Can sealing process optimization method, system and equipment based on artificial intelligence and medium

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a can sealing process optimization method, system and equipment based on artificial intelligence and a medium. The method comprises the following steps: carrying out time domain and frequency domain dual filtering preprocessing on original can sealing process sensing data to obtain preprocessed heterogeneous sensor data, and carrying out adaptive segmentation on a can sealing process parameter domain to obtain a plurality of parameter subintervals; constructing a quality evaluation function according to the effective process parameter state data set in the parameter subinterval; calculating a parameter logarithmic function gradient and performing adaptive step size control to obtain a parameter optimization direction field; performing feature extraction and feature fusion through a dual-channel parallel feature extraction network to obtain target fusion features; and performing variable genetic factor particle swarm optimization based on the parameter optimization direction field and the target fusion features to generate an optimal process parameter combination. According to the method, the self-adaptive optimization of the can sealing process parameters is realized, and the production process is ensured to be always operated in the optimal state.

Description

Technical Field

[0001] The present invention relates to the technical field of process optimization, and in particular to an artificial intelligence-based canning process optimization method, system, equipment and medium. Background Art

[0002] Despite the widespread adoption of automated canning technology, existing production lines still face issues with poor sealing and high product scrap rates. This is primarily due to inadequate process parameter optimization. Traditional canning process parameter optimization relies heavily on manual adjustments based on experience or simple single-factor experiments. This fails to fully account for the complex interactions between multiple parameters, often leading to parameter combinations stuck in local optimality and unable to adapt to changing production environments.

[0003] In recent years, artificial intelligence (AI) technology has made significant progress in industrial process optimization, with a variety of intelligent algorithms being applied to parameter optimization. However, existing AI approaches still face numerous challenges in applying them to canning process optimization. Complex model structures, such as neural networks, are computationally expensive and lack interpretability, making them difficult to deploy in real time in production environments. Furthermore, existing methods struggle to design specialized parameter optimization functions for different canning quality objectives, resulting in limited optimization accuracy. Furthermore, the effective fusion and processing of heterogeneous sensor data poses significant challenges for existing approaches. Summary of the Invention

[0004] The present invention provides a canning process optimization method, system, equipment and medium based on artificial intelligence. The present invention realizes adaptive optimization of canning process parameters and ensures that the production process always operates in an optimal state.

[0005] In a first aspect, the present invention provides an artificial intelligence-based canning process optimization method, the artificial intelligence-based canning process optimization method comprising: The original canning process sensor data is preprocessed by double filtering in the time and frequency domains to obtain preprocessed heterogeneous sensor data, and the canning process parameter domain is adaptively segmented to obtain multiple parameter sub-intervals; constructing a quality assessment function based on a valid process parameter state data set within the parameter subinterval; Based on the quality evaluation function, the parameter logarithmic function gradient is calculated and adaptive step size control is performed to obtain a parameter optimization direction field; Inputting the preprocessed heterogeneous sensor data into a dual-channel parallel feature extraction network for feature extraction and feature fusion to obtain target fusion features; Based on the parameter optimization direction field and the target fusion feature, variable genetic factor particle swarm optimization is performed to generate an optimal process parameter combination.

[0006] In a second aspect, the present invention provides an artificial intelligence-based canning process optimization system, the artificial intelligence-based canning process optimization system comprising: The preprocessing module is used to perform double filtering preprocessing in the time domain and frequency domain on the original canning process sensor data to obtain preprocessed heterogeneous sensor data, and to adaptively segment the canning process parameter domain to obtain multiple parameter sub-intervals; A construction module, configured to construct a quality assessment function based on a valid process parameter state data set within the parameter subinterval; A calculation module, configured to calculate the gradient of a parameter logarithmic function based on the quality evaluation function and perform adaptive step size control to obtain a parameter optimization direction field; A feature extraction module is used to input the pre-processed heterogeneous sensor data into a dual-channel parallel feature extraction network to perform feature extraction and feature fusion to obtain target fusion features; The particle swarm optimization module is used to perform variable genetic factor particle swarm optimization based on the parameter optimization direction field and the target fusion feature to generate an optimal process parameter combination.

[0007] In a third aspect, an artificial intelligence-based canning process optimization device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the artificial intelligence-based canning process optimization device executes the above-mentioned artificial intelligence-based canning process optimization method.

[0008] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned artificial intelligence-based canning process optimization method.

[0009] In the technical solution provided by the present invention, the temperature, pressure, speed and other heterogeneous sensor data are efficiently cleaned and aligned through the time-domain and frequency-domain dual filtering preprocessing technology, which significantly reduces the data noise interference. Based on the principle of maximizing information entropy, adaptive parameter domain segmentation is performed to accurately capture the boundary of the parameter sensitive interval, realize the refined modeling of the canning process parameter space, and avoid the loss of optimization accuracy caused by the overly coarse division of the parameter space in the traditional method. The multidimensional Gaussian distribution model and Gaussian mixture model are used to construct the quality evaluation function, which effectively describes the complex interaction relationship and boundary transition characteristics between parameters, and overcomes the technical defects of the traditional method that ignores the interaction of parameters. Based on the logarithmic gradient and adaptive step size control of the quality evaluation function, a parameter optimization direction field pointing to the direction of quality improvement is generated, providing a clear optimization path for parameter adjustment and avoiding the optimization process from falling into the local optimum. Through the dual-channel parallel feature extraction network, the nonlinear relationship features and structural features in the parameter data are captured simultaneously, and the deep fusion and attention weighting of features are realized, thereby enhancing the model's recognition ability for key features. By combining the parameter optimization direction field and target fusion features, a variable genetic factor particle swarm optimization algorithm is implemented to dynamically adjust the learning factor and mutation probability, achieving a balance between global search and local refinement to achieve adaptive optimization of canning process parameters. By automatically generating and implementing the optimal parameter combination, combined with real-time parameter deviation monitoring and correction, a complete closed-loop control system for canning process optimization is established to ensure that the production process always operates in an optimal state. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 Schematic diagram of an embodiment of a can sealing process optimization method based on artificial intelligence in an embodiment of the present invention; Figure 2 Schematic diagram of an embodiment of a can sealing process optimization system based on artificial intelligence in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of the can sealing process optimization equipment based on artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] Embodiments of the present invention provide a method, system, apparatus, and medium for optimizing a canning process based on artificial intelligence. The terms "first," "second," "third," "fourth," and so forth (if any) in the specification and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. Furthermore, the terms "including," "comprising," "having," and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, product, or apparatus.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In one embodiment of the present invention, an embodiment of a canning process optimization method based on artificial intelligence includes: Step S101: performing time-domain and frequency-domain double filtering preprocessing on the original canning process sensor data to obtain preprocessed heterogeneous sensor data, and adaptively segmenting the canning process parameter domain to obtain multiple parameter sub-intervals; It is understood that the execution subject of the present invention can be an artificial intelligence-based canning process optimization system, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0014] Specifically, various data acquisition devices, such as temperature sensors, pressure sensors, and speed sensors, are deployed throughout the canning production line to capture comprehensive data on the canning process status at high frequency and in real time. All collected data is synchronized with unified timestamps. Using a time alignment algorithm, the data from various sensors is mapped to a unified time base, forming a time-aligned, multi-source, heterogeneous data sequence. Z-score normalization is applied to the time-aligned data, mapping all data to a standard normal distribution space with mean zero and variance one, facilitating subsequent joint analysis and modeling. Missing points in the normalized heterogeneous data are filled using algorithms such as time series interpolation to obtain a complete data matrix. Furthermore, time-domain filtering is performed on key data points, such as temperature and pressure, within the matrix. Using a sliding average filter or other time-domain smoothing algorithms, high-frequency noise and sudden outliers are effectively removed, resulting in time-domain filtered data. The time-domain filtering results are then input into a fast Fourier transform module, which maps the signal from the time domain to the frequency domain, facilitating analysis and processing of periodic noise or interference signals within a specific frequency range. For interference components within a preset frequency range in the frequency domain signal, these unfavorable signals are accurately filtered out through a bandpass filter, retaining only the valid frequency band related to the process status. After completing the frequency domain filtering, the data is restored from the frequency domain to the time domain using an inverse Fourier transform to obtain preprocessed heterogeneous sensor data. Based on the preprocessed heterogeneous sensor data, the canning process parameter domain is adaptively segmented. Using clustering methods based on information entropy, information gain, or data-driven methods, the segmentation points of process parameters such as temperature, pressure, and speed are dynamically determined, and the overall parameter space is subdivided into several highly homogeneous sub-intervals, reflecting the differences and regularities in the canning effects under different process conditions.

[0015] In this embodiment, a canning process parameter space is constructed based on preprocessed heterogeneous sensor data. All intervals for key parameters, such as temperature, pressure, time, and speed, reflecting the state of the canning process, are combined into a multidimensional parameter domain, achieving a comprehensive description of the process state. By analyzing the distribution characteristics of each parameter dimension in this parameter domain and the correlation between historical production data, an information gain metric is used to mine each parameter dimension to quantify the contribution of parameter intervals to the determination of canning quality results at different segmentation points. By calculating information gain for all possible segmentation points, the optimal candidate segmentation points for process parameters such as temperature, pressure, time, and speed are screened. On this basis, and in accordance with the actual process control requirements, a three-point segmentation is implemented for the temperature parameter, and a two-point segmentation is implemented for each of the pressure, time, and speed parameters. This divides the original parameter space into a series of initial subintervals according to process patterns such as high and low, and fast and slow, thereby improving the discernibility and controllability of the parameter space. The initial subintervals obtained are evaluated using the principle of maximizing information entropy, that is, in the segmentation scheme, the homogeneity of the data within each subinterval is maximized, and the differences between different subintervals are strengthened, so as to ensure that the segmentation results have the best explanatory power in both statistical significance and actual process application. The conditional mutual information technology method is used to quantitatively model the interaction between different process parameters in the optimized subintervals. By calculating the joint contribution and influence of each parameter on the output quality under different subinterval combinations, a parameter interaction matrix is constructed to reveal the essential relationship of multi-parameter coupling in the canning process. The parameter interaction matrix and historical production data are input into machine learning models such as gradient boosting decision trees. Through the autonomous learning ability of the model, the division and weight distribution of parameter subintervals are optimized to achieve high-precision, data-driven intelligent segmentation of the process parameter space and obtain multiple parameter subintervals.

[0016] Step S102: constructing a quality assessment function based on the valid process parameter state data set within the parameter sub-interval; Specifically, a statistical analysis of historical production data within each parameter subinterval identifies all valid process parameter states that have actually occurred within that process interval. A dataset representing process fluctuations and quality characteristics for each interval is constructed, including typical values of parameters such as temperature, pressure, time, and speed across different production batches. Each parameter set is then associated with the actual product quality grade. Based on the valid datasets for each subinterval, a statistical method is used to calculate the four-dimensional mean vector for each parameter subinterval, identifying the optimal or most frequently occurring process parameter combination within the current interval. The covariance matrix between the parameter data is then calculated using maximum likelihood estimation to quantify the coupled fluctuations and correlations between the four parameters. Based on the parameter means and covariance moments, a multidimensional Gaussian distribution model is constructed for each subinterval. This model captures the typical distribution patterns of process parameters in probability space. To enhance the practical guidance of the model and enhance the sensitivity of process optimization, differential weighting is assigned to product data of different quality grades during the model parameter calculation process. Data points for high-quality products are given higher weights, allowing the model to focus on capturing the process parameter distribution characteristics that lead to optimal production results. In the boundary regions of the parameter space, due to the transition and superposition effects of process states between adjacent subintervals, a mixed modeling of the parameter distributions in these regions is performed. Through dynamic calculation of the mixing coefficients, the Gaussian models of adjacent intervals are linearly combined according to probability weights to establish a Gaussian mixture model. This ensures the continuity of the parameter space during modeling and more accurately reflects the production patterns of multi-parameter gradual changes in actual processes. By combining the quality-weighted Gaussian models of each subinterval with the Gaussian mixture model of the boundary region, a quality assessment function is formed from the process parameter vector to the target quality score.

[0017] Step S103: Based on the quality evaluation function, the gradient of the parameter logarithmic function is calculated and adaptive step size control is performed to obtain the parameter optimization direction field; Specifically, the quality assessment function is logarithmically transformed to obtain a logarithmic quality assessment function. The partial derivatives of the logarithmic quality assessment function with respect to the four parameter variables of temperature, pressure, time and speed are calculated respectively to obtain the gradient vector of each point in the parameter space. The gradient vector points to the direction that can most quickly improve the target quality score at the current point, so its physical meaning is equivalent to the "optimal process improvement direction". For all points in the parameter space that belong to a single parameter subinterval, the gradient vector within the subinterval is directly applied to form the initial direction field within the subinterval, so that each point can obtain an intelligent adjustment direction for the characteristics of this interval in subsequent optimization iterations. At the boundary and interaction area of the parameter space, due to the overlap and fusion of multiple subinterval models, the gradients of their respective logarithmic quality assessment functions are weightedly superimposed according to the weights of the previous Gaussian mixture model, and the optimization trends of adjacent subintervals are comprehensively considered to obtain the interaction area direction field that can reflect the transition characteristics. To improve optimization efficiency, an adaptive step-size control mechanism was introduced. By constructing a step-size control function with a base step size and a distance decay coefficient as variables, the intelligent dynamic adjustment of the parameter update amplitude at each step is achieved. Specifically, when the process parameter point is far from the optimal solution, the step size is automatically increased to achieve rapid global search. When approaching the optimal region, the step size is automatically reduced for detailed fine-tuning to avoid overshooting and oscillation. The initially generated sub-interval direction field, the interaction region direction field, and their respective gradient optimization step sizes are then subjected to high-dimensional mapping and nonlinear feature fusion through a fully connected neural network. The neural network input is the feature representation of any parameter point, and the output is the final optimized direction field vector for that point.

[0018] Step S104: input the pre-processed heterogeneous sensor data into a dual-channel parallel feature extraction network for feature extraction and feature fusion to obtain target fusion features; Specifically, all high-quality sensor data, after filtering and normalization in the time and frequency domains, is structured and segmented. Temperature and pressure, as the parameters most directly reflecting the physical environment during the canning process, are input into the first channel of the feature extraction network. This channel utilizes a multilayer perceptron architecture or a lightweight recurrent network to fully exploit the nonlinear interactions between temperature and pressure, outputting a feature vector representing the dynamic relationship between temperature and pressure. Simultaneously, time and speed parameters during the canning process, as key variables influencing process pacing and control, are input into the second channel of the feature extraction network. Within this channel, a convolutional neural network or a specific sequence modeling architecture is used to capture the temporal coordination and regulatory relationship between time and speed, generating a time-speed relationship feature vector. The temperature-pressure relationship feature vector and the time-speed relationship feature vector are concatenated to generate an initial fused feature vector. The importance weight of each dimension of the initial fused feature vector is evaluated. Through an attention mechanism, a learnable weight matrix is used to calculate the contribution of each dimension of the feature vector to the identification and optimization of the canning process, generating feature attention weights. The initial fusion feature vector is weighted based on the attention weight, and the feature components with high importance are given greater weights, while the relatively minor information is appropriately weakened, thereby outputting a target fusion feature that is highly focused on the target process state description and optimization target.

[0019] Step S105: performing variable genetic factor particle swarm optimization based on the parameter optimization direction field and the target fusion feature to generate an optimal process parameter combination.

[0020] Specifically, multiple initial parameter vectors are generated by random sampling in the multidimensional space of canning process parameters. Each vector corresponds to a particle, and each particle is assigned independent position and velocity attributes, forming an initial particle swarm with a certain diversity and spatial coverage. Combined with the pre-constructed parameter optimization direction field, a unique guidance vector is calculated for each particle based on its current position. This vector indicates the direction in which the particle converges towards the optimal quality interval and reflects the gradient trend of the local and global optimality in the parameter space. At the same time, the state characteristics of each particle are fused and analyzed with the target fusion characteristics. The degree of fit between the current environment and the particle parameters is evaluated through a deep model, and the environmental adaptation coefficient is dynamically calculated. This coefficient is combined with the particle update strategy to form a more intelligent particle iterative update rule. In each round of iteration, the positions of all particles in the initial particle swarm are input into the quality evaluation function for multi-dimensional quality scoring. At the same time, the particle parameter combination and the target fusion characteristics are input into the process quality prediction model together to obtain the fitness value of each particle. The algorithm globally ranks all particles based on their fitness values, selecting the particle with the highest fitness as a candidate for the global optimal solution. For particles with fitness below a preset threshold, a genetic algorithm-based mutation operation is applied, introducing random perturbations or recombination mechanisms to help them escape local optima, thereby enhancing the overall diversity of the particle swarm and its search space coverage. During each iteration, the position and velocity of each particle in the swarm are jointly updated according to the iterative update rule. This utilizes inertia factors and social cognitive learning factors, along with reference parameters to optimize the directional field and dynamically adjust the environmental adaptation coefficient. This ensures that the particle evolution path closely matches the optimal trend of the actual process. Furthermore, the mutation probability and learning factor are adaptively adjusted in real time based on global convergence, ensuring that the algorithm strikes an optimal balance between convergence speed and global exploration capability. The optimization process continues iteratively, repeating fitness evaluation, optimal selection, and mutation updates within each generation until a predetermined termination criterion is met, such as reaching the maximum number of iterations or the global optimal fitness plateauing. Finally, the particle with the highest global fitness is selected from the swarm as the optimal parameter combination for the canning process. In an embodiment of the present invention, through the time-domain and frequency-domain dual filtering preprocessing technology, the heterogeneous sensor data such as temperature, pressure, and speed are efficiently cleaned and aligned, which significantly reduces the data noise interference. Based on the principle of maximizing information entropy, adaptive parameter domain segmentation is performed to accurately capture the boundary of the parameter sensitive interval, and achieve refined modeling of the canning process parameter space, thus avoiding the loss of optimization accuracy caused by the overly coarse division of the parameter space in the traditional method. The multi-dimensional Gaussian distribution model and the Gaussian mixture model are used to construct the quality evaluation function, which effectively describes the complex interaction relationship and boundary transition characteristics between parameters, and overcomes the technical defects of the traditional method that ignores the interaction of parameters. Based on the logarithmic gradient and adaptive step size control of the quality evaluation function, a parameter optimization direction field pointing to the direction of quality improvement is generated, which provides a clear optimization path for parameter adjustment and avoids the optimization process from falling into the local optimum.Through a dual-channel parallel feature extraction network, the nonlinear relationship features and structural features in the parameter data are simultaneously captured, achieving deep feature fusion and attention weighting, enhancing the model's ability to identify key features. Combining the parameter optimization direction field and target fusion features, a variable genetic factor particle swarm optimization algorithm is executed to dynamically adjust the learning factor and mutation probability, achieving a balance between global search and local refinement to achieve adaptive optimization of canning process parameters. By automatically generating and implementing the optimal parameter combination, combined with real-time parameter deviation monitoring and correction, a complete closed-loop control system for canning process optimization is established, ensuring that the production process always operates in an optimal state.

[0021] In a specific embodiment, the process of executing step S101 may specifically include the following steps: The original can sealing process sensor data collected by temperature sensors, pressure sensors, and speed sensors on the can sealing production line are time-stamped and synchronized to obtain time-aligned multi-source heterogeneous data. Perform Z-score normalization transformation on the time-aligned multi-source heterogeneous data to obtain normalized heterogeneous data, and fill in the missing values in the normalized heterogeneous data to obtain complete normalized heterogeneous data; Performing time domain filtering on the temperature and pressure data in the complete standardized heterogeneous data to obtain time domain filtered data; The time domain filtered data is converted to the frequency domain through fast Fourier transform, and a bandpass filtering operation is performed on the interference signal of the preset frequency to obtain the frequency domain filtered data; Perform inverse Fourier transform on the frequency domain filtered data to restore it to the time domain to obtain the preprocessed heterogeneous sensor data; Based on the preprocessed heterogeneous sensor data, the canning process parameter domain is adaptively segmented to obtain multiple parameter sub-intervals.

[0022] Specifically, heterogeneous sensing devices such as temperature sensors, pressure sensors, and speed sensors are deployed at key locations on the canning production line to continuously and in real time collect multidimensional raw data reflecting process status. A timestamp synchronization algorithm is used to time-align the raw data collected by all sensors. By using a unified time base and interpolation completion techniques, different data sources can be completely and accurately mapped into a set of state parameters at the same physical moment, forming seamless, time-aligned, multi-source heterogeneous data. Once the time-aligned heterogeneous data is obtained, a Z-score normalization transformation is performed based on the dimensionality, distribution differences, and physical characteristics of each sensor data type. This transformation normalizes all data to a standard normal distribution with a mean of zero and a variance of one by calculating the mean and standard deviation of each parameter in historical samples, eliminating the effects of parameter units and scale. During the normalization process, some sensors may have missing values at certain moments due to device jitter, network latency, or sampling anomalies. For missing data points, data completion is performed using time series-based linear interpolation, K-nearest neighbor regression, or multivariate autoregression to ensure that all sample points have complete feature descriptions across all parameter dimensions. After this stage of processing, a complete, normalized heterogeneous dataset is obtained. Process data is filtered to address physical noise and environmental disturbances. Temperature and pressure, in particular, are highly susceptible to external disturbances and equipment micro-vibration during actual production. Time-domain filtering algorithms (such as moving average filtering, exponentially weighted moving average, and low-pass filtering) are used to smooth the temperature and pressure signals, effectively suppressing high-frequency noise, sudden spikes, and abnormal transitions. This ensures that data trends better reflect the inherent laws of the process rather than external interference. The data output from time-domain filtering is more continuous and physically interpretable. After time-domain filtering, the smoothed temperature and pressure data are input into the fast Fourier transform module, which converts the time-domain signals into frequency-domain signals. Frequency-domain analysis is advantageous in revealing implicit periodic variations and specific frequency components in the data. This information can help identify process anomalies, equipment aging, or external disturbances. Bandpass filtering of the frequency-domain signals specifically removes interference components within a preset frequency range, such as electromagnetic interference, mechanical resonance, or periodic noise, while retaining valid signal components relevant to the process. The design parameters of the bandpass filter (such as passband frequency and stopband attenuation) are dynamically adjusted based on the actual process and sensor characteristics to ensure maximum retention of authentic and effective information from the production process. After bandpass filtering, redundant noise in the frequency domain signal is significantly suppressed, and the signal energy is concentrated in the frequency band with process-critical significance. An inverse fast Fourier transform is used to restore the bandpass filtered frequency domain data to the time domain, generating preprocessed heterogeneous sensor data. Based on this preprocessed heterogeneous sensor data, the canning process parameter domain is adaptively segmented.By analyzing the distribution, variation, and joint characteristics of parameters such as temperature, pressure, and speed in historical samples, and using various algorithms based on information gain, entropy maximization, or cluster analysis, the optimal segmentation point for each parameter is dynamically determined, thereby achieving multi-level segmentation of the parameter space. For example, historical data is used to automatically identify the highly sensitive intervals, stable intervals, and transition intervals of process parameters, and the overall process parameter domain is divided into multiple sub-intervals. Each sub-interval corresponds to a group of process states with strong homogeneity and high controllability. The interaction between multiple parameters is taken into account during the segmentation process, and the mutual information matrix or parameter coupling model is used to accurately characterize the role of different parameter intervals in jointly regulating product quality, thereby ensuring that the segmentation results are consistent with physical reality and serve the theoretical needs of the intelligent optimization model.

[0023] In a specific embodiment, the step of adaptively segmenting the canning process parameter domain based on the preprocessed heterogeneous sensor data to obtain multiple parameter sub-intervals may specifically include the following steps: Based on the pre-processed heterogeneous sensor data, a canning process parameter space including temperature interval, pressure interval, time interval and speed interval is constructed to obtain the canning process parameter domain; Calculate the information gain index of each parameter dimension in the canning process parameter domain to obtain the parameter segmentation candidate points; Based on the parameter segmentation candidate points, three-point segmentation is performed on the temperature parameter, two-point segmentation is performed on the pressure parameter, two-point segmentation is performed on the time parameter, and two-point segmentation is performed on the speed parameter to obtain the initial sub-intervals; The information entropy maximization principle is used to evaluate the initial subinterval to obtain the optimized subinterval, and the conditional mutual information technology is used to calculate the mutual influence relationship between different parameters in the optimized subinterval to obtain the parameter interaction matrix; The parameter interaction matrix and historical production data are input into the gradient boosting decision tree model to obtain multiple parameter sub-intervals.

[0024] Specifically, a process parameter dataset is acquired from multi-source heterogeneous sensors. Based on this dataset, a multidimensional parameter space, encompassing temperature, pressure, time, and speed intervals, is constructed—the canning process parameter domain. During parameter domain construction, typical and abnormal operating conditions in historical production data are analyzed, and reasonable intervals for each parameter are set based on actual production practices. Information gain metrics are used to analyze each parameter dimension, such as temperature, pressure, time, and speed, to determine the most discriminative and process-sensitive split points within each dimension. The core of information gain calculations is to evaluate the contribution of different split points to production outcomes (e.g., canning product quality grade). Specifically, the degree to which the uncertainty of the process parameters regarding product quality classification is reduced at a given split point. By traversing and simulating the full interval of each parameter, candidate split points with the highest information gain are systematically identified. These candidate points represent the most significant boundaries between parameter value variations and process performance. Based on the calculated results, a three-point split is performed for the temperature parameter, subdividing the temperature interval into four subintervals. Two-point splits are performed for each of the three parameters, pressure, time, and speed, creating three subintervals at key points within their respective intervals. Through data-driven interval subdivision, a gridded partitioning of the multidimensional parameter space is achieved. The initial subintervals are re-evaluated using the principle of maximizing information entropy. By statistically analyzing the mass distribution of samples within each subinterval, the partitioning results with the strongest homogeneity and clearest boundaries are selected to ensure that the data distribution within the subintervals is consistent while maximizing the differences between different subintervals. Based on the completion of the optimized subinterval partitioning, the synergistic effects and interactions between the parameters are analyzed. Since parameters such as temperature and pressure, time and speed in actual production do not affect quality in isolation but rather jointly determine the final process outcome through complex interactions, conditional mutual information technology is used to quantitatively analyze the joint influence of various parameters within the optimized subintervals. The essence of conditional mutual information is to measure the additional information gain that a parameter brings to product quality given the known conditions of other parameters. The larger the value, the more significant the interaction between the parameter and other parameters. By systematically calculating the conditional mutual information between all parameters, a parameter interaction matrix is constructed to reflect the strength of multi-parameter coupling. The parameter interaction matrix, along with actual production data accumulated over past production runs, is fed into machine learning models such as the gradient boosting decision tree. Leveraging the model's high-dimensional nonlinear fitting and automatic feature selection capabilities, the logic for partitioning parameter subintervals is refined, and the weights of each subinterval within the overall parameter space are rationally assigned. The gradient boosting decision tree model exploits the complex interaction patterns between parameters to uncover the high-order influencing factors underlying multiple parameter intervals. It can then output an optimal set of parameter subinterval structures based on a trade-off between multiple objectives, such as product quality, energy efficiency, and cost.

[0025] In a specific embodiment, the process of executing step S102 may specifically include the following steps: Perform statistical analysis on historical production data within each parameter sub-interval to obtain an effective process parameter status data set; Based on the effective process parameter state data set, the four-dimensional mean vector of each parameter sub-interval is calculated to obtain the parameter mean, and at the same time, the maximum likelihood estimation of the effective process parameter state data set is performed to obtain the covariance matrix; A four-dimensional Gaussian distribution model of the parameter subinterval is constructed based on the parameter mean and covariance matrix, and product data points of different quality levels are assigned different weights to obtain a quality-weighted Gaussian model; The mixing coefficients of the boundary areas of adjacent parameter subintervals are calculated to obtain a Gaussian mixture model, and a quality assessment function from parameter vector to quality score is constructed based on the quality-weighted Gaussian model and the Gaussian mixture model.

[0026] Specifically, for each parameter subinterval obtained through adaptive parameter space segmentation, the process parameters and corresponding quality labels from historical production processes within that interval are systematically collected and organized, eliminating anomalous, invalid, and incomplete records. Statistical analysis of high-quality historical data yields a valid data set of process parameter states that fully represents the process state within each parameter interval. Statistical methods are used to characterize the data distribution characteristics of each subinterval. By calculating the mean of temperature, pressure, time, and velocity for all valid process states within each parameter subinterval, a four-dimensional mean vector is formed. This vector physically points to the most frequently occurring or highest-quality process parameter combination within that interval. Furthermore, maximum likelihood estimation is performed on the data samples within this subinterval to derive the covariance matrix between the four-dimensional parameters. This covariance matrix reflects the joint volatility and correlation between the parameters, characterizing the coupled effects of parameter changes on the overall process state. By calculating the mean and covariance, the "center" of the process state is determined, and a comprehensive understanding of the dispersion and correlation structure of the parameters within the interval is provided, providing important priors for establishing probabilistic models. Based on this, a multidimensional Gaussian distribution model is constructed for each parameter subinterval, using the four-dimensional mean vector and covariance matrix as core parameters. Gaussian distributions offer excellent mathematical interpretation and real-world applicability, effectively describing the probability and concentration of various typical process states within the parameter space. To enhance the model's process guidance and quality discrimination capabilities, product quality grade weighting is introduced. In actual production, each set of process parameters not only determines the equipment's operating status but also directly impacts the quality of the final product. Therefore, during the modeling process, historical data points of different quality grades are assigned different weights. For example, "excellent" samples are given higher weights, while "good," "acceptable," and "defective" samples receive decreasing weights. When calculating probability density and parameter expectations, the model favors process parameter distributions that produce high-quality products, effectively enhancing the model's ability to focus on high-quality process intervals and its optimization guidance. Each parameter subinterval forms a quality-weighted multidimensional Gaussian distribution model, associated with a quality grade and reflecting real-world production patterns. Building on the Gaussian modeling, Gaussian mixture modeling is performed on the boundary regions between adjacent parameter subintervals. According to the sample distribution and parameter statistical characteristics in the boundary area of the sub-intervals, the influence weight of each interval in the boundary area, that is, the mixing coefficient, is dynamically calculated. By weighted superposition of the Gaussian distributions of two adjacent sub-intervals according to the mixing coefficient, a Gaussian mixture model of the boundary interval is obtained. This model can continuously describe the probability distribution and quality level of the process state in the parameter transition area. Based on the quality-weighted multidimensional Gaussian model and the Gaussian mixture model of the boundary area constructed above, a quality assessment function from parameter vector to quality score is established. Any set of process parameter vectors is input into the Gaussian distribution model or mixture model of the corresponding sub-interval, and the corresponding quality score under the parameter combination is obtained through probability density calculation, weighted integration and distribution mapping.In practical implementation, the quality scoring function is designed as the product of the Gaussian probability density and the subinterval quality weight, or the components of the Gaussian mixture model are weighted according to their probability weights and then input into the scoring function. Through this step, every fine-tuning, optimization, and adjustment of process parameters can be accurately reflected in the quality assessment function at a theoretical level, thus providing quantifiable quality targets for subsequent intelligent optimization algorithms.

[0027] In a specific embodiment, the process of executing step S103 may specifically include the following steps: Performing logarithmic transformation on the quality assessment function to obtain a logarithmic quality assessment function, and solving the parameter partial derivative of the logarithmic quality assessment function to obtain a gradient vector; Apply the corresponding gradient vector to the points in the parameter subinterval according to the interval to which they belong, and obtain the initial direction field in the subinterval; Perform hybrid model gradient calculation on points in the interaction region in the parameter space to obtain the direction field in the interaction region; Construct an adaptive step-size control function of the basic step-size and the distance attenuation coefficient, and dynamically adjust the step-size according to the adaptive step-size control function to obtain the gradient optimized step-size; A fully connected neural network calculation is performed on the initial direction field in the subinterval, the direction field in the interaction region, and the gradient optimization step to obtain the parameter optimized direction field.

[0028] Specifically, the quality assessment function, a core metric for optimizing canning process parameters, is essentially a probabilistic scoring function derived from a multidimensional Gaussian distribution and a mixed weight mapping. It accurately maps any combination of process parameters to a corresponding product quality level. To facilitate subsequent gradient analysis and optimization calculations, this quality assessment function undergoes a logarithmic transformation, converting the exponential Gaussian distribution function into a quadratic form in the logarithmic domain. This simplifies the partial derivatives and makes the gradient representation more intuitive. Through the logarithmic transformation, the original quality assessment function, such as Q(x), is transformed into ln(Q(x)), where x is a four-dimensional process parameter vector. This step preserves the impact of parameter changes on the target quality and transforms the model from a probability density space to a gradient-controlled optimization space. The partial derivatives of the logarithmic quality assessment function with respect to the four core process parameters of temperature, pressure, time, and speed are calculated analytically. The gradient vector of the logarithmic Gaussian distribution function with respect to the parameters is derived. This vector points to the direction of fastest quality improvement within the parameter space. Specifically, fine-tuning the process parameters in this direction will result in the maximum rate of increase in the target quality score. For each point within a specific parameter subinterval, the Gaussian model gradient vector within that interval is applied to assign a unique parameter adjustment direction to that point, forming the initial direction field within the parameter subinterval. However, at the boundaries and interaction regions of the actual process parameter space, a single Gaussian gradient cannot fully characterize the parameter optimization direction due to process state transitions between adjacent subintervals and the combined effects of multiple parameters. For points within these boundaries or interaction regions, a Gaussian mixture model gradient calculation is introduced. The gradients of the logarithmic quality assessment functions from two or more subintervals are weighted and superimposed according to the mixing coefficient to obtain the optimized direction field within the interaction region. An adaptive step-size control function is constructed, combining a basic step-size and a distance decay coefficient. A basic step-size parameter α0 is set. The step-size is dynamically decayed based on the Euclidean distance ||x-μ|| from the parameter point to the center of the current interval μ, as follows: α(x)=α0·exp(-β||x-μ||²), where β is the step-size decay coefficient. When the parameter point is far from the optimal center, a larger step size is used to accelerate convergence and quickly cross the low-quality interval. When the parameter point is close to the optimal center, the step size is automatically reduced to improve the fine-tuning accuracy and prevent parameter oscillation, overshoot or falling into the gradient oscillation area. Dynamic step size control enables the parameter optimization process to achieve efficient local convergence while maintaining global exploration capabilities, effectively avoiding problems such as falling into local extremes or slow convergence. On this basis, in order to improve the expressiveness, generalization ability and reasoning efficiency of the direction field in multi-parameter high-dimensional space, the entire parameter optimization direction field generation and update process is introduced into the fully connected neural network architecture. The interval label of the parameter point, the initial direction field, the interaction area direction field and the current step size feature are used as input, and a fully connected network with more than three layers is designed. Through end-to-end training, the nonlinear mapping relationship and local structural characteristics of the direction field in the complex parameter space are learned.The training process uses the mean square error between the true direction field and the neural network prediction vector as the loss function, and combines a large number of historical optimization samples for batch training to obtain a neural network direction field model that has the ability to express local details and globally fit complex spatial features.

[0029] In a specific embodiment, the process of executing step S104 may specifically include the following steps: The temperature and pressure parameters in the preprocessed heterogeneous sensor data are input into the first channel of the dual-channel parallel feature extraction network for feature extraction to obtain the temperature-pressure relationship feature vector; The time and speed parameters in the preprocessed heterogeneous sensor data are input into the second channel of the dual-channel parallel feature extraction network for feature extraction to obtain the time-speed relationship feature vector; Perform a splicing operation on the temperature-pressure relationship feature vector and the speed relationship feature vector to obtain the initial fused feature vector, and calculate the importance weight of each dimension of the initial fused feature vector to obtain the feature attention weight; The initial fused feature vector is weighted based on the feature attention weight to obtain the target fused feature.

[0030] Specifically, the temperature and pressure parameters from the preprocessed heterogeneous sensor data are fed into the first channel of a dual-channel parallel feature extraction network for feature extraction. This channel utilizes deep learning architectures such as multi-layer perceptrons (MLPs) or lightweight recurrent neural networks (e.g., LSTMs). Leveraging nonlinear transformations and multi-layer feature representation, this network exploits the inherent correlations between temperature and pressure, including coupled changes, sudden changes, long-term drift, and typical operating ranges, thereby outputting a temperature-pressure relationship feature vector. Simultaneously, the time and speed parameters are fed into the second channel of the dual-channel parallel feature extraction network. This channel can utilize a one-dimensional convolutional neural network (1D-CNN), a gated recurrent unit (GRU), or the same MLP architecture as the first channel to capture high-level behavioral features such as production cycles, speed fluctuations, and rhythmic changes. This network effectively extracts dynamic responses between time and speed, process coordination, and process bottlenecks through time series modeling, sliding window analysis, and multi-scale feature fusion, ultimately outputting a time-speed relationship feature vector. While structurally independent of the temperature and pressure channels, this feature vector, through end-to-end training of a deep network at the data representation level, adaptively captures complex patterns related to production efficiency, abnormal downtime, and energy consumption fluctuations. The temperature-pressure relationship feature vector and the speed-speed relationship feature vector are concatenated to generate an initial fused feature vector. The importance of each dimension in the fused feature vector is calculated. A feature attention mechanism is introduced, employing a trainable weight matrix or self-attention network structure. Through mechanisms such as softmax activation or sigmoid gating, each dimension of the fused feature vector is assigned a weight score between 0 and 1. The essence of the attention weight is to measure the actual contribution of each feature component to the final process optimization goal (such as improved sealing, optimized efficiency, and reduced energy consumption). Through end-to-end training, the model adaptively adjusts the weight distribution of each feature dimension, focusing on key feature components such as temperature-pressure coordination, speed dynamics, and interactive response, while automatically weakening the contribution of features with large statistical fluctuations, weak correlation with the target, or susceptible to noise. The initial fused feature vector is weighted based on the feature attention weights to form the target fused feature. The target fused feature with adaptive focusing capabilities and high expressiveness is obtained by element-wise multiplication of the feature dimension with the corresponding attention weight, or by multiplying the fused feature with the attention weight as a gating parameter and then normalizing it through a nonlinear activation layer.

[0031] In a specific embodiment, the process of executing step S105 may specifically include the following steps: Randomly generate multiple initial parameter vectors as initial particles in the canning process parameter space, and assign position and velocity attributes to each particle to obtain an initial particle swarm; The guidance vector of each particle is calculated based on the parameter optimization direction field, and the environmental adaptation coefficient is calculated by combining the target fusion characteristics to obtain the particle iterative update rule; The quality evaluation function is input into each particle position in the initial particle swarm to perform scoring calculation, and the quality status is predicted in combination with the target fusion features to obtain the particle fitness value; The initial particle swarm is sorted based on the particle fitness value, the particle with the highest fitness is selected as the global optimal solution, and the particles with fitness below the threshold are mutated to obtain the optimized particle swarm; According to the particle iterative update rule, the position and velocity of the optimized particle swarm are updated, and the mutation probability and learning factor are dynamically adjusted to obtain the next generation of particle swarm; The fitness evaluation and iterative optimization are repeated for the next generation of particle swarm until the termination condition is reached, and the global optimal solution is selected as the optimal process parameter combination.

[0032] Specifically, based on the actual physical constraints of the parameter domain and engineering experience, the upper and lower bounds of key process parameters such as temperature, pressure, time, and speed are determined. Based on these bounds, a global sampling strategy, such as uniform distribution or Latin hypercube sampling, is employed within the parameter space to randomly generate multiple initial parameter vectors. Each parameter vector serves as an independent particle in the particle swarm optimization algorithm. These initial particles are widely distributed within the parameter space, balancing the diversity of the global search with the feasibility of subsequent local convergence. Each particle is assigned initial position and velocity attributes. The position vector represents the parameter group coordinate of the particle in the parameter space, while the velocity vector controls the parameter movement step and direction of the particle in each iteration. This provides a theoretical basis and optimization guidance for intelligent iteration and adaptive updating of particles. Based on the in-depth modeling results of the parameter optimization direction field described above, a unique guidance vector is calculated for each particle. The guidance vector is determined by the gradient information of the current position in the parameter space and reflects the direction and trend in which parameter adjustments at the current point will maximize process quality assessment. Furthermore, target fusion features are introduced into the optimization process. The feature fusion model evaluates the compatibility between the current environmental state and the particle parameter combination, thereby calculating the environmental adaptation coefficient. The environmental adaptability coefficient reflects the degree to which a parameter set matches multiple process quality objectives (such as sealing, stability, and energy consumption). It also adjusts the weights of particles in specific process scenarios, making the optimization algorithm more tailored to actual production needs. Combining these two factors, a systematic iterative update rule for each particle is derived: a particle swarm intelligent search equation that incorporates both directional field gradient drive and environmental adaptability. For each particle in the initial particle swarm, its current position parameter set is input into a quality assessment function based on Gaussian mixture modeling to obtain a comprehensive quality score corresponding to the multi-dimensional parameters. To overcome the limitations of a single score, the particle's state vector and target fusion features are input into a deep neural network discriminant model to achieve multi-feature, multi-dimensional process quality state prediction. This prediction result is quantified as the particle's fitness value during the process optimization process. A higher fitness value indicates that the particle parameter set more closely meets the multi-objective requirements of product quality, efficiency, and stability. After the fitness calculation for each generation of the particle swarm is completed, all particles are globally ranked based on their fitness values to quickly identify the globally optimal and locally excellent solutions. The particles with the highest fitness are temporarily retained as candidates for the current global optimal solution, while particles with fitness below a threshold trigger genetic mutation. This mutation utilizes various mechanisms, such as Gaussian perturbation, crossover recombination, or partial parameter reset, to maintain population diversity and avoid local extreme value traps. This enhances the algorithm's global search capabilities and solution space coverage in complex nonlinear process spaces. In each optimization iteration, all particles jointly update their position and velocity based on their guidance vector, inertia term, historical optimal information, environmental adaptation coefficient, and mutation genetic factor.The calculation of a particle's new position incorporates the current velocity, the gravitational forces of the global optimal particle, and the individual historically optimal particle, while simultaneously balancing individual exploration with swarm convergence under the control of weighting factors. The mutation probability and learning factor are dynamically adjusted within each generation based on swarm diversity, adaptive convergence trends, and the gradient of the objective function. When the swarm converges prematurely or reaches a local minimum, the mutation probability and exploration level are automatically increased. When swarm fitness improves and the solution stabilizes, the learning step size is gradually reduced to improve convergence efficiency. The entire optimization evolutionary process is cyclical, repeating fitness evaluation, global optimal selection, mutation, and parameter updates within each generation. As the number of iterations increases, the particle swarm gradually converges toward the high-quality process region in the parameter space. The particle fitness distribution evolves from initial dispersion and multi-peaks to convergence and unimodality until termination criteria are met—such as reaching the set maximum number of generations, no significant improvement in global fitness, or particle position change falling below a convergence threshold. At this point, the particle with the highest fitness value is selected as the final global optimal solution, representing the optimal parameter combination for the current canning process.

[0033] Among them, the guidance vector of each particle is calculated based on the parameter optimization direction field, and the environmental adaptation coefficient is calculated in combination with the target fusion feature to obtain the particle iterative update rule, including: applying a linear interpolation algorithm to the parameter optimization direction field, calculating the local gradient direction at each particle position, and obtaining the particle local optimization direction; weighting the particle local optimization direction and the particle current speed, and the weight coefficient depends on the historical iteration effect to obtain the basic speed update vector; decoding and analyzing the target fusion feature, extracting the canning quality state prediction information and parameter sensitivity information, and obtaining the quality prediction matrix; performing regional quality scoring on the parameter space based on the quality prediction matrix, constructing a parameter regional adaptability mapping, and obtaining the environmental adaptation coefficient matrix; implementing adaptive learning factor adjustment on the environmental adaptation coefficient matrix, dynamically calculating the cognitive learning factor and the social learning factor according to the particle position and iteration progress, and obtaining the variable genetic factor; combining the basic speed update vector, the global optimal position, the individual historical optimal position and the variable genetic factor to construct a complete speed and position update formula, and obtaining the particle iterative update rule.

[0034] The above describes the canning process optimization method based on artificial intelligence in the embodiment of the present invention. The following describes the canning process optimization system based on artificial intelligence in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an artificial intelligence-based canning process optimization system includes: The preprocessing module 201 is used to perform time-domain and frequency-domain double filtering preprocessing on the original canning process sensor data to obtain preprocessed heterogeneous sensor data, and to adaptively segment the canning process parameter domain to obtain multiple parameter sub-intervals; A construction module 202 is configured to construct a quality assessment function based on a valid process parameter state data set within a parameter subinterval; A calculation module 203 is used to calculate the gradient of the parameter logarithmic function based on the quality evaluation function and perform adaptive step size control to obtain a parameter optimization direction field; The feature extraction module 204 is used to input the pre-processed heterogeneous sensor data into a dual-channel parallel feature extraction network for feature extraction and feature fusion to obtain target fusion features; The particle swarm optimization module 205 is used to perform variable genetic factor particle swarm optimization based on the parameter optimization direction field and the target fusion feature to generate an optimal process parameter combination.

[0035] Through the collaborative efforts of these components, a dual filtering preprocessing technique in the time and frequency domains efficiently cleans and aligns heterogeneous sensor data, such as temperature, pressure, and speed, significantly reducing data noise. Adaptive parameter domain segmentation based on the principle of maximizing information entropy accurately captures the boundaries of parameter-sensitive intervals, enabling refined modeling of the canning process parameter space and avoiding the loss of optimization accuracy often associated with overly coarse parameter space partitioning in traditional methods. A quality assessment function is constructed using a multidimensional Gaussian distribution model and a Gaussian mixture model, effectively describing the complex interactions between parameters and boundary transition characteristics, overcoming the technical shortcomings of traditional methods that ignore parameter interactions. Based on the logarithmic gradient of the quality assessment function and adaptive step-size control, a parameter optimization direction field is generated that points to quality improvement, providing a clear optimization path for parameter adjustment and preventing the optimization process from falling into local optima. A dual-channel parallel feature extraction network simultaneously captures both nonlinear and structural features in the parameter data, enabling deep feature fusion and attention weighting, enhancing the model's ability to identify key features. By combining the parameter optimization direction field and target fusion features, a variable genetic factor particle swarm optimization algorithm is implemented to dynamically adjust the learning factor and mutation probability, achieving a balance between global search and local refinement to achieve adaptive optimization of canning process parameters. By automatically generating and implementing the optimal parameter combination, combined with real-time parameter deviation monitoring and correction, a complete closed-loop control system for canning process optimization is established to ensure that the production process always operates in an optimal state.

[0036] above Figure 2 The artificial intelligence-based can sealing process optimization system in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The artificial intelligence-based can sealing process optimization equipment in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0037] Figure 3The figure is a schematic diagram of the structure of an AI-based canning process optimization device provided by an embodiment of the present invention. The AI-based canning process optimization device 300 may vary significantly depending on configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instructions for operating on the AI-based canning process optimization device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, allowing the AI-based canning process optimization device 300 to execute the series of instructions stored in the storage medium 330 to implement the steps of the aforementioned AI-based canning process optimization method.

[0038] The artificial intelligence-based canning process optimization device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the artificial intelligence-based canning process optimization equipment shown does not constitute a limitation on the artificial intelligence-based canning process optimization equipment provided by the present invention, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0039] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, cause the computer to execute the steps of the artificial intelligence-based canning process optimization method.

[0040] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0041] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing an artificial intelligence-based canning process optimization device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0042] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A canning process optimization method based on artificial intelligence, characterized in that: include: The original canning process sensor data is preprocessed by double filtering in the time and frequency domains to obtain preprocessed heterogeneous sensor data, and the canning process parameter domain is adaptively segmented to obtain multiple parameter sub-intervals; constructing a quality assessment function based on a valid process parameter state data set within the parameter subinterval; Based on the quality evaluation function, the parameter logarithmic function gradient is calculated and adaptive step size control is performed to obtain a parameter optimization direction field; Inputting the preprocessed heterogeneous sensor data into a dual-channel parallel feature extraction network for feature extraction and feature fusion to obtain target fusion features; Based on the parameter optimization direction field and the target fusion feature, variable genetic factor particle swarm optimization is performed to generate an optimal process parameter combination.

2. The artificial intelligence-based canning process optimization method according to claim 1, characterized in that: The original canning process sensor data is preprocessed by dual filtering in the time and frequency domains to obtain preprocessed heterogeneous sensor data, and the canning process parameter domain is adaptively segmented to obtain multiple parameter sub-intervals, including: The original can sealing process sensor data collected by temperature sensors, pressure sensors, and speed sensors on the can sealing production line are time-stamped and synchronized to obtain time-aligned multi-source heterogeneous data. Performing Z-score normalization transformation on the time-aligned multi-source heterogeneous data to obtain normalized heterogeneous data, and filling missing values in the normalized heterogeneous data to obtain complete normalized heterogeneous data; performing time-domain filtering on the temperature and pressure data in the complete standardized heterogeneous data to obtain time-domain filtered data; Converting the time-domain filtered data to the frequency domain through fast Fourier transform, and performing a bandpass filtering operation on the interference signal of a preset frequency to obtain frequency-domain filtered data; Performing an inverse Fourier transform on the frequency domain filtered data to restore it to the time domain to obtain preprocessed heterogeneous sensor data; Based on the pre-processed heterogeneous sensor data, the canning process parameter domain is adaptively segmented to obtain a plurality of parameter sub-intervals.

3. The artificial intelligence-based canning process optimization method according to claim 2, characterized in that: Based on the pre-processed heterogeneous sensor data, the canning process parameter domain is adaptively segmented to obtain multiple parameter sub-intervals, including: Based on the preprocessed heterogeneous sensor data, a canning process parameter space including a temperature interval, a pressure interval, a time interval, and a speed interval is constructed to obtain a canning process parameter domain; Calculating the information gain index for each parameter dimension in the canning process parameter domain to obtain parameter segmentation candidate points; Based on the parameter segmentation candidate points, three-point segmentation is performed on the temperature parameter, two-point segmentation is performed on the pressure parameter, two-point segmentation is performed on the time parameter, and two-point segmentation is performed on the speed parameter to obtain initial subintervals; The initial subinterval is evaluated using the information entropy maximization principle to obtain an optimized subinterval, and the conditional mutual information technology is used to calculate the mutual influence relationship between different parameters in the optimized subinterval to obtain a parameter interaction matrix; The parameter interaction matrix and historical production data are input into a gradient boosting decision tree model to obtain a plurality of parameter subintervals.

4. The artificial intelligence-based canning process optimization method according to claim 1, characterized in that: The constructing of a quality assessment function according to the effective process parameter state data set within the parameter sub-interval includes: Perform statistical analysis on historical production data within each parameter sub-interval to obtain an effective process parameter status data set; Calculating a four-dimensional mean vector of each parameter subinterval based on the effective process parameter state data set to obtain a parameter mean, and performing maximum likelihood estimation on the effective process parameter state data set to obtain a covariance matrix; Constructing a four-dimensional Gaussian distribution model of the parameter subinterval according to the parameter mean and the covariance matrix, and assigning different weights to product data points of different quality levels to obtain a quality-weighted Gaussian model; Mixing coefficients are calculated for boundary areas of adjacent parameter subintervals to obtain a Gaussian mixture model, and a quality assessment function from a parameter vector to a quality score is constructed based on the quality-weighted Gaussian model and the Gaussian mixture model.

5. The artificial intelligence-based canning process optimization method according to claim 1, characterized in that: The method of calculating the parameter logarithmic function gradient and performing adaptive step size control based on the quality evaluation function to obtain the parameter optimization direction field includes: Performing logarithmic transformation on the quality assessment function to obtain a logarithmic quality assessment function, and solving parameter partial derivatives of the logarithmic quality assessment function to obtain a gradient vector; Applying corresponding gradient vectors to points within the parameter subinterval according to the interval to which they belong, to obtain an initial direction field within the subinterval; Perform hybrid model gradient calculation on points in the interaction region in the parameter space to obtain the direction field in the interaction region; Constructing an adaptive step length control function of a basic step length and a distance attenuation coefficient, and dynamically adjusting the step length according to the adaptive step length control function to obtain a gradient optimized step length; A fully connected neural network calculation is performed on the initial direction field in the subinterval, the direction field in the interaction area, and the gradient optimization step size to obtain a parameter optimized direction field.

6. The artificial intelligence-based canning process optimization method according to claim 1, characterized in that: The pre-processed heterogeneous sensor data is input into a dual-channel parallel feature extraction network for feature extraction and feature fusion to obtain target fusion features, including: Inputting the temperature and pressure parameters in the preprocessed heterogeneous sensor data into the first channel of a dual-channel parallel feature extraction network for feature extraction to obtain a temperature-pressure relationship feature vector; Inputting the time and speed parameters in the preprocessed heterogeneous sensor data into the second channel of the dual-channel parallel feature extraction network for feature extraction to obtain a time-speed relationship feature vector; Performing a splicing operation on the temperature-pressure relationship feature vector and the speed relationship feature vector to obtain an initial fused feature vector, and calculating the importance weight of each dimension of the initial fused feature vector to obtain a feature attention weight; The initial fused feature vector is weighted based on the feature attention weight to obtain a target fused feature.

7. The artificial intelligence-based canning process optimization method according to claim 1, characterized in that: The performing of variable genetic factor particle swarm optimization based on the parameter optimization direction field and the target fusion feature to generate an optimal process parameter combination includes: Randomly generate multiple initial parameter vectors as initial particles in the canning process parameter space, and assign position and velocity attributes to each particle to obtain an initial particle swarm; Calculating the guidance vector of each particle based on the parameter optimized direction field, calculating the environmental adaptation coefficient in combination with the target fusion feature, and obtaining the particle iterative update rule; Input the quality evaluation function into each particle position in the initial particle swarm to perform score calculation, and perform quality status prediction in combination with the target fusion feature to obtain the particle fitness value; sorting the initial particle swarm based on the particle fitness values, selecting the particle with the highest fitness as the global optimal solution, and performing a mutation operation on particles with fitness lower than a threshold to obtain an optimized particle swarm; The position and velocity of the optimized particle swarm are updated according to the particle iterative update rule, and the mutation probability and learning factor are dynamically adjusted to obtain the next generation particle swarm; Repeating fitness evaluation and iterative optimization on the next generation particle swarm until a termination condition is reached, and selecting a global optimal solution as the optimal process parameter combination.

8. An artificial intelligence-based canning process optimization system, characterized in that: For implementing the artificial intelligence-based canning process optimization method according to any one of claims 1 to 7, the artificial intelligence-based canning process optimization system comprises: The preprocessing module is used to perform double filtering preprocessing in the time domain and frequency domain on the original canning process sensor data to obtain preprocessed heterogeneous sensor data, and to adaptively segment the canning process parameter domain to obtain multiple parameter sub-intervals; A construction module, configured to construct a quality assessment function based on a valid process parameter state data set within the parameter subinterval; A calculation module, configured to calculate the gradient of a parameter logarithmic function based on the quality evaluation function and perform adaptive step size control to obtain a parameter optimization direction field; A feature extraction module is used to input the pre-processed heterogeneous sensor data into a dual-channel parallel feature extraction network to perform feature extraction and feature fusion to obtain target fusion features; The particle swarm optimization module is used to perform variable genetic factor particle swarm optimization based on the parameter optimization direction field and the target fusion feature to generate an optimal process parameter combination.

9. An artificial intelligence-based canning process optimization device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the artificial intelligence-based canning process optimization method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is enabled to execute the artificial intelligence-based canning process optimization method according to any one of claims 1 to 7.

Images