A method and system for online monitoring of production of diglyceride oil based on dielectric properties

Abstract

The present application relates to the technical field of edible oil processing, and specifically provides an online monitoring method and system for glyceride oil production based on dielectric properties, which comprises the following steps: collecting dielectric response time series signals at at least two different frequencies through a double-channel differential dielectric sensor installed on a glyceride oil production pipeline, and constructing a two-dimensional dielectric relaxation spectrum; inputting the two-dimensional dielectric relaxation spectrum into a one-dimensional convolutional neural network, and extracting local time series features through sliding convolution operation of a convolution kernel in the time dimension, wherein the convolutional neural network realizes time translation invariance through a weight sharing mechanism; based on the extracted local time series features, outputting real-time glyceride content estimates through a fully connected layer; and comparing the glyceride content estimates with a dynamic adaptive threshold, and generating and outputting monitoring information according to the comparison result. The present application realizes high-precision and fast-response industrial real-time online detection.

Description

Technical Field

[0001] This invention relates to the field of edible oil processing technology, specifically to an online monitoring method and system for diglyceride oil production based on dielectric properties. Background Technology

[0002] In the production of diglyceride oil, diglyceride content is a core indicator determining product quality. Traditional detection methods mainly rely on chromatography, which, while offering high accuracy, require offline sampling, involve cumbersome sample pretreatment, and take over 30 minutes per test. This results in significant latency, making it impossible to track production status in real time and failing to meet the demands of continuous industrial production for stable and controllable product quality.

[0003] Oil detection technology based on dielectric properties has been applied to the rapid detection of polar components in frying oils. Its principle is based on the property that polar substances undergo polarization in an electric field, causing a change in the dielectric constant. The quality of the oil is quickly determined by measuring the dielectric constant. However, existing technologies have significant shortcomings in monitoring diglyceride oil production: First, they only use a single frequency for detection, while the dielectric signal of the oil is simultaneously affected by multiple components such as diglycerides, moisture, and free fatty acids. Moisture increases the dielectric constant, while free fatty acids decrease it. A single frequency only yields a mixed superposition result, failing to accurately distinguish the contribution of each component. This results in poor targeting of diglyceride content detection and susceptibility to interference. Second, they rely on linear calibration models, assuming a simple linear relationship between the dielectric constant and the content of polar substances. However, in actual production, when the diglyceride content is in the high concentration range of 40% to 95%, the dielectric response exhibits a significant nonlinear relationship with the content. Temperature fluctuations further amplify this nonlinearity, leading to large fitting errors and low detection accuracy in traditional linear models, failing to meet the requirements of industrial online monitoring. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides an online monitoring method and system for diglyceride oil production based on dielectric properties. By installing a dual-channel differential dielectric sensor in the production pipeline, continuous online monitoring of diglyceride content without sampling is achieved. Combining multi-frequency two-dimensional dielectric relaxation spectroscopy enhances detection specificity and accuracy. Furthermore, a one-dimensional convolutional neural network is used to efficiently fit nonlinear relationships and optimize model parameters, achieving high-precision and fast-response industrial real-time online detection.

[0005] In a first aspect, the technical solution of the present invention provides an online monitoring method for the production of diglyceride oil based on dielectric properties, comprising the following steps: By using a dual-channel differential dielectric sensor installed on the diglyceride oil production pipeline, dielectric response time-series signals at at least two different frequencies are collected to construct a two-dimensional dielectric relaxation spectrum. Two-dimensional dielectric relaxation spectra are input into a one-dimensional convolutional neural network. Through the sliding convolution operation of the convolution kernel in the time dimension, local temporal features reflecting the polarization relaxation process of diglyceride molecules are extracted. The convolutional neural network achieves time translation invariance through a weight sharing mechanism, enabling the same convolution kernel to identify the same local temporal pattern appearing at any position on the time axis. The dielectric response features at different frequency channels are extracted through parallel processing of multiple convolution kernels. Based on the extracted local temporal features, the real-time diglyceride content estimate is output through direct regression via a fully connected layer; The estimated diglyceride content is compared with a dynamic adaptive threshold, and monitoring information is generated and output based on the comparison result.

[0006] Secondly, the technical solution of the present invention provides an online monitoring system for the production of diglyceride oil based on dielectric properties, comprising: The data acquisition and processing module is used to acquire dielectric response time-series signals at at least two different frequencies through a dual-channel differential dielectric sensor installed on the diglyceride oil production pipeline, and construct a two-dimensional dielectric relaxation spectrum. The diglyceride content estimation module is used to input a two-dimensional dielectric relaxation spectrum into a one-dimensional convolutional neural network. Through the sliding convolution operation of the convolution kernel in the time dimension, local temporal features reflecting the polarization relaxation process of diglyceride molecules are extracted. Based on the extracted local temporal features, the real-time diglyceride content estimate is output through direct regression by a fully connected layer. The convolutional neural network achieves time translation invariance through a weight sharing mechanism, enabling the same convolution kernel to identify the same local temporal pattern appearing at any position on the time axis. The dielectric response features at different frequency channels are extracted through parallel processing of multiple convolution kernels. The monitoring information generation module is used to compare the estimated diglyceride content with a dynamic adaptive threshold, and generate and output monitoring information based on the comparison result.

[0007] As can be seen from the above technical solutions, this application has the following advantages: First, by directly installing the dual-channel differential dielectric sensor on the production pipeline, no sampling or preprocessing is required throughout the process. It can output diglyceride content data in real time, enabling continuous online monitoring of the production process and rapidly reflecting dynamic changes in quality. Second, by collecting dielectric response time-series signals at multiple frequencies, a two-dimensional dielectric relaxation spectrum containing both frequency and time dimensions is constructed. Utilizing the differences in response of various polar substances at different frequencies, more comprehensive dielectric characteristic information is obtained. Compared to single-frequency detection, this significantly reduces interference and improves the specificity and accuracy of diglyceride detection. Third, by using a one-dimensional convolutional neural network to process the dielectric spectrum, it can automatically learn and fit the nonlinear relationship between dielectric response and diglyceride content, maintaining high-precision content estimation even in high-concentration ranges, thus solving the problem of insufficient accuracy in traditional linear models. Simultaneously, the network adopts a weight-sharing mechanism, significantly reducing model parameters while ensuring the effectiveness of time-series feature extraction. This allows for rapid operation on embedded terminals, fully meeting the speed requirements of real-time monitoring in industrial production lines. Attached Figure Description

[0008] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of an online monitoring method for the production of diglyceride oil based on dielectric properties, provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic block diagram of an online monitoring system for diglyceride oil production based on dielectric properties, provided as an embodiment of the present invention. Detailed Implementation

[0011] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0013] Figure 1 This is a schematic flowchart of an online monitoring method for diglyceride oil production based on dielectric properties, provided by an embodiment of the present invention. Figure 1 The executing entity can be an online monitoring system for diglyceride oil production based on dielectric properties. The online monitoring method for diglyceride oil production based on dielectric properties provided in this embodiment of the invention is executed by computer equipment; correspondingly, the online monitoring system for diglyceride oil production based on dielectric properties runs on the computer equipment. Depending on different requirements, the order of the steps in this flowchart can be changed, and some steps can be omitted.

[0014] like Figure 1 As shown, the method includes the following steps.

[0015] S1, by using a dual-channel differential dielectric sensor installed on the diglyceride oil production pipeline, collects dielectric response timing signals at at least two different frequencies to construct a two-dimensional dielectric relaxation spectrum.

[0016] S2, the two-dimensional dielectric relaxation spectrum is input into a one-dimensional convolutional neural network. Through the sliding convolution operation of the convolution kernel in the time dimension, local temporal features reflecting the polarization relaxation process of diglyceride molecules are extracted. Among them, the convolutional neural network achieves time translation invariance through a weight sharing mechanism, so that the same convolution kernel can identify the same local temporal pattern appearing at any position on the time axis, and the dielectric response features under different frequency channels are extracted through parallel processing of multiple convolution kernels.

[0017] S3, based on the extracted local time-series features, directly regresses through a fully connected layer to output a real-time estimate of diglyceride content.

[0018] S4, compare the estimated diglyceride content with the dynamic adaptive threshold, and generate and output monitoring information based on the comparison result.

[0019] As a refinement and extension of the specific implementation of the above embodiments, in order to fully explain the specific implementation process of this embodiment, the following will provide possible embodiments to describe the specific implementation of the above steps in a non-limiting manner.

[0020] In some optional embodiments, the dual-channel differential dielectric sensor used in step S1 includes a measuring channel exposed to the oil being tested and a reference channel sealed within the high-purity triglyceride oil. Specifically, a dual-channel differential dielectric sensor is installed on the discharge pipe of the enzymatic hydrolysis reactor in the diglyceride oil production line. The sensor has a concentric cylindrical structure and is fixed to the pipe using a flange connection, facilitating disassembly and cleaning.

[0021] The dual-channel differential dielectric sensor comprises two independent detection channels: a measurement channel and a reference channel. The measurement channel is exposed to the tested oil, with the oil flowing through the electrode gap to acquire the dielectric response signal. The reference channel is sealed and filled with high-purity triglyceride oil (purity not less than 98%), serving as a reference for the dielectric response. The reference channel is in the same temperature field as the tested oil but does not come into contact with it. The electrode surfaces of both channels are coated with a polytetrafluoroethylene (PTFE) insulating layer to ensure that the electrodes do not directly contact the oil, meeting food-grade production hygiene requirements.

[0022] In some optional implementations, step S1 involves acquiring dielectric response timing signals at at least two different frequencies to construct a two-dimensional dielectric relaxation spectrum, specifically including the following steps S1.1 to S1.4.

[0023] S1.1 Select N discrete frequency points at logarithmic intervals within a preset frequency range.

[0024] A frequency sweep excitation method is used to apply a sinusoidal AC excitation signal to the oil within a preset frequency range. Preferably, the preset frequency range is 10kHz to 1MHz, and N discrete frequency points are selected at logarithmic intervals.

[0025] S1.2, at each frequency point, an excitation signal is first applied and a first preset time is waited for the oil polarization to reach a steady state. Then, dielectric response signals at M time points are continuously acquired at a preset sampling frequency. The dielectric response signals include the capacitance values ​​acquired by each of the two channels.

[0026] For each discrete frequency point , and execute the operation.

[0027] The application frequency is A sinusoidal AC excitation signal is applied for 0.3 seconds. This stage is used to overcome electrode polarization effects and the initial setup process of dielectric relaxation, ensuring that the dipoles of oil molecules reach a stable orientation polarization state in the applied electric field. No data is collected during the excitation setup period.

[0028] After the excitation setup phase ends, the steady-state continuous sampling phase begins. The dielectric response signal at M time points is continuously acquired at a preset sampling frequency, where M is an even number. For example, M is 50, meaning 0.5 seconds of continuous data is acquired at each frequency point.

[0029] At each sampling time Synchronously execute the acquisition and measurement of capacitance values ​​of the channel. and the capacitance value of the reference channel. .

[0030] S1.3 For each sampling time, calculate the capacitance ratio between the measurement channel and the reference channel, and convert the capacitance ratio into a differential dielectric constant based on the dielectric properties of the high-purity triglyceride oil in the reference channel.

[0031] First, the capacitance ratio of the measurement channel to the reference channel is calculated, which eliminates common-mode interference from temperature drift and circuit drift:

[0032] Then, based on the known dielectric properties of the high-purity triglyceride oil in the reference channel, the capacitance ratio is converted into a differential dielectric constant:

[0033] in, The dielectric constant of the high-purity triglyceride oil in the reference channel at the current temperature T is obtained by looking up a pre-generated temperature-dielectric constant calibration table.

[0034] Complete the current frequency After collecting data, switch to the next frequency. Repeat steps S1.2 and S1.3 above until all N frequency points have been collected.

[0035] S1.4 Arrange the M differential dielectric constant values ​​continuously collected at the same frequency point in chronological order to form a dielectric response time series vector at that frequency point. Stack the time series vectors at N frequency points in chronological order to form an N-row, M-column two-dimensional matrix. This two-dimensional matrix is ​​the two-dimensional dielectric relaxation spectrum.

[0036] Same frequency M consecutively collected The values ​​are arranged in chronological order to form a time series vector of the dielectric response at that frequency:

[0037] The vector, with a length of M, reflects the evolution of the dielectric response of the oil over time under specific frequency excitation.

[0038] Time series vectors at N frequencies Stacked in frequency order, they form an N-row, M-column two-dimensional dielectric relaxation spectrum matrix:

[0039] This two-dimensional matrix is ​​a two-dimensional dielectric relaxation spectrum. The rows correspond to the frequency dimension, reflecting the variation of the dielectric constant with frequency; the columns correspond to the time dimension, reflecting the evolution of the dielectric constant over time.

[0040] In some optional implementations, while continuously acquiring dielectric response signals at M time points at a preset sampling frequency, the oil temperature is also simultaneously acquired at each sampling moment. Specifically, a PT100 platinum resistance temperature sensor is integrated into the dual-channel differential dielectric sensor probe for real-time acquisition of oil temperature.

[0041] Furthermore, the constructed two-dimensional dielectric relaxation spectrum is preprocessed before being input into the one-dimensional convolutional neural network to eliminate the influence of dimensional differences on subsequent neural network processing. Specifically, this includes the following steps S1.5 to S1.6.

[0042] S1.5, based on the synchronously collected oil temperature, corrects the differential dielectric constant at each time and frequency to the reference temperature.

[0043] Based on the synchronously collected temperature data The differential dielectric constant at each time and frequency is corrected to the reference temperature. Temperature correction employs a linear correction model, and the correction coefficients were determined through preliminary experiments. This is expressed as:

[0044] in, is the dielectric temperature coefficient.

[0045] S1.6, the temperature-corrected two-dimensional dielectric relaxation spectrum is normalized according to the frequency dimension.

[0046] The temperature-corrected two-dimensional dielectric relaxation spectrum is then normalized along its frequency dimension. Specifically, for each frequency... The following time series Calculate its mean and standard deviation, then subtract the mean from each element and divide by the standard deviation to make the dielectric constant sequence at each frequency after transformation have zero mean and unit variance.

[0047] In this embodiment, steps S2 and S3 use a one-dimensional convolutional neural network to output a real-time estimate of diglyceride content for subsequent estimation of production status. In some optional embodiments, a lightweight one-dimensional convolutional neural network model for processing two-dimensional dielectric relaxation spectra is provided. This model is designed for online monitoring of diglyceride oil production and can run in real time on embedded devices.

[0048] The one-dimensional convolutional neural network adopts an end-to-end regression architecture. The input is a preprocessed two-dimensional dielectric relaxation spectrum, and the output is the predicted value of diglyceride content. The overall network architecture consists of the following layers in sequence: input layer, first convolutional layer, first pooling layer, second convolutional layer, second pooling layer, flattening layer, and fully connected layer.

[0049] Both the first and second convolutional layers employ one-dimensional convolution operations, with the convolution kernel sliding only in the time dimension while maintaining the frequency dimension. This sliding convolution operation in the time dimension is designed based on the following physical understanding: in a dielectric relaxation spectrum, the frequency dimension reflects the amplitude distribution of the dielectric response at different frequencies, while the time dimension reflects the dynamic evolution of the polarization relaxation process. The convolution operation in this invention slides only in the time dimension to extract the dynamic characteristics of the polarization relaxation process, while simultaneously capturing the local correlation of the dielectric response at different frequency channels through a weight-sharing mechanism.

[0050] 1) Input layer The input data is a preprocessed two-dimensional dielectric relaxation spectrum, represented as a matrix. Where N is the number of frequency channels (i.e., the number of frequency sampling points), and M is the length of the time series (i.e., the number of sampling points at each frequency).

[0051] The input layer does not perform any transformation on the data; it simply organizes the two-dimensional matrix into N one-dimensional time series by row, with each time series corresponding to a frequency channel.

[0052] 2) First convolutional layer The first convolutional layer uses the following configuration: number of convolutional kernels K1 = 3, kernel size kernel_size1 = 3, stride1 = 1, padding = "same", that is, zeros are padded at both ends of the time series to keep the length unchanged before and after convolution, and the activation function is ReLU (corrected linear unit).

[0053] The calculation formula for convolution operation is as follows. For the k-th convolution kernel (k=1,2,3), its output feature sequence on the i-th frequency channel is... The j-th element is:

[0054] in, Given the element in the i-th row and j-th column of the input matrix, For the p-th weight parameter of the k-th convolutional kernel, For the bias term of the k-th convolution kernel, This is the ReLU activation function.

[0055] Because "same" padding is used, the length of the output feature sequence remains M. The outputs of the N frequency channels are stacked along the channel dimension to obtain the output tensor of the first convolutional module. .

[0056] 3) First pooling layer The first pooling module uses max pooling, configured as follows: pooling window size pool_size1 = 2, stride1 = 2. Pooling is performed along the time dimension, taking the maximum value between two adjacent time points of the feature sequence output by each convolutional kernel, thus halving the length of the time dimension. The formula for calculating the pooling operation is:

[0057] in, .

[0058] The output tensor dimension of the first pooling layer is: the number of frequency channels N remains unchanged, and the length of the time dimension becomes M / 2.

[0059] 4) Second convolutional layer The second convolutional layer is configured as follows: number of kernels K2 = 6, kernel size kernel_size2 = 3, stride2 = 1, padding = "same", and activation function ReLU (corrected linear unit).

[0060] This convolutional layer further extracts features from the output of the first pooling layer. Unlike the first convolutional layer, the input of the second convolutional layer already has multiple feature channels (K1=3 channels), so the dimension of the convolutional kernel is expanded to three dimensions: frequency channel dimension × kernel size × number of input channels.

[0061] Specifically, for the q-th convolutional kernel (q=1,…,K2), the j-th element of its output feature sequence is:

[0062] in, For the p-th weight parameter of the q-th convolutional kernel in the k-th input channel of the second convolutional layer, This is the bias term for the q-th convolutional kernel.

[0063] The output tensor dimension of the second convolutional layer is: K2×N×M / 2.

[0064] 5) Second pooling layer The second pooling layer employs global average pooling. Unlike the first pooling layer, global average pooling takes the average value over the entire time dimension, compressing the entire time series output by each frequency channel and each convolutional kernel into a single scalar value.

[0065] The formula for calculating global average pooling is: .

[0066] in, This represents the global average pooling result for the i-th frequency channel and the q-th convolutional kernel.

[0067] The output of the second pooling module is a two-dimensional matrix. .

[0068] 6) Flattening layer The flattening layer transforms the two-dimensional matrix G output by the second pooling layer into a one-dimensional feature vector. .

[0069] 7) Fully connected layer The fully connected layer maps the flattened one-dimensional feature vector to a single output value, namely the predicted diglyceride content. The calculation formula for the fully connected layer is:

[0070] in, The predicted diglyceride content output by the model (in %). Let r be the r-th element of the flattened eigenvector. Let r be the weight parameter of the fully connected layer. This is the bias term for the fully connected layer.

[0071] The one-dimensional convolutional neural network is trained using supervised learning. The training dataset is obtained by collecting dielectric response data on the production line under different operating conditions (DAG content range of 40%~95%, temperature range of 40℃~80℃), and simultaneously collecting oil samples and determining the DAG content using high performance liquid chromatography as the label value.

[0072] The collected samples were divided into training, validation, and test sets in an 8:1:1 ratio. In this embodiment, a total of 1000 valid data sets were collected, with 800 sets for training, 100 sets for validation, and 100 sets for test. Mean squared error was used as the loss function. The Adam optimizer was used for parameter updates, with a learning rate of 0.001 and a batch size of 32. Early stopping was employed during training; training was stopped when the validation set loss no longer decreased after 20 consecutive training iterations. After training, the model parameters (weights and biases of each layer) were saved as floating-point numbers and burned into the microcontroller of the field device for online inference.

[0073] The process of outputting real-time diglyceride content estimates based on a trained one-dimensional convolutional neural network includes the following steps.

[0074] Step A: Receive the input two-dimensional dielectric relaxation spectrum, which is represented as an N-row, M-column two-dimensional matrix, where N is the number of frequency channels and M is the time series length.

[0075] Step B: The first convolutional layer performs a one-dimensional convolution operation on the input two-dimensional dielectric relaxation spectrum. The input matrix is ​​organized into N one-dimensional time series by rows. K1 one-dimensional convolution kernels are used to perform sliding convolution operations in the time dimension. Zero padding is used to keep the length of the time dimension unchanged. The same weight parameters are used for the same convolution kernel at different time positions through a weight sharing mechanism to extract local temporal features that are independent of time position. The output of each convolution kernel is processed by the ReLU activation function to form a feature map. A total of K1 feature maps are output, and each feature map still maintains the dimension of N rows and M columns.

[0076] Step C: The first pooling layer performs max pooling on the output of the first convolutional layer.

[0077] Step D: The second convolutional layer performs a one-dimensional convolution operation on the output of the first pooling layer: the K1 feature maps output by the first pooling operation are used as multi-channel inputs, and K2 one-dimensional convolutional kernels are used to perform sliding convolution operations in the time dimension. Zero padding is used at the boundaries. Each convolutional kernel processes all K1 input channels at the same time, and local correlation features across channels are extracted through a weight sharing mechanism. The output of each convolutional kernel is processed by the ReLU activation function to form a feature map. A total of K2 feature maps are output, and the dimension of each feature map is N rows (M / 2) columns.

[0078] Step E: The second pooling layer performs a global average pooling operation on the output of the second convolutional layer, compressing each feature map into a single value in the time dimension, resulting in several feature values.

[0079] Step F: The flattening layer arranges the feature values ​​in order into a one-dimensional feature vector.

[0080] In step G, the fully connected layer maps the one-dimensional feature vector to a single output value, namely the diglyceride content estimate.

[0081] In this embodiment, the weight sharing mechanism is configured such that, for any convolution kernel, when it slides on the time series, the convolution calculation at each sliding position uses the same set of weight parameters of the convolution kernel, which does not change with the sliding position.

[0082] Specifically, for any convolution kernel in a one-dimensional convolutional neural network, when it slides across the time series, the convolution calculation at each sliding position uses the same set of weight parameters of the convolution kernel, which does not change with the sliding position.

[0083] Let the weight parameter sequence of a convolution kernel be... Where K is the kernel size (K=3 in this embodiment). This kernel operates on a time series of length M. When sliding upwards, at the j-th sliding position The output characteristic value h(j) is calculated according to the following formula:

[0084] Where b is the bias term and σ is the ReLU activation function.

[0085] In the above calculation, regardless of the value of j, the weight parameters w1, w2, w3 and the bias b remain unchanged. That is, the same convolution kernel uses the exact same parameters to perform convolution operations at different positions in the time series, which is called "weight sharing".

[0086] In the one-dimensional convolutional neural network of this embodiment, both the first and second convolutional layers employ the aforementioned weight-sharing mechanism. Taking the first convolutional layer as an example, the input consists of 20 frequency channels, each of which is a time series of length 50. For the k-th convolutional kernel (k=1,2,3), its weight parameters are as follows: , bias is As the convolutional kernel slides across the time series of a specific frequency channel, it sequentially calculates the output value at each sliding position: At sliding position 1 (corresponding time point) ): ; At sliding position 2 (corresponding time point) ): ; At sliding position j (corresponding time point) ): .

[0087] As can be seen, the weight parameters and bias It is the same in all sliding positions.

[0088] The weight sharing mechanism of the second convolutional layer is the same as that of the first convolutional layer. The difference is that each convolutional kernel of the second convolutional layer needs to process multiple input channels at the same time. Its convolution calculation at time position j also uses the same set of weight parameters, which do not change with the change of j.

[0089] The weight-sharing mechanism achieves time-shift invariance. Specifically, in the dielectric relaxation spectrum, the polarization relaxation process of diglyceride molecules manifests as a local waveform where the dielectric constant changes over time. The morphological characteristics of this waveform, such as the rising slope, peak height, and decay time constant, are determined by the physicochemical properties of the diglyceride molecule and are independent of its specific position on the time axis. Because the weight-sharing mechanism ensures that the weight parameters of the convolution kernel do not change with the sliding position, the convolution kernel actually learns a "position-independent local waveform pattern." Specifically: when the dielectric relaxation waveform appears in the early part of the time axis (e.g., ... When the convolution kernel produces a high-response output at sliding position 1, the dielectric relaxation waveform of the same shape appears in the middle of the time axis (e.g., When the convolution kernel is at the sliding position 25, it also produces a high response output.

[0090] Therefore, regardless of which stage of the sampling time window the polarization relaxation process occurs at, the same convolutional kernel can identify and extract the local temporal features of the process, demonstrating that the weight sharing mechanism achieves time translation invariance.

[0091] By adopting a weight-sharing mechanism, the total number of parameters in the one-dimensional convolutional neural network is small, and they can be stored in the microcontroller's internal flash memory in the form of floating-point numbers. During inference calculation, the microcontroller executes the operations of each layer in sequence according to the above convolution calculation formula, without the need for frequent reading and writing of external memory. The single inference time is short, which meets the real-time requirements of online monitoring of the production line.

[0092] In this embodiment, the first convolutional layer has three convolutional kernels, each independently learning a different type of local temporal pattern. The first convolutional kernel, after training, has weight parameters in a [positive, positive, negative] pattern to identify the polarization establishment process where the dielectric constant rises rapidly and then stabilizes. The second convolutional kernel, after training, has weight parameters in a [negative, negative, positive] pattern to identify the relaxation process where the dielectric constant decays slowly. The third convolutional kernel, after training, has weight parameters in a [positive, negative, positive] pattern to identify the steady-state response of dielectric constant fluctuations. These three convolutional kernels process the input time series in parallel, with each kernel sliding independently and calculating the output, collectively covering multiple local temporal patterns in the dielectric relaxation spectrum.

[0093] The second convolutional layer uses six convolutional kernels, each processing three feature maps output from the first convolutional layer simultaneously. The weight parameters of the q-th convolutional kernel (q=1,…,6) are a 3×3 matrix, where the row index corresponds to the input channel (three feature maps) and the column index corresponds to the time position (three time points). As this convolutional kernel slides along the time dimension, it simultaneously performs a weighted summation on the corresponding time windows of each input channel, outputting a scalar value. Through the parallel processing of the six convolutional kernels, the second convolutional layer can extract cross-channel correlation features between different frequency channels, such as the combined features when the rising pattern of the low-frequency channel and the stationary pattern of the high-frequency channel occur simultaneously.

[0094] In this embodiment, step S4 compares the estimated diglyceride content with a dynamic adaptive threshold and generates monitoring information based on the comparison result.

[0095] Dynamic adaptive thresholds refer to benchmark values ​​that are dynamically adjusted based on real-time production process status and historical data, adapting to changes in production conditions such as batch variations in raw materials, enzyme activity decay, and equipment drift. Dynamic adaptive thresholds include a first threshold, a third threshold, and a second threshold, which increase sequentially.

[0096] First threshold (lower threshold): The lower limit of the target diglyceride content. If the content is lower than this value, it is judged as insufficient reaction or incomplete separation. Second threshold (upper limit threshold): The upper limit of the target diglyceride content. If it is higher than this value, it is judged as excessive energy consumption or reduced production capacity. The third threshold (early warning threshold): an inner boundary value between the lower threshold and the upper threshold, used to trigger buffer adjustment.

[0097] The dynamic adaptive threshold is updated according to the following rules a) and b).

[0098] a) When the estimated diglyceride content of a predetermined number of consecutive test data is between the first threshold and the second threshold, the first threshold and the second threshold are recalculated based on the mean and standard deviation of the predetermined number of data.

[0099] For example, when the diglyceride content values ​​of 100 consecutive sets of test data are all between the current first threshold and the second threshold, the system automatically calculates the mean μ and standard deviation σ of the 100 sets of data, and updates the first threshold and the second threshold.

[0100] The new first threshold is μ - 2σ.

[0101] The new second threshold = μ + 2σ.

[0102] b) When the estimated diglyceride content shows a monotonically decreasing trend for a predetermined number of consecutive times, the third threshold is adjusted upward; when the estimated diglyceride content shows a monotonically increasing trend for a predetermined number of consecutive times, the third threshold is adjusted downward.

[0103] For example, when the diglyceride content is detected to show a monotonically decreasing trend for five consecutive sets of data, the system automatically adjusts the third threshold upward by 1 percentage point to trigger buffer adjustment in advance and prevent the content from falling below the lower limit. When the diglyceride content is detected to show a monotonically increasing trend for five consecutive sets of data, the system automatically adjusts the third threshold downward by 1 percentage point to trigger buffer adjustment in advance and prevent the content from exceeding the upper limit.

[0104] Preferably, after the system executes the process control command, the threshold is not updated within the preset response waiting time; after the response waiting time ends, if the diglyceride content returns to the target range, the third threshold is restored to the initial setting value to avoid over-adjustment.

[0105] Based on the comparison results, monitoring information is generated and output, specifically including: when the estimated diglyceride content is between the third and second thresholds, it is judged as a qualified state, and qualified monitoring information is generated; when the estimated diglyceride content is between the first and third thresholds, it is judged as a warning state, and warning monitoring information is generated. At this time, the diglyceride content is close to the lower limit boundary, and there may be a risk of insufficient reaction or enzyme activity decay, requiring slight intervention; when the estimated diglyceride content is lower than the first threshold or higher than the second threshold, it is judged as an abnormal state, and abnormal monitoring information is generated. If it is lower than the first threshold, it means that the diglyceride content is lower than the lower limit of the specification, the reaction is seriously insufficient or the separation is completely ineffective, and strong intervention is required. If it is higher than the second threshold but less than or equal to the second threshold + buffer width, it means that the diglyceride content is close to the upper limit boundary, and there may be problems with excessive energy consumption or reduced production capacity, requiring slight intervention; if it is higher than the second threshold and greater than the second threshold + buffer width, it means that the diglyceride content exceeds the upper limit of the specification, the distillation temperature is too high or the residence time is too long, and load reduction intervention is required.

[0106] Based on the above status determination results, the system can generate corresponding process control instructions and send the instructions to the production equipment actuators through the control output interface.

[0107] The foregoing has described in detail an embodiment of an online monitoring method for diglyceride oil production based on dielectric properties. Based on the online monitoring method for diglyceride oil production based on dielectric properties described in the above embodiment, this invention also provides an online monitoring system for diglyceride oil production based on dielectric properties corresponding to the method.

[0108] Figure 2 This is a schematic block diagram of an online monitoring system for diglyceride oil production based on dielectric properties, provided as an embodiment of the present invention. In this embodiment, the online monitoring system for diglyceride oil production based on dielectric properties can be divided into multiple functional modules according to the functions it performs. A module, as referred to in this invention, is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory.

[0109] The data acquisition and processing module is used to acquire dielectric response timing signals at at least two different frequencies through a dual-channel differential dielectric sensor installed on the diglyceride oil production pipeline, and construct a two-dimensional dielectric relaxation spectrum.

[0110] The diglyceride content estimation module is used to input a two-dimensional dielectric relaxation spectrum into a one-dimensional convolutional neural network. Through the sliding convolution operation of the convolution kernel in the time dimension, local temporal features reflecting the polarization relaxation process of diglyceride molecules are extracted. Based on the extracted local temporal features, the real-time diglyceride content estimate is output through direct regression by a fully connected layer. The convolutional neural network achieves time translation invariance through a weight sharing mechanism, enabling the same convolution kernel to identify the same local temporal pattern appearing at any position on the time axis. The dielectric response features at different frequency channels are extracted through parallel processing of multiple convolution kernels.

[0111] The monitoring information generation module is used to compare the estimated diglyceride content with a dynamic adaptive threshold, and generate and output monitoring information based on the comparison result.

[0112] The online monitoring system for diglyceride oil production based on dielectric properties in this embodiment is used to implement the aforementioned online monitoring method for diglyceride oil production based on dielectric properties. Therefore, the specific implementation of this system can be found in the embodiment section of the online monitoring method for diglyceride oil production based on dielectric properties mentioned above. Thus, the specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.

[0113] Furthermore, since the online monitoring system for diglyceride oil production based on dielectric properties in this embodiment is used to implement the aforementioned online monitoring method for diglyceride oil production based on dielectric properties, its function corresponds to the function of the above method, and will not be repeated here.

[0114] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for online monitoring of diglyceride oil production based on dielectric properties, characterized in that, Includes the following steps: By using a dual-channel differential dielectric sensor installed on the diglyceride oil production pipeline, dielectric response time-series signals at at least two different frequencies are collected to construct a two-dimensional dielectric relaxation spectrum. Two-dimensional dielectric relaxation spectra are input into a one-dimensional convolutional neural network. Through the sliding convolution operation of the convolution kernel in the time dimension, local temporal features reflecting the polarization relaxation process of diglyceride molecules are extracted. The convolutional neural network achieves time translation invariance through a weight sharing mechanism, enabling the same convolution kernel to identify the same local temporal pattern appearing at any position on the time axis. The dielectric response features at different frequency channels are extracted through parallel processing of multiple convolution kernels. Based on the extracted local temporal features, the real-time diglyceride content estimate is output through direct regression via a fully connected layer; The estimated diglyceride content is compared with a dynamic adaptive threshold, and monitoring information is generated and output based on the comparison result.

2. The online monitoring method for diglyceride oil production based on dielectric properties according to claim 1, characterized in that, The dual-channel differential dielectric sensor includes a measurement channel exposed to the oil being tested and a reference channel sealed in high-purity triglyceride oil.

3. The online monitoring method for diglyceride oil production based on dielectric properties according to claim 2, characterized in that, Acquire dielectric response time-series signals at at least two different frequencies to construct a two-dimensional dielectric relaxation spectrum, specifically including: Select N discrete frequency points at logarithmic intervals within a preset frequency range; At each frequency point, an excitation signal is first applied and a first preset time is waited for the oil polarization to reach a steady state. Then, dielectric response signals at M time points are continuously acquired at a preset sampling frequency. The dielectric response signals include the capacitance values ​​acquired by each of the two channels. For each sampling time, the capacitance ratio between the measurement channel and the reference channel is calculated, and the capacitance ratio is converted into a differential dielectric constant based on the dielectric properties of the high-purity triglyceride oil in the reference channel. M differential dielectric constant values ​​collected continuously at the same frequency point are arranged in chronological order to form a dielectric response time series vector at that frequency point. The time series vectors at N frequency points are stacked in chronological order to form an N-row, M-column two-dimensional matrix, which is the two-dimensional dielectric relaxation spectrum.

4. The online monitoring method for diglyceride oil production based on dielectric properties according to claim 3, characterized in that, When continuously collecting dielectric response signals at M time points at a preset sampling frequency, the oil temperature is also collected synchronously at each sampling time. The preprocessing operations on the two-dimensional dielectric relaxation spectrum before inputting it into the one-dimensional convolutional neural network include: Based on the synchronously collected oil temperature, the differential dielectric constant at each time and frequency is corrected to the reference temperature. The temperature-corrected two-dimensional dielectric relaxation spectrum is standardized according to the frequency dimension.

5. The online monitoring method for diglyceride oil production based on dielectric properties according to claim 3, characterized in that, A one-dimensional convolutional neural network consists of an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a flattening layer, and a fully connected layer, connected in sequence.

6. The online monitoring method for diglyceride oil production based on dielectric properties according to claim 5, characterized in that, A one-dimensional convolutional neural network outputs real-time estimates of diglyceride content, specifically including: The input is a two-dimensional dielectric relaxation spectrum, which is represented as an N-row, M-column two-dimensional matrix, where N is the number of frequency channels and M is the time series length; The first convolutional layer performs a one-dimensional convolution operation on the input two-dimensional dielectric relaxation spectrum: the input matrix is ​​organized into N one-dimensional time series by rows, and K1 one-dimensional convolution kernels are used to perform sliding convolution operations in the time dimension. Zero padding is used to keep the length of the time dimension unchanged, and a weight sharing mechanism is used to make the same convolution kernel use the same weight parameters at different time positions to extract local temporal features that are independent of time position. The output of each convolution kernel is processed by the ReLU activation function to form a feature map, and a total of K1 feature maps are output, each of which still maintains the dimension of N rows and M columns. The first pooling layer performs max pooling on the output of the first convolutional layer; The second convolutional layer performs a one-dimensional convolution operation on the output of the first pooling layer: the K1 feature maps output by the first pooling operation are used as multi-channel inputs, and K2 one-dimensional convolutional kernels are used to perform sliding convolution operations in the time dimension, with zero padding at the boundaries. Each convolutional kernel processes all K1 input channels simultaneously, and extracts cross-channel local correlation features through a weight sharing mechanism. The output of each convolutional kernel is processed by the ReLU activation function to form a feature map, resulting in a total of K2 feature maps, each with a dimension of N rows (M / 2) columns. The second pooling layer performs a global average pooling operation on the output of the second convolutional layer, compressing each feature map into a single value in the time dimension, resulting in several feature values; The flattening layer arranges the eigenvalues ​​in order into a one-dimensional eigenvector; The fully connected layer maps a one-dimensional feature vector to a single output value, namely the diglyceride content estimate.

7. The online monitoring method for diglyceride oil production based on dielectric properties according to claim 6, characterized in that, The weight sharing mechanism is configured such that, for any convolution kernel, as it slides across the time series, the convolution calculation at each sliding position uses the same set of weight parameters of the same convolution kernel, which does not change with the sliding position.

8. The online monitoring method for diglyceride oil production based on dielectric properties according to claim 1, characterized in that, The dynamic adaptive threshold consists of a first threshold, a third threshold, and a second threshold that increase sequentially; the dynamic adaptive threshold is updated according to the following rules: When the estimated diglyceride content of a predetermined number of consecutive test data is between the first threshold and the second threshold, the first threshold and the second threshold are recalculated based on the mean and standard deviation of the predetermined number of data. When the estimated diglyceride content shows a monotonically decreasing trend for a predetermined number of consecutive times, the third threshold is adjusted upward; when the estimated diglyceride content shows a monotonically increasing trend for a predetermined number of consecutive times, the third threshold is adjusted downward.

9. The online monitoring method for diglyceride oil production based on dielectric properties according to claim 8, characterized in that, Based on the comparison results, monitoring information is generated and output, specifically including: When the estimated diglyceride content is between the third threshold and the second threshold, it is judged to be in a qualified state, and qualified monitoring information is generated. When the estimated diglyceride content is between the first and third thresholds, it is determined to be in a warning state, and warning monitoring information is generated. When the estimated diglyceride content is lower than the first threshold or higher than the second threshold, it is determined to be an abnormal state, and abnormal monitoring information is generated.

10. An online monitoring system for diglyceride oil production based on dielectric properties, characterized in that, include: The data acquisition and processing module is used to acquire dielectric response time-series signals at at least two different frequencies through a dual-channel differential dielectric sensor installed on the diglyceride oil production pipeline, and construct a two-dimensional dielectric relaxation spectrum. The diglyceride content estimation module is used to input a two-dimensional dielectric relaxation spectrum into a one-dimensional convolutional neural network. Through the sliding convolution operation of the convolution kernel in the time dimension, local temporal features reflecting the polarization relaxation process of diglyceride molecules are extracted. Based on the extracted local temporal features, the real-time diglyceride content estimate is output through direct regression by a fully connected layer. The convolutional neural network achieves time translation invariance through a weight sharing mechanism, enabling the same convolution kernel to identify the same local temporal pattern appearing at any position on the time axis. The dielectric response features at different frequency channels are extracted through parallel processing of multiple convolution kernels. The monitoring information generation module is used to compare the estimated diglyceride content with a dynamic adaptive threshold, and generate and output monitoring information based on the comparison result.

Images