Deep learning-based method and system for predicting oil oxidation stability

Abstract

The present application relates to the technical field of data processing, and more particularly to a grease oxidation stability prediction method and system based on deep learning, which divides the multi-dimensional monitoring data sequences of different complete grease accelerated oxidation experiments at the same time into a plurality of monitoring data sequence sets; performs data fusion on the monitoring data of all monitoring data sequences in any monitoring data sequence set at the same time to obtain a fused data sequence, and divides the fused data sequence into three stage data sequences; obtains an adaptive smoothing window of each fused data in each stage data sequence, performs smoothing processing on the fused data sequence to obtain a smoothed data sequence; and uses the smoothed data sequences of all monitoring data sequence sets as the input of a CNN network, outputs the predicted value of the grease oxidation stability, effectively suppresses noise and stage differences through phased adaptive smoothing, strengthens the feature expression of the key change segment, and improves the prediction accuracy of the grease oxidation stability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for predicting the oxidation stability of oils and fats based on deep learning. Background Technology

[0002] As an important source of nutrition and energy for humans, oils not only provide essential fatty acids that the human body cannot synthesize (such as linoleic acid and alpha-linolenic acid), but also serve as important carriers of fat-soluble vitamins (VA, AD, AE, and VK), playing a crucial role in improving and enhancing the taste, flavor, and physical properties of food. However, oils are highly susceptible to oxidation during storage and use, leading to a decline in quality, the production of unpleasant odors and harmful substances, severely impacting the sensory flavor and shelf life of products, and potentially even endangering consumer health. Therefore, accurately predicting the oxidative stability of oils is of paramount importance for the production, storage, and processing of oils, as well as related industries such as food, cosmetics, and pharmaceuticals.

[0003] Traditionally, neural networks (CNNs) are used to predict the oxidative stability of lipids. However, since the input data for this prediction mainly comes from experimental methods, various factors inevitably influence the results, introducing noise. For example, the limited precision of instruments may lead to measurement errors, and fluctuations in environmental conditions such as temperature and humidity can affect the accuracy of the results. Therefore, existing technologies use moving average algorithms to smooth the collected monitoring data before predicting lipid oxidative stability. However, using a fixed window for smoothing is not suitable for all experimental data. For example, during the oxidation induction period, excessive smoothing may mask abrupt changes (such as a sudden increase in POV), leading to inaccurate predictions.

[0004] Therefore, how to dynamically adjust the window in the moving average algorithm to improve the data smoothing effect and thus improve the prediction effect of oil oxidation stability has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method and system for predicting the stability of lipid oxidation based on deep learning, in order to solve the problem of how to dynamically adjust the window in the moving average algorithm to improve the data smoothing effect and thus improve the prediction effect of lipid oxidation stability.

[0006] In a first aspect, embodiments of the present invention provide a deep learning-based method for predicting the oxidation stability of lipids, the method comprising the following steps:

[0007] Acquire multidimensional monitoring data sequences corresponding to different complete accelerated oil oxidation experiments at the same time, and divide the multidimensional monitoring data sequences into multiple monitoring data sequence sets according to the dimensions;

[0008] For any set of monitoring data sequences, data fusion is performed on the monitoring data of all monitoring data sequences in the set at the same time to obtain a fused data sequence; a monitoring data change curve is constructed for each monitoring data sequence, and the change trend of each monitoring data change curve is combined to obtain two stage division times. Using the two stage division times, the fused data sequence is divided into three stage data sequences.

[0009] For any stage of the data sequence, based on the data change trend and data volume of the data sequence, the corresponding initial smoothing window is obtained, and the initial smoothing window corresponding to each fused data in the data sequence is adaptively adjusted to obtain an adaptive smoothing window.

[0010] Based on the adaptive smoothing window of each fused data in each stage data sequence, the fused data sequence is smoothed to obtain the smoothed data sequence corresponding to any monitoring data sequence set; the smoothed data sequences corresponding to all monitoring data sequence sets are used as input to the CNN network, and the corresponding output is the predicted value of lipid oxidation stability.

[0011] In a second aspect, embodiments of the present invention provide a deep learning-based system for predicting the oxidation stability of lipids, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements a deep learning-based method for predicting the oxidation stability of lipids as described in the first aspect.

[0012] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows:

[0013] In this invention, monitoring data of any dimension at each time point is simultaneously collected by multiple sensors, and fused data corresponding to each time point is obtained through data fusion. This results in a fused data sequence characterizing a complete accelerated lipid oxidation experimental process, reducing sensor measurement errors. Furthermore, by analyzing the stage change characteristics of the lipid oxidation process under multiple sets of experimental data, two stage division points are obtained, which are used to divide the fused data sequence into three stages. For each stage, an initial smoothing window is first set according to the trend and volatility of the data sequence under each stage. To ensure the smoothing effect of noisy, abnormal, and normal data, the initial smoothing window of each data point is adaptively and dynamically adjusted. Thus, stage-wise smoothing is performed using the adaptive smoothing window of each data point under each stage. Stage-wise adaptive smoothing can effectively suppress noise and stage differences, strengthen the feature expression of key change segments, and improve the sensitivity of the subsequent neural network (CNN) to the stage inflection points and rates in the lipid oxidation process, thereby improving the prediction accuracy and robustness of lipid oxidation stability. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a deep learning-based method for predicting the oxidative stability of oils, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0016] Embodiments of this disclosure are described in detail below, with examples of these embodiments illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting it.

[0017] It should be noted that the terms "first," "second," etc., used in this disclosure and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0018] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0019] See Figure 1 This is a flowchart of a deep learning-based method for predicting the oxidative stability of lipids, as provided in Embodiment 1 of the present invention. Figure 1 As shown, the method may include:

[0020] Step S101: Obtain multidimensional monitoring data sequences corresponding to different complete accelerated oil oxidation experiments at the same time, and divide the multidimensional monitoring data sequences into multiple monitoring data sequence sets according to the dimensions.

[0021] The Schaal oven method simulates accelerated oxidation conditions (high temperature and constant temperature) to shorten the oxidation cycle of oils and fats. It assesses the oxidative stability of oils and fats by monitoring the changing trends of other indicators such as peroxide value (POV). Its core logic includes: 1. Accelerated oxidation: In a constant temperature oven at (63±1)℃ or (60±1)℃, the oxidation rate of oils and fats is significantly accelerated, completing an oxidation process that would take months or even years under natural conditions within days to weeks; 2. Quantification of indicators: The degree of oxidation is quantified by periodically measuring indicators such as POV (reflecting the accumulation of primary oxidation product hydroperoxides), acid value (AV, reflecting the content of free fatty acids), and anisidine value (P-AV, reflecting the content of secondary oxidation product aldehydes); 3. Trend analysis: The oxidative stability of oils and fats is determined based on the changing trends of these indicators over time.

[0022] Therefore, in a complete accelerated oxidation experiment of oils and fats, the peroxide value (POV value) is obtained by iodometric titration or automatic titration, the acid value (AV value) is obtained by titration, the anisidine value (P-AV value) is obtained by spectrophotometry by reacting with anisidine reagent, the polar compound (TPC) is obtained by infrared spectroscopy, and environmental data (temperature, oxygen, light) is obtained by sensors. Since the experiment involves chemical changes and the data changes slowly, the acquisition frequency is set to once every 10 minutes, so that the peroxide value, acid value, anisidine value, polar compound and environmental data at each time point can be obtained, forming multidimensional monitoring data at the corresponding time point.

[0023] Since the experimental and sampling processes are mainly carried out manually, and there may be fluctuations in the monitoring data due to operational errors when collecting relevant monitoring data, multidimensional monitoring data sequences corresponding to different (preferably set to 10 sets of experiments) complete accelerated oil oxidation experiments are obtained within the same experimental time. One complete accelerated oil oxidation experiment process corresponds to one multidimensional monitoring data sequence. Then, the multidimensional monitoring data sequence is divided into multiple monitoring data sequence sets according to the dimensions. One dimension corresponds to one monitoring data sequence set, and one monitoring data sequence corresponds to one complete accelerated oil oxidation experiment process. This is used for subsequent multi-source data fusion to obtain a sequence that can more accurately characterize the complete accelerated oil oxidation experiment process.

[0024] Step S102: For any set of monitoring data sequences, perform data fusion on the monitoring data of all monitoring data sequences in the set at the same time to obtain a fused data sequence; construct the monitoring data change curve of each monitoring data sequence, combine the change trends of each monitoring data change curve, obtain two stage division times, and use the two stage division times to divide the fused data sequence into three stage data sequences.

[0025] Taking the POV value in multidimensional monitoring data as an example, the analysis process for monitoring data in other dimensions is the same. The set of monitoring data sequences corresponding to the POV value is denoted as any monitoring data sequence set. It is known that the lower the POV value, the longer the oxidation induction period, indicating stronger antioxidant performance. Lipid oxidation is divided into three stages: induction period, propagation period, and termination period. The induction period is the initial stage of lipid oxidation reaction. At this time, the oxidation rate is slow, hydroperoxides (primary oxidation products) accumulate less, and the POV value does not change significantly. Generally, the first inflection point of the POV curve is taken as the end time of the induction period. The propagation period is the stage of accelerated oxidation reaction. Hydroperoxides decompose rapidly into free radicals, triggering a chain reaction, and the POV value rises sharply. Generally, the second derivative of the POV curve is taken, and the interval where the second derivative is positive and relatively large is regarded as the propagation period. The termination period is the stage where the oxidation reaction tends to end. Free radicals combine with each other to generate stable products, and the rate of increase of POV value slows down or reaches a plateau. When the change of POV value in two consecutive measurements is less than 5%, it can be determined that the termination period has begun.

[0026] Therefore, in this embodiment of the invention, based on the performance characteristics of POV values ​​under different experimental processes, the reliability of the POV values ​​corresponding to different groups at each moment is calculated to determine whether they conform to the data performance characteristics under the experimental stage. For example, during the induction period, the POV value rises slowly, the curve is flat, the slope is close to zero, the AV value changes little, it may decrease slightly or remain stable, and the P-AV value remains almost unchanged, reflecting that secondary oxidation products such as aldehydes have not accumulated; during the propagation period, the POV value rises rapidly, the curve is steep, the slope increases significantly, the AV value rises rapidly, reflecting the large generation of free fatty acids, and the P-AV value increases significantly; during the termination period, the rate of increase of the POV value slows down or reaches a plateau, the AV value continues to rise but at a slower rate, and the P-AV value tends to stabilize or rise slightly. If there is no noise interference or anomaly in the POV values ​​of multiple groups at the same moment, the difference between the multiple POV values ​​at the same moment and their adjacent data should be small or consistent. Since different sensors themselves have accuracy and error, in this embodiment of the invention, based on reliability, the monitoring data of all monitoring data sequences in any monitoring data sequence set at the same moment are fused to obtain a fused data sequence to eliminate measurement error.

[0027] The method for fusing monitoring data from all monitoring data sequences in any set of monitoring data sequences at the same time, based on reliability, to obtain a fused data sequence is as follows:

[0028] For the i-th monitoring data in any monitoring data sequence in any monitoring data sequence set, calculate the absolute value of the difference between the i-th monitoring data and the (i-1)-th monitoring data in any monitoring data sequence, denoted as the previous absolute value of the difference; calculate the absolute value of the difference between the i-th monitoring data and the (i+1)-th monitoring data, denoted as the subsequent absolute value of the difference; normalize the reciprocal of the absolute values ​​of the differences between the previous and subsequent absolute values ​​of the difference to obtain the similarity of the differences before and after.

[0029] Calculate the first ratio between the absolute value of the difference before and the (i-1)th monitoring data, calculate the second ratio between the absolute value of the difference after and the (i+1)th monitoring data, calculate the absolute value of the difference between the first ratio and the second ratio, and normalize the sum of the absolute value of the difference and the constant 1 to obtain the degree of normal variation.

[0030] Calculate the difference coefficient between the i-th monitoring data, the (i-1)-th monitoring data, and the (i+1)-th monitoring data, and denote it as the first difference coefficient. Calculate the difference coefficient of the data sequence of a preset length with the (i+1)-th monitoring data as the cutoff data, and denote it as the second difference coefficient. Normalize the absolute value of the difference between the first difference coefficient and the second difference coefficient to obtain the difference coefficient similarity of the i-th monitoring data.

[0031] The monitoring reliability of the i-th monitoring data is obtained by calculating the mean of the similarity between the before and after differences, the normal degree of change, and the similarity of the difference coefficient.

[0032] In one embodiment, taking the i-th monitoring data in the z-th monitoring data sequence of any monitoring data sequence set as an example, the formula for calculating the monitoring reliability of the i-th monitoring data is:

[0033]

[0034] in, This represents the monitoring reliability of the i-th monitoring data in the z-th monitoring data sequence. This represents the i-th monitoring data. This represents the (i-1)th monitoring data. This represents the (i+1)th monitoring data point, where | represents the absolute value sign. This represents the coefficient of variation among all monitoring data points from the (i-1)th monitoring data point to the (i+1)th monitoring data point. It represents the difference coefficient of a data sequence with a preset length (preferably set to 10) using the (i+1)th monitoring data as the cutoff data.

[0035] It should be noted that, The smaller the absolute value of the difference, the more similar the i-th monitoring data is to the differences between it and the monitoring data before and after it. Used to characterize the relative rate of change from the (i-1)th monitoring data point to the ith monitoring data point. The relative change rate is used to characterize the relative change rate from the i-th monitoring data to the (i+1)-th monitoring data. The relative change rate is more sensitive to abnormal data. It is further limited by the difference in the relative change rate between adjacent data. When the i-th monitoring data is abnormal, its relative change rate with the normal monitoring data before and after it will deviate significantly from the normal range. Therefore, the more similar the difference between the i-th monitoring data and the monitoring data before and after it, and the smaller the difference in the relative change rate, the greater the possibility that the i-th monitoring data is normal data, and the higher the monitoring reliability. The smaller the difference between the two difference coefficients, the more the local data of the i-th monitoring data is consistent with the data performance of the stage in which the monitoring data is located, and the higher the monitoring reliability of the i-th monitoring data.

[0036] Similarly, the monitoring credibility of each monitoring data in each monitoring data sequence within any set of monitoring data sequences is obtained. Then, for any given time point, based on the monitoring credibility of the monitoring data in each monitoring data sequence at that time point, the weight of the monitoring data in each monitoring data sequence at that time point is obtained; the higher the monitoring credibility, the greater the weight. The monitoring data at any given time point in all monitoring data sequences are then weighted and summed to obtain the fused data at that given time point. Finally, the fused data at each time point is obtained, resulting in a fused data sequence.

[0037] In one implementation, the calculation expression for the fused data at any given time is:

[0038]

[0039]

[0040] in, This represents the weight of the i-th monitoring data in the z-th monitoring data sequence. This represents the monitoring reliability of the i-th monitoring data in the z-th monitoring data sequence. This represents the number of sets of monitoring data sequences, that is, the number of sequences contained in any set of monitoring data sequences. This represents the fused data of the i-th monitoring data in all monitoring data sequences, that is, the fused data at the corresponding time point of the i-th monitoring data. This represents the i-th monitoring data in the z-th monitoring data sequence.

[0041] Similarly, weighted fusion is performed on the monitoring data at each same time point in all monitoring data sequences in any monitoring data sequence set to obtain the fused data at each time point, which in turn forms a fused data sequence, representing the POV value sequence in a complete oil accelerated oxidation experiment process.

[0042] It is known that the induction phase of the lipid oxidation process is characterized by a gradual rise in the monitoring data change curve in the initial stage, the propagation phase by a steep rise in the monitoring data change curve in the middle stage with a significantly increased slope, and the termination phase by a slowdown or horizontalization of the rise rate of the monitoring data change curve. Therefore, based on the monitoring data change curves of each monitoring data sequence and the combined trend of each monitoring data change curve, the phase division point can be determined to divide the fused data sequence into three phases: the induction phase, the propagation phase, and the termination phase.

[0043] Specifically, using monitoring data as the vertical axis and time as the horizontal axis, a monitoring data change curve, i.e., a POV value change curve, is constructed for each monitoring data sequence. For any monitoring data change curve, the probability that each data point is a stage point is obtained based on the fitting slope of the data points before and after each data point in the curve. All probabilities are then combined into a probability sequence. The formula for calculating the probability that the j-th data point in any monitoring data change curve is a stage point is as follows:

[0044]

[0045] in, This represents the probability that the j-th data point in any of the monitoring data change curves is a stage point. This represents the first weight, and 1 represents a constant. Represents the normalization function. This represents the slope of the fit between the j-th data point and the first m data points. This represents the fitting slope of the first m data points of the j-th data point. This represents the fitted slope of the m data points following the j-th data point. Let | represent the fitted slope of the m data points following the j-th data point, where || denotes the absolute value sign. This indicates the second weight.

[0046] It should be noted that the least squares method is used to fit the data points and obtain the corresponding fitting slope. This represents the difference in the trend of data performance before the j-th data point. The smaller the value, the more similar the trend of data performance before the j-th data point is, and the higher the probability that the data are from the same stage, which is consistent with the characteristics of data performance before the stage point. This value represents the difference in data performance trends after the j-th data point. The smaller the value, the more similar the data performance trends after the j-th data point, indicating a higher probability that the data belong to the same stage, which aligns with the consistency characteristic of data performance after a stage point. Therefore... This indicates the overall consistency of the trend of the data before and after the j-th data point. The smaller the overall consistency, the greater the probability that the j-th data point is a stage point. This represents the difference in performance trends between adjacent data points before and after the j-th data point. The larger this value, the greater the likelihood that the adjacent data points belong to different stages, and thus the greater the probability that the j-th data point is a stage point. Since inconsistent performance trends between adjacent data points can also occur in outlier or noisy data, the reference weight of the second term can be set relatively low. .

[0047] Since oxidation is divided into three stages, there should be two corresponding stage points: the stage point dividing the induction and propagation stages, and the stage point dividing the propagation and termination stages. Therefore, we first use the first-order difference method to obtain local extreme points in the probability sequence. This method is existing technology and will not be described in detail here. Then, we screen for effective extreme points among the local extreme points. Finally, based on the probability of the effective extreme points and the changes in local data, we determine the stage division time corresponding to the stage point.

[0048] The method for selecting valid extreme points among local extreme points includes: to eliminate false extreme points caused by data noise or small fluctuations, in this embodiment of the invention, a probability allowable fluctuation value is set, which is taken as an empirical value of 0.2; for any local extreme point, the average value of a preset number (set to 6) of probabilities closest to the local extreme point is obtained in the probability sequence, and the sum of the average value and the preset probability allowable fluctuation value is used as a local maximum threshold. If the probability corresponding to any local extreme point is greater than the local maximum threshold, the local extreme point is regarded as a valid local maximum point; the difference between the average value and the preset probability allowable fluctuation value is used as a local minimum threshold. If the probability corresponding to any local extreme point is less than the local minimum threshold, the local extreme point is regarded as a valid local minimum point; each local extreme point is traversed, and all valid local minimum points and valid local maximum points are combined to form a valid extreme point.

[0049] The method for determining the stage division time corresponding to a stage point based on the probability of effective extreme points and local data changes includes: for any effective extreme point, calculating the slope between the effective extreme point and the data points before and after it, obtaining the absolute value of the slope difference, and recording the product of the absolute value of the slope difference and the probability corresponding to the effective extreme point as the stage feature value of the effective extreme point; and based on the stage feature value of each effective extreme point, obtaining the two effective extreme points with the largest stage feature values ​​as the suspected stage points of any monitoring data change curve.

[0050] Similarly, obtain the suspected stage points of each monitoring data change curve, and number each suspected stage point of each monitoring data change curve according to the time series. The numbers include 1 and 2. Calculate the mean of the time corresponding to the first suspected stage point of all monitoring data change curves, and record it as the first stage division time. Calculate the mean of the time corresponding to the second suspected stage point of all monitoring data change curves, and record it as the second stage division time.

[0051] After obtaining the two stage division times, the fused data sequence is divided into three stage data sequences using the two stage division times. At this point, the purpose of dividing the fused data sequence into three stages is achieved, with one stage data sequence corresponding to one stage.

[0052] Step S103: For any stage data sequence, based on the data change trend and data volume of any stage data sequence, obtain the corresponding initial smoothing window, and adaptively adjust the initial smoothing window corresponding to each fused data in any stage data sequence to obtain an adaptive smoothing window.

[0053] After dividing the fused data sequence into stages, an initial smoothing window can be set for each stage. Taking any stage's data sequence as an example, since the maximum window length cannot exceed the number of data points within the corresponding stage, the number of data points in the statistical stage's data sequence is used as the maximum window length for that stage. The size of the smoothing window is usually affected by the data's trend, volatility, and noise level. Since the initial smoothing window is a general window for the corresponding stage, the noise level can be disregarded initially, and the data's trend and volatility should be prioritized. Furthermore, since smoothing should not be overdone to avoid masking the data's trend and volatility characteristics, if the data exhibits weak trend and weak volatility, the initial smoothing window can be set larger to capture long-term data performance; conversely, if the data exhibits strong trend and strong volatility, the initial smoothing window can be set smaller to avoid masking short-term data performance.

[0054] Based on the above features, in this embodiment of the invention, obtaining the corresponding initial smoothing window according to the data change trend and data volume of the data sequence at any stage includes:

[0055] Based on the ratio of data difference to time difference between adjacent fused data, the slope of change between every two fused data in any stage of the data sequence is calculated to obtain the average slope of change. The average slope of change is then normalized to obtain a data change trend index. The slope of change is a prior art and will not be described in detail here.

[0056] Polynomial fitting is performed on the data sequence at any stage to obtain the corresponding fitting curve, wherein the mathematical expression corresponding to the polynomial fitting is: Where x is the independent variable and y is the dependent variable. For constant terms, The coefficient of the linear term, The coefficient of the quadratic term is used to perform first-order differentiation on the fitted curve to obtain the number of data with negative derivatives and the number of data with positive derivatives. The absolute value of the difference between the number of data with negative derivatives and the number of data with positive derivatives is normalized to obtain a normalized value. The difference between the constant 1 and the normalized value is calculated and denoted as the fluctuation characteristic value.

[0057] Calculate the absolute value of the difference between every two adjacent fused data in any stage of the data sequence to obtain the average absolute value of the difference. Normalize the average absolute value of the difference to obtain the average fluctuation amplitude. Use the mean of the fluctuation characteristic value and the average fluctuation amplitude as the data fluctuation index.

[0058] The data change trend index and the data fluctuation index are weighted and summed to obtain a weighted summation result. The difference between the constant 1 and the weighted summation result is used as the window adjustment coefficient of the data sequence in any stage.

[0059] In one embodiment, the formula for calculating the window adjustment coefficient of the data sequence at any stage is:

[0060]

[0061] in, This represents the window adjustment coefficient for the data sequence in the r-th stage, where 1 indicates a constant. Indicates the third weight. Represents the normalization function. This represents the average slope of the data sequence in the r-th stage. Indicates the fourth weight. This indicates the number of data points where the derivative is negative. This indicates the number of data points where the derivative is positive; | represents the absolute value sign. This represents the absolute value of the average difference.

[0062] It should be noted that, The larger the value, the stronger the trend of the data sequence in the r-th stage. The initial smoothing window should be set to be smaller, and the corresponding window adjustment coefficient should be smaller. This value represents the volatility characteristics of the data sequence in the r-th stage. A smaller value indicates that the data sequence in the r-th stage exhibits a regular, recurring upward and downward pattern, corresponding to a more pronounced volatility characteristic. To differentiate between strong and weak volatility, we further consider the average volatility amplitude of the data sequence in the r-th stage. A larger average volatility amplitude indicates stronger volatility in the r-th stage, requiring a smaller initial smoothing window and a smaller corresponding window adjustment coefficient. Secondly, since smoothing has a stronger impact on volatility, we can set... This makes the results biased towards volatility.

[0063] After obtaining the window adjustment coefficient, the initial smoothing window of the data sequence at any stage is obtained by combining the maximum window length and the window adjustment coefficient. This initial smoothing window is used to initially set the smoothing window of each fused data in the data sequence at any stage as the initial smoothing window. The method for obtaining the initial smoothing window of the data sequence at any stage is as follows: the length of the data sequence at any stage is counted as the maximum window length, and the product between the maximum window length and the window adjustment coefficient is rounded up to obtain the initial smoothing window corresponding to the data sequence at any stage.

[0064] In one embodiment, the formula for calculating the initial smoothing window corresponding to the data sequence in the r-th stage is:

[0065]

[0066] in, This represents the size of the initial smoothing window corresponding to the data sequence in the r-th stage. This represents the maximum window length of the data sequence in the r-th stage. This represents the window adjustment coefficient of the data sequence in the r-th stage. The symbol indicates rounding up.

[0067] Since the POV value includes normal data, noisy data caused by random fluctuations or measurement errors in the experimental operation that do not significantly deviate from the overall trend but affect the smoothness of the data, as well as abnormal data that significantly deviates from the normal trend, in order to improve the smoothing effect of each fused data in any stage of the data sequence, the initial smoothing window of each fused data is further analyzed for normal data, noisy data and abnormal data. Then, based on the analysis results, the initial smoothing window of each fused data is adaptively adjusted to obtain an adaptive smoothing window.

[0068] Since noisy data fluctuates randomly but does not deviate from the normal trend, it is relatively inconspicuous in the initial smoothing window. However, it will fluctuate with adjacent fused data, and the fused data on both sides will be relatively unsmooth. On the other hand, abnormal data deviates significantly from the normal trend and has no change process. It is prominent in the initial smoothing window and may have the same trend as the fused data on both sides, appearing relatively smooth. Therefore, by analyzing the bias and window trend fluctuation of each fused data, the initial smoothing window is expanded for monitoring data that is biased towards noise and has low window trend fluctuation. This is to smooth the noise by capturing the long-term data change trend, thereby preventing the noise from affecting the subsequent prediction of oil oxidation stability. The initial smoothing window is reduced for monitoring data that is biased towards anomalies and has high window trend fluctuation to prevent the loss of abnormal details due to an excessively large window.

[0069] Based on the above feature analysis, taking the j-th fused data in any stage data sequence as an example, firstly, according to the difference between adjacent data in the initial smoothing window of the j-th fused data, the bias of the j-th fused data to the size of the initial smoothing window is obtained: the average data value in the initial smoothing window of the j-th fused data is calculated, the absolute value of the difference between the j-th fused data and the average data value is obtained, and the absolute value of the difference is normalized to obtain the local salient value;

[0070] Calculate the slope of change between the j-th fused data and the (j-1)-th fused data, and denote it as the first slope. Calculate the slope of change between the first two fused data of the (j-1)-th fused data, and denote it as the second slope. Normalize the absolute value of the difference between the first slope and the second slope to obtain the trend difference degree. Use the difference between the constant 1 and the trend difference degree as the data smoothness.

[0071] The mean between the local spur value and the data smoothness is calculated to obtain the data type bias of the j-th fused data.

[0072] In one implementation, the formula for calculating the data type bias of the j-th fused data is:

[0073]

[0074] in, This indicates the degree of data type bias in the j-th fused data. This represents the normalization function, and 1 represents a constant. This represents the j-th fused data. This represents the average data value within the initial smoothing window of the j-th fused data, where || denotes the absolute value sign. This represents the slope of the change between the j-th fused data and the (j-1)-th fused data. This represents the slope of change between the first two fused data points of the (j-1)th fused data point.

[0075] It should be noted that, The value is used to characterize the local salience of the j-th fused data. The larger the value, the greater the amplitude and the more prominent the j-th fused data is, and the larger the corresponding local salience value. The value of is used to characterize the trend difference between the j-th fused data and the preceding adjacent fused data. The greater the trend difference, the less smooth the j-th fused data is, and the lower the corresponding data smoothness. If the local salient value is small, the data smoothness is large, and the probability that this type of data is normal data is relatively high. If the local salient value is large, the data smoothness is large, and the probability that this type of data is abnormal data is relatively high. If the local salient value is small, the data smoothness is small, and the probability that this type of data is noisy data is relatively high. Therefore, the data type bias corresponding to abnormal data is large, the data type bias corresponding to noisy data is small, and the data type bias corresponding to normal data is moderate.

[0076] Then, considering that in practical applications, the surrounding data analysis of the j-th fused data is not a single data type, it is necessary to further consider the data volatility within the initial smoothing window of the j-th fused data. Therefore, volatility analysis is performed on the data within the initial smoothing window of the j-th fused data to obtain the window trend volatility of the j-th fused data. This includes: calculating the data variance within the initial smoothing window of the j-th fused data, normalizing the data variance to obtain a first normalized value; performing linear fitting on the data within the initial smoothing window of the j-th fused data to obtain a fitting slope, normalizing the absolute value of the fitting slope to obtain a second normalized value; and calculating the mean of the first normalized value and the second normalized value to obtain the window trend volatility of the j-th fused data.

[0077] In one embodiment, the formula for calculating the window trend fluctuation of the j-th fused data is:

[0078]

[0079] in, This indicates the degree of fluctuation in the window trend of the j-th fused data. This represents the data variance within the initial smoothing window of the j-th fused data. This represents the slope of the linear fit performed on the data within the initial smoothing window of the j-th fused data. The symbol represents the normalization function, and | represents the absolute value symbol.

[0080] It should be noted that the larger the data variance, the greater the data volatility within the initial smoothing window of the j-th fused data. Conversely, the smaller the fitting slope, the smaller the data volatility trend within the initial smoothing window of the j-th fused data. Therefore, if the data variance is large and the fitting slope is small, it means that although the data within the initial smoothing window of the j-th fused data fluctuates, it may be random up and down fluctuations with no obvious trend, and the corresponding window trend fluctuation is low. On the contrary, if the data variance is large and the fitting slope is large, the corresponding window trend fluctuation is also large.

[0081] Finally, the weighted sum of the data type bias and the window trend fluctuation is used to obtain the window adjustment feature value of the j-th fused data, wherein the formula for calculating the window adjustment feature value is:

[0082]

[0083] in, This indicates the window adjustment feature value of the j-th fused data. This indicates the degree of data type bias in the j-th fused data. This indicates the degree of fluctuation in the window trend of the j-th fused data. This represents the first fusion weight coefficient. This represents the second fusion weight coefficient.

[0084] It should be noted that since smoothing mainly targets the j-th monitoring data point itself, the data type itself is more important, and can be set as follows: , .

[0085] Similarly, the data type bias and window trend fluctuation of each fused data in any stage of the data sequence are obtained, as well as the corresponding window adjustment feature value. Since abnormal and noisy data are random, the number of normal data is often much greater than the number of abnormal and noisy data. Therefore, a statistical histogram is constructed based on the data type bias of each fused data in any stage of the data sequence. The horizontal axis of the statistical histogram is the data type bias, and the vertical axis is the quantity. The maximum peak point in the statistical histogram is obtained, and the data type bias corresponding to the maximum peak point is recorded as the target data type bias. Then, the fused data corresponding to the target data type bias is marked as normal fused data.

[0086] If the data type of the fused data is highly biased and the window trend fluctuation is also large, the probability that the fused data is anomalous is high. Therefore, the initial smoothing window needs to be reduced to capture the changing characteristics of the anomalous data. Conversely, if the data type of the fused data is relatively unbiased and the window trend fluctuation is also small, the probability that the fused data is noisy is high. Therefore, the initial smoothing window needs to be increased to accurately smooth the noisy data. Thus, in this embodiment of the invention, the adaptive window adjustment for each fused data is performed by combining the window adjustment feature value obtained by fusing the data type bias and the window trend fluctuation, as detailed below:

[0087] The mean value of the window adjustment feature value of the normal fused data in any stage data sequence is obtained and denoted as the window adjustment reference value. The window adjustment feature value of the j-th fused data is compared with the window adjustment reference value to obtain the adaptive smoothing window of the j-th fused data. The calculation formula of the adaptive smoothing window is as follows:

[0088]

[0089] in, This represents the size of the adaptive smoothing window for the j-th fused data in any stage of the data sequence. This represents the size of the initial smoothing window corresponding to any stage of the data sequence, where 1 represents a constant. This represents the window adjustment feature value of the j-th fused data in any stage of the data sequence. This indicates the window adjustment reference value; | represents the absolute value sign. The symbol indicates rounding up.

[0090] Similarly, by obtaining the adaptive smoothing window of each fused data in any stage of the data sequence, we can obtain the adaptive smoothing window of each fused data in each stage of the data sequence, which is equivalent to obtaining the adaptive smoothing window of each fused data in the fused data sequence.

[0091] Step S104: Based on the adaptive smoothing window of each fused data in each stage data sequence, smooth the fused data sequence to obtain the smoothed data sequence corresponding to any monitoring data sequence set; use the smoothed data sequences corresponding to all monitoring data sequence sets as input to the CNN network, and output the predicted value of lipid oxidation stability.

[0092] Based on the adaptive smoothing window of each fused data in each stage data sequence, the fused data sequence is smoothed using a moving average algorithm to obtain a smoothed data sequence corresponding to any set of monitoring data sequences, which is also the smoothed data sequence corresponding to the POV value. The general method for smoothing the fused data sequence is as follows: the moving average algorithm is used to smooth each stage data sequence separately to obtain the corresponding smoothed sequence, and all smoothed sequences are combined into a smoothed data sequence according to time sequence. It should be noted that the moving average algorithm is an existing technology and will not be described in detail here.

[0093] Following the method for obtaining smoothed data sequences corresponding to any set of monitoring data sequences, smoothed data sequences corresponding to each set of monitoring data sequences are obtained. This results in a complete set of smoothed data sequences under different monitoring indicators during the accelerated lipid oxidation experiment. Then, the smoothed data sequences corresponding to all sets of monitoring data sequences are used as input to a CNN network, which outputs a predicted value for lipid oxidation stability. Here, the CNN network refers to a trained neural network used to obtain the predicted value for lipid oxidation stability. The prediction process of the CNN network is roughly as follows:

[0094] The shape of the input layer is determined based on the sequence length of the smoothed data sequence and the dimensions of the monitoring data. For example, if the sequence length of the smoothed data sequence is T and there are n dimensions of monitoring data, then the shape of the input layer is (batch_size, T, n). An appropriate convolutional kernel (such as (3,1) or (5,1)) is selected to extract local features in the time series. Multiple convolutional layers are stacked, and max pooling or average pooling methods are selected to reduce the dimension of the feature map, reduce the amount of computation, and enhance the translation invariance of the model. An appropriate pooling window size is selected based on the size of the feature map output by the convolutional layer, which is generally smaller than the size of the convolutional kernel. The outputs of the convolutional and pooling layers are flattened to convert the multidimensional feature map into a one-dimensional vector. Finally, multiple fully connected layers are used to further extract features for regression, that is, the last fully connected layer outputs a continuous value to represent the stability of the oil.

[0095] It should be noted that using CNN networks to predict the stability of lipid oxidation is an existing technology, and will not be elaborated on here.

[0096] Based on the same inventive concept as the above methods, this embodiment of the invention also provides a deep learning-based lipid oxidation stability prediction system, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described deep learning-based lipid oxidation stability prediction methods.

[0097] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A deep learning-based method for predicting the oxidative stability of lipids, characterized in that, The method includes: Acquire multidimensional monitoring data sequences corresponding to different complete accelerated oil oxidation experiments at the same time, and divide the multidimensional monitoring data sequences into multiple monitoring data sequence sets according to the dimensions; For any set of monitoring data sequences, data fusion is performed on the monitoring data of all monitoring data sequences in the set at the same time to obtain a fused data sequence; a monitoring data change curve is constructed for each monitoring data sequence, and the change trend of each monitoring data change curve is combined to obtain two stage division times. Using the two stage division times, the fused data sequence is divided into three stage data sequences. For any stage of the data sequence, based on the data change trend and data volume of the data sequence, the corresponding initial smoothing window is obtained, and the initial smoothing window corresponding to each fused data in the data sequence is adaptively adjusted to obtain an adaptive smoothing window. Based on the adaptive smoothing window of each fused data in each stage data sequence, the fused data sequence is smoothed to obtain the smoothed data sequence corresponding to any monitoring data sequence set; the smoothed data sequences corresponding to all monitoring data sequence sets are used as input to the CNN network, and the corresponding output is the predicted value of lipid oxidation stability.

2. The deep learning-based method for predicting the oxidation stability of lipids according to claim 1, characterized in that, The step of fusing the monitoring data of all monitoring data sequences in any given set at the same time to obtain a fused data sequence includes: For the i-th monitoring data in any monitoring data sequence in any monitoring data sequence set, calculate the absolute value of the difference between the i-th monitoring data and the (i-1)-th monitoring data in any monitoring data sequence, denoted as the previous absolute value of the difference; calculate the absolute value of the difference between the i-th monitoring data and the (i+1)-th monitoring data, denoted as the subsequent absolute value of the difference; normalize the reciprocal of the absolute values ​​of the differences between the previous and subsequent absolute values ​​of the difference to obtain the similarity of the differences before and after. Calculate the first ratio between the absolute value of the difference before and the (i-1)th monitoring data, calculate the second ratio between the absolute value of the difference after and the (i+1)th monitoring data, calculate the absolute value of the difference between the first ratio and the second ratio, and normalize the sum of the absolute value of the difference and the constant 1 to obtain the degree of normal variation. Calculate the difference coefficient between the i-th monitoring data, the (i-1)-th monitoring data, and the (i+1)-th monitoring data, and denote it as the first difference coefficient. Calculate the difference coefficient of the data sequence of a preset length with the (i+1)-th monitoring data as the cutoff data, and denote it as the second difference coefficient. Normalize the absolute value of the difference between the first difference coefficient and the second difference coefficient to obtain the difference coefficient similarity of the i-th monitoring data. The monitoring reliability of the i-th monitoring data is obtained by calculating the mean of the similarity between the before and after differences, the degree of normal change, and the similarity of the difference coefficient. For any given time, based on the monitoring reliability of the monitoring data at any given time in each monitoring data sequence, obtain the weight of the monitoring data at any given time in each monitoring data sequence. The greater the monitoring reliability, the greater the weight. Perform a weighted summation on the monitoring data at any given time in all monitoring data sequences to obtain the fused data at any given time. Obtain the fused data at each time to obtain a fused data sequence.

3. The deep learning-based method for predicting the oxidation stability of lipids according to claim 1, characterized in that, The changing trends of each monitoring data curve are analyzed to obtain the two-stage division time points, including: For any monitoring data change curve, based on the fitting slope of the data points before and after each data point in the curve, the probability of each data point being a stage point is obtained, resulting in a probability sequence; local extreme points in the probability sequence are obtained, and effective extreme points are selected from the local extreme points; For any valid extreme point, calculate the slope between the valid extreme point and the data points before and after it, and obtain the absolute value of the slope difference. The product of the absolute value of the slope difference and the probability corresponding to the valid extreme point is recorded as the stage feature value of the valid extreme point. Based on the stage feature value of each valid extreme point, the two valid extreme points with the largest stage feature values ​​are obtained as the suspected stage points of the monitoring data change curve. Obtain the suspected stage point of each monitoring data change curve, calculate the mean of the time corresponding to the first suspected stage point of all monitoring data change curves, and record it as the first stage division time. Calculate the mean of the time corresponding to the second suspected stage point of all monitoring data change curves, and record it as the second stage division time.

4. The deep learning-based method for predicting the oxidation stability of lipids according to claim 3, characterized in that, The step of obtaining the probability that each data point is a stage point based on the fitting slope of the data points before and after each data point in the change curve of any monitoring data includes: ； in, This represents the probability that the j-th data point in any of the monitoring data change curves is a stage point. This represents the first weight, and 1 represents a constant. Represents the normalization function. This represents the slope of the fit between the j-th data point and the first m data points. This represents the fitting slope of the first m data points of the j-th data point. This represents the fitted slope of the m data points following the j-th data point. Let || denote the fitted slope of the m data points following the j-th data point, where || denotes the absolute value sign. This indicates the second weight.

5. The deep learning-based method for predicting the oxidation stability of lipids according to claim 3, characterized in that, The process of selecting effective extreme points from local extreme points includes: For any local extreme point, the average value of a preset number of probabilities that are closest to the local extreme point is obtained in the probability sequence. The sum of the average value and the preset probability allowable fluctuation value is used as the local maximum threshold. If the probability corresponding to any local extreme point is greater than the local maximum threshold, the local extreme point is regarded as a valid local maximum point. The difference between the average value and the preset probability allowable fluctuation value is used as the local minimum threshold. If the probability corresponding to any local extreme point is less than the local minimum threshold, the local extreme point is used as a valid local minimum point. All effective local minima and effective local maxima are combined to form effective extreme points.

6. The deep learning-based method for predicting the oxidation stability of lipids according to claim 1, characterized in that, The step of obtaining the corresponding initial smoothing window based on the data change trend and data volume of the data sequence at any stage includes: Calculate the slope of change between every two fused data points in any stage of the data sequence to obtain the average slope of change. Normalize the average slope of change to obtain a data change trend index. A polynomial fit is performed on the data sequence of any stage to obtain the corresponding fitting curve. The first derivative of the fitting curve is calculated to obtain the number of data with negative derivatives and the number of data with positive derivatives. The absolute value of the difference between the number of data with negative derivatives and the number of data with positive derivatives is normalized to obtain a normalized value. The difference between the constant 1 and the normalized value is calculated and denoted as the fluctuation characteristic value. Calculate the absolute value of the difference between every two adjacent fused data in any stage of the data sequence to obtain the average absolute value of the difference. Normalize the average absolute value of the difference to obtain the average fluctuation amplitude. Use the mean of the fluctuation characteristic value and the average fluctuation amplitude as the data fluctuation index. The data change trend index and the data fluctuation index are weighted and summed to obtain a weighted summation result. The difference between the constant 1 and the weighted summation result is used as the window adjustment coefficient of the data sequence in any stage. The length of the data sequence in any stage is counted as the maximum window length. The product between the maximum window length and the window adjustment coefficient is rounded up to obtain the initial smoothing window corresponding to the data sequence in any stage.

7. The deep learning-based method for predicting the oxidation stability of lipids according to claim 1, characterized in that, The adaptive adjustment of the initial smoothing window corresponding to each fused data in the data sequence of any stage to obtain an adaptive smoothing window includes: For any fused data in any stage of the data sequence, the data type bias of any fused data is obtained based on the differences between adjacent data within the initial smoothing window of any monitoring data; volatility analysis is performed on the data within the initial smoothing window of any fused data to obtain the window trend volatility of any fused data; the data type bias and the window trend volatility are weighted and summed to obtain the window adjustment feature value of any fused data. The mean value of the window adjustment feature value of the normal fused data in any stage data sequence is obtained and denoted as the window adjustment reference value. The window adjustment feature value of any fused data is compared with the window adjustment reference value to obtain the adaptive smoothing window of the fused data. The calculation formula of the adaptive smoothing window is as follows: ； in, This represents the size of the adaptive smoothing window for the j-th fused data in any stage of the data sequence. This represents the size of the initial smoothing window corresponding to any stage of the data sequence, where 1 represents a constant. This represents the window adjustment feature value of the j-th fused data in any stage of the data sequence. This indicates the window adjustment reference value; | represents the absolute value sign. The symbol indicates rounding up.

8. The deep learning-based method for predicting the oxidation stability of lipids according to claim 7, characterized in that, The step of obtaining the data type bias of any fused data based on the differences between adjacent data within the initial smoothing window of any fused data includes: Calculate the average data value within the initial smoothing window of any fused data, obtain the absolute value of the difference between any fused data and the average data value, and normalize the absolute value of the difference to obtain the local salient value; Calculate the slope of change between any fused data and its previous fused data, denoted as the first slope. Calculate the slope of change between the two fused data of the previous fused data, denoted as the second slope. Normalize the absolute value of the difference between the first slope and the second slope to obtain the trend difference degree. Use the difference between the constant 1 and the trend difference degree as the data smoothness. The mean between the local spur value and the data smoothness is calculated to obtain the data type bias of any fused data.

9. The deep learning-based method for predicting the oxidation stability of lipids according to claim 7, characterized in that, The step of performing volatility analysis on the data within the initial smoothing window of any fused data to obtain the window trend volatility of any fused data includes: Calculate the data variance within the initial smoothing window of any fused data, normalize the data variance to obtain a first normalized value; perform linear fitting on the data within the initial smoothing window of any fused data to obtain a fitting slope, normalize the absolute value of the fitting slope to obtain a second normalized value; calculate the mean of the first normalized value and the second normalized value to obtain the window trend fluctuation degree of any fused data.

10. A deep learning-based system for predicting the oxidative stability of lipids, 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 deep learning-based method for predicting the oxidation stability of oils as described in any one of claims 1-9.

Images