A method for detecting the quality of oil-in-water emulsion based on spectral technology

Abstract

This invention relates to the field of optical detection technology, specifically a method for quality detection of oil-in-water emulsions based on spectral technology. The method includes: acquiring transmitted light intensity across all angles to determine the macroscopic state of the emulsion; scanning vertically along the emulsion height by height to acquire backscattered light intensity, and determining local features and their effective feature counts at different liquid layer heights based on the changing trends of backscattered light intensity at different liquid layer heights; using the effective feature counts at different liquid layers as constraints to constrain the search range for local feature matching from top to bottom, and constructing layered features for each liquid layer; performing multi-component mapping between the layered features and the target components of the emulsion to be detected, and clarifying the correlation characteristics of different target components in the emulsion based on the update time of the layered features; and determining the quality risk index of the target components based on the correlation characteristics corresponding to the target components. This method improves the detection accuracy and stability of emulsions.

Description

Technical Field

[0001] This invention relates to the field of optical detection technology, specifically a method for quality detection of oil-in-water emulsions based on spectral technology. Background Technology

[0002] Oil-in-water emulsions are multiphase dispersion systems composed of an oil phase, an aqueous phase, emulsifiers, and additives. Their storage stability and component homogeneity directly affect the product's performance. During storage, emulsions are prone to problems such as water flotation and droplet agglomeration. Therefore, emulsions require quality testing and stability monitoring. Current monitoring methods for emulsions often include single-spectral detection and laser particle size analysis. However, these methods struggle to distinguish localized structural degradation within the emulsion, leading to reduced testing accuracy.

[0003] For example, Chinese Patent Publication No. CN121027015A discloses a method and system for quality detection of dairy products based on spectral analysis, belonging to the field of optical detection technology. The method includes: introducing the dairy product sample to be tested into a constant-temperature colorimetric cell; emitting incident light using a modulated light source; simultaneously acquiring original spectral signals in the sample channel and reference channel; generating a correction spectrum based on the original spectral signals; inputting the correction spectrum into an integrating sphere to collect omnidirectional transmitted light to obtain the total transmission spectrum; acquiring scattering distribution data at different angles based on a ring-shaped multi-angle detector array deployed on the outer wall of the constant-temperature colorimetric cell; constructing a scattering compensation factor based on the scattering distribution data and the total transmission spectrum; applying the scattering compensation factor to the total transmission spectrum to obtain an effective absorption spectrum; inputting the effective absorption spectrum into a component calculation model to output the content of each target component in the dairy product to be tested.

[0004] Existing technologies employ the Mie scattering principle, assuming the presence of isolated, single spherical particles in dairy products. They focus on the influence of individual particles on scattering patterns and absorbance changes, emphasizing the quality detection of highly diluted samples. However, considering the multiple scattering interactions within a liquid at a specific concentration, it is necessary to detect changes in light intensity over time and height to identify the quality stability within the emulsion. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for quality detection of oil-in-water emulsion based on spectral technology, comprising: S1, introducing the emulsion to be tested into a constant temperature test scene, collecting the transmitted light intensity at all angles, and determining the macroscopic state of the emulsion accordingly.

[0006] S2, scans vertically along the emulsion height by height, collecting backscattered light intensity as it changes with height and time. Based on the trend of backscattered light intensity at different liquid layer heights, the local features and the number of effective features under the current macroscopic state are determined.

[0007] S3 uses the number of effective features at different liquid layers as a constraint to constrain the search range of local feature matching from top to bottom, and constructs the layered features of each liquid layer accordingly.

[0008] S4. Multi-component mapping is performed between the layered features and the target components of the emulsion to be detected. Combined with the update time of the layered features, the correlation characteristics of different target components in the emulsion are clarified.

[0009] S5. Based on the correlation characteristics corresponding to the target components, determine the quality risk indicators of the target components.

[0010] The beneficial effects of this invention are as follows: First, this invention constructs a global transmission baseline by collecting transmitted light intensity at a constant temperature and calculating the global average transmittance, corrects the backscattered spectral signal, eliminates the systematic amplitude deviation caused by overall substrate drift and temperature interference, and achieves the accuracy of the overall macroscopic state definition.

[0011] Second, this invention constructs local features of temporal evolution and local features of vertical spatial gradient by using vertical height-sequential backscattered signals; effective liquid layers are screened by variance, and the number of effective features per liquid layer is obtained by statistical analysis of effective extreme points and effective principal components. This quantifies the non-uniformity, flocculation and other problems of different vertical liquid layers in microscopic scenes, thereby improving the accuracy and processing effect of vertical spatial resolution.

[0012] Third, this invention constrains the cross-liquid layer feature matching search range and similarity threshold by using the number of effective features per liquid layer, and completes vertical feature alignment matching layer by layer from top to bottom; then, it suppresses false abnormal signal fluctuations by using residual regularization constraints, and completes iterative constraints in conjunction with the temporal constraints of emulsion evolution, so as to obtain the feature sequence of vertical layering, reduce noise interference in the vertical direction of emulsion, and further improve the continuity and accuracy of feature output.

[0013] Fourth, this invention achieves effective tracking of the entire emulsion detection process by performing multi-component regression mapping between vertical stratification features and target components, screening the associated features of the components, and recording the safety thresholds of relevant components, thereby further improving the accuracy and processing efficiency of online quality control of emulsions. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Figure 1 This is a flowchart illustrating a method for quality detection of oil-in-water emulsions based on spectral technology. Figure 2 This is a flowchart illustrating step S2 of a method for quality detection of oil-in-water emulsions based on spectral technology. Figure 3This is a flowchart illustrating step S3 of a method for quality detection of oil-in-water emulsions based on spectral technology. Figure 4 This is a profile of the longitudinal backscattered light intensity of a water-in-oil emulsion quality detection method based on spectral technology. Figure 5 This is a schematic diagram of the spatial gradient corresponding to the backscattered light intensity in a water-in-oil emulsion quality detection method based on spectral technology. Detailed Implementation

[0016] The embodiments of the present invention are described in detail below. The embodiments described below are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Where specific techniques or conditions are not specified in the embodiments, they shall be performed in accordance with the techniques or conditions described in the literature in the art or in accordance with the product manual.

[0017] See Figure 1 A method for quality detection of oil-in-water emulsions based on spectral technology includes: S1, introducing the emulsion to be tested into a constant temperature test environment, collecting the transmitted light intensity from all angles, and using this to determine the macroscopic state of the emulsion.

[0018] S2, scans vertically along the emulsion height by height, collecting backscattered light intensity as it changes with height and time. Based on the trend of backscattered light intensity at different liquid layer heights, the local features and the number of effective features under the current macroscopic state are determined.

[0019] S3 uses the number of effective features at different liquid layers as a constraint to constrain the search range of local feature matching from top to bottom, and constructs the layered features of each liquid layer accordingly.

[0020] S4. Multi-component mapping is performed between the layered features and the target components of the emulsion to be detected. Combined with the update time of the layered features, the correlation characteristics of different target components in the emulsion are clarified.

[0021] S5. Based on the correlation characteristics corresponding to the target components, determine the quality risk indicators of the target components.

[0022] The method of this invention is applicable to emulsions such as milk, soy milk, and cream, as well as dairy products such as ice cream, yogurt, and cream cheese. It assesses sample stability by identifying the stability of the emulsion at corresponding particle size distributions using spectroscopic techniques, thereby determining the relative quality of the dairy product.

[0023] Furthermore, using a constant-temperature environment as the sampling basis, near-infrared spectroscopy is employed to acquire the absorption spectrum corresponding to transmission and the scattering spectrum corresponding to backscattering. Optical field signals are collected from multiple angles to obtain raw local spectral information. This spectral information is then correlated and mapped with macroscopically observed instability phenomena such as stratification, demulsification, flocculation, and sedimentation to determine the initial data format. At this point, macroscopic observations will determine whether the current emulsion exhibits stratification, demulsification, flocculation, or sedimentation phenomena through changes in spectral information. This information is then combined with the acquired spectral data to determine the main data content to be analyzed in the current scenario.

[0024] Then, based on the absorption intensity and scattering characteristics of the spectrum, a quantitative mapping was performed on the content of emulsion components such as oil content, water content, emulsifier content, and oxidation by-product content. Then, based on the spectral waveform, peak shift, and spectral response corresponding to turbidity, the correlation characteristics with oil droplet size, viscosity, color, emulsification uniformity, etc. were determined. Finally, based on the various characteristics mapped by the spectral waveform, the relative mass of the current emulsion was obtained.

[0025] It should be noted that during the acquisition of transmitted light intensity, the light is processed by a near-infrared continuous light source to keep the incident angle and incident light power constant. The emitted incident light is incident perpendicularly from one side of the sample, passes through the entire emulsion sample layer, and a transmission detector is set up on the other side to collect the transmitted light intensity and full-band transmission spectrum after passing through the entire emulsion. These data will be continuously collected at fixed time intervals.

[0026] During backscattered light intensity acquisition, multiple light scattering (MLS) is used. The light source is moved vertically downwards along the longitudinal direction of the sample tube, emitting beams into the emulsion at progressively higher heights. The intensity of the backscattered light returned after multiple angular scattering by particles within the emulsion is detected to capture the local backscattering intensity differences caused by variations in oil droplet size, flocculation, and local concentration in different liquid layers. In other words, by collecting the light signals that return at approximately 180° angles after single or multiple scattering at the oil droplet interface, these signals are used to quantify particle changes in local areas.

[0027] In one embodiment of the present invention, the physical state of the emulsion is initially screened by the characteristic form presented by the transmitted light intensity, and unstable phenomena such as layering or demulsification on the overall macroscopic level are identified. Based on this, it is determined whether the aqueous phase and oil phase conform to the description of a homogeneous O / W type emulsion, thereby clarifying how to perform data verification of other parts in subsequent processes based on the macroscopic performance of the emulsion.

[0028] Since transmitted light intensity does not have a relative spatial dimension, the average value of transmitted light intensity collected at the same time will be used to characterize the overall transmittance, and the change form of emulsion will be preliminarily defined based on the change in transmittance.

[0029] Specifically, one implementation of step S1 includes: averaging the collected transmitted light intensity to calculate the overall average transmittance; and determining the macroscopic state of the emulsion based on the trend of the average transmittance. Here, transmittance represents the proportion of transmitted light intensity.

[0030] Furthermore, when the average transmittance shows a continuous downward trend, there are related problems such as increased overall turbidity, overall flocculation, coarsening of oil droplets, and decreased overall emulsion uniformity; when the average transmittance shows a continuous upward trend, there are problems such as overall free water precipitation, increased macroscopic supernatant, and overall demulsification phase separation. These potential problems will serve as the current macroscopic state of the emulsion to explain the preliminary description of the tested sample.

[0031] Furthermore, the compositional content relative to the global context can be quantified by establishing a fitting relationship between average transmittance and parameters such as water separation rate, water content, and oil content. These global compositional contents also represent the specific circumstances of the macroscopic state.

[0032] If there are fluctuations in average transmittance, it is necessary to mark the corresponding time period, record the macroscopic state of the current scene based on the range of average transmittance values ​​corresponding to the trend of change.

[0033] It should be noted that the macroscopic state only represents the mean characteristics of the sample and cannot distinguish the specific internal problems. Backscattered light intensity needs to be introduced to quantify the local microscopic characteristics.

[0034] In one embodiment of the present invention, the problem of no spatial resolution of transmitted light intensity is compensated by capturing the differences between different liquid layers of the emulsion. Essentially, it is described in layers according to different liquid layers. After explaining the relative thickness of each liquid layer, the local backscattering rate of change and the evolution of particle size at local locations, it is output as local feature forms under specific spatial differences according to each liquid layer.

[0035] In other words, by extracting the changes in backscattered light intensity of each liquid layer of the emulsion, we can interpret microscopic aspects such as particle size evolution and flocculation behavior, thereby explaining the spatial differences at different heights. Each liquid layer represents a thin liquid segment corresponding to a height sampling point, representing the current relative position of the emulsion in the vertical direction.

[0036] At the same time, since the backscattered light intensity is easily affected by the overall concentration and turbidity, causing signal drift, it is necessary to use the recorded transmitted light intensity to determine the transmission baseline of the current scene and use it to remove interference from the overall water phase.

[0037] Specifically, such as Figure 2 As shown, one implementation of step S2 includes: S21, construct a global transmission baseline using the transmitted light intensity, and determine the corrected backscattered light intensity; the correction process follows the following implementation method: The arithmetic mean of the transmitted light intensity at any given time is used as the global transmission baseline, and the initial global transmission baseline is used as the reference value. A compensation coefficient is constructed, which is based on the moment when the sample is just mixed and no stratification has occurred, determining the relative change ratio between the current time and the reference time. The compensation coefficient is set based on the ratio of the reference value to the current global transmission baseline. When the global transmission baseline at any time is less than the reference value, the overall turbidity increases, leading to a decrease in transmission. In this case, the compensation coefficient is greater than 1 to increase the overall backscattered light intensity and enhance the relative resolution of the overall signal. When the global transmission baseline at any time is greater than the reference value, large particles precipitate out of the optical path, causing an increase in transmission. In this case, the accuracy of the value at the corresponding position is improved by reducing the backscattered light intensity.

[0038] For the obtained compensation coefficient, the product of the backscattered light intensity and the compensation coefficient is regarded as the corrected backscattered light intensity.

[0039] Current correction methods emphasize the overall ratio relative to the reference state to eliminate global shifts caused by overall solvent evaporation and ambient temperature leading to an overall increase or decrease in backscattering intensity; and enhance the variation of backscattered light intensity at different locations.

[0040] S22, for each spatial location corresponding to the liquid layer height, the slope change rate of the backscattered light intensity is used as a local feature in the time dimension. S23, for each fixed time, the spatial gradient of backscattered light intensity at adjacent liquid layer heights is used as a local feature of the spatial dimension.

[0041] In time-dimensional analysis, the slope change rate represents the first derivative of backscattered light intensity with time. Its value directly reflects the relative phenomena of flocculation and layer separation at different liquid layer heights, and can demonstrate the dynamic activity of the liquid layer. If the slope change rate is always close to zero in the time dimension, it indicates that the corresponding liquid layer height exhibits the local characteristic of an inert mean.

[0042] In spatial dimension analysis, the spatial gradient can indicate the relative level representing particle distribution, such as the interface of the water separation layer or the edge of the flocculation enrichment layer, and explain the liquid layer structure boundary corresponding to the emulsion. If the spatial gradient is close to zero, the spatial interval being analyzed can be regarded as the overall stable liquid layer range.

[0043] In other words, the slope change rate represents the specific change pattern of the height of each liquid layer, while the spatial gradient is used to verify the spatial gradient of adjacent liquid layers within the same time period, thereby identifying local features that show stratification or obvious spatial differences, which serve as the basis for subsequent stratification matching.

[0044] Furthermore, in the current scenario, the liquid layer height will be divided into multiple equal-sized liquid segments based on experimental requirements, with at least three segments, to compare the states of different liquid segments.

[0045] Finally, after summarizing the extreme points and rates of change in the spatial and temporal dimensions, the local features of the output are combined with the descriptions of different liquid layers, the backscattered light intensity, the backscattered light intensity, and the rate of change of the backscattered light intensity at each location.

[0046] S24, based on local features in spatial and temporal dimensions, filters the number of effective features under different macroscopic states.

[0047] Furthermore, when quantifying the local differences at different liquid layer heights, the local features output by each liquid layer are also screened to determine the number of valid output features. This avoids bringing invalid boundary liquid layers into the subsequent calculation process and prevents the overall feature layering from being dominated by extreme value features, which would make the internal evolution of the emulsion invisible.

[0048] For the backscattering time series variance of each liquid layer under different macroscopic conditions, the liquid layer that satisfies the variance contribution rate is regarded as the effective liquid layer.

[0049] At this point, the backscattering time series variance represents the variance of the backscattered light intensity at each liquid layer as a function of time. After statistically analyzing the variance of each liquid layer over multiple time periods, the variance of each liquid layer is compared with the sum of the variances of all liquid layers to obtain the variance contribution rate. Then, all variance contribution rates are sorted from largest to smallest, and the top 80% of the data are selected as the data for the current valid analysis. These data are considered to meet the variance contribution rate to reduce input interference from the boundary liquid layer.

[0050] At the same time, for the effective liquid layer obtained, feature filtering is performed based on the distribution of extreme points to determine the number of effective extreme points corresponding to the spatial gradient at the same time.

[0051] Singular value decomposition is performed on the slope change rate of backscattered light intensity of each liquid layer to select the number of effective principal components corresponding to each liquid layer; the number of effective extrema and the number of effective principal components are used together as the number of effective features in the output.

[0052] Among them, the number of effective features represents the feature distribution of the current emulsion in a specific dimension, including the number of effective extrema points after calculating the spatial gradient along the vertical direction at the same time; and the number of principal components after singular value decomposition of the backscattered light intensity at different liquid layers. These numbers represent the distribution ratio of effective features at each liquid layer. They are scalarized to describe all local features according to the analysis region of the spatiotemporal combination corresponding to the current macroscopic state.

[0053] It should be noted that an effective extreme point needs to have an absolute value greater than the mean of the global spatial gradient at the corresponding moment plus twice its standard deviation to obtain the relative position of the prominent overall change, and synchronize this position to the corresponding liquid layer height. If the extreme point corresponding to the liquid layer is 0 at a certain moment, it means that the homogeneity is stable and there is no visible stratification; if the corresponding extreme point is 1, it means that a clear oil-water separation layer appears at this position; if the extreme point is 2 or above, it indicates the existence of multiple flocculation interfaces and extremely uneven emulsion distribution.

[0054] For singular value decomposition in the time dimension, singular value decomposition is performed on the matrix composed of the slope change rate of backscattered light intensity over multiple time periods. The number of principal components with singular values ​​greater than 5% of the maximum singular value is selected as the effective principal components to mark the distinction between different liquid layers. The emulsion states of different samples vary greatly. Homogeneous and stable samples have very few dynamic patterns, and the number of principal components is usually only 1-2. Samples that have experienced floating, water separation, and multiple instabilities due to aggregation have complex dynamic patterns, and the number of principal components can reach 3-5. The more principal components identified, the more obvious the compositional changes at the corresponding location, and the more significant the local differences.

[0055] Table 1. Longitudinal Schematic Diagram of Backscattered Light Intensity

[0056] As shown in Table 1, the relative distribution of backscattered light intensity and its spatial gradient in the longitudinal direction over a 60-minute period was obtained. The overall trend exhibits an irreversible decreasing pattern, representing the vertical density gradient caused by the upward movement of oil droplets and the downward movement of the water phase under gravity. Combined with... Figure 4 As shown, the relative intensity is high at the top and low at the bottom. The steeper the curve, the greater the concentration change per unit height. In some areas, there is a local concave shape, indicating that the oil droplets are aggregated but contain water, resulting in a reduced scattering cross section per unit volume. The red dashed line marked in the image represents the relative distinction between different concentrations, defining the boundary of light scattering by the emulsion.

[0057] After that, from Figure 5 As can be seen, extremely high peak values ​​indicate the presence of sharp physical interfaces, suggesting macroscopic gravitational stratification of the emulsion. The higher the peak, the greater the density difference between the two phases, and the worse the stability. Conversely, lower peak values ​​indicate microscopic flocculation, where the emulsion distribution breaks down at corresponding locations. For regions with gradients close to zero, the corresponding area shows uniform distribution; however, the relative situation at each location needs to be determined by considering the backscattered light intensity, thus revealing the relative stratification pattern at different locations. By highlighting the local spatial differences at different locations, a data foundation and basis are provided for subsequent feature verification of stratification, improving the visibility and interpretability of emulsion detection.

[0058] In one embodiment of the present invention, the effective feature count represents the number of independent spaces that can be distinguished in the current emulsion and the number of related features. The more features there are, the stronger the microscopic inhomogeneity of the vertical liquid layer of the emulsion and the richer the stratification differences, exhibiting multiple local flocculation and particle size abrupt change problems. If the feature count is smaller, it means that the backscattering evolution of each liquid layer is highly consistent, the spatial difference is small, and the emulsion is evenly distributed. In other words, the local feature count directly quantifies the richness of the emulsion in the microscopic dimension to characterize the spatial differences under stratification treatment.

[0059] The layered feature here represents the layered structure of multiple liquid layers after being associated and grouped layer by layer, characterizing the differences between each layer.

[0060] Considering the number of valid features at the current liquid layer, the search range for local feature alignment will be constrained layer by layer based on the number of valid features currently entered, as well as the relative content under local feature similarity matching, to fit the application processing of local features.

[0061] like Figure 3 As shown, one implementation of step S3 includes: S31, constraining the search range for local feature alignment layer by layer according to the number of valid features entered.

[0062] When constraining local feature search, the more effective features a liquid layer has, the higher the microscopic differences and the more significant the instability of the layer. It is necessary to expand the candidate search range and use refined similarity matching. Similarly, the fewer effective features there are, the smaller the search range is, and the more the matched content is biased towards the overall average calculation content.

[0063] Furthermore, when constraining the search range for local feature alignment, the implementation method includes: mapping the number of effective features passed from the previous layer to liquid layer state labels; the liquid layer state labels include stable layers, metastable transition layers, and unstable abrupt change layers; these labels are obtained based on the sample mean mapping under different scenarios. Freshly prepared emulsions and emulsions exhibiting significant instability after centrifugation or high-temperature damage are selected as samples. The effective feature counts of these two samples are converted into probability distribution curves. The effective sample count value at the 95th percentile of the probability distribution of the fresh sample, i.e., when the cumulative probability reaches 95%, ensures that 95% of the parameters are not considered non-steady, thus distinguishing between the stable layer and the metastable transition layer; then, using the 5th percentile of the probability distribution of the significantly unstable samples to ensure that most data are not considered non-instable, the effective feature count at the corresponding quantile is used as the distinguishing value between the metastable transition layer and the unstable abrupt change layer, thus completing the label mapping of the effective feature count. It should be noted that this mapping process is based on the user-verified emulsion sample requirements, and the relative quantity value is adjusted accordingly for each mapping.

[0064] Furthermore, the constraint search range is the total number of all local features extracted from the current liquid layer, from which a finite subset is selected for matching. Simultaneously, when completing the liquid layer state label mapping, it is also necessary to determine the optimal index position for matching in the previous layer. It should be noted that the optimal index position represents the feature with the highest similarity between the previous liquid layer and all features of the current layer; its index is considered the optimal index position. Here, it is only for quickly locating the anchor point corresponding to the search window. Each local feature will be used as the value of any index in the feature set, in the form of a combination of backscattered light intensity, slope change rate, and spatial gradient. Since each liquid layer is a continuously set relative form, when an optimal index exists in the previous layer, a corresponding index position will also appear in subsequent layers. The relevant data to be analyzed in the current scene can be inferred directly based on the optimal index position of the previous layer's matching.

[0065] Based on the optimal index position during the previous layer matching, the size of the subset that matches the liquid layer state label is divided, and the similarity threshold of local feature matching is adjusted synchronously to complete the search range constraint.

[0066] Specifically, the implementation of subset partitioning is as follows: ; in, Let represent the radius value corresponding to the search range of the i-th liquid layer, so that the search range of the constraint forms an interval in the form of the optimal index position ± radius value, to illustrate the subset index range involved in the current matching process. This represents the basic search radius and the minimum computational cost per operation. It can be set based on the average number of local features in each similarity match in historical data. This represents the maximum expansion radius, indicating the maximum expandable radius in each calculation. The value of this maximum expandable radius does not exceed 30% of the number of local features in the current liquid layer. This represents the number of valid features corresponding to the search range of the (i-1)th liquid layer. The current calculation process is based on adaptive scaling of the state of the previous liquid layer, so that each similarity operation can be compared layer by layer based on the state of the previous layer. This represents the number of effective features used to distinguish between the steady-state layer and the metastable transition layer. This represents the number of effective features used to distinguish between metastable transition layers and unstable abrupt change layers.

[0067] The similarity threshold is calculated in the same way as the subset partitioning process in terms of the effective feature count, and it is expressed as follows: ; in, This represents the similarity threshold when comparing the i-th liquid layer. This represents the base similarity. Setting the base similarity to 0.75 during similarity adjustment will yield higher precision. This indicates the threshold adjustment range, with a value of 0.2, to slightly adjust the matching content in different scenarios.

[0068] S32 uses cosine similarity as the main factor to perform local feature matching between adjacent liquid layers, and determines the matching features that meet the similarity test layer by layer from top to bottom.

[0069] When performing similarity matching, the feature range that matches the current liquid layer with the previous layer is determined based on the search range of the constraints, and features that meet the adjusted similarity threshold are determined based on the adjusted similarity threshold and regarded as the output matching features.

[0070] S33 concatenates the matched features after layer-by-layer alignment to form a vertically layered feature sequence, thus obtaining the output layered features.

[0071] Specifically, the hierarchical feature will first calculate the average value relative to the global value for the feature dimensions of similar matching between each layer, and then subtract this average value from the matching feature at each liquid layer to obtain the relative residual signal at each position; or directly introduce the matching feature and splice it in multiple dimensions according to the feature dimensions of the current analysis to form a feature sequence spliced ​​under multiple feature dimensions, thereby showing the relative differences in the height of different liquid layers.

[0072] In one embodiment of the present invention, the residuals of the matching features are used to determine the regularization terms that include all the matching features, and according to the state update process of the regularization terms, the push constraints are set at each position to limit the variation of behaviors such as flocculation in the emulsion.

[0073] Specifically, the methods for forming a vertically layered feature sequence include: constructing a regularization constraint term that matches the feature residual distribution, and using the loss of the regularization constraint term to constrain the fluctuation amplitude of the backscattered features.

[0074] By combining the relative laws of emulsion evolution, a layered boundary under time constraints is constructed, and the layered features of the vertical distribution are obtained by iterative solution.

[0075] Specifically, the regularization constraint term represents the offset of light intensity on the time axis. It is used to explain the drift process of light intensity corresponding to each matching feature, preventing instantaneous changes in light intensity caused by transient electrical disturbances, which could lead to misidentification of matching features at corresponding positions. In other words, the calculation of the residual distribution includes the changes in multi-dimensional feature vectors such as the average backscattering intensity, particle size index, and spatial gradient magnitude of the same physical layer at two adjacent sampling times. These changes are then subtracted from the changes under noise-free conditions to form the current regularization constraint term.

[0076] The regularity constraint term is expressed as: ;in, This represents the loss caused by regularization constraints, indicating the loss due to characteristic fluctuations at consecutive time points; This represents the change in the characteristic of liquid layer i at time t, used to illustrate the change in the corresponding characteristic value of the matching feature at adjacent time points; This represents the predicted change in liquid layer i at time t, and represents the change without noise interference. This represents the calculation method of Mahalanobis distance, used to assign different tolerances to fluctuations in different feature dimensions, avoiding dimensional dominance; This represents the regularization coefficient for liquid layer i, used to assign values ​​based on the steady-state layer, metastable transition layer, and unstable abrupt change layer indicated by the liquid layer state label. n represents the current number of liquid layers. This constraint term is used to determine the degree to which the measured fluctuation deviates from the expected fluctuation, and when the loss increases sharply at a certain moment, it limits the relative fluctuation of the overall loss based on the total loss generated at the current moment. This represents the covariance matrix, used to normalize fluctuations in different dimensions.

[0077] Furthermore, during the evolution of the emulsion, the oil droplets migrate upward as a whole, while the aqueous phase migrates downward as a whole. Therefore, the backscattered light intensity at the top decreases monotonically with time, while the backscattered light intensity at the bottom increases monotonically with time. Moreover, as the overall process changes, the migration process gradually changes. At this point, an additional time decay term due to the emulsion evolution will be introduced, thereby constraining the output of the regularization constraint term.

[0078] The regularization constraint term after introducing time decay then transforms into: ; in, This represents the optimal hierarchical feature at the current moment. This represents the hierarchical feature output at time t. This feature is obtained by concatenating and matching features after layer-by-layer alignment without regularization constraints. It is the quantized entity of the decay at the current time t. This represents the initial regularization coefficient, with an initial value of 0.5. Represents an exponential constant. This represents the decay rate, and denoted by a positive real number greater than 0.

[0079] Meanwhile, the optimal hierarchical features at each time step are affected by the currently introduced time decay and regularization coefficients.

[0080] ;in, This represents the optimal hierarchical feature at the previous time step; This represents the regularization coefficient after time decay, which is... When the current layering feature represents the values ​​of multiple liquid layers, the regularization coefficients are summed to obtain the regularization coefficients relative to the multiple liquid layers. It should be noted that if time t is 0, i.e., the initial time, since there is no optimal layering feature from the previous time, the layering feature initially concatenated in step S3 is directly taken as the optimal layering feature for the initial time. .

[0081] Then, the loss values ​​at multiple time points are used to minimize the total loss as the iteration objective. The iteration process stops when the total loss is minimized. This iteration process does not involve direct modification of the hierarchical features. Instead, it is used to determine the time decay rate and the corresponding regularization coefficient that are consistent with the current changes, so that the overall process can suppress short-term high-frequency noise and make the overall process closer to the trend of physical changes. The final output hierarchical features are the optimal hierarchical features calculated at each time point, which maximizes the relative expressive power of each feature and provides an accurate data foundation for subsequent component mapping.

[0082] Furthermore, during the iteration process, when the optimal hierarchical features are obtained in two adjacent iterations... When the relative change is less than 1%, it is judged as iterative convergence, the iteration is stopped, and the current optimal hierarchical feature is set. Hierarchical features as output.

[0083] In one embodiment of the present invention, the layering features are mapped to the target components detected in the emulsion to determine the corresponding chemical components under different feature layers, and the coupling of the relative contents of these components under the current detection process.

[0084] Specifically, the target components include the dispersed phase (oil phase), the continuous phase (aqueous phase), surfactants (emulsifiers), and trace mass indicators. The dispersed phase mainly includes triglycerides, unsaturated fatty acids in oils, and volatile flavor oils. The continuous phase mainly includes deionized water, soluble polysaccharides, and buffer salts. The surfactants mainly include small molecule surfactants such as monoglycerides, large molecule proteins, and hydrophilic colloids. The trace mass indicators mainly include peroxides and carbonyl compounds. These components directly indicate the current production quality of the emulsion. Based on the regression calculation of the collected spectra, the main components will be identified to explain the oil content, water content, and relative concentration of emulsifiers within the layer, enabling these components to be physically aligned with the layered characteristics. Then, based on the relative time points of component changes, the dominant characteristics during component changes will be calculated, thereby identifying the relative characteristics of the emulsion under multiple tests.

[0085] One implementation of step S4 includes: S41, during the target component mapping process, calling the regression equation corresponding to the target component, using the stratification features as input values, and determining the quantitative values ​​of each target component; wherein, the regression equation is established using partial least squares method. Specifically, at least 30 sets of standard emulsion samples with known component contents are collected, their stratification feature data are acquired, and the stratification features are used as independent variables, with the component contents obtained from chemical analysis as dependent variables, and the regression coefficient matrix is ​​obtained by fitting using the partial least squares algorithm. During online detection, the measured stratification features are input into the model, and the quantitative values ​​of each target component are obtained by multiplying the regression coefficient matrix and the independent variables by the dependent variable.

[0086] S42, when the layered features are updated, the similarity between the rate of change of the target component and the rate of change of the physical features of each liquid layer is calculated, and the physical feature with the highest similarity is used as the output associated feature.

[0087] The physical characteristics of each liquid layer are represented by the changes in particle size growth corresponding to the current component, the steepening of the spatial gradient of backscattered light intensity, and the overall change in backscattered light intensity. These three physical characteristics indicate changes in the physical properties of the emulsion. For example, particle size is obtained by fitting backscattered data from multiple geometric angles using Mie scattering theory. Changes in this parameter directly indicate the physical stability of the emulsion, representing problems such as insufficient emulsifier coverage. The spatial gradient of backscattered light intensity indicates gravitational stratification, suggesting problems such as excessively low emulsion viscosity or excessive density difference. Spatial gradients are needed to verify the relative changes in the mapping of different components. The overall backscattered light intensity describes the process of chemical self-oxidation or overall hydrolysis to determine the changes in the relative properties of the emulsion.

[0088] If the stratification features are updated after optimization, it indicates that some data content has relevant differences at different times. It is necessary to determine the correlation between the quantitative value of the target component and the relevant physical features in the relative data at the time of update to explain the relative problem of emulsion quality. At this time, according to the calculation method of Pearson correlation coefficient, the physical feature with the largest similarity is selected as the current output correlation feature. Each correlation feature is bound to the physical location of the stratification feature to obtain the relative correlation at multiple liquid layer heights.

[0089] Furthermore, in Mie scattering theory, the size parameter is the product of the oil droplet diameter and pi, divided by the incident light wavelength, which is approximately equal to 1. This means that the particle size can be inverted using the backscattered light intensity reflected by the incident light wavelength. During the inversion process, the measured backscattered light intensity is compared with the intensity value from Mie scattering theory. The particle size is then calculated using the least squares method, and the particle size is obtained by measuring the same liquid layer multiple times to obtain the output particle size value. When performing similarity comparison, the influence of changes in the target composition in the liquid layer on the output particle size value is determined by repeatedly inverting the particle size value.

[0090] In one embodiment of the present invention, the target components are summarized into multidimensional features according to their liquid surface height based on the characteristic forms associated with the target components, thereby improving the accuracy of microscopic detection of emulsions.

[0091] One implementation of step S5 includes: comparing the quantitative value of the target component with the corresponding safety threshold to determine the relative score of the target component, and weighting the relative score according to the duration and type of the associated features of the target component to obtain the quality risk index of the target component.

[0092] In this embodiment, the quality risk index is used to characterize the degree to which a target component in the emulsion under test approaches or exceeds its corresponding safety threshold, as well as the cumulative risk formed by the continuous action of physical changes associated with the target component within a preset detection period. The quality risk index ranges from 0 to 1. The larger the quality risk index, the higher the degree of abnormality corresponding to the target component, and the higher the degree to which the emulsion under test approaches instability or exceeds its quality limit.

[0093] Specifically, for each continuously set liquid layer, the quantitative value of the target component at the current liquid layer height and the corresponding safety threshold are obtained. The safety threshold is retrieved from a preset database according to the product type of the emulsion and the corresponding production quality standards, and is used to characterize the critical value that the emulsion still meets the preset quality requirements when it undergoes a certain degree of flocculation, gravity stratification, or component change.

[0094] The ratio of the quantitative value of a target component to its corresponding safety threshold is defined as the relative score of that target component. The relative score characterizes how close the target component is to the safety boundary; a higher relative score indicates that the target component is closer to the safety boundary, and its corresponding quality risk is higher. When the relative score of any target component is greater than 1, it indicates that the quantitative value of that target component has exceeded the corresponding safety threshold. The target component is then classified as exceeding the limit, and the emulsion to be tested is deemed unqualified. Simultaneously, the corresponding quality risk index is set to 1.

[0095] When the relative score of the target component is not greater than 1, the relative score is used as the static score of the target component, and the time weight and importance weight of the associated features corresponding to the target component are further obtained to perform weighted correction on the static score.

[0096] Specifically, for each associated feature corresponding to the target component, the ratio of the duration of the associated feature's continuous occurrence within a preset detection period to the total duration of the preset detection period is determined as the time weight of that associated feature. The longer the associated feature continues to function, the higher the degree of accumulation of the corresponding physical change in the emulsion, and the greater its impact on quality risk indicators.

[0097] The importance weights are set based on the emulsion instability processes and their impact as characterized by different correlation features. Changes in particle size usually characterize oil droplet coalescence or decreased interfacial film stability, and have a relatively large impact on the microstructural stability of the emulsion; changes in overall backscattered light intensity can characterize the continuous change in the overall scattering properties of the emulsion, and may correspond to oxidation, hydrolysis, or changes in the overall component state; steepening of the spatial gradient of backscattered light intensity mainly characterizes the degree of component migration and stratification between different liquid layers under the influence of gravity.

[0098] As an example, the importance weights corresponding to changes in particle size growth, steepening of the spatial gradient of backscattered light intensity, and overall changes in backscattered light intensity can be set to 0.45, 0.20, and 0.35, respectively. The static scores of the target component are weighted and summarized according to the time and importance weights of the corresponding associated features to obtain a quality risk index for the target component in the current detection scenario. The closer the quality risk index is to 1, the more obvious the abnormal changes corresponding to the target component, and the higher the degree to which the emulsion is close to instability or exceeding quality limits.

[0099] It should be noted that the above-mentioned importance weights are only used to illustrate the calculation method of the quality risk indicators in this embodiment and do not constitute a limitation on the specific values ​​of the weights. In practical applications, the importance weights of each associated feature can be calibrated or adjusted according to the product type, production process requirements, and historical testing data of the emulsion.

[0100] Furthermore, when it is necessary to evaluate the relative impact of different target components on the overall quality of the emulsion, the component weights corresponding to each target component can be obtained from the database, and a weighted summary can be performed based on the quality risk index of each target component and its component weight to obtain the comprehensive quality risk index of the emulsion to be tested. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention, which are still covered within the protection scope of the present invention.

Claims

1. A method for quality detection of oil-in-water emulsions based on spectral technology, characterized in that, include: S1, the emulsion to be tested is introduced into a constant temperature test environment, the transmitted light intensity is collected from all angles, and the macroscopic state of the emulsion is determined accordingly. S2, scan vertically along the emulsion to collect backscattered light intensity as it changes with height and time. Based on the trend of backscattered light intensity at different liquid layer heights, determine the local features and the number of effective features under the current macroscopic state. S3 uses the number of effective features at different liquid layers as a constraint to constrain the search range of local feature matching from top to bottom, and constructs the layered features of each liquid layer accordingly. S4. Multi-component mapping is performed between the layered features and the target components of the emulsion to be detected. Combined with the update time of the layered features, the correlation characteristics of different target components in the emulsion are clarified. S5. Based on the correlation characteristics corresponding to the target components, determine the quality risk indicators of the target components.

2. The method for quality detection of oil-in-water emulsions based on spectral technology according to claim 1, characterized in that, The implementation methods of the macroscopic state in step S1 include: The average transmittance of the collected transmitted light intensity is calculated, and the overall average transmittance is determined based on the trend of the average transmittance.

3. The method for quality detection of oil-in-water emulsions based on spectral technology according to claim 1, characterized in that, The implementation methods of local features in step S2 include: S21, construct a global transmission baseline using transmitted light intensity, and determine the corrected backscattered light intensity; S22, for each spatial location corresponding to the liquid layer height, the slope change rate of the backscattered light intensity is used as a local feature in the time dimension. S23, for each fixed time, the spatial gradient of backscattered light intensity at adjacent liquid layer heights is used as a local feature of the spatial dimension. S24, based on local features in spatial and temporal dimensions, filters the number of effective features under different macroscopic states.

4. The method for quality detection of oil-in-water emulsions based on spectral technology according to claim 3, characterized in that, The methods for determining the corrected backscattered light intensity include: The arithmetic mean of the transmitted light intensity at any time is used as the global transmission baseline, and the global transmission baseline at the initial time is used as the reference value to construct the compensation coefficient. For the obtained compensation coefficient, the product of the backscattered light intensity and the compensation coefficient is regarded as the corrected backscattered light intensity.

5. The method for quality detection of oil-in-water emulsions based on spectral technology according to claim 3, characterized in that, The effective number of features can be achieved in the following ways: For the backscattering time series variance of each liquid layer under different macroscopic conditions, the liquid layer that satisfies the variance contribution rate is regarded as the effective liquid layer; At the same time, for the effective liquid layer obtained, feature filtering is performed based on the distribution of extreme points to determine the number of effective extreme points corresponding to the spatial gradient at the same time. Singular value decomposition is performed on the slope change rate of backscattered light intensity of each liquid layer to select the number of effective principal components corresponding to each liquid layer; the number of effective extrema and the number of effective principal components are used together as the number of effective features in the output.

6. The method for quality detection of oil-in-water emulsions based on spectral technology according to claim 1, characterized in that, The implementation methods of hierarchical features in step S3 include: S31, Based on the number of valid features entered, constrain the search range for local feature alignment layer by layer; S32, using cosine similarity as the main factor to perform local feature matching between adjacent liquid layers, and determining matching features that meet the similarity test layer by layer from top to bottom; S33 concatenates the matched features after layer-by-layer alignment to form a vertically layered feature sequence, thus obtaining the output layered features.

7. The method for quality detection of oil-in-water emulsions based on spectral technology according to claim 6, characterized in that, When constraining the search range for local feature alignment, the implementation methods include: Based on the number of valid features passed from the previous layer, the number of valid features is mapped to liquid layer state labels, and the optimal index position for matching in the previous layer is determined. Based on the optimal index position during the previous layer matching, the size of the subset that matches the liquid layer state label is divided, and the similarity threshold of local feature matching is adjusted synchronously to complete the search range constraint.

8. The method for quality detection of oil-in-water emulsions based on spectral technology according to claim 6, characterized in that, Methods for forming vertically hierarchical feature sequences include: Construct a regularization constraint term for the residual distribution of matching features, and use the loss of the regularization constraint term to constrain the fluctuation amplitude of the backscattered features. By combining the relative laws of emulsion evolution, a layered boundary under time constraints is constructed, and the layered features of the vertical distribution are obtained by iterative solution.

9. The method for quality detection of oil-in-water emulsions based on spectral technology according to claim 1, characterized in that, The methods for implementing the associated features in step S4 include: S41, In the process of target component mapping, the regression equation corresponding to the target component is called, and the hierarchical features are used as input values ​​to determine the quantitative values ​​of each target component. S42, when the layered features are updated, the similarity between the rate of change of the target component and the rate of change of the physical features of each liquid layer is calculated, and the physical feature with the highest similarity is used as the output associated feature.

10. The method for quality detection of oil-in-water emulsions based on spectral technology according to claim 1, characterized in that, The implementation methods of the quality risk indicators in step S5 include: The ratio of the target component's corresponding component value to the safety threshold is used as the relative score for the target component. Combined with the duration of the effects of each associated feature, the output quality risk index is obtained.

Images