such as the accumulation of high-concentration milk fat in specific areas of the separation discs, or poor milk fat dispersion due to improper flow rates. These localized phenomena are often early signs of decreased separation efficiency or even equipment blockage, but monitoring strategies based on limited measurement points are insufficient to effectively capture this crucial information.

[0004] Existing monitoring strategies typically employ fixed sampling frequencies and scanning paths, indiscriminately collecting data from all areas. This uniform monitoring model leads to low resource utilization in real-world industrial scenarios with limited computing resources and communication bandwidth. A large amount of data comes from areas with gradual changes and stable conditions, while the data collection density for rapidly changing and fault-prone critical areas may be insufficient. This results in monitoring systems that are either slow to respond due to excessive data processing burdens or have blind spots due to incomplete collection of key information. Therefore, existing technologies for monitoring the milk fat separation process generally suffer from insufficient real-time performance, low spatial resolution, and rigid monitoring strategies, making it difficult to achieve precise perception and adaptive control of the internal dynamics of the separation process. There is an urgent need for a method that can perceive the internal component distribution of equipment in real time and comprehensively, and intelligently optimize the allocation of monitoring resources. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring the milk fat separation process, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for monitoring the milk fat separation process, the method comprising:

[0007] Optical property data of the milk fat fluid are captured by a multispectral sensor installed at the inlet of the separation equipment;

[0008] Based on the optical property data, the composition change pattern of the milk fat fluid is identified, and a dynamic distribution cloud map of the milk fat in the separation device is drawn according to the composition change pattern.

[0009] High-concentration milk fat aggregation areas and low-concentration milk fat dispersion areas are extracted from the dynamic distribution cloud map, and monitoring priority areas are defined based on the high-concentration milk fat aggregation areas and low-concentration milk fat dispersion areas.

[0010] The scanning frequency of the sensor array is adjusted according to the monitoring priority area, and the data correlation between each area is calculated;

[0011] Based on the data correlation, a monitoring network topology is constructed, and a graph traversal algorithm is used to generate a monitoring node activation sequence;

[0012] The data acquisition operation of the sensor array is controlled according to the activation sequence of the monitoring node.

[0013] Preferably, the step of identifying the compositional change pattern of milk fat fluid based on the optical property data includes: performing wavelet transform processing on the optical property data to extract frequency features at different scales; matching the frequency features with a pre-stored milk fat composition database to identify the distribution pattern of fat particles in milk fat; calculating the compositional change gradient according to the distribution pattern, and determining the compositional change pattern based on the compositional change gradient.

[0014] Preferably, the step of drawing a dynamic distribution cloud map of milk fat within the separation device based on the component change pattern includes: mapping the component change gradient onto the three-dimensional coordinate space of the separation device to generate an initial distribution point cloud; performing spatial interpolation calculations on the initial distribution point cloud to fill the gaps between the point clouds and form a continuous distribution surface; and superimposing the continuous distribution surface onto the device structure model to output a dynamic distribution cloud map.

[0015] Preferably, the step of extracting high-concentration milk fat aggregation regions and low-concentration milk fat dispersion regions from the dynamic distribution cloud map includes: setting a concentration threshold range and dividing the dynamic distribution cloud map into regions; calculating the average concentration value of each segmented region, marking regions with average concentration values ​​greater than the upper threshold as high-concentration milk fat aggregation regions, and marking regions with average concentration values ​​less than the lower threshold as low-concentration milk fat dispersion regions.

[0016] Preferably, defining the monitoring priority area based on the high-concentration milk fat aggregation area and the low-concentration milk fat dispersion area includes: assigning the highest priority to the high-concentration milk fat aggregation area and assigning the lowest priority to the low-concentration milk fat dispersion area; for the transition zone between adjacent areas, calculating the concentration change rate and adjusting the priority level according to the concentration change rate.

[0017] Preferably, adjusting the scanning frequency of the sensor array according to the monitoring priority area includes: acquiring the position information of each sensor in the sensor array; setting a scanning frequency coefficient for each sensor according to the priority level of the monitoring priority area, wherein the higher the priority, the larger the scanning frequency coefficient; and multiplying the scanning frequency coefficient by the base frequency to obtain the adjusted scanning frequency.

[0018] Preferably, the calculation of the data correlation degree between each region includes: extracting feature vectors from the time series data collected by the sensor array; calculating the cosine similarity between the feature vectors of different regions, and using the cosine similarity as the data correlation degree.

[0019] Preferably, the step of constructing the monitoring network topology based on the data correlation degree includes: representing each monitoring area as a network node, representing the data correlation degree as the edge weight between nodes; using an adjacency matrix to store node and edge weight information, and constructing the monitoring network topology;

[0020] The method of generating the monitoring node activation sequence using the graph traversal algorithm includes: selecting the node with the highest correlation from the monitoring network topology as the starting node; applying a depth-first search algorithm to traverse all nodes and recording the access order; and generating the monitoring node activation sequence according to the access order.

[0021] Preferably, the data acquisition operation of controlling the sensor array according to the monitoring node activation sequence includes: parsing the node number in the monitoring node activation sequence; querying the corresponding sensor identifier according to the node number; and sending an acquisition command to the sensor identifier to trigger data acquisition.

[0022] Preferably, the present invention also includes a milk fat separation process monitoring system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the milk fat separation process monitoring method described above.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] This invention utilizes multispectral data to create dynamic distribution cloud maps, providing a clear visual representation of the spatial distribution and temporal evolution of milk fat within the separation equipment. This overcomes the limitations of traditional single-point measurements, which can only acquire localized information, and enables more comprehensive and visualized monitoring of the separation process. By automatically identifying and defining priority monitoring areas from the cloud maps, monitoring attention is focused on key areas such as high-concentration aggregation and low-concentration dispersion, making monitoring more targeted.

[0025] By dynamically adjusting the scanning frequency of the sensor array based on priority areas, the allocation of limited monitoring resources was optimized. Higher data acquisition density was allocated to more critical and volatile areas, thereby significantly improving the ability to capture key process parameters while keeping the overall data volume under control. Calculating the data correlation between areas and constructing the monitoring network topology revealed the intrinsic connections and influence paths between state parameters in different areas, enhancing the system's depth of understanding of the process's operational status.

[0026] By utilizing a graph traversal algorithm to generate the activation sequence of monitoring nodes, data acquisition activities follow an orderly and efficient pattern based on the inherent logical relationships within the process, avoiding the blindness and redundancy of sensor operation. This method enhances the ability to dynamically track the separation process, enabling more sensitive detection of early signs of anomalies and creating conditions for early intervention in the control system. The entire solution improves the intelligence and adaptability of the monitoring system, allowing it to better cope with changes in operating conditions and raw material fluctuations in the production process, ultimately promoting stable operation of the separation process and consistent product quality. Attached Figure Description

[0027] Figure 1 This is a schematic diagram illustrating the working principle of the milk fat separation process monitoring method described in this invention.

[0028] Figure 2 A flowchart illustrating the working principle of identifying milk fat fluid composition variation patterns based on optical property data;

[0029] Figure 3 A flowchart illustrating the working principle of adjusting the sensor array scanning frequency based on monitoring priority areas;

[0030] Figure 4 Heatmap of the correlation matrix of data in the milk fat separation monitoring area;

[0031] Figure 5 A graph showing the correlation between activation sequences and priorities of milk fat separation monitoring nodes. Detailed Implementation

[0032] 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.

[0033] Please see Figure 1 This invention provides a method and system for monitoring the milk fat separation process. The method comprises the following steps: installing a multispectral sensor at the inlet of the separation equipment. This sensor captures optical property data of the milk fat fluid during its flow, including parameters such as transmittance, reflectance, and scattering. Based on the optical property data, the compositional change patterns of the milk fat fluid are identified. These patterns reflect the dynamic behavior of fat particles in the fluid. A dynamic distribution cloud map of the milk fat within the separation equipment is plotted based on the compositional change patterns, visually representing the spatial variation of milk fat concentration. High-concentration milk fat aggregation regions and low-concentration milk fat dispersion regions are extracted from the dynamic distribution cloud map, with the high-concentration regions representing fat-rich areas. Low-concentration areas represent diluted fat regions. Monitoring priority areas are defined based on high-concentration milk fat accumulation areas and low-concentration milk fat dispersion areas. These priority areas guide the allocation of monitoring resources. The scanning frequency of the sensor array is adjusted according to the monitoring priority areas, with the adjustment based on the importance of each area. Simultaneously, the data correlation degree between each area is calculated, which quantifies the similarity of data between areas. A monitoring network topology is constructed based on the data correlation degree, representing areas as nodes and connecting them with edges. A graph traversal algorithm is used to generate a monitoring node activation sequence, which specifies the order in which sensors are activated. The data acquisition operation of the sensor array is controlled according to the monitoring node activation sequence to achieve efficient and targeted monitoring.

[0034] Example 1: See Figure 2Identifying the compositional variation patterns of milk fat fluid based on optical property data involves wavelet transform processing of optical property data captured by multispectral sensors. Wavelet transform processing employs multi-resolution analysis to decompose the optical signal into frequency components at different scales, extracting frequency features at different scales, including high-frequency details and low-frequency approximations. High-frequency details correspond to microscopic changes in fat particles, while low-frequency approximations reflect overall concentration trends. When matching these frequency features with a pre-stored milk fat component database, which stores frequency patterns corresponding to fat particle size and concentration obtained from historical experiments, the matching process uses pattern recognition algorithms such as support vector machines or neural networks to compare real-time features with database entries, identifying the distribution patterns of fat particles in milk fat. These distribution patterns describe the aggregation or dispersion dynamics of particles in the fluid. The compositional variation gradient is calculated based on the distribution patterns using spatial derivative operations. The compositional variation gradient represents the rate of change of concentration values ​​in the spatial direction, and the compositional variation pattern is determined based on the gradient. The compositional variation patterns are classified into uniform distribution, gradient change, or abrupt change types to guide subsequent visualization. In practical implementation, drawing a dynamic distribution cloud map of milk fat within the separation equipment based on the composition change pattern involves mapping the composition change gradient to the three-dimensional coordinate space of the separation equipment. The three-dimensional coordinate space is defined by the geometric model of the separation equipment. The mapping process assigns gradient values ​​to spatial points through a coordinate transformation algorithm, generating an initial distribution point cloud. The initial distribution point cloud consists of a discrete set of points, each carrying a concentration value and three-dimensional position information. Spatial interpolation calculations are performed on the initial distribution point cloud. Spatial interpolation uses the Kriging algorithm or the inverse distance weighting method to fill the gaps between the point clouds, forming a continuous distribution surface. The continuous distribution surface smoothly represents the concentration change trend. When the continuous distribution surface is superimposed on the equipment structure model, the equipment structure model is a computer-aided design model or a three-dimensional mesh representation. The superposition is completed through three-dimensional rendering techniques such as ray casting or texture mapping. The output dynamic distribution cloud map is displayed on the monitoring interface in the form of color coding or contour lines.

[0035] In practical implementation, the specific operations of wavelet transform processing include selecting wavelet basis functions such as Daubechies wavelet or Haar wavelet, performing discrete wavelet transform on optical property data to decompose it into multiple scale layers, each scale layer corresponding to a specific frequency bandwidth, high-frequency features capturing the rapid fluctuations of fat particles, low-frequency features revealing the macroscopic concentration distribution, matching the extracted features with a milk fat composition database using similarity calculations such as Euclidean distance or Pearson correlation coefficient, applying clustering algorithms such as K-means or hierarchical clustering to divide particle groups when identifying distribution patterns, calculating the gradient of component changes using numerical differencing methods such as central difference or forward difference, and determining the component change pattern through threshold comparison or decision tree classifiers. In practical implementation, the mapping steps for drawing dynamic distribution cloud maps involve assigning component change gradient values ​​to three-dimensional grid points. The grid points are based on the internal space of the separated equipment. The initial distribution point cloud is generated by projecting sensor data points onto a three-dimensional coordinate system. Spatial interpolation calculations ensure surface continuity and avoid holes. After the continuous distribution surface is generated, it is aligned with the equipment structure model using a coordinate matching algorithm. The output dynamic distribution cloud map supports real-time updates and historical data backtracking.

[0036] In some embodiments, continuous wavelet transform can be used as an alternative method for wavelet transform processing. Continuous wavelet transform provides finer frequency resolution and is suitable for non-stationary signal analysis. The extracted frequency features include instantaneous frequency and amplitude information. When matching with a milk fat composition database, dynamic time warping algorithms can be used to process time series changes. The distribution pattern is identified by introducing a time dimension to analyze particle motion trajectories. The gradient of composition change is calculated by combining spatial and temporal derivatives, and the composition change pattern is determined and incorporated into a machine learning model such as random forest for pattern classification. In some embodiments, the three-dimensional coordinate space for drawing dynamic distribution cloud maps can be based on polar coordinates or cylindrical coordinates to adapt to the shape of the separated equipment. The mapping process uses spherical interpolation to process the surface space. Dense sampling technology is used to improve the resolution of the initial distribution point cloud generation. Spatial interpolation calculation uses natural neighborhood interpolation or radial basis function interpolation. The continuous distribution surface is superimposed with the equipment structure model to achieve mixed reality visualization using augmented reality technology.

[0037] Optionally, signal preprocessing steps such as denoising and normalization can be added to the wavelet transform process. Denoising uses filters such as Wiener filters or wavelet thresholding. Normalization scales the optical characteristic data to a standard range, improving the accuracy of feature extraction. When matching with a milk fat composition database, an online learning mechanism can be implemented to update the database content, adapting to changes in different milk fat types. When identifying distribution patterns, multi-sensor data is fused to improve robustness. An adaptive step-size difference algorithm is used to optimize computational efficiency when calculating the gradient of composition changes. Fuzzy logic is introduced to handle uncertainties when determining the composition change pattern. Optionally, when drawing a dynamic distribution cloud map, the initial distribution point cloud can be compressed to reduce memory usage. Spatial interpolation calculations are parallelized to accelerate surface generation. Continuous distribution surfaces are dynamically simplified using multi-level detail technology. The output dynamic distribution cloud map integrates interactive functions, allowing users to adjust view parameters.

[0038] It is understandable that wavelet transform processing is a standard technique in the field of signal processing; multi-resolution analysis can effectively separate noise and useful information; frequency feature extraction provides input data for pattern recognition; the construction of a milk fat composition database relies on the accumulation of a large amount of experimental data; distribution pattern identification helps to understand fluid dynamics behavior; component change gradient calculation is the foundation of spatial analysis; and component change pattern classification simplifies the interpretation of complex data. Similarly, it is understandable that the drawing of dynamic distribution cloud maps relies on computer graphics principles; three-dimensional coordinate space mapping ensures spatial accuracy; the initial distribution point cloud represents discrete sampled data; spatial interpolation compensates for insufficient sampling; continuous distribution surfaces provide intuitive visualization; overlaying equipment structure models enhances context awareness; and the output cloud map serves monitoring and decision-making.

[0039] In practical implementation, the process of identifying the compositional variation patterns of milk fat fluid based on optical property data includes data acquisition, preprocessing, transformation, feature extraction, matching, pattern recognition, gradient calculation, and pattern determination. Data acquisition is completed by a multispectral sensor. Preprocessing includes calibration and formatting. The transformation stage performs wavelet decomposition. Feature extraction selects salient feature vectors. The matching process queries the database. Pattern recognition outputs statistical descriptions. Gradient calculation generates a scalar field. Pattern determination outputs classification labels. In practical implementation, the process of drawing a dynamic distribution cloud map of milk fat within the separation device based on the compositional variation patterns includes spatial mapping, point cloud generation, interpolation calculation, surface formation, model overlay, and cloud map output. Spatial mapping projects data onto a 3D mesh. Point cloud generation creates a discrete dataset. Interpolation calculation fills the data gaps. Surface formation constructs a geometric surface. Model overlay merges virtual and real models. Cloud map output generates a visualization result.

[0040] In practical implementation, the parameters of wavelet transform processing, such as wavelet type and number of decomposition levels, are dynamically adjusted according to the characteristics of the optical signal. Too many decomposition levels may lead to overfitting, while too few may result in the loss of details. Feature extraction selects features with high discriminative power, such as energy entropy or spectral centroid. The algorithm for matching with the milk fat composition database optimizes the index structure to improve query speed. When identifying distribution patterns, the stationarity of the time series is considered. The gradient of component change is calculated to verify numerical stability, and the component change pattern is determined to evaluate classification accuracy. In practical implementation, the 3D coordinate space resolution of the dynamic distribution cloud map is set according to the device size. The mapping process handles coordinate transformation errors. The initial distribution point cloud density affects the interpolation accuracy. The spatial interpolation calculation algorithm is selected based on the data distribution characteristics. Continuous distribution surface smoothing avoids jaggedness. The device structure model is superimposed to ensure consistent proportions. The output dynamic distribution cloud map format supports standard image files.

[0041] It is understandable that the accuracy of component change pattern recognition directly affects the subsequent monitoring effect. Wavelet transform processing can capture multi-scale features, the quality of the milk fat component database determines the matching reliability, distribution pattern recognition provides a basis for process optimization, component change gradient quantifies the change intensity, and component change patterns simplify the monitoring logic. It is also understandable that the intuitiveness of dynamic distribution cloud maps improves monitoring efficiency, three-dimensional coordinate space mapping maintains spatial relationships, the initial distribution point cloud is the foundation of digital twins, spatial interpolation calculation enhances data continuity, continuous distribution surfaces facilitate concentration analysis, equipment structure model overlay provides a reference framework, and the output cloud map supports multi-terminal access.

[0042] In practical implementation, recognizing the compositional variation patterns of milk fat fluid based on optical property data requires hardware support such as a high-performance processor to perform wavelet transforms. Software modules include a signal processing library and a database management system. Feature extraction algorithms optimize computational complexity, the matching process achieves real-time response, pattern recognition outputs structured data, gradient calculations are processed in parallel, and pattern determination is integrated into the monitoring pipeline. Furthermore, generating dynamic distribution cloud maps of milk fat within the separation equipment based on compositional variation patterns requires a graphics processing unit to accelerate rendering, a 3D engine to handle coordinate transformations, a point cloud processing library to manage data, interpolation algorithms to optimize numerical calculations, surface generation using meshing technology, model overlay calling CAD interfaces, and cloud map output driving the display device.

[0043] In practical implementation, the details of wavelet transform processing include selecting wavelet function families such as Symlet or Coiflet wavelets, determining the decomposition level based on the signal sampling rate, using symmetric expansion or zero-filling for boundary processing, dimensionality reduction processing after feature extraction such as principal component analysis, implementing a caching mechanism to improve efficiency by matching with a milk fat composition database, eliminating baseline drift during distribution pattern identification, using a Gaussian derivative filter to smooth noise when calculating component change gradients, and using cross-validation to evaluate the model when determining component change patterns. In practical implementation, the details of drawing dynamic distribution cloud maps include setting the step size for discretization in the three-dimensional coordinate space, using the world coordinate system for the mapping process, using an octree data structure for initial distribution point cloud storage, fitting spatial interpolation parameters such as the variation model in the Kriging algorithm, applying the marchingcubes algorithm to generate continuous distribution surfaces, handling occlusion relationships by overlaying equipment structure models, and selecting a rainbow color system or a heatmap system for the output dynamic distribution cloud map color mapping.

[0044] In some embodiments, composition change pattern recognition can incorporate deep learning models such as convolutional neural networks to automatically learn features, avoiding manual feature engineering. Wavelet transform is used as a preprocessing step input to the network. The milk fat composition database is expanded into an online knowledge base. Distribution pattern recognition uses graph neural networks to capture spatial dependencies. Composition change gradient calculation is combined with physical constraints, and composition change pattern determination is trained end-to-end. In some embodiments, dynamic distribution cloud map drawing can employ point cloud direct rendering technology to omit interpolation steps. The initial distribution point cloud is completed through deep learning. Spatial interpolation uses neural radiation fields to achieve light field rendering. Continuous distribution surfaces are replaced with probability distribution maps. Device structure model overlay uses virtual-real registration technology, and the output cloud map supports immersive viewing on virtual reality devices.

[0045] Optionally, the wavelet transform processing can incorporate quality control steps to monitor transform quality, such as signal-to-noise ratio calculation, feature extraction incorporating domain adaptation techniques to handle distribution shifts, version control of the milk fat composition database to track changes, uncertainty quantification for distribution pattern identification, composition change gradient calculation to verify conservation laws, and confidence scores for composition change pattern determination. Optionally, the dynamic distribution cloud map rendering can optimize computational resource allocation, stream processing of initial distribution point clouds to handle large datasets, incremental updates via spatial interpolation to reduce repetitive work, compressed transmission of continuous distribution surfaces to save bandwidth, lightweight loading of device structure models, and adaptive output cloud maps to screen resolution.

[0046] In practical implementation, the overall architecture for identifying the compositional variation patterns of milk fat fluid based on optical property data includes a data input layer, a processing layer, and an output layer. The data input layer receives sensor stream data, the processing layer sequentially performs wavelet transform, feature extraction, matching, pattern recognition, gradient calculation, and pattern determination modules, and the output layer publishes the pattern results to the monitoring system. In practical implementation, the overall architecture for drawing a dynamic distribution cloud map of milk fat within the separation device based on the compositional variation patterns includes a data conversion layer, a geometry processing layer, and a rendering layer. The data conversion layer processes gradients to point cloud mapping, the geometry processing layer performs interpolation and surface generation, and the rendering layer handles overlay and output.

[0047] It is understood that the implementation method of Example 1 lays the foundation for monitoring the milk fat separation process. Composition change pattern recognition transforms raw data into knowledge, and dynamic distribution cloud mapping provides spatial insights. The combination of these two methods achieves a closed loop from data to decision-making. It is also understood that optimizing implementation details can improve system performance. Wavelet transform balances time-frequency resolution, feature extraction focuses on relevant information, database matching ensures historical consistency, pattern recognition reveals hidden patterns, gradient calculation highlights areas of change, pattern determination classifies key monitoring areas, cloud mapping enhances contextual awareness, spatial mapping maintains realism, point cloud generation captures details, interpolation fills gaps, curved surfaces form a smooth representation, model overlay integrates the environment, and the output cloud map drives action.

[0048] Example 2: Extracting high-concentration milk fat aggregation regions and low-concentration milk fat dispersion regions from a dynamic distribution cloud map includes setting concentration threshold ranges. The concentration threshold ranges are determined based on statistical analysis of historical separation process data, with an upper and lower threshold. The upper threshold represents the critical value for high-concentration regions, while the lower threshold represents the critical value for low-concentration regions. The dynamic distribution cloud map is segmented into continuous and non-overlapping sub-regions using image segmentation algorithms such as the watershed algorithm or region growing algorithm. The average concentration value of each segmented region is calculated by summing the concentration values ​​of all pixels or voxels within the region and dividing by the total area or volume of the region. Regions with an average concentration value greater than the upper threshold are marked as high-concentration milk fat aggregation regions, indicating a highly enriched state of fat particles. Regions with an average concentration value less than the lower threshold are marked as low-concentration milk fat dispersion regions, reflecting a sparsely distributed state of fat particles. Based on the high-concentration milk fat accumulation area and the low-concentration milk fat dispersion area, the monitoring priority area is defined. The highest priority is assigned to the high-concentration milk fat accumulation area, which corresponds to the densest allocation of monitoring resources. The lowest priority is assigned to the low-concentration milk fat dispersion area, which corresponds to the most basic monitoring coverage. For the transition zone between adjacent areas, which is the boundary area where the concentration value changes continuously, the concentration change rate is calculated by solving the gradient amplitude of the concentration field in space or calculating the concentration difference per unit distance. The priority level is adjusted according to the concentration change rate. The higher the concentration change rate, the higher the priority level is assigned to capture the rapidly changing fluid dynamics.

[0049] In practical implementation, setting concentration threshold ranges involves collecting a large amount of cloud map data during normal separation processes, calculating the statistical quantiles of the overall concentration distribution (e.g., the 95th percentile as the upper threshold and the 5th percentile as the lower threshold), preprocessing the dynamic distribution cloud map during region segmentation using Gaussian filtering to remove noise, pre-calculating the concentration gradient map and determining marker points for the watershed algorithm, merging adjacent pixels based on concentration similarity using the region growing algorithm starting from the seed point, employing numerical integration to ensure accuracy in calculating the average concentration value, and generating binary images or label matrices to identify different region types. In practical implementation, when defining monitoring priority areas, a priority mapping table is established to map concentration ranges to priority levels, with the highest priority set to level 1 and the lowest to level 5. Transition zone detection uses edge detection algorithms such as the Canny operator to locate region boundaries, concentration change rate calculation uses the Sobel or Prewitt operator to calculate the spatial derivative, and priority adjustment is based on predefined functional relationships such as linear or piecewise functions to convert change rate values ​​into level adjustment amounts.

[0050] In some embodiments, the concentration threshold range can be dynamically adjusted using an adaptive thresholding method. This method calculates the upper and lower thresholds based on real-time cloud map statistical characteristics such as mean and variance. Region segmentation can utilize superpixel segmentation algorithms like the SLIC algorithm to generate more natural region boundaries. The average concentration value calculation can incorporate a weighting factor to consider the complexity of the region's shape. When labeling regions, a medium-concentration region category can be added as a supplement to high-concentration milk fat aggregation regions and low-concentration milk fat dispersion regions. In some embodiments, defining a monitoring priority region can establish separate priority calculation rules for transition zones. Priority level division can employ more levels, such as a ten-level system, to improve resolution. The concentration change rate calculation can combine the time dimension to calculate the spatiotemporal gradient. Priority adjustment can incorporate machine learning models to optimize the adjustment strategy based on historical performance.

[0051] Optionally, a safety margin can be added when setting the concentration threshold range, appropriately lowering the upper limit threshold and appropriately raising the lower limit threshold to avoid false judgments. After region segmentation, morphological operations such as opening or closing operations can be performed to optimize the region shape. Robust statistics such as the median can be used to calculate the average concentration value to resist the influence of outliers. When marking regions, a region merging operation can be performed to merge small fragmented regions into adjacent large regions. Optionally, when defining monitoring priority regions, a region area factor can be considered, assigning higher priority to large areas with high concentrations. Priority adjustment in transition zones can introduce a smoothing function to prevent level jumps. The monitoring priority region results can be post-processed, such as priority diffusion, to smoothly transition the priority between adjacent regions.

[0052] It is understandable that extracting regions from dynamic distribution cloud maps is a fundamental step in spatial analysis. Setting concentration threshold ranges discretizes continuous concentration values ​​into categories, region segmentation algorithms divide the image into semantic units, average concentration value calculation provides overall regional features, and labeling operations achieve region classification. Identifying high-concentration milk fat accumulation areas and low-concentration milk fat dispersion areas focuses on key monitoring areas. It is also understandable that defining monitoring priority areas is the core of resource optimization and allocation. Priority allocation establishes a hierarchy of monitoring importance, transition zones handle boundary phenomena of concern, concentration change rate quantifies the intensity of change, and priority level adjustment enables dynamic resource scheduling.

[0053] In practical implementation, the complete process for extracting high-concentration milk fat aggregation regions and low-concentration milk fat dispersion regions from dynamic distribution cloud maps includes image preprocessing, threshold calculation, segmentation execution, concentration calculation, and region labeling stages. Image preprocessing enhances cloud map quality, threshold calculation determines classification criteria, segmentation execution divides the image space, concentration calculation quantifies regional characteristics, and region labeling outputs classification results. In practical implementation, the process for defining monitoring priority regions based on high-concentration milk fat aggregation regions and low-concentration milk fat dispersion regions includes priority initialization, transition zone identification, change rate calculation, and level adjustment steps. Priority initialization assigns basic levels, transition zone identification locates boundary regions, change rate calculation assesses the degree of change, and level adjustment outputs the final priority mapping.

[0054] In practical implementation, setting the concentration threshold range needs to consider the separation process requirements. The upper and lower thresholds should be updated regularly to adapt to changes in raw materials. Parameters of the region segmentation algorithm, such as the flooding level of the watershed algorithm or the similarity tolerance of the region growing algorithm, need careful adjustment. The average concentration value calculation verifies numerical stability. After marking regions, a region validity check is performed to remove invalid small regions. In practical implementation, when defining monitoring priority regions, the absolute values ​​of the highest and lowest priorities are set based on the total monitoring resources. The uncertainty of the region boundary is addressed by identifying transition zones. The concentration change rate calculation verifies the numerical accuracy of the gradient calculation. The priority level adjustment results generate a priority mapping table for subsequent module calls.

[0055] It is understandable that the accurate extraction of high-concentration milk fat accumulation areas and low-concentration milk fat dispersion areas directly affects the monitoring effect. The concentration threshold range needs to balance sensitivity and specificity. The quality of region segmentation affects the integrity of the region. The representativeness of the average concentration value determines the classification accuracy. The region labeling results drive subsequent decisions. It is understandable that defining monitoring priority areas enables intelligent monitoring scheduling, priority allocation optimizes resource utilization, handling transition zones avoids monitoring blind spots, concentration change rate senses sensitive areas, and priority level adjustment enhances system adaptability.

[0056] In practical implementation, extracting regions from dynamically distributed cloud maps requires image processing libraries such as OpenCV; threshold calculation uses statistical function libraries; segmentation algorithms optimize computational efficiency for handling large-size cloud maps; average concentration value calculation is parallelized; and region labeling results are stored in a GIS-compatible format. Furthermore, defining monitoring priority areas requires a priority management module; transition zone detection calls image edge detection functions; concentration change rate calculation uses a numerical differentiation library; and level adjustment logic is implemented using a configuration rule engine.

[0057] In practical implementation, the details of setting concentration threshold ranges include data standardization to ensure threshold comparability, the use of a dual-threshold method to generate strong and weak edges, post-processing of region segmentation including region contour extraction and area filtering, calculation of average concentration values ​​using different integration methods for different region shapes, and labeling operations to generate labeled raster data. The details of defining monitoring priority areas include the design of a priority level coding scheme, determining the width of the transition zone based on the equipment size ratio, selecting a gradient operator in the concentration change rate calculation considering computational efficiency, calibrating the priority adjustment function parameters experimentally, and finally outputting a spatial priority distribution map.

[0058] In some embodiments, region extraction can incorporate multi-threshold segmentation techniques to handle complex concentration distributions. The Otsu algorithm is used to automatically determine the optimal threshold, and the graph cut algorithm is employed to optimize the global energy function for region segmentation. Spatial autocorrelation correction is added to the calculation of the average concentration value, and region labeling is combined with a convolutional neural network for semantic segmentation. In some embodiments, priority definition can incorporate a multi-objective optimization model that considers concentration value, rate of change, and regional location factors simultaneously. The level set method is used to describe boundary evolution in transition zone processing, temporal smoothing filtering is added to the concentration change rate calculation, and a fuzzy inference system is used to handle uncertainty issues in priority level adjustment.

[0059] Optionally, a quality assessment step can be added during region extraction, calculating segmentation evaluation indicators such as the Descein coefficient to verify the segmentation effect, incorporating a manual review mechanism for threshold range settings, providing uncertainty estimates for average concentration calculations, and supporting manual correction of region labeling results. Optionally, a priority consistency check can be added during priority definition implementation to avoid excessive differences in priorities between adjacent regions, incorporating multi-scale analysis for transition zone identification, verifying physical rationality through concentration change rate calculations, and performing smoothing filtering on priority level adjustment results.

[0060] In practical implementation, the system architecture for extracting high-concentration milk fat aggregation areas and low-concentration milk fat dispersion areas from dynamic distribution cloud maps includes a data input module, a processing module, and an output module. The data input module receives cloud map data, the processing module performs threshold setting, region segmentation, concentration calculation, and labeling operations, and the output module generates a region classification map. In practical implementation, the system architecture for monitoring priority regions based on region definitions includes a priority initialization module, a transition processing module, and a level output module. The priority initialization module assigns basic priorities, the transition processing module identifies boundaries and calculates the rate of change, and the level output module generates the final priority distribution.

[0061] It is understood that the implementation method of Example 2 establishes a conversion pipeline from cloud map to priority, region extraction transforms visualized data into structured information, and priority definition provides intelligent guidance for monitoring behavior. It is also understood that the rigor of implementation details ensures system reliability; the setting of concentration threshold ranges requires a scientific statistical basis; the selection of region segmentation algorithms affects the morphology of the regions; the calculation of average concentration values ​​requires numerical accuracy; region marking requires clear classification standards; priority allocation requires a reasonable hierarchical system; transition zone processing requires fine boundary analysis; concentration change rate calculation requires numerical stability; and priority adjustment requires smooth transition logic.

[0062] Example 3: See Figure 3 The process of adjusting the scanning frequency of the sensor array based on the monitoring priority area includes acquiring the position information of each sensor in the sensor array. The position information is stored in the system configuration database or acquired in real time through the global positioning system module. The sensor position information includes three-dimensional coordinates and azimuth data. Based on the priority level of the monitoring priority area, the priority level is predefined as a discrete value, such as level 1 to level 5, where level 1 corresponds to the highest priority and level 5 corresponds to the lowest priority. A scanning frequency coefficient is set for each sensor. The scanning frequency coefficient is a scaling factor determined by linear interpolation or lookup table method based on the priority level. The higher the priority, the larger the scanning frequency coefficient. The coefficient value ranges from 0.5 to 2.0, corresponding to priority levels from low to high. The scanning frequency coefficient is multiplied by the base frequency, which is the default sampling frequency of the sensor array, such as 10 Hz. The product operation yields the adjusted scanning frequency, which is used to dynamically configure the data acquisition cycle of the sensor. Calculating the data correlation between different regions involves extracting feature vectors from time-series data collected by the sensor array. The time-series data is a sequence of concentration values ​​recorded by each sensor at consecutive time points. Feature vector extraction is performed by processing the time series through a sliding window. Within each window, statistical features such as the arithmetic mean, standard deviation, and frequency domain energy obtained from the fast Fourier transform are calculated. The feature vector dimension is consistent with the number of features. The cosine similarity between feature vectors from different regions is calculated. The cosine similarity measures the degree of similarity between two vectors in a direction, with a value ranging from negative one to positive one. The cosine similarity is used as the data correlation degree, which is used to quantify the strength of the data correlation between monitored areas.

[0063] In practical implementation, the position information of each sensor in the sensor array is obtained by reading the configuration file. The configuration file stores the sensor identifier and coordinate mapping relationship in JSON or XML format. The position information update mechanism supports the dynamic addition or removal of sensors. Priority level allocation is based on the spatial distribution map of the monitoring priority area. The scanning frequency coefficient setting uses a piecewise linear function to map the priority level to the coefficient value. The function parameters are determined through calibration experiments. The base frequency is set according to hardware limitations and power consumption requirements. The multiplication operation is performed in the arithmetic logic unit of the embedded system. The adjusted scanning frequency is written to the sensor control register. In the feature vector extraction stage for calculating data correlation, the sliding window size is set to 100 sampling points with an overlap rate of 50%. Statistical feature calculation calls mathematical library functions, such as cosine similarity. The calculation uses the following formula:

[0064]

[0065] in and These are the feature vectors of two regions. and It is the i-th component of the vector. It represents the dimension of the eigenvectors. The numerator represents the vector dot product, and the denominator represents the product of the vector Euclidean norms. The calculation result is normalized to the interval [-1,1]. The data correlation is stored as a symmetric matrix for subsequent use.

[0066] In practice, the scanning frequency adjustment process includes location information loading, priority query, coefficient calculation, frequency update, and verification steps. Location information is loaded from non-volatile memory; priority query accesses the monitoring priority area database; coefficient calculation uses floating-point arithmetic units; frequency update is implemented through system calls; and the verification step checks whether the frequency value is within the allowed range. In practice, the data correlation calculation process includes data buffering, feature extraction, vector normalization, similarity calculation, and matrix construction stages. Data buffering manages the real-time data stream; feature extraction processes multiple regions in parallel; vector normalization eliminates the influence of amplitude; similarity calculation optimizes computational efficiency; and matrix construction fills the correlation data structure.

[0067] In some embodiments, sensor location information can be acquired using distributed positioning protocols such as the IEEE 1588 Precision Time Protocol to synchronize coordinate data. Priority level allocation can incorporate fuzzy logic to handle boundary cases. Scan frequency coefficient settings can use nonlinear functions such as exponential functions to adapt to different monitoring needs. The base frequency can be dynamically adjusted based on system load balancing, and the adjusted scan frequency is smoothed using filtering to avoid frequent switching. In some embodiments, feature vector extraction can incorporate time-domain features such as autocorrelation functions or complexity features. Cosine similarity calculation can be replaced with other similarity measures such as the Pearson correlation coefficient. The data correlation matrix can be filtered using thresholding to remove weakly correlated connections.

[0068] In some embodiments, sensor location information management can be integrated into an IoT platform for remote configuration; priority level mapping can support user-defined rules; temperature compensation factors can be incorporated into the calculation of scanning frequency coefficients to adapt to environmental changes; the selection of the base frequency can consider signal bandwidth requirements; and the adjusted scanning frequency can be logged for performance analysis. In some embodiments, the feature vector dimension can be reduced by principal component analysis to decrease computational load; cosine similarity calculation can be accelerated using approximation algorithms or hardware accelerators; and data correlation can be calculated using time aggregation to obtain moving average similarity.

[0069] Optionally, an exception handling mechanism can be added during scanning frequency adjustment. This includes using default values ​​when location information is lost, reverting to the base frequency when the priority level is invalid, clamping coefficients to a reasonable range when calculations overflow, and implementing a retry mechanism for failed frequency updates. Verification steps include boundary value testing. Optionally, a quality control step can be added during data correlation calculation. This includes checking data integrity during feature vector extraction, handling the special case of zero vectors during cosine similarity calculation, verifying the symmetry of the data correlation matrix, and caching results to improve query efficiency.

[0070] It is understandable that scanning frequency adjustment is a core aspect of resource allocation. Location information provides spatial context, priority levels reflect monitoring importance, scanning frequency coefficients enable flexible scaling, base frequency sets a benchmark, and the adjusted scanning frequency optimizes data acquisition efficiency. Similarly, it is understandable that data correlation calculations reveal interactions between regions, feature vectors capture the essence of the data, cosine similarity provides a measure of directional consistency, and the data correlation matrix supports network analysis.

[0071] In practical implementation, the hardware implementation of scanning frequency adjustment relies on a microcontroller unit, position information is stored in flash memory, priority levels are read from shared memory, scanning frequency coefficients are calculated using a digital signal processor, the fundamental frequency is generated by a crystal oscillator, and the adjusted scanning frequency is output via pulse width modulation. In practical implementation, the software module for data correlation calculation includes a data preprocessing library, a feature extraction algorithm library, a linear algebra library, and a similarity calculation library. Feature vectors are stored in an array structure, cosine similarity is optimized using the BLAS library, and the data correlation matrix uses a sparse matrix format to save space.

[0072] In practical implementation, the detailed steps for acquiring sensor location information include parsing configuration file syntax, verifying coordinate value validity, handling coordinate system transformations such as from local coordinates to world coordinates, and an update mechanism that listens for configuration change events. The detailed logic for priority level allocation includes looking up the regional priority mapping table, handling priority conflicts, and supporting real-time priority update requests. The detailed method for setting the scan frequency coefficients includes initializing the coefficient lookup table, calculating the table entry index, performing interpolation calculations using Lagrange interpolation or linear interpolation, and rounding the coefficient values ​​to the specified precision. Factors considered in determining the base frequency include sensor response time, signal-to-noise level, and power consumption constraints; the base frequency value is determined experimentally to achieve the optimal value. The application of the adjusted scan frequency includes configuring the timer interrupt frequency, adjusting sample-and-hold circuit parameters, and synchronizing the acquisition timing of multiple sensors.

[0073] In practical implementation, the detailed operations for feature vector extraction include time series segmentation, window size selection based on signal period estimation, overlap rate setting to balance redundancy and continuity, statistical feature calculation using a recursive algorithm to reduce memory usage, and frequency domain features calculated using Fast Fourier Transform to calculate power spectral density. The detailed process for cosine similarity calculation includes vector preprocessing such as centering or normalization, dot product calculation using loop expansion optimization, norm calculation using square root operations, division operations to handle division-to-zero anomalies, and result cropping to the domain.

[0074] It is understandable that adjusting the scanning frequency requires comprehensive consideration of accuracy and real-time performance. The accuracy of location information affects spatial alignment, the granularity of priority levels affects resource allocation resolution, the design of the scanning frequency coefficient affects response speed, the selection of the base frequency affects data quality, and the adjusted scanning frequency affects system stability. Similarly, it is understandable that calculating data correlation requires balancing accuracy and complexity. The selection of feature vectors affects representational capability, cosine similarity calculation affects correlation assessment, and the application of data correlation affects monitoring strategies.

[0075] In some embodiments, scanning frequency adjustment can be combined with machine learning to predict priority changes, adjust the frequency in advance to reduce latency, fuse position information with inertial measurement unit data to improve accuracy, adaptively learn historical patterns for priority levels, dynamically optimize scanning frequency coefficients using control theory, and adapt the base frequency to signal characteristics. In some embodiments, data correlation calculation can introduce deep learning to automatically learn feature representations, extend cosine similarity to multi-kernel similarity to handle nonlinear relationships, and use data correlation for anomaly detection and root cause analysis.

[0076] In practical implementation, the entire implementation of Example 3 is integrated into the main loop of the monitoring system. The scanning frequency adjustment module and the data correlation calculation module are executed in parallel, intermediate results are transmitted through a message queue, and the final output is used to adjust sensor operations and build the monitoring network. In practical implementation, performance optimization measures include memory pool management of sensor data, multi-threaded processing of correlation in different areas, overlapping execution of pipelined calculation steps, and resource monitoring to prevent overload.

[0077] It is understood that the implementation of Example 3 achieves intelligent sensor scheduling and data correlation analysis, scan frequency adjustment ensures high sampling in key areas, and data correlation reveals the inherent connections in the process, providing input for subsequent monitoring network construction. It is also understood that rigorous implementation of technical details ensures system robustness, location information processing ensures spatial consistency, priority mapping ensures rational resource allocation, coefficient calculation ensures the accuracy of frequency adjustment, base frequency setting ensures signal fidelity, feature extraction ensures information integrity, and similarity calculation ensures the reliability of correlation.

[0078] See Figure 4 This graph constructs a cosine similarity matrix between monitoring areas A, B, C, D, and E, using the monitoring areas as row and column dimensions. A yellow-green-blue-purple color gradient visually quantifies the strength of data associations between different monitoring areas. The values ​​in the matrix represent cosine similarity, reflecting the directional similarity of milk fat data feature vectors between areas. Its core value lies in providing a quantitative basis for constructing a monitoring network topology. By clarifying the strength of data associations between areas, technicians can identify closely related clusters of areas, and then abstract these areas into network nodes, using association as edge weights. This lays a crucial data foundation for subsequent use of graph traversal algorithms to generate monitoring node activation sequences and achieve intelligent data acquisition and scheduling. Ultimately, it helps upgrade milk fat separation process monitoring from single-point data collection to global resource optimization based on association logic.

[0079] Example 4: Constructing a monitoring network topology based on data correlation includes representing each monitoring area as a network node. A network node is a basic element in graph theory, and each node corresponds to a physical area defined in the monitoring priority area. Node attributes include area identifier, spatial coordinates, and priority level. Data correlation is represented as edge weights between nodes. Edge weights are the numerical values ​​of the edges connecting two nodes, derived from the previously calculated data correlation matrix. The data correlation matrix stores the cosine similarity values ​​between areas. An adjacency matrix is ​​used to store node and edge weight information. The adjacency matrix is ​​a two-dimensional array data structure, where row and column indices correspond to node numbers, and matrix elements store the edge weight values ​​between corresponding node pairs. A monitoring network topology is then constructed. The monitoring network topology is a weighted undirected graph used to visualize the data relationships between monitoring areas. The process of generating a monitoring node activation sequence using a graph traversal algorithm involves selecting the node with the highest correlation from the monitoring network topology as the starting node. Highest correlation means the node has the largest sum of edge weights with its neighboring nodes. This selection is achieved by calculating the weighted degree of each node. A depth-first search algorithm is then applied to traverse all nodes. Depth-first search is a graph traversal method that recursively or iteratively visits unvisited neighboring nodes starting from the starting node, recording the visit order as a time sequence of node visits. Based on this visit order, a monitoring node activation sequence is generated. This sequence is an ordered list where each element is a node number, used to control the activation order of the sensor array.

[0080] In practical implementation, representing each monitoring area as a network node involves assigning a unique node identifier to each monitoring area, mapping the node identifier to the area name, storing node attributes in a node attribute table (containing the area center coordinates, area area, and priority level), and initializing the node list after node creation. When representing data correlation as edge weights between nodes, edge weight values ​​are extracted from the data correlation matrix. This matrix is ​​symmetric, with matrix elements representing cosine similarity values. Edge weights are normalized to between zero and one using a min-max scaling method. Edge weights are stored in an edge list, with each edge recording the source node, target node, and weight value. When using an adjacency matrix to store node and edge weight information, the size of the adjacency matrix is ​​determined by the number of nodes. During matrix initialization, all elements are set to zero. When filling the matrix, the edge list is traversed, and edge weight values ​​are filled into the corresponding rows and columns. The diagonal elements of the matrix are kept zero to indicate the absence of self-loops. The monitoring network topology is constructed using a graph data structure class, which contains a set of nodes and a set of edges, providing graph operation interfaces such as adding nodes and edges.

[0081] In practical implementation, when selecting the node with the highest correlation from the monitoring network topology as the starting node, the weighted degree of each node is calculated. The weighted degree is the sum of the weights of all edges connecting to that node. All nodes are traversed, and the weighted degree values ​​are calculated. A sorting algorithm, such as quicksort, is used to sort the nodes in descending order of weighted degree, and the node ranked first is selected as the starting node. When applying a depth-first search algorithm to traverse all nodes, a stack data structure is used for iterative traversal. Initially, the starting node is pushed onto the stack, marking it as visited. The process of popping the top node from the stack, recording the visit order, and pushing unvisited adjacent nodes of that node onto the stack is repeated until the stack is empty. The visit order is then recorded in the visit order list. When generating the monitoring node activation sequence based on the visit order, the node numbers in the visit order list are mapped to sensor identifiers. The sensor identifiers are retrieved from the sensor configuration database, generating a monitoring node activation sequence file. The file format is either text or a binary sequence. Refer to Table 1 for the adjacency matrix of the monitoring network topology.

[0082] Table 1: Edge weight values ​​of nodes in the three monitoring areas

[0083]

[0084] In practical implementation, the detailed steps for constructing the adjacency matrix include: allocating matrix memory after determining the number of nodes; verifying the integrity of the data association matrix before filling matrix elements to ensure that each node pair has a corresponding value; whether the matrix storage uses a dense or sparse matrix format depends on the edge density; checking the connectivity of the graph after constructing the monitoring network topology; verifying the reachability of all nodes using a connected component algorithm; and logging any isolated nodes found. The implementation details of the depth-first search algorithm include setting a recursion depth limit to prevent stack overflow; fast access to adjacent nodes via the adjacency matrix or adjacency list; adding timestamps to the access order records for debugging; and verifying that the sequence length matches the number of nodes after the monitoring node activation sequence is generated.

[0085] In practical implementation, overlapping or boundary conditions are handled during node representation. A globally unique identifier algorithm is used for node identifier allocation, and the node attribute table is periodically updated synchronously. Missing data is handled during edge weight calculation, using interpolation to estimate missing correlation values. Weak connections below a threshold are filtered out to simplify the graph structure. Adjacency matrix storage is optimized for row-wise compressed storage to reduce memory usage, and matrix operations are accelerated using a linear algebra library. When selecting the starting node, cases with the same weight are handled using auxiliary rules such as node number order or random selection. The depth-first search algorithm can be configured for preorder or postorder traversal, and the access order list is stored in a circular buffer for real-time updates. The transmission of the monitoring node activation sequence to the sensor control unit is achieved through a message queue.

[0086] In some embodiments, the monitoring network topology can be constructed using a directed graph model, with edge weights allocated according to the directionality of data correlation, for example, from high-priority areas to low-priority areas. The adjacency matrix is ​​an asymmetric matrix, and the graph traversal algorithm uses breadth-first search instead of depth-first search. Breadth-first search uses a queue data structure to implement hierarchical traversal, and the generation of monitoring node activation sequences is based on distance priority.

[0087] Optionally, node representation can incorporate dynamic attributes such as real-time concentration values; edge weight calculation can combine the average correlation degree of a time-series sliding window; adjacency matrix storage can be expanded to a three-dimensional matrix for processing the time dimension; and the network topology can be visualized and rendered as a network graph after construction. Optionally, starting node selection can be based on node betweenness centrality indicators; the depth-first search algorithm can incorporate heuristic rules to prioritize accessing high-weight edges; and the sequence can be optimized after the monitoring node activation sequence is generated to reduce sensor switching overhead.

[0088] It is understandable that the construction of the monitoring network topology transforms the abstract data correlation into an intuitive graph structure, which facilitates the analysis of regional interactions. The graph traversal algorithm ensures systematic coverage of all monitoring areas, and the monitoring node activation sequence provides a scientific basis for sensor scheduling. The entire implementation method enhances the intelligence and adaptability of the monitoring system.

[0089] See Figure 5 This diagram uses the node activation order as the horizontal axis and the total node correlation as the vertical axis, distinguishing the priority levels of monitoring nodes with four colors: red, orange, green, and blue. It intuitively presents the monitoring node activation logic based on data correlation. Its core value lies in transforming abstract regional data correlation into an executable monitoring node activation sequence: based on a calculated regional data correlation matrix, each monitoring area is abstracted as a network node, with correlation as the edge weight between nodes. After constructing the monitoring network topology, a depth-first search algorithm prioritizes the node with the highest total correlation as the activation starting point, traversing all nodes sequentially to generate this ordered activation sequence. This sequence breaks the rigid pattern of traditional uniform sampling in monitoring, allowing sensor acquisition to follow a logic of high correlation priority and high priority priority. This ensures high-frequency monitoring of key areas of milk fat accumulation while reducing redundant acquisition through data correlation logic. It provides a direct execution basis for controlling sensor acquisition according to the monitoring node activation sequence, ultimately upgrading milk fat separation process monitoring from indiscriminate data acquisition to intelligent scheduling based on data correlation, achieving precise allocation of monitoring resources and early detection of process anomalies.

[0090] Example 5: The data acquisition operation of the sensor array controlled according to the monitoring node activation sequence includes parsing the node numbers in the monitoring node activation sequence. The monitoring node activation sequence is an ordered list structure, and the list elements are node numbers in string or integer format. The node number is a unique identifier assigned to each monitoring area when constructing the monitoring network topology. The parsing process reads the list elements through a sequence parser and verifies the legality of the number format. The corresponding sensor identifier is queried according to the node number. The sensor identifier is the logical name or address code of the physical sensor device. The query operation accesses the sensor mapping database. The sensor mapping database stores the correspondence between node numbers and sensor identifiers. The database structure is in key-value pair form, with the node number as the key and the sensor identifier as the value. A acquisition command is sent to the sensor identifier. The acquisition command is a control signal containing acquisition parameters, including sampling duration, sampling accuracy, and data format, etc. The data acquisition is triggered. The data acquisition starts the analog-to-digital conversion circuit inside the sensor, converts the analog signal into a digital reading, and transmits it to the storage system.

[0091] In practical implementation, the specific operations for parsing the node numbers in the monitoring node activation sequence include reading the sequence file or the sequence data structure in memory. The sequence file format is JSON, XML, or plain text. The parser parses the content line by line or element by element, and uses regular expressions or a syntax analyzer to verify that the node numbers conform to a predetermined pattern, such as "Node_001" or a numeric code. The verification process checks whether the number contains illegal characters, whether the length is consistent, and whether it is within the valid number range. The parsed node numbers are temporarily stored in the parsing result queue. When querying the corresponding sensor identifier based on the node number, a connection is established with the sensor mapping database. The database can be a relational database such as MySQL or an embedded database such as SQLite. An SQL query statement is executed or a database application programming interface is called. The query statement is "SELECT sensor_id FROM mapping_table WHERE node_id=?", with the current node number as the parameter. The query result returns a sensor identifier string such as "SENSOR_01_A". Cases of empty or duplicate query results are handled, and query logs are recorded for auditing. When sending a data acquisition command to the sensor identifier, the acquisition command is encapsulated in a specific protocol format. The protocol can be Modbus, OPCUA, or a custom binary protocol. The command content includes parameters such as the start acquisition command, acquisition duration, and gain setting. The command transmission is achieved through industrial Ethernet, RS-485, or a wireless communication module. After triggering data acquisition, the sensor returns an acknowledgment signal, and the system monitors the acquisition status until it is completed.

[0092] In practical implementation, the detailed implementation of the parsing step includes error handling mechanisms during sequence file loading, such as enabling a default sequence when the file does not exist or has an incorrect format, skipping invalid entries and recording alarms when node number verification fails, and using a producer-consumer model to balance parsing and processing speeds in the parsing result queue. Database connection management in the query step includes connection pool optimization to reduce latency, query statement pre-compilation to improve efficiency, and a sensor mapping database update mechanism that supports dynamic mapping relationship adjustments, automatically remapping to backup sensors when a sensor fails. Communication protocol processing in the command sending step includes packet assembly, checksum calculation, timeout retransmission mechanisms, priority setting for acquisition commands to ensure high-priority nodes acquire data first, and status monitoring after triggering data acquisition, including data integrity checks and transmission confirmation.

[0093] In some embodiments, parsing the activation sequence of monitoring nodes can employ a streaming parsing method to process extremely long sequences, reducing memory consumption by parsing and executing simultaneously, and adding checksum verification to node number verification improves reliability. In some embodiments, querying sensor identifiers can introduce a caching mechanism to store frequently used mapping relationships, reducing the number of database queries, and the sensor mapping database can adopt a distributed architecture to improve scalability.

[0094] Optionally, sequence compression and decompression functions can be added during the parsing process, supporting binary format sequence files to reduce storage space, and Huffman coding can be used to further compress data for node numbers. Optionally, the query operation can be implemented as a fault-tolerant query, switching to a backup mapping table when the main database is unavailable, and fuzzy matching can be supported to handle minor differences in numbers for sensor identifier parsing. Optionally, command sending can use multicast to send acquisition commands to multiple sensors simultaneously, improving initialization efficiency, and acquisition parameters can be dynamically adjusted based on real-time load conditions.

[0095] It's understandable that parsing the activation sequence of monitoring nodes is the starting point for control execution. Node numbers serve as the bridge connecting the logical sequence and physical devices, and the accuracy of parsing directly affects the correctness of subsequent operations. Querying sensor identifiers to map logical nodes to physical sensors is the core step in resource scheduling. Sending acquisition commands is the final execution action, and the completeness and timeliness of the commands determine the quality of data acquisition. Triggering data acquisition completes the closed loop from control to data acquisition.

[0096] In practice, the entire process is integrated through the workflow engine of the monitoring system. The workflow engine sequentially calls the parsing module, query module, and command sending module, with data exchanged between modules via message brokers. Performance optimization measures include using efficient string processing algorithms in the parser, optimizing database queries with indexes, and employing an asynchronous non-blocking mode for command sending to improve concurrency.

[0097] In practical implementation, when parsing the node number in the monitoring node activation sequence, dynamic updates to the sequence are handled. If the sequence is modified during execution, copy-on-write or version control mechanisms are used to ensure consistency. When querying the corresponding sensor identifier based on the node number, dynamic changes to the sensor identifier are handled, such as sensor hot-plugging, and a listener mode is used to update the mapping relationship in real time. When sending acquisition commands to the sensor identifier, network latency or packet loss is handled, a heartbeat mechanism is used to detect the sensor's online status, and retry logic or an alarm mechanism is activated when the command transmission fails.

[0098] In practical implementation, the parser can be implemented using recursive descent parsing or state machine parsing of complex sequence formats. Query operations support transaction processing to ensure the consistency of mapping relationships. The instruction sending module is implemented as a pluggable protocol adapter to support multiple industrial protocols. In practical implementation, the error handling mechanism includes classified handling of abnormal situations such as parsing errors, no query results, and instruction sending timeouts to ensure system robustness.

[0099] It is understandable that the implementation of Example 5 transforms the logical monitoring sequence into physical acquisition actions. Parsing ensures the accuracy of instruction interpretation, querying guarantees the precision of resource location, and instruction sending ensures the reliability of control. The entire process demonstrates the execution layer capabilities of the monitoring system. The rigorous implementation of technical details ensures a seamless transition from decision-making to action.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the milk fat separation process, characterized in that, The method performs the following steps: Optical property data of the milk fat fluid are captured by a multispectral sensor installed at the inlet of the separation equipment; Based on the optical property data, the composition change pattern of the milk fat fluid is identified, and a dynamic distribution cloud map of the milk fat in the separation device is drawn according to the composition change pattern. High-concentration milk fat aggregation areas and low-concentration milk fat dispersion areas are extracted from the dynamic distribution cloud map, and monitoring priority areas are defined based on the high-concentration milk fat aggregation areas and low-concentration milk fat dispersion areas. The scanning frequency of the sensor array is adjusted according to the monitoring priority area, and the data correlation between each area is calculated; Based on the data correlation, a monitoring network topology is constructed, and a graph traversal algorithm is used to generate a monitoring node activation sequence; The data acquisition operation of the sensor array is controlled according to the activation sequence of the monitoring node; The step of extracting high-concentration milk fat aggregation regions and low-concentration milk fat dispersion regions from the dynamic distribution cloud map includes: setting a concentration threshold range and performing region segmentation on the dynamic distribution cloud map; Calculate the average concentration value of each segmented region, mark the region with the average concentration value greater than the upper threshold as a high-concentration milk fat aggregation region, and mark the region with the average concentration value less than the lower threshold as a low-concentration milk fat dispersion region.

2. The method for monitoring the milk fat separation process according to claim 1, characterized in that, The method of identifying the compositional variation pattern of milk fat fluid based on the optical property data includes: performing wavelet transform processing on the optical property data to extract frequency features at different scales; matching the frequency features with a pre-stored milk fat composition database to identify the distribution pattern of fat particles in milk fat; calculating the compositional variation gradient according to the distribution pattern, and determining the compositional variation pattern based on the compositional variation gradient.

3. The method for monitoring the milk fat separation process according to claim 2, characterized in that, The step of drawing a dynamic distribution cloud map of milk fat in the separation device based on the component change pattern includes: mapping the component change gradient to the three-dimensional coordinate space of the separation device to generate an initial distribution point cloud; performing spatial interpolation calculation on the initial distribution point cloud to fill the gaps between the point clouds and form a continuous distribution surface; and superimposing the continuous distribution surface with the device structure model to output a dynamic distribution cloud map.

4. The method for monitoring the milk fat separation process according to claim 1, characterized in that, The definition of monitoring priority areas based on the high-concentration milk fat aggregation area and the low-concentration milk fat dispersion area includes: assigning the highest priority to the high-concentration milk fat aggregation area and the lowest priority to the low-concentration milk fat dispersion area; for the transition zone between adjacent areas, calculating the concentration change rate and adjusting the priority level according to the concentration change rate.

5. The method for monitoring the milk fat separation process according to claim 1, characterized in that, The step of adjusting the scanning frequency of the sensor array according to the monitoring priority area includes: obtaining the position information of each sensor in the sensor array; setting a scanning frequency coefficient for each sensor according to the priority level of the monitoring priority area, with a larger scanning frequency coefficient for higher priority; and multiplying the scanning frequency coefficient by the base frequency to obtain the adjusted scanning frequency.

6. The method for monitoring the milk fat separation process according to claim 5, characterized in that, The calculation of the data correlation between different regions includes: extracting feature vectors from the time series data collected by the sensor array; calculating the cosine similarity between the feature vectors of different regions, and using the cosine similarity as the data correlation.

7. The method for monitoring the milk fat separation process according to claim 1, characterized in that, The construction of the monitoring network topology based on the data correlation degree includes: representing each monitoring area as a network node, representing the data correlation degree as the edge weight between nodes; using an adjacency matrix to store node and edge weight information, and constructing the monitoring network topology; The method of generating the monitoring node activation sequence using the graph traversal algorithm includes: selecting the node with the highest correlation from the monitoring network topology as the starting node; applying a depth-first search algorithm to traverse all nodes and recording the access order; and generating the monitoring node activation sequence according to the access order.

8. The method for monitoring the milk fat separation process according to claim 1, characterized in that, The data acquisition operation of the sensor array controlled according to the monitoring node activation sequence includes: parsing the node number in the monitoring node activation sequence; querying the corresponding sensor identifier according to the node number; and sending an acquisition command to the sensor identifier to trigger data acquisition.

9. A monitoring system based on the milk fat separation process, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for monitoring the milk fat separation process as described in any one of claims 1 to 8.

Images