A bottle nitrogen filling process quality evaluation method applying artificial intelligence

Abstract

The application provides a bottle nitrogen filling process quality evaluation method applying artificial intelligence, comprising: acquiring airflow velocity distribution, vortex intensity and streamline form at a bottle opening of a nitrogen filling equipment, performing feature extraction and time series analysis, and generating an airflow feature dataset; if airflow feature stability is lower than a preset threshold, predicting short-term change trend of the airflow velocity distribution, the vortex intensity and the streamline form through a long short-term memory network algorithm to obtain a future evolution state of the airflow feature; adjusting a nitrogen injection rate and an injection angle of the nitrogen filling equipment according to the future evolution state of the airflow feature to generate optimized nitrogen filling control parameters; and adjusting the nitrogen injection parameters according to the optimized nitrogen filling control parameters, updating the running state of the nitrogen filling equipment in real time, and generating an updated airflow feature dataset.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and particularly relates to a bottle nitrogen filling process quality evaluation method applying artificial intelligence. BACKGROUND

[0002] Nitrogen filling quality evaluation is a key link in the packaging process of food, medicine and other industries, and is directly related to the shelf life and safety of products. The nitrogen filling process replaces oxygen in a sealed container by injecting nitrogen gas to inhibit oxidation reactions, and its quality evaluation is crucial to production efficiency and product quality. Traditional nitrogen filling quality evaluation methods rely mainly on pressure detection or gas concentration analysis. Although these methods are simple to operate, they are difficult to accurately capture the subtle changes in gas flow at the bottle opening when the internal pressure approaches the external pressure, resulting in inaccurate judgment of the completion state of nitrogen filling. In addition, existing methods lack analysis of dynamic flow field characteristics such as gas flow velocity distribution, vortex intensity and streamline morphology, limiting the optimization control of the nitrogen filling process. The core challenge lies in the complex changes in gas flow characteristics at the bottle opening when the pressure is balanced, including changes in the uniformity of gas flow velocity distribution, decay characteristics of vortex intensity and stability fluctuations of streamline morphology. These factors are difficult to quantify in real time through traditional detection methods, directly affecting the accurate judgment of the completion state of nitrogen filling. Existing technologies lack dynamic analysis methods for these flow field characteristics, making it difficult to reduce energy consumption and improve production efficiency while ensuring quality. Therefore, how to accurately judge the completion state of the nitrogen filling process based on the gas flow characteristics at the bottle opening and combined with artificial intelligence flow field analysis algorithms has become a key problem in optimizing the energy consumption and production efficiency of the nitrogen filling system. SUMMARY

[0003] The present application provides a bottle nitrogen filling process quality evaluation method applying artificial intelligence, mainly including:

[0004] Obtain the gas flow velocity distribution, vortex intensity and streamline morphology at the bottle opening of the nitrogen filling equipment, perform feature extraction and time series analysis, and generate a gas flow feature dataset;

[0005] Extract the uniformity feature of the gas flow velocity distribution, the decay feature of the vortex intensity and the stability feature of the streamline morphology in the gas flow feature dataset, and perform feature fusion processing to obtain a multi-dimensional vector representation of the bottle opening gas flow feature;

[0006] Perform multi-dimensional vector representation on the gas flow velocity distribution, vortex intensity and streamline morphology, construct a pre-trained flow field dynamic analysis model, predict the dynamic variation law of the gas flow feature under pressure balance, and obtain a stability index of the bottle opening gas flow feature through stability analysis;

[0007] If the gas flow feature stability is lower than a preset threshold, predict the short-term variation trend of the gas flow velocity distribution, vortex intensity and streamline morphology through a long short-term memory network algorithm, and obtain the future evolution state of the gas flow feature.

[0008] adjusting the nitrogen injection rate and angle of the nitrogen filling equipment according to the future evolution state of the gas flow characteristics, generating optimized nitrogen filling control parameters;

[0009] adjusting the nitrogen injection parameters according to the optimized nitrogen filling control parameters, updating the running state of the nitrogen filling equipment in real time, and generating an updated gas flow characteristic data set;

[0010] inputting the updated gas flow characteristic data set into the flow field dynamic analysis model, recalculating the characteristic vectors of the gas flow velocity distribution, vortex intensity and streamline shape, outputting the updated gas flow characteristic stability index, and determining the completion state of the nitrogen filling process;

[0011] If the gas flow characteristic stability index reaches a threshold value, then classify the update index by a logistic regression algorithm, output the nitrogen filling completion probability, generate a termination signal according to the probability, control the nitrogen filling equipment to stop, record the gas flow characteristic data set, and determine the production efficiency improvement amplitude.

[0012] Further, the gas flow velocity distribution, vortex intensity and streamline shape at the bottle opening of the nitrogen filling equipment are obtained, feature extraction and time series analysis are performed, and a gas flow characteristic data set is generated, including: arranging a gas sensor array at the grid points divided by the bottle opening diameter, measuring the gas flow velocity data at each point in the array using an ultrasonic gas flowmeter, collecting the inlet pressure data through a gas pressure sensor, and obtaining a first gas flow data set by denoising and standardizing the sensor data. According to the first gas flow data set, a gas flow velocity field distribution map is constructed, a gas flow probe is used to scan the bottle opening plane to obtain streamline trajectory data, the gas flow field vorticity value is calculated, and a first vortex distribution map is obtained by superimposing the vorticity contour lines on the gas flow velocity field distribution map. According to the first vortex distribution map, the vortex intensity gradient is calculated, the vortex intensity gradient data is hierarchically clustered to obtain the vortex intensity interval division result, and the curvature feature parameters are extracted from the streamline trajectory data. The sensor array is used to collect the gas concentration data during the nitrogen filling process, and a Bayesian probability network is constructed combining the first gas flow data set to extract the gas concentration distribution feature vector from the gas flow velocity field distribution map. The gas flow fluctuation frequency signal sequence is decomposed by db4 wavelet transform to extract the detail coefficient sequence, and the gas flow fluctuation feature data is constructed according to the coefficient sequence. The streamline trajectory curve is superimposed on the first vortex distribution map, the gas diffusion coefficient is calculated according to the gas concentration distribution feature vector, and the flow field structure shape is quantitatively described using the diffusion coefficient to obtain the gas flow characteristic time series data set, forming the gas flow characteristic data set.

[0013] Further, the uniformity of the airflow velocity distribution, the decay of the vortex intensity and the stability of the streamline morphology in the airflow feature data set are extracted, and feature fusion processing is performed to obtain a multi-dimensional vector representation of the bottle mouth airflow feature, including: data cleaning is performed on the airflow feature data set, the median filter is used to remove outliers, and the sliding window method is used to segment the time series data to obtain the preprocessed airflow feature data set. The airflow velocity field variance distribution map of the bottle mouth area is calculated according to the preprocessed airflow feature data set, the first velocity field fluctuation value is calculated by using the gridding variance, and the first airflow uniformity index is obtained by performing maximum and minimum value normalization processing on the fluctuation value. The vortex intensity time series is extracted from the preprocessed airflow feature data set, the difference value of the vortex intensity at adjacent time points is calculated to obtain the first vortex intensity change curve, and the vortex decay parameter is obtained by fitting the change curve using an exponential decay function. The streamline trajectory curve is extracted from the preprocessed airflow feature data set, the curvature change rate is calculated to obtain the first curvature sequence, and the three-dimensional principal feature vector is obtained by dimension reduction processing of the curvature sequence using principal component analysis. An eight-layer recurrent neural network is constructed, the number of input layer nodes is set to three, the number of hidden layer nodes is set to sixteen, the first curvature sequence is used to train the network, and the streamline morphology stability feature vector is obtained by extracting the output layer data of the network. A feature correlation matrix is constructed, the matrix dimension is the same as the sum of the dimensions of the first airflow uniformity index, the vortex decay parameter and the streamline morphology stability feature vector, the correlation coefficient between each two features is calculated to obtain a feature weight vector. The first airflow uniformity index, the vortex decay parameter and the streamline morphology stability feature vector are weighted and summed according to the feature weight vector to obtain a multi-dimensional vector representation of the bottle mouth airflow feature.

[0014] Further, the airflow velocity distribution, vortex intensity and streamline shape are represented by multi-dimensional vectors, a pre-trained flow field dynamic analysis model is constructed, the dynamic change law of the airflow characteristics under pressure balance state is predicted, and the stability index of the bottle mouth airflow characteristics is obtained through stability analysis, including: the normalized method is used to preprocess the airflow velocity distribution data, the median filter is used to remove outliers, and the first airflow data set is obtained by dividing the time window of the preprocessed data. According to the first airflow data set, the velocity distribution feature vector is constructed, the discrete Fourier transform is used to extract the airflow fluctuation frequency spectrum feature, the vortex intensity data is superimposed on the basis of the frequency spectrum feature, the curvature change feature is extracted from the streamline shape parameter to obtain the first multi-dimensional feature vector. The deep autoencoder is pre-trained by using the labeled data, the number of input layer nodes is the same as the dimension of the first multi-dimensional feature vector, the number of intermediate layer nodes is halved, and the second multi-dimensional feature vector is obtained by feature dimension reduction of the pre-trained autoencoder. The residual neural network predictor is constructed, the second multi-dimensional feature vector is used as the input layer, the time sequence feature is extracted by stacking the residual block, the predictor is trained by using the historical data, the flow field dynamic analysis model is obtained, and the airflow dynamic prediction result is output. According to the airflow dynamic prediction result, the bottle mouth pressure change sequence is calculated, the sliding average method is used to smooth the pressure change sequence, and the third multi-dimensional feature vector under the pressure balance state is extracted. The fluctuation amplitude and fluctuation frequency are calculated for the third multi-dimensional feature vector, the phase space reconstruction method is used to calculate the maximum Lyapunov index of the feature sequence. According to the maximum Lyapunov index, the stability interval is divided, the stability threshold is determined by the variance of the feature sequence, and the airflow characteristic stability index is extracted from the third multi-dimensional feature vector.

[0015] Further, if the airflow characteristic stability is lower than a preset threshold, a long short-term memory network algorithm is used to predict short-term change trends of airflow velocity distribution, vortex intensity and streamline shape, to obtain a future evolution state of the airflow characteristic, including: comparing the airflow characteristic stability index with the preset stability threshold, if the stability index is lower than the preset threshold value 0.75, the airflow characteristic data is preprocessed by a standardization method, and first preprocessing data set is obtained by removing abnormal values through median filtering. The first preprocessing data set is decomposed by four layers using wavelet transform, airflow velocity distribution features are extracted from the wavelet coefficients, vortex intensity time series features are calculated by autocorrelation function, and a first feature vector is constructed combining the streamline curvature change data. A long short-term memory network predictor is constructed, the number of input layer nodes is the same as the dimension of the first feature vector, the hidden layer contains three memory units, the number of each memory unit is 64, and the number of output layer nodes is the same as the dimension of the prediction target. The first feature vector is divided into sequences by a sliding window method, the window length is set to 50 sampling points, the sliding step is 10 sampling points, and training data sequences are generated. The long short-term memory network predictor is trained according to the training data sequences, the network parameters are optimized by a back propagation algorithm, the prediction error is calculated on the validation data set to obtain a first prediction result. The future 100 sampling points of the airflow velocity distribution trend, the vortex intensity change law and the streamline shape evolution feature are extracted from the first prediction result, the features are fused by a weighted average method to obtain a second feature vector. The airflow characteristic evolution sequence is constructed according to the second feature vector, the evolution sequence is processed by an exponential smoothing method, and the future evolution state of the airflow characteristic is obtained from the smoothed sequence.

[0016] Further, the nitrogen injection rate and injection angle of the nitrogen filling equipment are adjusted according to the future evolution state of the gas flow characteristics, and optimized nitrogen filling control parameters are generated, including: normalizing the gas flow characteristic evolution state data, removing abnormal values by using a median filter, and obtaining a first evolution data set by segmenting the processed data by a sliding window method. According to the first evolution data set, a gas flow velocity distribution sequence is extracted, and the velocity distribution frequency spectrum characteristics are calculated by using Fourier transform. The gas flow fluctuation period data are obtained from the frequency spectrum characteristics, and the nitrogen injection rate range is constructed for the fluctuation period. The vortex intensity sequence is reconstructed by using a phase space reconstruction method, the embedding dimension is set to 3, the time delay is set to 5, the first vortex evolution law is obtained by calculating the curvature radius of the reconstructed trajectory, and the injection angle range is constructed according to the first vortex evolution law. The angle adjustment sequence is generated by using a uniform division method, and the multi-objective optimization function is constructed in combination with the injection rate range. The particle swarm optimizer parameters are set, the particle number is 50, the maximum iteration number is 100, the inertia weight is 0.7, the learning factor is 2, the multi-objective optimization function is solved, and the first control parameter combination is obtained. The first control parameter combination is predicted for gas flow stability by using a recurrent neural network, and the parameter adjustment sequence is generated by using a control parameter update rule. The gas flow stability index is calculated according to the parameter adjustment sequence, the stability index is dynamically corrected by using a Kalman filter, and the optimized nitrogen filling control parameters are extracted from the corrected data.

[0017] Further, the nitrogen injection parameters are adjusted according to the optimized nitrogen filling control parameters, the running state of the nitrogen filling equipment is updated in real time, and an updated airflow characteristic data set is generated, including: generating an injection parameter adjustment instruction according to the optimized nitrogen filling control parameters, adjusting the nitrogen injection rate using a proportional-integral regulator, setting the integral time constant to one quarter of the response period, collecting bottle mouth airflow data in real time through a sensor array, and obtaining a first airflow data set by performing Gaussian filtering on the collected data. A grid variance distribution is calculated for the first airflow data set, a regional block statistical method is used to obtain a first velocity field uniformity index, and a difference value of adjacent sampling points is extracted from a vortex intensity sequence to obtain a first decay rate curve. A curve slope sequence is calculated according to the first decay rate curve, a least squares method is used to fit the curve slope trend, and a first response characteristic vector is constructed in combination with a streamline curvature parameter. A support vector regressor is used to extract features from the first response characteristic vector, the input feature dimension is the same as the vector dimension, a radial basis function is selected as the kernel function, and a second response characteristic vector is obtained from the regression result. Dynamic response judgment rules are set for the second response characteristic vector, if the velocity field uniformity index is greater than a preset threshold, it is determined that the velocity distribution is uniform, and if the vortex intensity decay rate is within a preset interval, it is determined that the vortex decay is stable. A sliding window method is used to update the streamline morphology parameters in real time, the window length is set to twice the sampling period, a first stability index is obtained by calculating the parameter fluctuation amplitude in the window. A Kalman filter is constructed according to the first stability index, the measurement noise covariance matrix is obtained by statistical analysis of historical data, and an updated airflow characteristic data set is obtained by filtering.

[0018] Further, the updated airflow characteristic data set is input into a flow field dynamic analysis model, the characteristic vectors of airflow velocity distribution, vortex intensity and streamline morphology are recalculated, and an updated airflow characteristic stability index is output, so as to determine the completion state of the nitrogen charging process, including: performing outlier detection according to the updated airflow characteristic data set, removing abnormal data points by adopting the three standard deviation principle, and obtaining a first processed data set by smoothing the data through a median filter. The first processed data set is decomposed by four layers by adopting wavelet transform, the low-frequency component and the high-frequency component of the airflow velocity distribution data are extracted, and a 16-dimensional velocity distribution characteristic vector is constructed from the decomposition coefficients. The vortex intensity change rate is calculated for the 16-dimensional velocity distribution characteristic vector, the change rate sequence is smoothed by adopting an exponential weighting method, the upper limit value and the lower limit value of the adaptive threshold interval are set, and a 32-dimensional vortex characteristic vector is obtained. A long short-term memory network is constructed, the number of input layer nodes is 32, the hidden layer contains three memory units, the number of units in each layer is 64, and the number of output layer nodes is 16, so as to dynamically predict the vortex characteristic vector. The streamline morphology parameters are extracted according to the prediction result, the principal component analysis method is adopted to reduce the dimension of the parameters, the principal components with a contribution rate higher than a preset threshold value are retained to obtain an 8-dimensional streamline characteristic vector. The velocity distribution characteristic vector, the vortex characteristic vector and the streamline characteristic vector are fused by adopting a weighted summation method, the weight coefficients are obtained by statistical analysis of historical data, and an airflow characteristic stability index is obtained. A five-layer recurrent neural network is constructed to track the stability index change, and if the stability index is higher than a preset completion threshold value in ten consecutive sampling periods, it is determined that the nitrogen charging process is completed.

[0019] Further, if the airflow characteristic stability index reaches the threshold value, the update index is classified by a logistic regression algorithm, the nitrogen filling completion probability is output, the termination signal is generated according to the probability, the nitrogen filling equipment is controlled to stop, the airflow characteristic data set is recorded, and the production efficiency improvement range is determined, including: comparing the airflow characteristic stability index with the preset threshold value, if the stability index reaches the preset threshold value, a sliding window with a length of 100 sampling points is used to segment the stability index, a window overlap rate is set to 50% to obtain a first index sequence. The first index sequence is subjected to feature extraction, and the mean, variance, skewness, kurtosis, maximum value, minimum value, and peak-valley difference of the sequence are calculated to construct a feature vector. The seven-dimensional feature vector is classified by using a logistic regression algorithm, and the classification result is mapped to a probability value by using a sigmoid function. If the probability value exceeds the preset probability threshold value for five consecutive sampling periods, a termination signal is generated. The device control instruction is constructed according to the termination signal, and a proportional-integral controller is used for execution. The integral time constant is set to one-fourth of the sampling period, and the proportional coefficient is obtained by statistical analysis of historical data. The airflow characteristic data is recorded by the data acquisition unit, the sampling frequency is set to 200 Hz, the recording time covers the complete nitrogen filling process, and the process parameter data set is generated. The airflow velocity uniformity index, vortex intensity stability index, and flow line shape continuity index are extracted from the process parameter data set to construct an eight-dimensional feature vector. The eight-dimensional feature vector is classified by using a support vector machine, a radial basis function is selected as the kernel function, the kernel parameters are optimized by using a grid search method, and the efficiency improvement index is extracted from the classification result.

[0020] The technical scheme provided by the embodiment of the application can include the following beneficial effects:

[0021] The application discloses a bottle nitrogen filling process quality evaluation method applying artificial intelligence, obtains airflow velocity distribution, vortex intensity and flow line shape data at a bottle opening, extracts features and performs multi-dimensional vector representation, constructs a flow field dynamic analysis model to predict airflow characteristic change law. When the airflow characteristic stability is lower than a threshold value, a long short-term memory network algorithm is used to predict a short-term change trend, and nitrogen injection parameters are adjusted accordingly. The application updates the airflow characteristic data set in real time, recalculates the feature vector and outputs the stability index until the preset threshold value is reached. Finally, the nitrogen filling completion probability is judged by a logistic regression algorithm, and a termination signal is generated. The method can realize dynamic optimization control of the nitrogen filling process, improve production efficiency and ensure stability of the nitrogen filling quality. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A flowchart of the bottle nitrogen filling process quality evaluation method applying artificial intelligence.

[0023] Figure 2A schematic diagram of a bottle nitrogen filling process quality evaluation method using artificial intelligence according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to further understand the content of the present application, the present application will be described in detail with reference to the accompanying drawings and examples. It can be understood that the specific embodiments described herein are only used to explain the related application, and not to limit the application. In addition, it should be noted that, in order to facilitate description, only the parts related to the application are shown in the drawings.

[0025] In the embodiment of the present application, the nitrogen filling equipment is used to inject nitrogen before sealing the bottle, which replaces oxygen to prolong the shelf life of the product. The airflow characteristics at the bottle opening, including airflow velocity distribution, vortex intensity and streamline shape, directly affect the nitrogen filling effect. The traditional method is difficult to quantize these characteristics in real time, while the present application realizes dynamic analysis through artificial intelligence algorithm, which improves the evaluation accuracy. The following embodiments do not make too many limitations on the implementation details of the technical solutions, which can be determined by the technical personnel according to the actual scene.

[0026] Figure 1 The flowchart of the bottle nitrogen filling process quality evaluation method using artificial intelligence provided by the embodiment of the present application is shown, and the details are as follows:

[0027] S101 obtains the airflow velocity distribution, vortex intensity and streamline shape data at the bottle opening, performs feature extraction and time series analysis, and generates airflow feature data set containing time series.

[0028] In the embodiment of the present application, after the nitrogen filling equipment is started, the airflow dynamic characteristics at the bottle opening need to be monitored in real time to evaluate the nitrogen filling process quality. The generation of airflow feature data set is the basis for subsequent analysis, involving multi-dimensional sensor data acquisition and intelligent processing. The specific implementation of this step includes the following sub-steps:

[0029] As Figure 2 S1011, deploy sensor array to collect airflow velocity and pressure data at the bottle opening, perform noise reduction and standardization processing, and generate first airflow data set.

[0030] In the embodiment of the present application, ultrasonic gas flow sensors and air pressure sensors are arranged at the 4x4 grid points divided according to the diameter of the bottle opening, the sampling frequency is set to 200 Hz, and the airflow velocity and inlet pressure data are collected. The original data is affected by environmental noise and equipment vibration, and a Butterworth low-pass filter is used for noise reduction, with a cutoff frequency of 80 Hz to filter out high-frequency interference. The data after noise reduction is processed by zero-mean standardization to generate the first airflow data set, ensuring data consistency and reliability. The sensor array covers all grid points of the bottle opening, ensuring the comprehensiveness of data acquisition.

[0031] S1012 construct a flow velocity field distribution map based on the first airflow data set, generate a first vortex distribution map in combination with streamlines and vorticity data, and extract flow field features.

[0032] In the embodiment of the present application, based on the first airflow data set, the Kriging interpolation method is used to perform spatial interpolation on the airflow velocity data, the interpolation grid density is 0.5 millimeters, a flow velocity field distribution map is generated to reflect the uniformity of the velocity distribution at the bottle mouth. The gas flow probe is scanned in a plane 5 millimeters away from the bottle mouth with an interval of 1 millimeter, and the streamlines data are recorded to generate streamlines. The airflow field vorticity value is calculated, ranging from-50 to 50, 10 contour lines are set with an interval of 10, and are superimposed on the velocity field distribution map to form a first vortex distribution map. For the vortex intensity data, the improved K-means clustering algorithm is used for hierarchical clustering, the number of cluster centers is 5, and the vortex intensity interval division result is obtained. The curvature feature is extracted from the streamlines, the curvature radius ranges from 2 to 20 millimeters, and the airflow deflection degree is reflected.

[0033] S1013 collect gas concentration data, construct a Bayesian probability network, extract a concentration distribution feature vector, and generate an airflow feature time series data set in combination with a diffusion coefficient.

[0034] In the embodiment of the present application, the gas concentration data during the nitrogen filling process is collected by the sensor array, the center concentration of the bottle mouth is as high as 95%, and the edge area decreases to 75%. In combination with the first airflow data set, a Bayesian probability network is constructed, a gas concentration distribution feature vector is extracted, the dimension is 8, including mean, variance, skewness and other statistical features. Based on the concentration distribution feature vector, the gas diffusion coefficient is calculated, and the diffusion coefficient reaches 2.5 when the vortex intensity is 30. The db4 wavelet transform is performed on the airflow pulsation frequency signal, 4 layers are decomposed, the detail coefficient sequences of 0-25 Hz, 25-50 Hz, 50-75 Hz and 75-100 Hz are extracted, and an airflow pulsation feature data set is constructed. The streamlines curve is superimposed on the first vortex distribution map, and the flow field structure is quantitatively described in combination with the diffusion coefficient to generate a time series data set containing velocity field, vortex intensity, concentration distribution and other features.

[0035] In the embodiment of the present application, through multi-dimensional sensor data acquisition and intelligent processing, the generated airflow feature data set comprehensively reflects the dynamic characteristics of the bottle mouth airflow. Compared with the traditional method, this step can capture the uniformity of the airflow velocity distribution, the spatial layering of the vortex intensity and the curvature change of the streamlines in real time, providing high-quality data support for subsequent analysis. The finally generated time series data set lays a foundation for dynamic optimization and completion state judgment of the nitrogen filling process, and significantly improves the evaluation accuracy and production efficiency.

[0036] S102 extracts the uniformity feature of the airflow velocity distribution, the decay characteristic of the vortex intensity, and the stability feature of the streamline form in the airflow feature data set, performs feature fusion processing, and generates a multi-dimensional vector representation of the bottle mouth airflow feature.

[0037] In the embodiment of the application, the airflow feature data set contains time series data of airflow velocity, vortex intensity and streamline form at the bottle mouth. To achieve accurate evaluation of the nitrogen filling process, key features need to be extracted and fused to generate a multi-dimensional vector representation to reflect the airflow dynamic characteristics. This step comprehensively quantifies the airflow features through data cleaning, feature extraction and weighted fusion, providing high-quality input for subsequent analysis. The specific implementation of this step includes the following sub-steps:

[0038] S1021 cleans and segments the airflow feature data set to generate a preprocessed airflow feature data set.

[0039] In the embodiment of the application, the airflow feature data set is affected by equipment vibration and environmental noise, and needs to be cleaned to ensure analysis accuracy. A 5-point median filter is used to remove outliers, with a filter window length of 5 sampling points to effectively smooth short-term fluctuations. Based on the time series characteristics, the data set is segmented by the sliding window method, with a window length of 100 milliseconds and an overlap rate of 50%, generating a preprocessed airflow feature data set. The segmentation process preserves the temporal continuity of the data while reducing the interference of noise on feature extraction, providing a reliable data foundation for subsequent analysis.

[0040] S1022 calculates the airflow velocity field uniformity feature based on the preprocessed airflow feature data set to generate a first airflow uniformity index.

[0041] In the embodiment of the application, the uniformity of airflow velocity distribution is a key indicator for evaluating the stability of nitrogen filling. The bottle mouth area is divided into an 8x8 grid, the variance of airflow velocity in each grid is calculated, a variance distribution map is generated to reflect the local velocity fluctuation characteristics. The variance value ranges from 0.2 to 2.5, and the maximum and minimum value normalization processing is used to map the variance value to the 0 to 1 interval to obtain the first airflow uniformity index. The closer the uniformity index value is to 0, the more uniform the airflow distribution, and the higher the stability of the nitrogen filling process. This step quantifies the spatial distribution characteristics of the velocity field through gridding calculation and normalization processing, providing an important input for feature fusion.

[0042] S1023 extracts the vortex intensity time series, fits the decay characteristic, and generates a vortex decay parameter.

[0043] In the embodiment of the present application, the attenuation characteristics of vortex intensity reflect the dynamic evolution of the nitrogen filling process. The time series of vortex intensity is extracted from the preprocessed airflow characteristic data set. Within a 200-millisecond time window, the vortex intensity gradually decreases from the initial value of 85 to 25. The difference in vortex intensity between adjacent sampling points is calculated to generate a first vortex intensity change curve. An exponential decay function is used to fit the curve, with the function form being I(t) = I0*e (-kt) , where I0 is the initial intensity and k is the attenuation coefficient. By least squares optimization fitting, the attenuation coefficient k = 0.015 is obtained, and the goodness of fit is 0.92, indicating that the vortex intensity presents a significant exponential decay trend. The vortex decay parameter k is used as a characteristic input, which quantifies the dynamic change law of the vortex.

[0044] S1024 extracts the curvature feature of the streamline trajectory, and generates a streamline morphology stability feature vector combining a recurrent neural network.

[0045] In the embodiment of the present application, the stability of the streamline morphology reflects the complexity of the airflow movement, which directly affects the nitrogen filling effect. The streamline trajectory curve is extracted from the preprocessed airflow characteristic data set, and the curvature change rate is calculated to generate a first curvature sequence. The curvature change rate in the center area of the bottle opening is small, with a maximum value of not more than 0.05; the curvature change in the edge area is severe, with a maximum value of 0.15. Principal component analysis is used to reduce the dimension of the curvature sequence, and the cumulative contribution rate of the first three principal components is 85%, generating a three-dimensional principal feature vector. An eight-layer recurrent neural network is constructed, with 3 nodes in the input layer to receive the principal feature vector; the number of nodes in each hidden layer is 16, and a long short-term memory unit structure is used to capture the time dependence. 250 groups of labeled data are used to train the network, with a learning rate of 0.001 and 1000 iterations of training. After the network converges, an 8-dimensional streamline morphology stability feature vector is extracted from the output layer. This step accurately describes the dynamic stability of the streamline morphology through neural network deep learning.

[0046] S1025 constructs a feature correlation matrix, calculates a feature weight vector, performs weighted fusion, and generates a multi-dimensional vector representation of the bottle opening airflow feature.

[0047] In the embodiment of the present application, in order to synthesize the airflow characteristics, the uniformity index, the vortex attenuation parameter and the flow line stability characteristics need to be fused. A 12x12 feature correlation matrix is constructed, and the matrix elements are the Pearson correlation coefficients between the first airflow uniformity index, the vortex attenuation parameter and the flow line shape stability characteristic vector. The features with an absolute correlation coefficient greater than 0.7 have significant correlation, indicating that there is a strong interaction between the features. Based on the correlation coefficient, the feature weight vector is calculated, the uniformity index weight is 0.35, the vortex attenuation parameter weight is 0.25, and the flow line stability characteristic weight is 0.4. The three types of feature vectors are fused by weighted summation method according to the weight, and a 12-dimensional airflow feature vector is generated. The vector comprehensively describes the uniformity of the airflow velocity distribution, the attenuation characteristics of the vortex intensity and the stability of the flow line shape, and provides a high-quality feature representation for subsequent nitrogen charging state evaluation.

[0048] In the embodiment of the present application, through multi-dimensional feature extraction and fusion, the generated bottle mouth airflow feature multi-dimensional vector can accurately reflect the dynamic characteristics of the nitrogen charging process. Compared with the traditional method, this step significantly improves the accuracy and robustness of feature extraction through median filtering, gridding calculation and neural network analysis. The application of the feature correlation matrix further optimizes the feature weight distribution, ensures the representativeness of the fused features, lays a solid foundation for subsequent dynamic analysis and device optimization, and finally improves the reliability and production efficiency of the nitrogen charging quality evaluation.

[0049] S103 Based on the multi-dimensional vector representation of the bottle mouth airflow characteristics, a pre-trained flow field dynamic analysis model is constructed to predict the dynamic change law of the airflow characteristics under pressure balance state, and an airflow characteristic stability index is generated through stability analysis.

[0050] In the embodiment of the present application, based on the multi-dimensional vector representation of the bottle mouth airflow characteristics, a flow field dynamic analysis model needs to be further constructed to predict the dynamic evolution trend of the airflow velocity distribution, the vortex intensity and the flow line shape under the pressure balance state, and to quantify the stability of the airflow characteristics through stability analysis. This step realizes the accurate evaluation of the dynamic characteristics of the nitrogen charging process through feature preprocessing, deep learning model construction and stability index calculation, and provides a scientific basis for subsequent optimization control. The specific implementation of this step includes the following sub-steps:

[0051] S1031 The airflow characteristic data is preprocessed, the fluctuation spectrum features are extracted and the vortex and flow line features are fused to generate a first multi-dimensional feature vector.

[0052] In the embodiment of the present application, the airflow velocity distribution data contains noise and abnormal fluctuations, and needs to be preprocessed to ensure the quality of the model input. The data is mapped to the 0-1 interval using the normalization method, and the short-term fluctuations are removed by using a 5-point median filter with a window length of 200 milliseconds. The sliding window interval is set to 50 milliseconds, and the first airflow data set containing 2000 sampling points is generated by segmentation. For this data set, the airflow fluctuation spectrum features are extracted by using a 1024-point fast Fourier transform, and three main frequency peaks of 15 Hz, 35 Hz and 65 Hz are identified in the range of 0-100 Hz. A 16-dimensional spectrum component vector is generated. In order to integrate the flow field characteristics, the vortex intensity data and the curvature change characteristics of the streamline shape are fused. The vortex intensity sequence reflects the dynamic decay trend, and the streamline curvature change rate is less than 0.05 in the central region and as high as 0.15 in the edge region. Finally, a 32-dimensional first multi-dimensional feature vector is generated, which contains spectrum components, vortex intensity and curvature features, and comprehensively describes the airflow dynamic characteristics.

[0053] S1032 utilizes the deep autoencoder to perform dimension reduction processing on the first multi-dimensional feature vector to generate a second multi-dimensional feature vector.

[0054] In the embodiment of the present application, in order to reduce the feature dimension and retain the core information, a deep autoencoder is used for feature compression. The deep autoencoder is designed as a symmetrical structure, the input layer has 32 nodes corresponding to the dimension of the first multi-dimensional feature vector, the middle encoding layer has 16 nodes to realize feature compression, and the decoding layer has 32 nodes for feature reconstruction. 1500 labeled data are used for pre-training, the training round is set to 500, the optimization target is to minimize the mean square error, and the error after training is reduced to 0.015, indicating that the feature extraction effect is good. After pre-training is completed, the first multi-dimensional feature vector is input into the autoencoder, and a 16-dimensional second multi-dimensional feature vector is extracted through the middle encoding layer. This vector retains the main information of the airflow feature, while reducing the computational complexity, and provides an efficient input for the subsequent prediction model.

[0055] S1033 constructs a residual neural network predictor to predict the dynamic change trend of the airflow and generates an airflow dynamic prediction result.

[0056] In the embodiment of the present application, to capture the time dependence of the airflow characteristics, a residual neural network predictor is constructed. The network contains 8 residual blocks, each consisting of two layers of convolutional layers and a shortcut connection. The input layer receives a 16-dimensional second multi-dimensional feature vector, and the output layer predicts the airflow change trend within the next 100 milliseconds. 1000 groups of historical data covering different working conditions are used for training, the learning rate is set to 0.001, the training iteration is 2000 times, and the prediction mean square error is controlled within 0.025. After training is completed, the network can accurately predict the short-term evolution trend of the airflow velocity distribution, vortex intensity and streamline shape, and generate the airflow dynamic prediction result. The result reflects the dynamic change law of the airflow characteristics under pressure balance state, providing data support for stability analysis.

[0057] S1034 Calculate the pressure change sequence and stability index based on the airflow dynamic prediction result, and generate the airflow characteristic stability index.

[0058] In the embodiment of the present application, the pressure balance state is the key to evaluate the stability of nitrogen filling. Based on the airflow dynamic prediction result, the pressure change sequence of the bottle mouth is calculated, and the 15-point sliding average method is used for smoothing processing with a window length of 150 milliseconds. When the pressure fluctuation amplitude is less than 0.05 bar and the duration is more than 300 milliseconds, it is determined that the pressure balance state is reached, and the 24-dimensional feature vector under this state is extracted to describe the airflow velocity, vortex and streamline characteristics. The phase space reconstruction method is used to analyze the dynamic characteristics of the feature vector, the embedding dimension is set to 3, the time delay is 5 sampling points, the maximum Lyapunov exponent is calculated, and the value is 0.15, indicating that the system has certain instability. To quantify the stability, the variances of the feature sequence at different time scales are calculated, and the stability threshold is set to 0.08. When the variance of the feature sequence is less than the threshold and the maximum Lyapunov exponent is less than 0.1, it is determined that the airflow is in a stable state. Finally, the airflow characteristic stability index is generated, with a value range of 0 to 1, and in actual application, the value is distributed between 0.75 and 0.95. The closer the value is to 1, the more stable the airflow characteristics are.

[0059] In the embodiment of the present application, through the combination of deep autoencoder and residual neural network, the accurate prediction of the dynamic change of airflow characteristics is realized. The phase space reconstruction and Lyapunov exponent analysis further quantify the stability of the airflow system, which significantly improves the evaluation accuracy compared with traditional methods. The generated stability index provides a reliable basis for the completion state judgment of the nitrogen filling process, and lays a foundation for optimizing the nitrogen injection parameters, helping to improve production efficiency and reduce energy consumption.

[0060] S104 If the airflow characteristic stability index is lower than the preset threshold, use the long short-term memory network algorithm to predict the short-term change trend of the airflow velocity distribution, vortex intensity and streamline shape, and generate the future evolution state of the airflow characteristics.

[0061] In the embodiment of the present application, when the gas flow characteristic stability index is lower than the preset threshold value, it indicates that there is fluctuation in the gas flow state during the nitrogen charging process, and the short-term evolution trend needs to be predicted to optimize the equipment control. This step generates the future evolution state of the gas flow characteristics through data preprocessing, feature extraction, long short-term memory network prediction and feature fusion, to provide basis for dynamically adjusting the nitrogen injection parameters. The specific implementation of this step includes the following sub-steps:

[0062] S1041, the gas flow characteristic data is preprocessed, and the wavelet transform is used to extract the features to generate the first feature vector.

[0063] In the embodiment of the present application, the gas flow characteristic data is affected by measurement noise and environmental interference, and needs to be preprocessed to improve the data quality. The standardized method is used to map the data to the interval of-1 to 1, and the 15-point median filter is used to remove abnormal fluctuations to generate the first preprocessed data set. For this data set, db4 wavelet basis function is used for four-layer wavelet transform, and four frequency scales of 0-25 Hz, 25-50 Hz, 50-75 Hz and 75-100 Hz are obtained. The detail coefficients and an approximate coefficient are obtained. The detail coefficients capture the dynamic change characteristics of the gas flow velocity distribution in different frequency bands. The correlation coefficient of the vortex intensity time series is calculated by the autocorrelation function, and the correlation length is set to 50 sampling points to quantify the time correlation of the vortex intensity. Combined with the curvature change rate of the streamline trajectory, the curvature change in the central region is less than 0.05, and the edge region is as high as 0.12, a 48-dimensional first feature vector is constructed, which contains the gas flow velocity distribution, vortex intensity and streamline shape characteristics, to provide comprehensive input for subsequent prediction.

[0064] S1042, the training data sequence is generated based on the first feature vector, the long short-term memory network predictor is trained, and the first prediction result is generated.

[0065] In the embodiment of the present application, in order to capture the time sequence dependence of the airflow characteristics, a long short-term memory network predictor is constructed. The network includes an input layer, three hidden layers and an output layer, the number of nodes of the input layer is 48, which is consistent with the dimension of the first feature vector; each hidden layer includes 64 memory units, which uses a forgetting gate, an input gate and an output gate to control the information flow, the threshold of the forgetting gate is set to 0.5, the activation function is sigmoid, and the state update uses the tanh function; the output layer has 32 nodes corresponding to the dimension of the prediction target. The sliding window method is used to divide the sequence of the first feature vector, the window length is 50 sampling points, the step length is 10 sampling points, and the adjacent windows overlap by 40 sampling points, generating 2000 training samples. The network is trained using the stochastic gradient descent method, the batch size is 32, the initial learning rate is 0.01, and the learning rate is reduced by 0.9 times every 50 rounds. After 1000 rounds of training, the mean square error of the validation set is reduced to 0.015, indicating that the network converges. After training is completed, the network predicts the trend of airflow characteristic changes in the next 100 sampling points to generate the first prediction result, and the prediction standard deviation is controlled within 0.08, and the relative error of the vortex intensity is less than 5%.

[0066] S1043 extracts the trend characteristics of the first prediction result, fuses to generate a second feature vector, and constructs and smooths the airflow characteristic evolution sequence.

[0067] In the embodiment of the present application, the first prediction result includes airflow velocity distribution trend, vortex intensity change rule and flow line shape evolution characteristics. The standard deviation of the velocity distribution, the periodic variation amplitude of the vortex intensity and the flow line curvature change rate of the future 100 sampling points are extracted, the prediction accuracy rate of the velocity distribution is 90%, and the prediction coincidence degree of the flow line shape is 92%. The features are fused by weighted average method, and the weights are distributed as follows: velocity distribution 0.4, vortex intensity 0.3, and flow line shape 0.3, to generate a 32-dimensional second feature vector. Based on the second feature vector, an airflow characteristic evolution sequence is constructed, and an exponential smoothing method is used for processing, with a smoothing coefficient of 0.85 to enhance the continuity of the sequence trend. The smoothed sequence shows that the airflow velocity fluctuation gradually increases within 300 milliseconds, the vortex intensity presents periodic oscillation, and the flow line shape appears distorted in the edge area. The prediction result is consistent with the actual observation with a coincidence degree of 88%, providing a reliable basis for evaluating the future evolution state of the airflow.

[0068] In the embodiment of the present application, the stability threshold is set to 0.75, and based on the statistical data, when the index is lower than the value, the prediction process is triggered to deal with the airflow fluctuation. The application of wavelet transform and long short-term memory network significantly improves the ability to capture the dynamic characteristics of the airflow, and compared with the traditional method, the prediction result more accurately reflects the short-term change trend of the airflow velocity, vortex intensity and flow line shape. The generated evolution state provides data support for optimizing the operating parameters of the nitrogen charging equipment, which helps to reduce energy consumption and improve production efficiency.

[0069] S105 adjusts the nitrogen injection rate and angle of the nitrogen filling equipment according to the future evolution state of the airflow characteristics, and generates optimized nitrogen filling control parameters.

[0070] In the embodiment of the application, the future evolution state of the airflow characteristics reflects the dynamic trend of the airflow velocity, vortex intensity and streamline shape during the nitrogen filling process, and the nitrogen injection parameters need to be optimized accordingly to ensure process stability. The specific implementation of this step includes the following sub-steps:

[0071] S1051 preprocesses the airflow characteristic evolution state data, extracts the airflow fluctuation period and vortex evolution law, and generates a first evolution data set and related features.

[0072] In the embodiment of the application, the airflow characteristic evolution state data includes airflow velocity, vortex intensity and streamline shape features, which are affected by measurement noise and need to be preprocessed to improve data quality. The data is mapped to the 0-1 interval using the normalization method, and the 11-point median filter is used to smooth abnormal fluctuations. The sliding window method is used for segmentation, the window length is 100 sampling points, and the overlap rate is 50%, to generate a first evolution data set. The airflow velocity distribution sequence is extracted from the data set, the 1024-point fast Fourier transform is used to calculate the frequency spectrum features, the main frequency components are identified at 15 Hz, 35 Hz and 55 Hz, the amplitude is 0.8, 0.6 and 0.4 respectively, the airflow fluctuation period is determined, and the nitrogen injection rate range is set to 2-8 meters per second. To analyze the vortex dynamic characteristics, the phase space reconstruction method is used to process the vortex intensity sequence, the embedding dimension is set to 3, the time delay is 5 sampling points, the reconstructed trajectory is spiral-shaped, the curvature radius range is 5-20 mm, reflecting the periodic change of the vortex intensity, and a first vortex evolution law is generated. This step accurately captures the airflow dynamic characteristics through frequency spectrum analysis and phase space reconstruction, providing a data basis for parameter optimization.

[0073] S1052 constructs a multi-objective optimization function, uses a particle swarm algorithm to optimize the injection rate and angle, and generates a first control parameter combination.

[0074] In the embodiment of the present application, based on the first vortex evolution law and the fluctuation period data, the nitrogen injection angle range is 15 to 45 degrees, the step is 5 degrees, and the injection rate range is 2 to 8 meters per second. A multi-objective optimization function is designed, including three sub-targets of airflow uniformity, vortex stability and streamline continuity, with weights of 0.4, 0.3 and 0.3 respectively. Particle swarm optimization algorithm is used for solving, setting particle number as 50, maximum iteration number as 100, inertia weight as 0.7 and learning factor as 2. The algorithm updates the particle position through iteration, optimizes the objective function value, and converges to 0.92, obtaining the optimal injection rate of 5.5 meters per second and the injection angle of 32 degrees, forming the first control parameter combination. This step balances the mutual restraint relationship of airflow characteristics through multi-objective optimization, ensuring the overall performance of the control parameters.

[0075] S1053 predicts the airflow stability of the control parameters using a recurrent neural network, combines Kalman filter to correct the stability index, and generates optimized nitrogen charging control parameters.

[0076] In the embodiment of the present application, in order to verify the effect of the first control parameter combination, a three-layer recurrent neural network is constructed to predict the airflow stability. The network input layer has 16 nodes to receive control parameters and airflow characteristics, the hidden layer has 32 nodes to capture the time sequence relationship, and the output layer has 8 nodes to predict the stability trend. Using 1000 sets of labeled data for training, the learning rate is set to 0.001, and the prediction accuracy reaches 90%. When the predicted stability is lower than 0.8, adaptive adjustment is triggered, the injection rate adjustment step is 0.5 meters per second, the angle adjustment step is 3 degrees, and the parameter adjustment sequence is generated. Based on the sequence, the airflow stability index is calculated, and the Kalman filter is used for dynamic correction, with the measurement noise variance set to 0.01 and the process noise variance set to 0.005. After correction, the stability index shows that the airflow velocity fluctuation amplitude is reduced by 45%, the vortex intensity change amplitude is reduced by 38%, the streamline form maintains continuity, and the index value is maintained above 0.85. The optimized nitrogen charging control parameters extracted from the corrected data are suitable for various working conditions.

[0077] In the embodiment of the present application, the optimized control parameters accurately capture the dynamic changes of airflow characteristics, ensuring the stability of the nitrogen charging process. Compared with the traditional method of relying on experience adjustment, this step combines Fourier transform, particle swarm optimization and neural network prediction, improving the scientificity and adaptability of parameter optimization. The optimization result effectively reduces the system fluctuation, improves the production efficiency, and reduces the nitrogen consumption, providing an efficient solution for the nitrogen charging process in food, medicine and other industries.

[0078] S106 adjusts the nitrogen injection parameters according to the optimized nitrogen filling control parameters, updates the running state of the nitrogen filling equipment in real time, generates an updated airflow characteristic data set, and analyzes the dynamic response characteristics of the nitrogen filling process, including airflow velocity distribution uniformity, vortex intensity decay rate, and flow line stability.

[0079] In the embodiment of the present application, the optimized nitrogen filling control parameters are used to dynamically adjust the nitrogen injection rate and angle to ensure the stability of the nitrogen filling process. By real-time collection of bottle mouth airflow data, characteristic analysis and dynamic response evaluation, an updated airflow characteristic data set is generated to quantify the airflow dynamic characteristics and provide a basis for continuous optimization of equipment operation. The specific implementation of this step includes the following sub-steps:

[0080] S1061 collects bottle mouth airflow data and performs preprocessing to generate a first airflow data set, extracts velocity field uniformity indicators and decay rate characteristics.

[0081] In the embodiment of the present application, when the nitrogen filling equipment is running, the sensor array collects bottle mouth airflow velocity, pressure and flow line data in real time at a sampling frequency of 200 Hz. The original data is affected by random noise, and a Gaussian filter is used for processing, with a standard deviation of 1.5 and a window length of 15 sampling points. This effectively smooths short-term fluctuations and generates a first airflow data set. The bottle mouth area is divided into an 8x8 grid with a grid size of 5mm x 5mm and an overlap rate of 25%. The standard deviation of airflow velocity in each grid is calculated to generate a grid variance distribution. When the standard deviation of more than 90% of the grids is less than 0.2, the velocity field distribution is determined to be uniform, and a first velocity field uniformity indicator is generated. The difference between adjacent sampling points is extracted from the vortex intensity sequence with a sampling interval of 5 milliseconds to generate a first decay rate curve reflecting the dynamic change rate of vortex intensity. This step provides a reliable foundation for dynamic response characteristic evaluation through high-precision data collection and grid analysis.

[0082] S1062 constructs a response characteristic vector, extracts features using a support vector regressor, generates a second response characteristic vector, and judges the dynamic response characteristics.

[0083] In the embodiment of the application, based on the first attenuation rate curve, a 5-order polynomial is used to fit the slope change trend, the goodness of fit reaches 0.95, and a 16-dimensional first response characteristic vector is constructed by combining the streamline curvature parameter, which contains the uniformity of the velocity field, the vortex attenuation and the streamline shape characteristics, and each dimension is normalized to 0 to 1. A support vector regressor is used for feature extraction, a radial basis kernel function is set, the gamma parameter is 0.1, the relaxation factor C is 10, 1000 groups of training samples are used, the cross-validation error is less than 0.05, and an 8-dimensional second response characteristic vector is generated. The dynamic response judgment rule is set: the velocity field uniformity index is greater than 0.85, the velocity distribution is judged to be uniform; the vortex intensity attenuation rate is between 0.2 and 0.4, the vortex attenuation is judged to be stable; and the streamline shape stability index is greater than 0.8, the streamline shape is judged to be stable. Through regression analysis and rule judgment, the airflow dynamic response characteristics are accurately quantified in this step.

[0084] S1063 updates the airflow characteristic data set in real time, optimizes the data by combining Kalman filtering, and generates the airflow characteristic data set.

[0085] In the embodiment of the application, the sliding window method is used to update the airflow velocity distribution and streamline shape parameters in real time, the window length is 400 milliseconds, which is about twice the sampling period, the parameter fluctuation amplitude in the window is calculated, and when the amplitude is less than 0.15, the streamline shape is judged to be stable, and the first stability index is generated. A Kalman filter is constructed to optimize the data, the state transition matrix is derived based on the physical model, the measurement noise covariance matrix is determined by statistical analysis of 1000 groups of historical data, and the process noise covariance matrix is a diagonal matrix with diagonal elements of 0.01. The filter tracks the airflow characteristic changes in real time, outputs the updated airflow characteristic data set, and accurately reflects the dynamic response characteristics of the nitrogen charging process. Through dynamic updating and filtering processing, the quality and real-time performance of the data set are ensured in this step.

[0086] In the embodiment of the application, a proportional-integral regulator is used to adjust the nitrogen injection parameters in real time, the proportional coefficient is 0.8, the integral time constant is 50 milliseconds, and the response speed is fast. When the injection rate is adjusted from 5 meters per second to 7 meters per second, the velocity field uniformity index rises from 0.75 to 0.88 within 150 milliseconds, the vortex intensity attenuation rate stabilizes at 0.35 within 200 milliseconds, and the streamline shape stability remains above 0.85, indicating that the system quickly reaches a stable state. Compared with the traditional fixed parameter control, this step significantly improves the response performance and stability of the nitrogen charging process through real-time monitoring and dynamic adjustment, is suitable for various working conditions, reduces nitrogen waste, and improves production efficiency.

[0087] S107 inputs the updated airflow characteristic data set into the flow field dynamic analysis model, recalculates the characteristic vectors of the airflow velocity distribution, the vortex intensity and the streamline shape, generates the updated airflow characteristic stability index, and judges the completion state of the nitrogen charging process.

[0088] In the embodiment of the application, the updated airflow characteristic data set reflects the adjusted airflow dynamic characteristics of the nitrogen charging equipment. The characteristics need to be extracted again through the flow field dynamic analysis model and the stability needs to be evaluated to determine whether the nitrogen charging process is completed. The specific implementation of this step includes the following sub-steps:

[0089] S1071 performs outlier detection and preprocessing on the airflow characteristic data set, extracts the velocity distribution characteristics using wavelet transform, and generates a first processed data set and a characteristic vector.

[0090] In the embodiment of the application, the airflow characteristic data set may contain measurement noise or abnormal fluctuations, and needs to be cleaned to improve data quality. The three-sigma rule is used to detect outliers, the data mean and standard deviation are calculated, the boundary is set to the mean plus or minus three times the standard deviation, the airflow velocity standard deviation range is 0.5 to 1.2, and the data points outside the boundary are removed. A 15-point median filter is used to smooth the data and eliminate short-term fluctuations to generate a first processed data set. For this data set, a db4 wavelet basis function is used for four-layer wavelet transform, and decomposition coefficients of 0-25 Hz, 25-50 Hz, 50-75 Hz and 75-100 Hz are obtained. The low-frequency component reflects the overall trend of the airflow velocity, and the high-frequency component captures local disturbances. A 16-dimensional velocity distribution characteristic vector is extracted from the decomposition coefficients, which contains multi-scale velocity distribution characteristics. This step ensures the accuracy of feature extraction through outlier detection and wavelet transform, and lays a foundation for subsequent analysis.

[0091] S1072 calculates the vortex intensity change rate and performs smoothing processing to generate a vortex characteristic vector, and combines a long short-term memory network for dynamic prediction.

[0092] In the embodiment of the application, the dynamic change of vortex intensity is the key to evaluating the stability of nitrogen charging. Based on the 16-dimensional velocity distribution characteristic vector, the 5-point difference method is used to calculate the vortex intensity change rate to generate a change rate sequence. The sequence is smoothed using the exponential weighted average method, the weight decay coefficient is 0.85, the adaptive threshold interval is set according to historical data, the upper limit is the mean plus 1.5 times the standard deviation, and the lower limit is the mean minus 1.5 times the standard deviation, to generate a 32-dimensional vortex characteristic vector. A long short-term memory network is constructed for dynamic prediction, the network includes an input layer, three hidden layers and an output layer, the input layer has 32 nodes to receive the vortex characteristic vector, each hidden layer has 64 memory cells, the forgetting gate is used to control information retention, the threshold is set to 0.5, the input gate and the output gate use the sigmoid activation function, and the output layer has 16 nodes to predict the future trend. Using 1500 training samples, the learning rate is 0.001, and the training is 1000 rounds, the prediction accuracy is 92%. This step accurately captures the evolution law of the vortex intensity through smoothing processing and time series prediction.

[0093] S1073 extracts the streamline shape feature and performs dimension reduction, fuses the multi-dimensional feature vector, and generates a gas flow feature stability index.

[0094] In the embodiment of the present application, the stability of the streamline shape directly affects the nitrogen charging effect. The streamline shape parameters are extracted from the long short-term memory network prediction result, the curvature change rate is calculated, the curvature change in the central region is less than 0.05, and the edge region is as high as 0.15. Principal component analysis is used for dimension reduction, the first 8 principal components with a contribution rate of more than 85% are retained, and an 8-dimensional streamline feature vector is generated. The velocity distribution feature vector, vortex feature vector and streamline feature vector are fused, and a weighted summation method is used, with weights of 0.4, 0.35 and 0.25 respectively, which are determined by 1000 groups of historical data statistics. The generated gas flow feature stability index value is between 0 and 1, reflecting the comprehensive stability of the gas flow. Through dimension reduction and feature fusion, the dynamic characteristics of the gas flow are quantitatively combined.

[0095] S1074 constructs a recurrent neural network to track the change of the stability index, and judges the completion state of the nitrogen charging process.

[0096] In the embodiment of the present application, in order to monitor the stability trend in real time, a five-layer recurrent neural network is constructed, with layer node numbers of 32, 64, 64, 32 and 16 respectively, and a tanh activation function is used. The network inputs the stability index at the current time, outputs the predicted value at the next time, the training samples cover 1000 groups of different working conditions, and the prediction error is less than 0.05. The completion threshold is set to 0.85, and when the stability index is higher than the value for 10 consecutive sampling periods, it is determined that the nitrogen charging process is completed. In practical application, the index fluctuates between 0.6 and 0.7 in the initial stage, and after adjustment, it stabilizes at 0.88 to 0.92 within 800 milliseconds, indicating that the system has reached a stable state. Through dynamic tracking, the accuracy of the completion state judgment is ensured.

[0097] In the embodiment of the present application, compared with the traditional method relying on pressure detection, this step significantly improves the accuracy of gas flow feature analysis through multi-scale feature extraction and neural network prediction. The generated stability index accurately reflects the dynamic evolution of the nitrogen charging process, judges the completion state in real time, avoids resource waste caused by premature or late termination, optimizes production efficiency, and is suitable for high-standard nitrogen charging demand in food, medicine and other industries.

[0098] S108 If the gas flow feature stability index reaches the preset threshold, the updated index is classified by a logistic regression algorithm, the nitrogen charging completion probability is calculated, a termination signal is generated to control the nitrogen charging equipment to stop, and the gas flow feature dataset is recorded to evaluate the production efficiency improvement amplitude.

[0099] In the embodiment of the present application, when the airflow characteristic stability index reaches the preset threshold value, it indicates that the nitrogen charging process has entered a stable state, and the classification algorithm is used to confirm the completion state and generate a control instruction, and the process data is recorded to quantify the efficiency improvement. The specific implementation of this step includes the following sub-steps:

[0100] S1081 segment the airflow characteristic stability index, extract statistical and time sequence features, and construct a seven-dimensional feature vector.

[0101] In the embodiment of the present application, the airflow characteristic stability index reflects the dynamic stability of the nitrogen charging process, which needs to be quantified by feature extraction. The sliding window method is used to segment the index sequence, the window length is 100 sampling points, covering 500 milliseconds, and the overlap rate is 50%, generating the first index sequence. For this sequence, statistical features are calculated, including mean, variance, skewness and kurtosis, where the mean is in the range of 0.85 to 0.95, the variance is less than 0.02, the skewness is in the range of -0.5 to 0.5, and the kurtosis is close to 3, indicating that the data distribution is stable and approximately normal. At the same time, time sequence features are extracted, including the maximum value, the minimum value and the peak-valley difference of the sequence, and the difference is less than 0.15, indicating that the fluctuation is small. Fusion of statistical and time sequence features, a seven-dimensional feature vector is constructed, providing high-quality input for subsequent classification. Through multi-dimensional feature extraction, the dynamic characteristics of the stability index are fully captured.

[0102] S1082 uses a logistic regression algorithm to classify the seven-dimensional feature vector, generates a nitrogen charging completion probability, and outputs a termination signal.

[0103] In the embodiment of the present application, the logistic regression algorithm is used to determine the completion state of the nitrogen charging. The seven-dimensional feature vector is input, the model is trained using 1000 groups of historical data, the optimization goal is to minimize the cross-entropy loss, the learning rate is set to 0.01, the training is 1000 rounds, and the classification accuracy is 95%. The classification result is mapped to a probability value through the sigmoid function, and the probability threshold is set to 0.9. When the probability value of five consecutive sampling periods exceeds the threshold value, it indicates that the nitrogen charging process is stable, and a termination signal is generated. The termination signal triggers the device control instruction, which is executed through a proportional-integral controller, with a proportional coefficient of 0.8 and an integral time constant of 12.5 milliseconds, which is about one-fourth of the sampling period, ensuring smooth stopping and avoiding airflow disturbance. This step realizes automatic termination through probability classification and precise control, improving the operation reliability.

[0104] S1083 records the process parameter data set, extracts a multi-dimensional feature vector, and uses a support vector machine to evaluate the production efficiency improvement.

[0105] In the embodiment of the application, the data acquisition unit records the nitrogen filling process data at a sampling frequency of 200 Hz, covering about 800 milliseconds, to generate a process parameter data set. The uniformity index of gas flow velocity, the stability index of vortex intensity and the continuity index of streamline shape are extracted from the data set, which respectively reflect the consistency of velocity field, the smoothness of vortex decay and the smoothness of streamline trajectory, and the values are respectively above 0.9, above 0.85 and above 0.88, to construct an eight-dimensional feature vector. Support vector machine is used for classification, radial basis kernel function is selected, and the kernel parameter is optimized by grid search, the search range is 0.01 to 10, the optimal parameter is 0.5, and the classification accuracy is 93%. The classification results show that, compared with the traditional method, the optimized nitrogen filling process improves the uniformity of gas flow velocity by 15%, the stability of vortex intensity by 12%, the continuity of streamline shape by 18%, and the comprehensive efficiency by about 15%. This step quantifies the optimization effect through multi-dimensional feature analysis, and verifies the practical value of the method.

[0106] In the embodiment of the application, compared with the traditional manual judgment, this step significantly improves the accuracy and automation level of the nitrogen filling completion state judgment through sliding window analysis, logistic regression classification and support vector machine evaluation. Real-time data recording and efficiency analysis provide data support for process optimization, reduce resource waste, and improve the production efficiency and stability of the nitrogen filling process in the food, pharmaceutical and other industries.

[0107] The above is only a specific embodiment of the present specification, and those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, module and unit can refer to the corresponding process in the foregoing method embodiment, which will not be described here. It should be understood that the protection scope of the present specification is not limited to this, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed in the present specification, and these modifications or replacements should be covered within the protection scope of the present specification.

Claims

1. A quality assessment method for the nitrogen filling process of bottles using artificial intelligence, characterized in that, The method includes: The airflow velocity distribution, eddy intensity, and streamline morphology at the nozzle of the nitrogen filling equipment are obtained, and feature extraction and time series analysis are performed to generate an airflow feature dataset. The uniformity of airflow velocity distribution, the attenuation of eddy intensity, and the stability of streamline morphology in the airflow feature dataset are extracted and fused to obtain a multidimensional vector representation of the airflow features at the bottle opening. Multidimensional vector representations of airflow velocity distribution, eddy intensity, and streamline morphology are used to construct a pre-trained dynamic analysis model of the flow field, predict the dynamic change law of airflow characteristics under pressure equilibrium, and obtain the stability index of airflow characteristics at the bottle opening through stability analysis. If the stability of airflow characteristics is lower than a preset threshold, the short-term change trends of airflow velocity distribution, eddy intensity and streamline morphology are predicted by the Long Short Time Memory Network algorithm to obtain the future evolution state of airflow characteristics. The nitrogen injection rate and injection angle of the nitrogen filling equipment are adjusted according to the future evolution of the airflow characteristics to generate optimized nitrogen filling control parameters; Adjust the nitrogen injection parameters according to the optimized nitrogen filling control parameters, update the operating status of the nitrogen filling equipment in real time, and generate an updated airflow characteristic dataset; The updated airflow characteristic dataset is input into the flow field dynamic analysis model to recalculate the characteristic vectors of airflow velocity distribution, eddy intensity and streamline morphology, and output the updated airflow characteristic stability index to determine the completion status of the nitrogen filling process. If the airflow characteristic stability index reaches the threshold, the index is updated by classifying using a logistic regression algorithm, the probability of nitrogen filling completion is output, a termination signal is generated based on the probability, the nitrogen filling equipment is stopped, the airflow characteristic dataset is recorded, and the improvement in production efficiency is determined.

2. The method for quality assessment of the nitrogen filling process of a bottle using artificial intelligence as described in claim 1, characterized in that, The process involves acquiring the airflow velocity distribution, eddy intensity, and streamline morphology at the nozzle of the nitrogen filling device, performing feature extraction and time series analysis, and generating an airflow feature dataset, including: A gas sensor array was used to collect airflow velocity and inlet pressure data at grid points divided by the bottle opening diameter. Noise reduction and standardization were then performed to obtain the first airflow dataset. Based on the first airflow dataset, an airflow velocity field distribution map is constructed. After obtaining streamline trajectory data by scanning the bottle opening plane with a gas flow probe, vorticity contour lines are superimposed to obtain the first vortex distribution map. A Bayesian probability network was constructed based on the first airflow dataset and the gas concentration data collected during the nitrogen filling process to extract the gas concentration distribution feature vector from the airflow velocity field distribution map. Based on the first vortex distribution map, streamline trajectory curves are superimposed, and the gas diffusion coefficient is calculated according to the gas concentration distribution characteristic vector. The diffusion coefficient is used to quantitatively describe the flow field structure morphology to obtain a time series dataset of airflow characteristics, thus forming an airflow characteristic dataset.

3. The method for quality assessment of the nitrogen filling process of a bottle using artificial intelligence as described in claim 1, characterized in that, The extracted airflow feature dataset contains features such as the uniformity of airflow velocity distribution, the attenuation of eddy intensity, and the stability of streamline morphology. These features are then fused to obtain a multidimensional vector representation of the airflow features at the bottle opening, including: The airflow feature dataset is cleaned using a median filter, and then segmented using a sliding window method to obtain a preprocessed airflow feature dataset. The variance distribution of the airflow velocity field in the bottle mouth region is calculated based on the preprocessed airflow feature dataset. The variance distribution is then gridded to obtain the first velocity field fluctuation value. The fluctuation value is then normalized by the maximum and minimum values ​​to obtain the first airflow uniformity index. Extract the vortex intensity time series from the preprocessed airflow feature dataset, calculate the difference in vortex intensity between adjacent moments to obtain the first vortex intensity change curve, and use an exponential decay function to fit the change curve to obtain the vortex decay parameter. A feature correlation matrix is ​​constructed based on the first airflow uniformity index and eddy attenuation parameter. The pairwise correlation coefficients between features are calculated on the feature correlation matrix to obtain the feature weight vector. The feature weight vector is used to perform a weighted summation of the airflow features to obtain a multidimensional vector representation of the airflow features at the bottle opening.

4. The method for quality assessment of the nitrogen filling process of a bottle using artificial intelligence as described in claim 1, characterized in that, The process involves multi-dimensional vector representation of airflow velocity distribution, eddy intensity, and streamline morphology to construct a pre-trained dynamic flow field analysis model. This model predicts the dynamic changes in airflow characteristics under pressure equilibrium conditions. Stability analysis yields stability indices for the airflow characteristics at the bottle opening, including: The airflow velocity distribution data is preprocessed using a normalization method. The airflow pulsation spectrum features are extracted by median filtering and discrete Fourier transform. The first multidimensional feature vector is obtained by superimposing the vortex intensity data on the spectrum features. The first multidimensional feature vector is input into a deep autoencoder, and the second multidimensional feature vector is obtained by halving the number of intermediate layer nodes; A residual neural network predictor is constructed based on the second multidimensional feature vector. Temporal features are extracted by stacking residual blocks to form a dynamic flow field analysis model and obtain dynamic airflow prediction results. The pressure change sequence at the bottle opening is calculated based on the airflow dynamic prediction results. The maximum Lyapunov exponent is calculated based on the pressure change sequence. The stability index of the airflow characteristics at the bottle opening is obtained by dividing the stability interval through the variance of the characteristic sequence.

5. The method for quality assessment of the nitrogen filling process of a bottle using artificial intelligence as described in claim 1, characterized in that, If the stability of the airflow characteristics is lower than a preset threshold, the short-term trends of airflow velocity distribution, eddy intensity, and streamline morphology are predicted using a long short-term memory network algorithm to obtain the future evolution state of the airflow characteristics, including: If the airflow characteristic stability index is lower than the preset threshold, median filtering is used to preprocess the airflow characteristic data, and wavelet transform is used to extract the airflow velocity distribution characteristics and eddy intensity time series characteristics to obtain the first feature vector. Based on the first feature vector, a training data sequence is generated using the sliding window method, and a first prediction result is obtained by training a long short-term memory network predictor. Based on the first prediction result, the airflow velocity distribution trend and eddy intensity variation law are extracted, and the second feature vector is obtained by fusion using a weighted average method; An airflow feature evolution sequence is constructed based on the second feature vector, and the evolution sequence is processed using an exponential smoothing method to obtain the future evolution state of the airflow features from the smoothed sequence.

6. The method for quality assessment of the nitrogen filling process of a bottle using artificial intelligence according to claim 1, characterized in that, The process of adjusting the nitrogen injection rate and injection angle of the nitrogen filling equipment based on the future evolution of airflow characteristics to generate optimized nitrogen filling control parameters includes: The airflow characteristic evolution state data is acquired, and the airflow characteristic evolution state data is normalized and median filtered. The first evolution dataset is obtained by using the sliding window method. Based on the first evolution dataset, an airflow velocity distribution sequence is extracted, and a Fourier transform operation is performed on the airflow velocity distribution sequence to obtain airflow fluctuation period data from the Fourier transform operation result. The phase space reconstruction method is used to reconstruct the airflow fluctuation period data, and the evolution law of the first vortex is obtained by calculating the radius of curvature of the reconstructed trajectory. Based on the first eddy current evolution law, an injection angle range and an injection rate range are constructed. The particle swarm optimization algorithm is used to optimize the injection angle range and the injection rate range to obtain optimized nitrogen filling control parameters.

7. The method for quality assessment of the nitrogen filling process of a bottle using artificial intelligence as described in claim 1, characterized in that, The process of adjusting nitrogen injection parameters based on optimized nitrogen filling control parameters, updating the operating status of the nitrogen filling equipment in real time, and generating an updated airflow characteristic dataset includes: Collect airflow data at the bottle opening, and perform Gaussian filtering on the airflow data to obtain a first airflow dataset; The gridded variance distribution is calculated based on the first airflow dataset. The first velocity field uniformity index is obtained by using a regional block statistical method. The first attenuation rate curve is obtained by extracting the difference between adjacent sampling points from the first velocity field uniformity index. A first response characteristic vector is constructed based on the first decay rate curve, and a second response characteristic vector is obtained by using a support vector regressor to extract features from the first response characteristic vector. If the velocity field uniformity index in the second response characteristic vector is greater than a preset threshold, then the velocity distribution is determined to be uniform, and the velocity distribution data is updated in real time using the sliding window method to obtain the airflow characteristic dataset.

8. The method for quality assessment of the nitrogen filling process of a bottle using artificial intelligence according to claim 1, characterized in that, The updated airflow characteristic dataset is input into the flow field dynamic analysis model to recalculate the characteristic vectors of airflow velocity distribution, eddy intensity, and streamline morphology, and outputs the updated airflow characteristic stability index to determine the completion status of the nitrogen filling process, including: Outlier detection is performed on the airflow feature dataset using the three-standard-deviation principle, and the airflow feature dataset is smoothed using a median filter to obtain the first processed dataset; Based on the first processed dataset, perform wavelet transform four-level decomposition, and construct a multidimensional velocity distribution feature vector from the wavelet transform decomposition coefficients; The eddy current intensity change rate is calculated for the multidimensional velocity distribution feature vector, and the eddy current feature vector is obtained by smoothing the eddy current intensity change rate using an exponential weighting method. A long short-term memory network is constructed based on the eddy current feature vector to perform dynamic prediction. Principal component analysis is performed on the prediction results to obtain the airflow feature stability index. If the airflow feature stability index is higher than the preset completion threshold for ten consecutive sampling periods, the nitrogen filling process is determined to be complete.

9. The method for quality assessment of the nitrogen filling process of a bottle using artificial intelligence according to claim 1, characterized in that, If the airflow characteristic stability index reaches a threshold, the index is updated by classifying using a logistic regression algorithm, the probability of nitrogen filling completion is output, a termination signal is generated based on the probability, the nitrogen filling equipment is stopped, the airflow characteristic dataset is recorded, and the improvement in production efficiency is determined, including: The stability index of the airflow characteristics is compared with a preset threshold, and the stability index is segmented by a sliding window. A seven-dimensional feature vector is obtained by calculating statistical features and time series features. The seven-dimensional feature vector is classified using a logistic regression algorithm, and a probability value is obtained through a sigmoid function. A termination signal is obtained when the probability value continuously exceeds a preset probability threshold. The device control command is constructed based on the termination signal, and the control command is executed using a proportional-integral controller to record the airflow characteristic dataset; Airflow velocity uniformity index, eddy intensity stability index, and streamline morphology continuity index are extracted from the airflow feature dataset to construct an eight-dimensional feature vector. The eight-dimensional feature vector is then classified using a support vector machine, and efficiency improvement index is extracted from the classification results.

Images