Online monitoring method and system for food frying process

Abstract

The invention discloses a food frying process online monitoring method and system, and relates to the technical field of food processing monitoring. The method comprises the following steps: transmitting a frequency-adjustable ultrasonic signal to a frying medium, receiving a response signal after the medium acts on food, and filtering; according to the comparison result of the amplitude attenuation rate of the response signal and the threshold value, the ultrasonic emission frequency is dynamically adjusted; performing feature extraction on the preprocessed signal to form a feature vector containing amplitude, frequency and energy features; inputting the feature vector into a judgment model trained by a dynamic frequency sample, and outputting a current frying state; and a control instruction of the frying equipment is generated so as to realize process regulation and control. Through combination of frequency-controlled ultrasonic excitation and acoustic feature extraction, the method can adapt to different grease environments and food types, gives consideration to the signal penetration depth and the signal-to-noise ratio, achieves non-destructive, real-time and quantifiable monitoring of the frying process, and remarkably improves the cooking state recognition precision and monitoring consistency.

Description

Technical Field

[0001] This application relates to the field of food processing monitoring technology, specifically to an online monitoring method and system for the food frying process. Background Technology

[0002] With the continuous improvement of automation in the food industry and the increasingly stringent quality control standards, the processing of fried foods is gradually shifting from traditional experience-based control to precise and intelligent monitoring. Frying, as a crucial step in food heat processing, directly affects the product's color, crispness, texture stability, and food safety due to its process parameters (including oil temperature, time, oil quality, and internal moisture content). Precise monitoring of the frying process is not only essential for consistent product quality but also crucial for energy conservation, emission reduction, and equipment maintenance. Currently, monitoring of industrial frying processes mainly relies on temperature sensors, visual inspection, or sampling analysis. Temperature sensors can only reflect changes in oil temperature and cannot reveal the internal cooking state and moisture migration patterns of the food; while visual inspection methods can provide surface image information, they are easily affected by factors such as oil fumes, bubbles, and changes in light, making stable and continuous process monitoring difficult; sampling inspection is destructive and cannot reflect the dynamic frying process in real time, thus limiting its application in continuous production lines.

[0003] To overcome the limitations of traditional monitoring methods, existing research has attempted to incorporate acoustic or ultrasonic signals into the monitoring of the frying process to detect bubble formation, oil deterioration, and internal changes in food. However, current technologies mostly employ fixed-frequency ultrasonic excitation, failing to fully consider the differences in acoustic propagation characteristics caused by changes in oil viscosity, bubble formation rate, and internal food structure over time during frying. Since different stages require different signal penetration depths and signal-to-noise ratios, fixed-frequency methods often lead to signal distortion and increased reflection interference, thus affecting monitoring accuracy and result repeatability. Summary of the Invention

[0004] The purpose of this application is to provide an online monitoring method and system for the food frying process to solve the problems mentioned in the background art.

[0005] In a first aspect, one embodiment of this application provides an online monitoring method for the food frying process. The method includes: transmitting an ultrasonic signal to a frying medium and receiving a response signal from the frying medium and the food, wherein the frequency of the ultrasonic signal is adjustable; filtering the response signal to obtain a pre-processed response signal; adjusting the transmission frequency of the ultrasonic signal based on a comparison between the amplitude attenuation rate of the pre-processed response signal and a preset threshold; performing signal analysis on the pre-processed response signal to extract feature parameters characterizing the frying state; forming a feature vector based on the feature parameters, wherein the feature parameters include amplitude features, frequency features, and energy features; inputting the feature vector into a pre-trained judgment model, outputting the current frying state through the judgment model, wherein the judgment model is obtained by training a machine learning algorithm using training samples collected under a dynamic frequency adjustment strategy, and the judgment model is adaptable to signals collected at different frequencies; and generating control commands for controlling the frying equipment based on the current frying state to control the food frying process.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the transmission frequency of the ultrasonic signal is adjusted based on the comparison result of the amplitude attenuation rate of the preprocessed response signal and a preset threshold, including: calculating the amplitude attenuation rate of the preprocessed response signal; increasing the transmission frequency of the ultrasonic signal when the amplitude attenuation rate is lower than a first preset threshold; decreasing the transmission frequency of the ultrasonic signal when the amplitude attenuation rate is greater than or equal to the first preset threshold and less than a second preset threshold; and switching the transmission frequency of the ultrasonic signal to a low-frequency band when the amplitude attenuation rate is greater than or equal to the second preset threshold; wherein the second preset threshold is greater than the first preset threshold.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, signal analysis is performed on the preprocessed response signal to extract characteristic parameters representing the frying state, including: performing time-domain analysis on the preprocessed response signal to extract the amplitude attenuation rate; performing frequency-domain analysis on the preprocessed response signal to extract the spectral centroid and phase delay; and performing energy analysis on the preprocessed response signal to extract energy distribution parameters.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, a feature vector is formed based on feature parameters, including: combining amplitude features, frequency features and energy features to form a feature vector; wherein, the amplitude feature is the amplitude attenuation rate, the frequency feature includes the spectral centroid and phase delay, and the energy feature is the energy distribution parameter.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the feature vector is input into a pre-trained decision model, and the decision model outputs the current frying state, including: inputting the feature vector into the decision model to obtain the decision result and confidence level; performing state smoothing processing on multiple consecutive decision results, the state smoothing processing including setting a state buffer queue of length N, adding the new decision result to the tail of the queue and removing the oldest result from the head of the queue; when the N decision results in the state buffer queue are consistent, the current frying state is confirmed.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the decision model is obtained by training a machine learning algorithm using training samples collected under a dynamic frequency adjustment strategy. The decision model is adapted to signals collected at different frequencies, including: conducting frying experiments under a dynamic frequency adjustment strategy for different types of food and simultaneously collecting training samples, which include multi-dimensional feature vectors collected at different frying stages and corresponding cooking status labels; extracting feature vectors from the labeled training samples to construct a training set; using the training set to train the machine learning algorithm to generate an initial decision model; and using a validation set collected under the dynamic frequency adjustment strategy to validate the initial decision model, ultimately obtaining a decision model adapted to signals of different frequencies.

[0011] In conjunction with the first aspect, in certain implementations of the first aspect, transmitting an ultrasonic signal to a frying medium and receiving a response signal from the frying medium and the food includes: transmitting an ultrasonic pulse signal at a preset pulse repetition frequency; and receiving a response signal reflected or transmitted through the frying medium and the food.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the response signal is filtered to obtain a preprocessed response signal, including: removing the DC component from the response signal to obtain a response signal after removing the DC component; The response signal after removing the DC component is subjected to noise suppression processing to obtain the noise-suppressed response signal; the noise-suppressed response signal is then subjected to baseline calibration processing to obtain the preprocessed response signal.

[0013] In conjunction with the first aspect, in some implementations of the first aspect, the response signal after noise suppression is subjected to baseline calibration processing to obtain a preprocessed response signal, including: acquiring a reference response signal before frying begins; performing differential processing on the response signal during frying and the reference response signal to obtain a difference signal; and using the difference signal as the response signal after baseline calibration processing.

[0014] Secondly, one embodiment of this application provides an online monitoring system for the food frying process. The system includes: an ultrasonic signal transceiver module for transmitting ultrasonic signals to the frying medium and receiving response signals from the frying medium and food, wherein the frequency of the ultrasonic signals is adjustable; a signal preprocessing module for filtering the response signals to obtain preprocessed response signals; a frequency adjustment module for adjusting the transmission frequency of the ultrasonic signals based on a comparison between the amplitude attenuation rate of the preprocessed response signals and a preset threshold; a feature extraction module for performing signal analysis on the preprocessed response signals, extracting feature parameters characterizing the frying state, and forming a feature vector based on the feature parameters, wherein the feature parameters include amplitude features, frequency features, and energy features; a state determination module for inputting the feature vectors into a pre-trained determination model, outputting the current frying state through the determination model, wherein the determination model is obtained by training a machine learning algorithm using training samples collected under a dynamic frequency adjustment strategy, and the determination model is adaptable to signals collected at different frequencies; and a control command generation module for generating control commands for controlling the frying equipment based on the current frying state, so as to control the food frying process.

[0015] Compared with the prior art, the beneficial effects of this application are: 1. This application combines frequency-controlled ultrasonic signal excitation with acoustic feature extraction to achieve real-time sensing of changes in oil viscosity, bubble formation, and dynamic changes in the internal moisture content of food during frying. Compared with traditional fixed-frequency monitoring methods, this application can adaptively adjust the frequency range of the ultrasonic signal to balance signal penetration depth and signal-to-noise ratio, significantly improving the accuracy of identifying the internal cooking state of food. This enables non-destructive, real-time, and quantifiable monitoring of the frying process, improving monitoring accuracy and result consistency.

[0016] 2. This application constructs an intelligent judgment model based on acoustic features, which can adaptively learn and optimize parameters according to the differences in acoustic response of different oil types, food materials, and processing stages, achieving compatibility and generalization for multiple frying scenarios. This model possesses good versatility and scalability, maintaining stable performance under different equipment and process conditions, thereby achieving intelligent control of the frying process and ensuring consistent product quality, effectively reducing human intervention and improving the automation level of the production line. Attached Figure Description

[0017] Figure 1 A schematic flowchart of an online monitoring method for food frying process provided in an embodiment of this application; Figure 2 This is a flowchart illustrating how the transmission frequency of an ultrasonic signal is adjusted based on a comparison between the amplitude attenuation rate of the preprocessed response signal and a preset threshold, according to an embodiment of this application. Figure 3A schematic flowchart of an online monitoring method for food frying process provided in another embodiment of this application; Figure 4 This is a schematic diagram illustrating the process of inputting a feature vector into a pre-trained decision model and outputting the current frying state through the decision model, as provided in one embodiment of this application. Figure 5 This is a flowchart illustrating the process of training a machine learning algorithm to obtain a judgment model using training samples collected under a dynamic frequency adjustment strategy, as provided in an embodiment of this application. Figure 6 A schematic diagram illustrating the process of transmitting ultrasonic signals to a frying medium and receiving response signals from the frying medium and food, according to an embodiment of this application. Figure 7 This is a flowchart illustrating the filtering process of a response signal to obtain a preprocessed response signal, according to an embodiment of this application. Figure 8 This is a schematic flowchart illustrating the baseline calibration process performed on the noise-suppressed response signal to obtain a preprocessed response signal, as provided in one embodiment of this application. Figure 9 This is a schematic diagram of the structure of an online monitoring system for the food frying process provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Figure 1 This is a schematic flowchart illustrating an embodiment of the online monitoring method for the food frying process provided in this application. Figure 1 As shown in the embodiments of this application, the online monitoring method for the food frying process includes the following steps: Step 100: An ultrasonic signal is emitted to the frying medium and the response signal from the frying medium and the food is received. The frequency of the ultrasonic signal is adjustable.

[0020] It should be understood that ultrasonic signals refer to sound wave signals with frequencies higher than the upper limit of human hearing (usually above 20 kHz). The ultrasonic frequency range used in this application is 50 kHz to 2 MHz, which is used to detect the acoustic properties of frying media and food.

[0021] Step 101: Filter the response signal to obtain the preprocessed response signal.

[0022] It should be understood that the response signal refers to the echo signal that returns to the receiving transducer after the ultrasonic signal propagates in the frying medium and food, through reflection, transmission or scattering. This signal carries information about the characteristics of the medium along the propagation path.

[0023] Step 102: Based on the comparison between the amplitude attenuation rate of the preprocessed response signal and the preset threshold, adjust the transmission frequency of the ultrasonic signal.

[0024] It should be understood that the amplitude attenuation rate refers to the rate at which ultrasonic waves lose energy during propagation. It is calculated by comparing the initial amplitude with the current amplitude and reflects the acoustic characteristics of the frying medium and the internal structural state of the food.

[0025] Step 103: Perform signal analysis on the preprocessed response signal to extract characteristic parameters representing the frying state.

[0026] It should be understood that characteristic parameters refer to acoustic parameters extracted from the response signal that can quantitatively characterize the frying state, including multi-dimensional parameters such as amplitude attenuation rate, spectral centroid, phase delay, energy distribution entropy, and low-frequency to high-frequency energy ratio.

[0027] Step 104: Form a feature vector based on the feature parameters, which include amplitude features, frequency features, and energy features.

[0028] Step 105: Input the feature vector into the pre-trained judgment model, and output the current frying state through the judgment model. The judgment model is obtained by training the machine learning algorithm using training samples collected under the dynamic frequency adjustment strategy. The judgment model is adapted to signals collected at different frequencies.

[0029] It should be understood that the judgment model refers to a mathematical model trained by a machine learning algorithm. This model learns the mapping relationship between feature vectors and fried state labels, and can perform state classification prediction on new feature vectors, outputting fried state labels and confidence scores.

[0030] Step 106: Generate control instructions for controlling the frying equipment based on the current frying state, so as to control the food frying process.

[0031] It should be understood that control commands refer to command signals generated based on the current frying state to control the operating parameters and actions of the frying equipment, including discharge commands, temperature adjustment commands, time extension commands, and abnormal alarm commands.

[0032] The beneficial effects of this embodiment are: by dynamically adjusting the ultrasonic frequency and combining it with a machine learning model, accurate real-time monitoring and intelligent control of the frying process are achieved. It can adapt to the signal acquisition requirements under different frying conditions, thereby improving the accuracy of frying state determination and the level of automation of equipment control.

[0033] Figure 2 This is a flowchart illustrating how the transmission frequency of an ultrasonic signal is adjusted based on a comparison between the amplitude attenuation rate of the preprocessed response signal and a preset threshold, according to an embodiment of this application. Figure 2 As shown in the embodiments of this application, the online monitoring method for the food frying process adjusts the transmission frequency of the ultrasonic signal based on the comparison result of the amplitude attenuation rate of the preprocessed response signal and a preset threshold, including the following steps: Step 200: Calculate the amplitude attenuation rate of the preprocessed response signal.

[0034] Specifically, the initial amplitude of the echo signal is recorded at the moment frying begins. At the current moment during the frying process Measuring echo amplitude Through calculation and Obtain the amplitude attenuation rate : ; in, The amplitude attenuation rate represents the propagation distance of the ultrasonic wave in the frying medium. This rate reflects the degree of attenuation of the ultrasonic energy during propagation and can characterize the changes in the acoustic properties of the frying medium and the internal structural state of the food.

[0035] Step 201: When the amplitude attenuation rate is lower than the first preset threshold, increase the transmission frequency of the ultrasonic signal.

[0036] Specifically, when the amplitude attenuation rate is lower than the first preset threshold, it indicates that the attenuation of the ultrasonic signal at the current frequency is small, the propagation conditions of the frying medium are good, the bubble density is low, or the internal structure of the food changes gently. At this time, the transmission frequency of the ultrasonic signal can be appropriately increased by increasing the transmission frequency by one frequency step, with the frequency step being 10 to 50 kHz, and the preferred transmission frequency being 20 to 30 kHz.

[0037] Step 202: When the amplitude attenuation rate is greater than or equal to the first preset threshold and less than the second preset threshold, reduce the transmission frequency of the ultrasonic signal.

[0038] Specifically, when the amplitude attenuation rate is greater than or equal to the first preset threshold and less than the second preset threshold, it indicates that the attenuation of the ultrasonic signal has increased but is still within an acceptable range. This may be due to the rapid evaporation of moisture on the surface of the food during the initial frying stage, which leads to an increase in the density of bubbles in the frying medium, or changes in the internal structure of the food. In this case, the transmission frequency of the ultrasonic signal is reduced by one frequency step. Reducing the frequency can increase the wavelength of the ultrasonic wave, reduce the influence of bubble scattering, improve the signal penetration ability, and ensure that the received echo signal has sufficient amplitude intensity and signal-to-noise ratio for accurate feature extraction and state determination.

[0039] Step 203: When the amplitude attenuation rate is greater than or equal to the second preset threshold, the transmission frequency of the ultrasonic signal is switched to the low frequency band.

[0040] Specifically, when the amplitude attenuation rate is greater than or equal to the second preset threshold, it indicates that the ultrasonic signal has been severely attenuated. This may be due to the extremely high density of air bubbles in the frying medium or the formation of a dense crust layer on the surface of the food, which hinders the propagation of ultrasonic waves. At this time, the transmission frequency of the ultrasonic signal is switched to the low-frequency band working mode. The frequency range of the low-frequency band is 50 to 150 kHz, preferably 80 to 120 kHz.

[0041] The second preset threshold is greater than the first preset threshold. Preferably, the first preset threshold is 0.8 to 1.2 dB / cm, and the second preset threshold is 1.5 to 2.0 dB / cm.

[0042] The beneficial effects of this embodiment are: by setting graded thresholds and multi-level frequency adjustment strategies, the ultrasonic frequency can be adaptively optimized according to the signal attenuation characteristics at different stages of frying, thereby maintaining stable signal quality and monitoring accuracy in the complex environment of bubble density changes and food structure evolution.

[0043] Figure 3 This is a schematic flowchart illustrating an online monitoring method for the food frying process, provided as another embodiment of this application. Figure 3 As shown in the embodiments of this application, the online monitoring method for the food frying process performs signal analysis on the pre-processed response signal and extracts characteristic parameters representing the frying state, including the following steps: Step 300: Perform time-domain analysis on the preprocessed response signal and extract the amplitude attenuation rate.

[0044] Specifically, the amplitude attenuation rate quantitatively reflects the degree of attenuation of ultrasonic energy during propagation. By continuously monitoring the change curve of the amplitude attenuation rate, the dynamic evolution of the food's cooking state during frying can be tracked.

[0045] Step 301: Perform frequency domain analysis on the preprocessed response signal to extract the spectral centroid and phase delay.

[0046] Specifically, the preprocessed echo signal is first subjected to a Fast Fourier Transform to obtain the frequency domain power spectrum. ,in Represents frequency. Calculates the spectral centroid. The formula is: ; Simultaneously, calculate the phase delay. Phase delay is defined as the phase difference between the echo signal and the transmitted signal. The phase delay is obtained by performing Hilbert transform on the transmitted and echo signals respectively to obtain the analytical signals, extracting their instantaneous phases, and calculating the phase difference.

[0047] Step 302: Perform energy analysis on the preprocessed response signal and extract energy distribution parameters.

[0048] Specifically, energy analysis includes calculating two parameters: energy distribution entropy and the low-frequency to high-frequency energy ratio. For the calculation of energy distribution entropy, the preprocessed echo signal is divided along the time axis. A time window of equal length The value range is 5 to 20, preferably 8 to 12. Calculate each time window. Energy of internal signals: ; in, This represents the amplitude of the sampling points within the given time window. The energy of each time window is normalized to obtain the energy percentage. Calculate the energy distribution entropy. The formula is: ; Entropy of energy distribution reflects the degree of dispersion of echo signal energy along the time axis. When the signal energy is highly concentrated in a few time periods, the entropy value is small; when the energy is evenly distributed across various time periods, the entropy value is large. In the early stages of frying, the internal structure of the food is uneven, the moisture distribution is inconsistent, and different parts reflect and absorb sound waves differently, resulting in a relatively dispersed echo energy distribution and a large entropy value. As cooking progresses, the internal structure tends to be more uniform, the energy distribution becomes more regular, and the entropy value changes accordingly. This parameter can quantitatively reflect the degree of homogenization within the food.

[0049] To calculate the low-frequency to high-frequency energy ratio, the preprocessed echo signal is decomposed into low-frequency and high-frequency components using wavelet packet decomposition or short-time Fourier transform. The low-frequency range is 0.5 to 1 times the transmission frequency, and the high-frequency range is 1 to 2 times the transmission frequency. The energy of the low-frequency band is then calculated separately. and high-frequency energy Calculate the energy ratio Low-frequency energy mainly comes from reflected and transmitted signals inside the food, carrying information about its internal cooking state; high-frequency energy reflects more information about the food's surface structure and surface roughness.

[0050] The beneficial effects of this embodiment are: by combining multi-dimensional analysis of time-domain amplitude attenuation rate, frequency-domain spectral centroid and phase delay, and energy distribution parameters, the quantitative characterization and dynamic tracking of the cooking state during food frying can be achieved, which significantly improves the accuracy and reliability of monitoring.

[0051] In another embodiment, the amplitude features, frequency features, and energy features are further combined to form a feature vector.

[0052] Specifically, the feature vector is a multi-dimensional vector composed of multiple acoustic feature parameters extracted from the echo signal arranged in a predetermined order, used to comprehensively characterize the acoustic properties of the current frying state. The feature vector is constructed as follows: ; in, For amplitude characteristics, and Frequency characteristics and It is an energy characteristic.

[0053] Amplitude characteristics Amplitude attenuation rate This parameter, obtained through time-domain analysis, reflects the attenuation of ultrasonic energy during propagation. Its value ranges from 0.3 to 3.0 dB / cm. This characteristic is highly sensitive to changes in bubble concentration and food moisture content in the frying medium.

[0054] Frequency characteristics include spectral centroid and phase delay Spectral centroid Obtained through frequency domain analysis, it reflects the center position of signal energy distribution in the frequency domain. Its numerical range depends on the transmission frequency, fluctuating between 0.7 and 1.3 times the transmission frequency. This characteristic can characterize the density and uniformity of the internal microstructure of food. Phase delay It is obtained by calculating the phase difference between the echo signal and the transmitted signal, reflecting the change in the propagation speed of ultrasound in the medium, and its numerical range is 0 to... The radius, or equivalent to 0 to 360 degrees, is sensitive to changes in the density and elastic modulus of the medium and can reflect the physical property transformation process of food tissue.

[0055] Energy characteristics include energy distribution entropy. The energy ratio of low frequency to high frequency Entropy of energy distribution It is obtained by dividing the echo signal into multiple time windows and calculating the information entropy of the energy distribution; its value ranges from 0 to... ,in The number of time windows reflects the dispersion of echo energy along the time axis, quantifying the homogenization process of the food's internal structure. Low-frequency to high-frequency energy ratio. The ratio of low-frequency energy to high-frequency energy is obtained through frequency domain energy analysis. The value range is usually 0.5 to 5.0. This feature reflects the difference in acoustic properties between the surface and interior of the food and can be used to assess the uniformity of cooking.

[0056] Figure 4 This is a schematic diagram illustrating a process in one embodiment of the present application whereby a feature vector is input into a pre-trained decision model, and the decision model outputs the current frying state. For example... Figure 4 As shown in the embodiments of this application, the online monitoring method for the food frying process inputs the feature vector into a pre-trained judgment model and outputs the current frying state through the judgment model, including the following steps: Step 400: Input the feature vector into the decision model to obtain the decision result and confidence level.

[0057] Specifically, the decision model is a trained machine learning model used for pattern recognition and state classification of the input feature vector. Types of decision models include, but are not limited to, gradient boosting decision trees, support vector machines, or small fully connected neural networks. Preferably, a gradient boosting decision tree is used as the decision model.

[0058] The normalized feature vector As input to the decision model, the model performs a forward inference computation process. For gradient boosting decision tree models, the input feature vector passes through multiple decision trees sequentially. Each decision tree selects a branch based on the judgment conditions of the feature value, eventually reaching a leaf node and outputting the prediction result of that tree. The prediction results of all decision trees are summed according to their weights to obtain the final decision result. For neural network models, the input feature vector is computed layer by layer through the input layer, hidden layer, and output layer. Nonlinear transformations are performed through activation functions, and the output layer uses the softmax function to convert the output into the probability distribution of each class.

[0059] The determination result is a classification label for the current frying state, which is divided into at least three categories: uncooked, semi-cooked, and cooked.

[0060] For gradient boosting decision tree models, confidence is calculated by statistically analyzing the consistency of voting results across all decision trees. The higher the proportion of votes cast by all decision trees for the same category, the greater the confidence. The calculation formula is: ,in, The number of decision trees that vote for the category of the decision result. This represents the total number of decision trees. For neural network models, the confidence score directly corresponds to the maximum probability value output by the softmax function of the output layer, i.e. ,in, For the number of categories, For the first The predicted probability of each category. High confidence indicates that the current feature vector is located in the core position of a certain category region in the feature space, and the judgment result is reliable and stable; low confidence indicates that the feature vector is near the boundary region of different categories, and the judgment result is uncertain, which may be due to the difference between the characteristics of the current food and the training samples, or in the transition stage of state transition.

[0061] Step 401: Perform state smoothing on the multiple consecutive judgment results. State smoothing includes setting a state buffer queue of length N, adding the new judgment result to the tail of the queue and removing the oldest result from the head of the queue.

[0062] Specifically, a first-in-first-out state buffer queue is set up, with a queue length of [value missing]. , The value range is 3 to 10, preferably 3 to 5. When When the value is 3, the system response speed is fast but the anti-interference ability is relatively weak; when A value of 5 results in better system stability but a slight delay in response. In practical applications, The value should be selected based on a trade-off between the dynamic characteristics of the frying process and the requirements for real-time performance. For the initial stage of frying when state changes are rapid, a smaller value can be set. Value; for the later stages of frying where the state changes slowly, a larger value can be set. Values ​​that enable adaptive smoothing.

[0063] Whenever the decision model outputs a new decision, that decision is added to the tail of the queue, while the oldest decision at the head of the queue is removed, maintaining the queue length at all times. Initially, all positions in the queue are filled with initial state labels, typically indicating an uncooked state. As the frying process progresses, new judgment results are continuously added, dynamically updating the queue content. The queue stores the most recent... The sequence of judgment results.

[0064] To further improve the smoothing effect, the decision results in the queue are weighted. Newer decisions are assigned higher weights, while older decisions are assigned lower weights, reflecting the timeliness of the decisions. The weight coefficients decrease linearly, i.e. ,in, To determine the position number of the result in the queue, the position numbers are sequentially from 1 to 10. Alternatively, an exponential decay method can be used, i.e. ,in, The attenuation coefficient ranges from 0.1 to 0.5.

[0065] Step 402: When the N judgment results in the state buffer queue are consistent, confirm the current frying state.

[0066] Specifically, consistent judgment results refer to all states in the state buffer queue. All judgment results have the same classification label. The system checks the queue in real time. Each judgment result is compared with its classification label. A decision is made if and only if all... When all judgment results are for the same category label, the queue state is considered consistent. At this point, the system confirms that a valid state transition has occurred and officially updates the current frying state to the consistent category label. For example, if the current frying state is "half-cooked," the state buffer queue length is... The setting is 3. At a certain moment, the judgment model outputs a mature state, and this result is added to the queue. However, the queue still contains the previous semi-ripe state judgment result, so the queue content is [semi-ripe, semi-ripe, mature]. The three results are inconsistent, and no state confirmation is triggered. At the next moment, the judgment model outputs a mature state again, and the queue is updated to [semi-ripe, mature, mature], which is still inconsistent. At the third moment, the judgment model continues to output a mature state, and the queue is updated to [mature, mature, mature]. At this time, the three judgment results in the queue are completely consistent, and the system officially confirms that the food has changed from a semi-ripe state to a mature state, triggering the corresponding control action, such as sending a discharge command.

[0067] The beneficial effects of this embodiment are: by introducing a state buffer queue and multi-level verification, the instantaneous fluctuations and accidental misjudgments of a single judgment are effectively suppressed, the stability and reliability of the frying state judgment are significantly improved, and the accuracy of the state transition judgment is ensured.

[0068] Figure 5 This is a schematic diagram illustrating the process of training a machine learning algorithm to obtain a judgment model using training samples collected under a dynamic frequency adjustment strategy, as provided in one embodiment of this application. Figure 5 As shown in the embodiments of this application, the online monitoring method for the food frying process determines that the model is obtained by training a machine learning algorithm using training samples collected under a dynamic frequency adjustment strategy. The determination model is adapted to signals collected at different frequencies, and includes the following steps: Step 500: For different types of food, frying experiments are conducted under a dynamic frequency adjustment strategy, and training samples are collected simultaneously. The training samples include multi-dimensional feature vectors collected at different frying stages and corresponding cooking state labels.

[0069] Specifically, training samples were collected under standard laboratory conditions or on a pilot production line. Systematic frying experiments were conducted on various typical fried food types, including but not limited to French fries, chicken nuggets, fish fillets, meatballs, fried dough sticks, and tempura, covering different food matrix characteristics, including starchy, protein-based, and mixed types, ensuring the representativeness and diversity of the training samples. Samples were collected under different oil conditions, including fresh oil, oil that had been used for a certain period, and mixed oil. By collecting samples under different oil conditions, the model could adapt to the dynamic changes in oil quality in actual production, improving the model's generalization ability and robustness. Before each frying experiment, the initial parameters of the food were strictly controlled, including size specifications, initial moisture content, and initial temperature, to ensure the repeatability of the experiment. Simultaneously, frying process parameters were strictly controlled, including frying temperature setpoint, oil depth, food input amount, and stirring method, and detailed experimental condition information was recorded to provide a reference for sample labeling and data analysis. The frying process is monitored under a dynamic frequency adjustment strategy. The frequency-controlled ultrasonic transducer operates under adaptive frequency adjustment as described in steps 200 to 203, automatically adjusting the transmission frequency of the ultrasonic signal based on the real-time monitored amplitude attenuation rate. Throughout the frying process, echo signals are continuously acquired at set time intervals, ranging from 1 to 5 seconds, preferably 2 to 3 seconds, to ensure the capture of dynamic characteristics of state changes during frying. Each acquired echo signal is preprocessed and feature extracted. Feature vectors are extracted according to the methods described in steps 300 to 302, including amplitude attenuation rate, spectral centroid, phase delay, energy distribution entropy, and low-frequency to high-frequency energy ratio.

[0070] During the frying process, samples are taken periodically for objective measurements and subjective evaluations. Objective measurements include using a moisture meter to measure the food's moisture content, a colorimeter to measure color parameters, a texture analyzer to measure mechanical properties such as hardness and crispness, and a temperature probe to measure the food's core temperature. Subjective evaluations are conducted by trained professional sensory evaluators, and the evaluation content includes indicators such as appearance, color, aroma, taste, and degree of cookedness. Combining the objective measurement data and subjective evaluation results, professionals accurately label the feature vectors collected at each time point with four categories: uncooked, partially cooked, fully cooked, and overcooked.

[0071] The training samples are structured as binary tuples. ,in The normalized multidimensional feature vector, This corresponds to the cooking state label. All collected and enhanced training samples are categorized and stored according to information such as food type, oil quality, and frying stage to establish a complete training sample database.

[0072] Step 501: Extract feature vectors from the labeled training samples to construct a training set, use the training set to train the machine learning algorithm, and generate an initial judgment model.

[0073] Specifically, all labeled samples in the training sample database are divided into a training set and a validation set according to a certain ratio. The preferred ratio is 80% training set and 20% validation set. Feature vectors are extracted from the training set to construct a feature matrix. Each row corresponds to a feature vector of a sample, and each column corresponds to a feature dimension. The corresponding ripening state labels are extracted to form a label vector. The feature matrix and label vector Used as input for model training in machine learning algorithms.

[0074] Preferably, the gradient boosting decision tree algorithm is used to train the machine learning algorithm. After training, an initial decision model is obtained. This model has learned the mapping relationship between feature vectors and mature state labels and can perform state classification prediction for new feature vectors.

[0075] Step 502: The initial decision model is validated using the validation set collected under the dynamic frequency adjustment strategy, and a decision model adapted to different frequency signals is finally obtained.

[0076] Specifically, the feature vectors in the validation set are input into the initial decision model to obtain the model's prediction results. The prediction results are then compared with the true labels to calculate various performance evaluation indicators.

[0077] Performance evaluation metrics include classification accuracy, precision for each category, recall, and Classification accuracy is defined as the proportion of correctly predicted samples out of the total number of samples, and is calculated using the following formula: ,in To predict the correct number of samples, Let be the total number of samples. The classification accuracy should be at least 90%, preferably at least 93%. For each maturation state category, calculate precision and recall separately. Precision is defined as the proportion of samples predicted to belong to that category that actually do not, calculated using the following formula: ,in This represents the number of true positive samples. This represents the number of false positive samples. Recall is defined as the proportion of samples that correctly belong to the category and are correctly predicted; the formula is... ,in This represents the number of false negative samples. The value is the harmonic mean of precision and recall, calculated using the following formula: This comprehensively reflects the model's predictive performance for that category.

[0078] The recall rate for mature product categories should be no less than 92%, and for preferred categories no less than 95%, to ensure that products are not prematurely released due to missed detections, thus affecting product quality. The validation set is divided into three subsets based on the frequency of the ultrasonic signal collected: high-frequency samples, mid-frequency samples, and low-frequency samples. High-frequency samples correspond to those with emission frequencies above 300 kHz, mid-frequency samples to those with emission frequencies between 150 and 300 kHz, and low-frequency samples to those with emission frequencies below 150 kHz. The classification accuracy and the scores for each category are calculated for each of the three frequency subsets. The value is used to evaluate the model's adaptability to signals of different frequencies. If the model's performance in a certain frequency band is significantly lower than in other frequency bands, it indicates that the training samples in that frequency band are insufficient or the feature representation is inadequate. It is necessary to supplement the training samples in that frequency band or adjust the feature extraction method and perform iterative optimization.

[0079] If the initial decision model meets the preset performance requirements on the validation set, it indicates that the model has good decision-making and generalization abilities and can adapt to different frequency signals collected under the dynamic frequency adjustment strategy. This model is then selected as the final decision model. If the initial decision model does not meet the performance requirements, model optimization and iterative improvement are necessary.

[0080] The beneficial effects of this embodiment are: by collecting diverse training samples and performing frequency band verification under the dynamic frequency adjustment strategy, the judgment model can be accurately adapted to different frequency signals, which significantly improves the model's generalization ability and robustness to different food types and oil conditions.

[0081] Figure 6 This is a schematic diagram illustrating a process for transmitting ultrasonic signals to a frying medium and receiving response signals from the frying medium and food, according to an embodiment of this application. Figure 6 As shown in the embodiments of this application, the online monitoring method for the food frying process transmits ultrasonic signals to the frying medium and receives response signals from the frying medium and the food, including the following steps: Step 600: Transmit an ultrasonic pulse signal at a preset pulse repetition frequency.

[0082] Specifically, the ultrasonic pulse signal employs narrow pulse modulation, modulating the continuous wave signal into short-duration pulses. This ensures sufficient excitation energy while avoiding signal aliasing and multipath interference, thus improving time-domain resolution. The pulse width of the narrow pulse is set to 3 to 10 times the transmission frequency period, preferably 5 to 8 times. For example, when the transmission frequency is 200 kHz, the corresponding period is 5 microseconds, and the pulse width is set to 25 to 40 microseconds. The pulse repetition frequency is set from 10 Hz to 1000 Hz, preferably 50 to 200 Hz. In practical applications, the pulse repetition frequency can be dynamically adjusted according to different stages of the frying process. In the early stages of frying, when the state changes drastically, a higher pulse repetition frequency, such as 100 to 200 Hz, is set to increase the sampling density and capture state changes promptly. In the middle and later stages of frying, when the state changes become more gradual, the pulse repetition frequency can be appropriately reduced to 50 to 100 Hz to reduce system load and optimize resource utilization while maintaining monitoring effectiveness.

[0083] Step 601: Receive the response signal reflected or transmitted through the frying medium and food.

[0084] Specifically, the receiving transducer is arranged at an appropriate position in the frying tank. Depending on the relative position of the transmitting and receiving transducers, either a face-to-face receiving or a reflective receiving method can be used.

[0085] The beneficial effects of this embodiment are: by using narrow pulse modulation and adaptive pulse repetition frequency adjustment, combined with high-speed and high-precision digital sampling, signal aliasing and multipath interference are effectively avoided, the time domain resolution and signal acquisition quality are significantly improved, and accurate monitoring of different frying stages is ensured.

[0086] Figure 7 This is a schematic flowchart illustrating the filtering process of a response signal to obtain a preprocessed response signal, as provided in one embodiment of this application. Figure 7 As shown in the embodiments of this application, the online monitoring method for the food frying process filters the response signal to obtain a pre-processed response signal, including the following steps: Step 700: Perform DC component removal processing on the response signal to obtain the response signal after removing the DC component.

[0087] Specifically, during the acquisition and digitization of the response signal, a DC component is superimposed on the digitized response signal due to factors such as the bias voltage of the preamplifier, the reference voltage deviation of the analog-to-digital converter, and the thermal drift of electronic components. The DC component removal is achieved using a high-pass filter, preferably a second- or third-order Butterworth high-pass filter. The cutoff frequency of the high-pass filter is set to 1 / 5 to 1 / 10 of the ultrasonic signal transmission frequency, preferably 1 / 10. For example, when the transmission frequency is 200 kHz, the cutoff frequency of the high-pass filter is set to 20 kHz. High-pass filtering can be implemented in the time or frequency domain. The response signal after removing the DC component is symmetrically distributed around zero, and the average value of the signal is close to 0.

[0088] Step 701: Perform noise suppression processing on the response signal after removing the DC component to obtain the noise-suppressed response signal.

[0089] Specifically, an adaptive denoising algorithm based on wavelet transform is used for noise suppression. First, a suitable wavelet basis function is selected, preferably the Daubechies 4 wavelet or the Symlet 8 wavelet. The response signal after removing the DC component undergoes multi-level wavelet decomposition, with the decomposition level set to 4 to 6 levels, preferably 5 levels. The decomposition process progressively decomposes the signal into low-frequency approximation coefficients and high-frequency detail coefficients. The low-frequency approximation coefficients mainly contain the main features and gradual change components of the signal, while the high-frequency detail coefficients contain the subtle changes in the signal and noise components.

[0090] Thresholding is applied to the wavelet coefficients obtained from the decomposition. By setting an appropriate threshold, wavelet coefficients corresponding to noise are set to zero or reduced, while wavelet coefficients corresponding to useful signals are retained. Coefficients with absolute values ​​less than the threshold are set to zero, while coefficients with absolute values ​​greater than the threshold are reduced by the threshold value towards zero. The threshold is determined using an adaptive strategy, dynamically calculated based on the statistical characteristics of the detail coefficients at each decomposition level. Different thresholds can be used for detail coefficients at different decomposition levels. High-frequency detail coefficients mainly contain noise, so smaller thresholds are used for strong suppression; mid-to-low-frequency detail coefficients contain more useful signals, so larger thresholds are used for weak suppression.

[0091] Step 702: Perform baseline calibration on the noise-suppressed response signal to obtain the preprocessed response signal.

[0092] Specifically, before the frying operation begins, when the frying tank contains only oil and no food, the system enters the reference signal acquisition mode. At this time, the oil temperature has reached the set value and stabilized, and the oil is in a preheated state. The system emits ultrasonic pulse signals and receives response signals according to the normal operating mode, collecting signal data for a period of time, ranging from 10 to 60 seconds, preferably 30 seconds. During this period, multiple sets of response signals are continuously acquired, and the acquired signals are averaged to obtain the reference signal. During the actual frying process, once the food is placed in the frying tank, the system acquires response signals in real time. The current response signal and the reference signal are differentially processed to calculate the difference signal. : ; in, This is a calibration coefficient, ranging from 0.8 to 1.2, used to compensate for the amplitude of the reference signal caused by changes in oil temperature or deterioration in oil quality. To accommodate the gradual deterioration of oil quality and the slow changes in equipment condition during long-term continuous production, the reference signal is periodically updated. The update cycle is set after each batch of frying, before the start of each shift, or at regular intervals, such as every 2 to 4 hours. The update method involves re-collecting the reference signal when there is no food, replacing or weighting the original reference signal.

[0093] The beneficial effects of this embodiment are as follows: By using a multi-level signal preprocessing method, including DC component removal, wavelet transform-based adaptive noise suppression, and dynamic baseline calibration, the interference effects of system bias, environmental noise, and oil quality changes are effectively eliminated, significantly improving the signal-to-noise ratio and feature extraction accuracy of the response signal.

[0094] Figure 8 This is a schematic flowchart illustrating the baseline calibration process performed on the noise-suppressed response signal to obtain a preprocessed response signal, as provided in one embodiment of this application. Figure 8 As shown in the embodiments of this application, the online monitoring method for the food frying process performs baseline calibration processing on the noise-suppressed response signal to obtain a preprocessed response signal, including the following steps: Step 800: Acquire a baseline response signal before frying begins.

[0095] Specifically, the baseline response signal is acquired before the frying operation officially begins, serving as a reference benchmark for the system's inherent response. Before acquiring the baseline response signal, it is necessary to ensure that the frying system is in standard operating condition. First, the frying tank is heated to the set operating temperature, which is determined according to the type of food to be processed, within the range of 160 to 190 degrees Celsius, and the temperature is kept stable, with temperature fluctuations controlled within ±2 degrees Celsius. After the oil temperature stabilizes, the frying tank contains only the oil medium, without any food or other objects, ensuring that the acquired signal only reflects the ultrasonic propagation characteristics in a pure oil environment.

[0096] The frequency-controlled ultrasonic transducer transmits ultrasonic pulse signals according to a preset pulse repetition frequency and transmission frequency, and the receiving transducer receives the response signals after propagation through the oil medium. The transmission frequency is selected as an initial frequency pre-calculated for the current food type and oil conditions to ensure the comparability of the reference signal with the signal during the actual frying process. The system continuously acquires multiple sets of response signals, with the acquisition duration set to 10 to 60 seconds, preferably 20 to 40 seconds. During this time, continuous sampling is performed at a constant pulse repetition frequency to obtain multiple sets of time-series response signal data.

[0097] Assuming a total of [number] data were collected within the collection period. A group of response signal samples, denoted as The average reference response signal is calculated. Through averaging, random noise components cancel each other out, and the deterministic components in the reference signal are enhanced. The value is determined based on the acquisition duration and pulse repetition frequency, ranging from 500 to 5000 sets. For example, when the pulse repetition frequency is 100 Hz and the acquisition duration is 30 seconds, 3000 sets of response signals can be obtained and averaged. The acquired reference response signal records the propagation path characteristics of ultrasound under food-free conditions, including the propagation attenuation of ultrasound in pure oil medium, reflection at the inner and outer walls of the frying tank, reflection at the interface between the transducer protective layer and the acoustic coupling layer, reflection on the oil surface, and all other inherent system responses.

[0098] Step 801: Perform differential processing on the response signal during the frying process and the reference response signal to obtain the difference signal.

[0099] Specifically, once the food is placed in the frying tank and the frying process begins, the system collects the current response signal in real time. Current response signal This is the echo signal after ultrasonic waves propagate in the presence of food. This signal contains two components: one is the inherent system response, identical to the reference response signal in step 800; the other is an additional response caused by the presence of food and its acoustic properties. The additional response reflects the food's effects on ultrasonic wave reflection, absorption, and scattering, including information about the food's size, shape, location, internal structure, moisture content, and density.

[0100] The inherent system response is removed from the current response signal, i.e., the difference between the current response signal and the reference response signal is calculated. The basic formula for differential processing is: ; in, It is a difference signal. For time variables, The reference signal amplitude is used. This differential operation is performed point-by-point in the time domain, subtracting the signal amplitudes at corresponding times to obtain the time-domain sequence of the difference signal. In practical applications, small fluctuations in oil temperature, slow deterioration of oil quality, and changes in ambient temperature may cause slight changes in the system's inherent response. A calibration coefficient can be introduced into the differential processing to compensate for these changes. The improved differential processing formula is: ; in, This is the calibration coefficient. The temperature-compensated method corrects for the temperature difference between the current oil temperature and the oil temperature at the time the reference signal was acquired using a temperature coefficient. The temperature coefficient is determined based on the thermophysical properties of the oil. For commonly used edible oils, the calibration coefficient is calculated based on the temperature difference. ,in For the overall temperature correction factor, This represents the temperature difference.

[0101] Step 802: Use the difference signal as the response signal after baseline calibration.

[0102] The beneficial effects of this embodiment are as follows: by collecting the reference response signal under pure oil environment and performing differential processing, the interference of the inherent response of the frying tank, transducer and oil medium and other systems is effectively removed, the food characteristic signal is accurately extracted, and the accuracy and sensitivity of the characteristic parameters in representing the food cooking state are significantly improved.

[0103] Figure 9 This is a schematic diagram of the structure of an online monitoring system for the food frying process provided in an embodiment of this application. Figure 9As shown in the embodiment of this application, the online monitoring system for the food frying process includes an ultrasonic signal transceiver module 900 whose output is electrically connected to the input of a signal preprocessing module 901. This allows the received response signal to be transmitted to the signal preprocessing module 901 for filtering. The output of the signal preprocessing module 901 is simultaneously connected to the inputs of both the frequency adjustment module 902 and the feature extraction module 903, ensuring that the preprocessed response signal can be used by both modules simultaneously without affecting the normal operation of the other module due to data reading from one. After completing processes such as DC component removal, noise suppression, and baseline calibration, the signal preprocessing module 901 outputs the preprocessed response signal as a data stream containing a complete time-domain signal sequence and corresponding timestamp information. The control output of the frequency adjustment module 902 is connected to the frequency control input of the ultrasonic signal transceiver module 900, forming a feedback control loop for adaptive frequency adjustment. This connection is achieved through a control signal line or communication bus, transmitting frequency adjustment commands and parameter setting commands. The frequency adjustment module 902 receives the preprocessed response signal from the signal preprocessing module 901, calculates the amplitude attenuation rate and compares it with a first preset threshold and a second preset threshold, and generates a frequency adjustment control command based on the comparison result. The frequency adjustment control command includes a frequency increase command, a frequency decrease command, or a frequency band switching command, as well as a corresponding frequency step or target frequency value.

[0104] The input of the feature extraction module 903 is connected to the output of the signal preprocessing module 901, receiving the preprocessed response signal to extract multi-dimensional feature parameters. The output of the feature extraction module 903 is connected to the input of the state determination module 904, used to transmit the constructed feature vector to the state determination module 904 for state classification. The connection is implemented using a data interface or shared memory. The feature vector is transmitted in a standardized data structure, including the values ​​of each feature parameter, the normalized feature values, and the corresponding timestamps and metadata information.

[0105] The input of the state determination module 904 is connected to the output of the feature extraction module 903, receiving feature vectors as input data for the determination model. The state determination module 904 integrates a pre-trained machine learning model, which performs forward inference calculations on the input feature vectors and outputs the classification label and corresponding confidence score of the current frying state. The output of the state determination module 904 is connected to the input of the control command generation module 905, transmitting the confirmed current frying state information to the control command generation module 905. The transmitted data includes the frying state label, confidence score, state transition flag, and historical state sequence information.

[0106] The input terminal of the control command generation module 905 is connected to the output terminal of the state determination module 904, receiving the current frying state information as the basis for control decisions. The control command generation module 905 generates corresponding control commands based on the current frying state. These control commands are used to control the operating parameters and actions of the frying equipment. The output terminal of the control command generation module 905 is connected to the control system of the frying equipment, using industrial standard communication protocols, including but not limited to Modbus TCP, OPC UA, Profinet, or EtherCAT. Control commands include discharge commands, temperature adjustment commands, time extension commands, and abnormal alarm commands. When the state determination module 904 confirms that the food has reached a cooked state, the control command generation module 905 generates a discharge command and sends it to the control system of the frying equipment, triggering the discharge mechanism. When an abnormality in the cooking process is detected, the control command generation module 905 generates a temperature adjustment command or an alarm command to control the frying process.

[0107] The beneficial effects of this embodiment are: through modular system architecture design and standardized data interface, seamless integration and real-time collaboration of signal acquisition, adaptive frequency adjustment, feature extraction, status determination and equipment control are achieved, which significantly improves the system's reliability, scalability and industrial application adaptability.

[0108] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for online monitoring of the food frying process, characterized in that, include: An ultrasonic signal is emitted to the frying medium, and a response signal is received from the frying medium and the food, wherein the frequency of the ultrasonic signal is an adjustable frequency. The response signal is filtered to obtain a preprocessed response signal; Based on the comparison between the amplitude attenuation rate of the preprocessed response signal and a preset threshold, the transmission frequency of the ultrasonic signal is adjusted. Signal analysis is performed on the preprocessed response signal to extract characteristic parameters representing the frying state; A feature vector is formed based on the feature parameters, which include amplitude features, frequency features, and energy features; The feature vector is input into a pre-trained judgment model, and the current frying state is output by the judgment model. The judgment model is obtained by training a machine learning algorithm using training samples collected under a dynamic frequency adjustment strategy. The judgment model is adapted to signals collected at different frequencies. Based on the current frying state, control commands are generated to control the frying equipment in order to control the food frying process.

2. The method according to claim 1, characterized in that, The step of adjusting the transmission frequency of the ultrasonic signal based on the comparison result of the amplitude attenuation rate of the preprocessed response signal and a preset threshold includes: Calculate the amplitude attenuation rate of the preprocessed response signal; When the amplitude attenuation rate is lower than a first preset threshold, the transmission frequency of the ultrasonic signal is increased; When the amplitude attenuation rate is greater than or equal to a first preset threshold and less than a second preset threshold, the transmission frequency of the ultrasonic signal is reduced. When the amplitude attenuation rate is greater than or equal to the second preset threshold, the transmission frequency of the ultrasonic signal is switched to the low frequency band. Wherein, the second preset threshold is greater than the first preset threshold.

3. The method according to claim 1, characterized in that, The step of performing signal analysis on the preprocessed response signal to extract characteristic parameters representing the frying state includes: Perform time-domain analysis on the preprocessed response signal to extract the amplitude attenuation rate; Frequency domain analysis is performed on the preprocessed response signal to extract the spectral centroid and phase delay; Energy analysis is performed on the preprocessed response signal to extract energy distribution parameters.

4. The method according to claim 3, characterized in that, The process of forming a feature vector based on the feature parameters includes: The amplitude feature, frequency feature, and energy feature are combined to form the feature vector; Wherein, the amplitude characteristic is the amplitude attenuation rate, the frequency characteristic includes the spectral centroid and phase delay, and the energy characteristic is the energy distribution parameter.

5. The method according to any one of claims 1 to 4, characterized in that, The step of inputting the feature vector into a pre-trained decision model and outputting the current frying state through the decision model includes: The feature vector is input into the decision model to obtain the decision result and confidence level; The state smoothing process is performed on the results of multiple consecutive judgments. The state smoothing process includes setting a state buffer queue of length N, adding the new judgment result to the tail of the queue and removing the oldest result from the head of the queue. When N determination results in the state buffer queue are consistent, the current frying state is confirmed.

6. The method according to any one of claims 1 to 4, characterized in that, The determination model is obtained by training a machine learning algorithm using training samples collected under a dynamic frequency adjustment strategy. The determination model is adaptable to signals collected at different frequencies, including: Frying experiments were conducted on different types of food under the dynamic frequency adjustment strategy, and training samples were collected simultaneously. The training samples included multi-dimensional feature vectors collected at different frying stages and corresponding cooking state labels. Feature vectors are extracted from labeled training samples to construct a training set, and the machine learning algorithm is trained using the training set to generate an initial judgment model; The initial decision model is validated using the validation set collected under the dynamic frequency adjustment strategy, and finally the decision model adapted to different frequency signals is obtained.

7. The method according to any one of claims 1 to 4, characterized in that, The method of emitting ultrasonic signals to the frying medium and receiving response signals from the frying medium and the food includes: The ultrasonic pulse signal is emitted at a preset pulse repetition frequency; Receive response signals reflected or transmitted through the frying medium and the food.

8. The method according to claim 1, characterized in that, The step of filtering the response signal to obtain a preprocessed response signal includes: The response signal is subjected to DC component removal processing to obtain the response signal after DC component removal; The response signal after removing the DC component is subjected to noise suppression processing to obtain the noise-suppressed response signal; The noise-suppressed response signal is subjected to baseline calibration to obtain the preprocessed response signal.

9. The online monitoring method for the food frying process according to claim 8, characterized in that, The baseline calibration process performed on the noise-suppressed response signal to obtain the preprocessed response signal includes: A baseline response signal was acquired before frying began; The response signal during the frying process is differentially processed with the reference response signal to obtain the difference signal; The difference signal is used as the response signal after baseline calibration.

10. An online monitoring system for the food frying process, characterized in that, include: An ultrasonic signal transceiver module is used to transmit ultrasonic signals to the frying medium and receive response signals from the frying medium and food. The frequency of the ultrasonic signal is adjustable. A signal preprocessing module is used to filter the response signal to obtain a preprocessed response signal; The frequency adjustment module is used to adjust the transmission frequency of the ultrasonic signal based on the comparison result of the amplitude attenuation rate of the preprocessed response signal and a preset threshold. The feature extraction module is used to perform signal analysis on the preprocessed response signal, extract feature parameters that characterize the frying state, and form a feature vector based on the feature parameters. The feature parameters include amplitude features, frequency features, and energy features. The state determination module is used to input the feature vector into a pre-trained determination model and output the current frying state through the determination model. The determination model is obtained by training a machine learning algorithm using training samples collected under a dynamic frequency adjustment strategy. The determination model is adapted to signals collected at different frequencies. The control instruction generation module is used to generate control instructions for controlling the frying equipment based on the current frying state, so as to control the food frying process.

Images