Method and system for separating and controlling an aluminium-plastic composite packaging material

Abstract

The application provides an aluminum-plastic composite packaging material separation control method and system, relates to the technical field of composite material processing, and comprises the following steps: solving a microwave energy transmission equation to realize directional energy focusing to form an initial peeling area, collecting temperature distribution data to determine mechanical separation action parameters in combination with deep reinforcement learning, and evaluating a separation state through wavelet transform and support vector regression analysis of a strain signal and real-time adjustment of a control strategy, so that efficient and accurate separation of aluminum-plastic composite materials is realized, and resource recycling efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of composite material processing, and in particular to an aluminum-plastic composite packaging material separation control method and system. BACKGROUND

[0002] Aluminum-plastic composite packaging materials are widely used in the fields of food, medicine, cosmetics, etc., and have good barrier properties, freshness preservation and moisture resistance. The aluminum-plastic composite packaging material is usually composed of an aluminum foil layer and multiple layers of plastic, which are tightly combined through an adhesive. The multi-layer structure makes the material separation process face great challenges.

[0003] Traditional aluminum-plastic composite material separation methods mainly include chemical dissolution, thermal decomposition and mechanical peeling. The chemical dissolution method uses a specific solvent to dissolve the adhesive layer to achieve separation, but it will produce harmful chemical waste liquid, causing secondary pollution. The thermal decomposition method decomposes the material through high temperature, which has high energy consumption and produces harmful gases. The pure mechanical peeling method is difficult to precisely control the separation process due to the strong bonding between the layers of the aluminum-plastic composite material, resulting in low separation efficiency and serious material damage.

[0004] The existing technology lacks precise recognition and analysis of the structure of aluminum-plastic composite materials, cannot develop individualized separation schemes according to the specific structural characteristics of the material, is difficult to achieve precise energy control of the interlayer interface at the microscopic level, cannot form directional energy focusing between material layers, and lacks real-time monitoring and intelligent control mechanisms, cannot dynamically adjust the separation strategy according to the material state changes during the separation process, resulting in uncontrollable separation process, unstable separation quality and other problems. SUMMARY

[0005] The embodiments of the present application provide an aluminum-plastic composite packaging material separation control method and system, which can at least solve some of the problems in the prior art.

[0006] In a first aspect of the embodiments of the present application, an aluminum-plastic composite packaging material separation control method is provided, comprising:

[0007] Collecting image information of the aluminum-plastic composite packaging material, and extracting and identifying features of the image information through a convolutional neural network to obtain a first feature map and extract interlayer bonding parameters;

[0008] Constructing a microwave energy transfer equation based on the interlayer bonding parameters, iteratively solving the microwave energy transfer equation through a particle swarm optimization algorithm to obtain optimal distribution data of the microwave energy, and controlling the microwave radiation antenna array to form directional energy focusing at the interlayer interface according to the optimal distribution data to obtain an initial peeling area;

[0009] Collecting temperature distribution data of the initial peeling area to extract a second feature map and a heat distribution rule, calculating a material stress distribution trend according to the heat distribution rule, solving an optimal mechanical separation action sequence by combining a deep reinforcement learning algorithm, determining the force direction and force size of the mechanical separation device based on the optimal mechanical separation action sequence and executing the separation action;

[0010] Collecting strain signals of the mechanical separation device during the execution of the separation action, performing time-frequency analysis on the strain signals by wavelet transform to obtain a separation characteristic frequency band and calculate a separation state evaluation value by combining a support vector regression algorithm, predicting separation trend data based on the separation state evaluation value, and obtaining an optimal control strategy by combining the separation state evaluation value to calculate a microwave energy compensation value and a mechanical force compensation value in real time and executing the separation action.

[0011] In an alternative embodiment,

[0012] Collecting image information of the aluminum-plastic composite packaging material, extracting a first feature map and extracting interlayer bonding parameters by feature extraction and hierarchical recognition of the image information through a convolutional neural network, including:

[0013] Collecting image information of the aluminum-plastic composite packaging material, the image information including material surface topography features and material internal structure features;

[0014] Inputting the image information into a pre-trained convolutional neural network, and performing feature extraction and hierarchical recognition of the image information through multi-layer convolution and pooling operations of the convolutional neural network to obtain a first feature map containing material interlayer structure features;

[0015] Performing feature analysis on the first feature map to extract interlayer bonding parameters of the aluminum-plastic composite packaging material, the interlayer bonding parameters including interlayer bonding strength, interlayer bonding area, and interlayer bonding defect distribution.

[0016] In an alternative embodiment,

[0017] Based on the interlayer bonding parameters, a microwave energy transfer equation is constructed, and the microwave energy transfer equation is iteratively solved by a particle swarm optimization algorithm to obtain optimal distribution data of the microwave energy, and the microwave radiation antenna array is controlled to form directional energy focusing at the interlayer interface based on the optimal distribution data to obtain an initial peeling area, including:

[0018] Based on the interlayer bonding parameters, a microwave energy transfer equation is constructed, and the microwave energy transfer equation is iteratively solved by a particle swarm optimization algorithm to obtain optimal distribution data of the microwave energy, and the microwave radiation antenna array is controlled to form directional energy focusing at the interlayer interface based on the optimal distribution data to obtain an initial peeling area, including:

[0019] The particle swarm is divided into competitive subgroups and collaborative subgroups based on initial microwave energy distribution data, and an optimization objective function corresponding to each subgroup is set, the optimal solution of each subgroup is determined, the optimal solution of each subgroup is constructed into a target vector, and collaborative subgroup optimization is performed to obtain a global optimal solution set, the degree of conflict between targets is calculated according to the optimal solution of each subgroup and the global optimal solution set, and a comprehensive optimization result is obtained by fusing through a weight coefficient;

[0020] The distance between antenna units, the main lobe direction angle and the phase compensation value are calculated based on the comprehensive optimization result, the phase control parameter and the power distribution parameter are obtained, the phase control parameter and the power distribution parameter are used to control the microwave radiation antenna array to form a directional energy focusing region at the interlayer interface, and the energy distribution characteristics and the corresponding initial peeling region are obtained.

[0021] In an alternative embodiment,

[0022] The optimal solution of each subgroup is constructed into a target vector, and collaborative subgroup optimization is performed to obtain a global optimal solution set, the degree of conflict between targets is calculated according to the optimal solution of each subgroup and the global optimal solution set, and a comprehensive optimization result is obtained by fusing through a weight coefficient, which includes:

[0023] The component values of the target vector are calculated based on the optimal solution of each subgroup and the initial weight coefficient corresponding to each objective function, differential evolution optimization is performed based on the target vector, an exponential decay function is constructed according to the ratio of the current iteration number to the maximum iteration number, an adaptive scaling factor is constructed based on the exponential decay function and a pre-set initial scaling factor, and an optimal search step is calculated;

[0024] The correlation coefficient between the objective functions in the target vector is calculated based on the optimal search step, the adaptive crossover operator is constructed based on the correlation coefficient, and the difference vector generated by the random walk strategy is combined with the adaptive crossover operator and the optimal solution of each subgroup to calculate the mutation vector;

[0025] The extended subgroups are generated by expanding each subgroup based on the mutation vector, and non-dominated sorting is performed to obtain individual ranks, the crowding distance value is calculated for individuals with the same individual rank to determine high-quality individuals, and the high-quality individuals in each extended subgroup are obtained to obtain the global optimal solution set;

[0026] The function values of the global optimal solution set on each objective function are calculated, the degree of conflict between targets is calculated in combination with the optimal solution of each subgroup, the weight coefficients of each objective function are updated in combination with the exponential decay function, and the comprehensive optimization result is calculated based on the updated weight coefficients and the degree of conflict between targets.

[0027] In an alternative embodiment,

[0028] Collecting temperature distribution data of the initial peeling area to extract a second feature map and a heat distribution law, and calculating a material stress distribution trend according to the heat distribution law, including:

[0029] A multi-band infrared thermal imaging array is used to dynamically scan the initial peeling area to obtain temperature distribution data, and a second feature map is obtained by enhancing the temperature distribution data, and the second feature map is input into a spatio-temporal graph neural network. The node feature matrix and edge feature update equation of the spatio-temporal graph neural network are used to extract the spatio-temporal characteristics of heat conduction, and the multi-head self-attention mechanism is used to capture long-range heat conduction dependencies to obtain the heat distribution law.

[0030] Based on the heat distribution law and the pre-acquired energy distribution characteristics, a capsule network is trained, a coupling coefficient is calculated through a dynamic routing mechanism, and a multi-level stress field is constructed according to the coupling coefficient. The multi-level stress field is input into a variational autoencoder to reduce the dimensionality of the features to obtain the material stress distribution trend.

[0031] In an alternative embodiment,

[0032] An optimal mechanical separation action sequence is determined based on the optimal mechanical separation action sequence, and a separation action is performed, including:

[0033] Topological features are extracted from the material stress distribution trend using a dual-channel spatial transformation network and input into a fractional order neural network, a state transition equation is constructed based on a non-integer order differential operator, a non-uniform sampling is performed on a continuous state space through a fractional order dynamic programming method, and a nonlinear state obtained by sampling is mapped to the Fourier domain for fast solving through a Mellin transform, to obtain a separation action strategy.

[0034] An optimal mechanical separation action sequence is generated according to the separation action strategy, and a separation action is performed by determining the direction and size of the force of the mechanical separation device based on a pre-set action evaluation index.

[0035] In an alternative embodiment,

[0036] Strain signals of the mechanical separation device during the execution of the separation action are collected, time-frequency analysis of the strain signals is performed through wavelet transform, and a separation feature frequency band is obtained, and a separation state evaluation value is calculated by combining a support vector regression algorithm, including:

[0037] Strain signals of the mechanical separation device during the execution of the separation action are collected;

[0038] The strain signal is segmented according to a preset scale interval by using a box dimension calculation formula, the minimum number of boxes required to cover the strain signal is calculated in each scale segment, the self-similarity feature is obtained based on the change rate of the minimum number of boxes with the scale, the strain signal is constructed into a delay sequence according to a preset time delay interval to obtain a state trajectory, and the distance change rate between adjacent sampling time points of the state trajectory is calculated to obtain a maximum Lyapunov exponent;

[0039] The strain signal is wavelet transformed to obtain a time-frequency distribution, a first threshold is set according to the self-similarity feature, a feature scale interval is determined as a region in the time-frequency distribution that exceeds the first threshold, a second threshold is set according to the maximum Lyapunov exponent, a dynamic unstable region is determined as a region in the time-frequency distribution that exceeds the second threshold, and an overlap part of the feature scale interval and the dynamic unstable region is extracted to obtain a separation feature frequency band;

[0040] The separation feature frequency band is input into a support vector regression algorithm, and a separation state evaluation value is calculated.

[0041] In an optional implementation,

[0042] Based on the separation state evaluation value, separation trend data is predicted, and a microwave energy compensation value and a mechanical action force compensation value are calculated in real time in combination with the separation state evaluation value to obtain an optimal control strategy and execute, which includes:

[0043] A separation state evaluation value sequence is input into a long short-term memory network, and separation trend data is predicted by the long short-term memory network;

[0044] A deviation value between the separation trend data and a target separation state is calculated, and a microwave energy compensation value and a mechanical action force compensation value are calculated according to the deviation value and a change rate of the deviation value;

[0045] The microwave energy compensation value is added to a pre-set microwave energy reference value to obtain a microwave energy control instruction, and the mechanical action force compensation value is added to a pre-set mechanical action force reference value to obtain a mechanical action force control instruction;

[0046] The separation control is performed according to the microwave energy control instruction and the mechanical action force control instruction.

[0047] In a second aspect of the embodiment of the application, an aluminum-plastic composite packaging material separation control system is provided, which includes:

[0048] A first unit is configured to acquire image information of the aluminum-plastic composite packaging material, extract features and perform hierarchical recognition on the image information by using a convolutional neural network to obtain a first feature map and extract inter-layer joint parameters;

[0049] A second unit is configured to construct a microwave energy transmission equation based on the interlayer bonding parameters, iteratively solve the microwave energy transmission equation by using a particle swarm optimization algorithm to obtain optimal distribution data of the microwave energy, and control the microwave radiation antenna array to form directional energy focusing at the interlayer interface according to the optimal distribution data to obtain an initial separation region.

[0050] A third unit is configured to collect temperature distribution data of the initial separation region, extract a second feature spectrum and heat distribution law, calculate a material stress distribution trend according to the heat distribution law, solve an optimal mechanical separation action sequence by using a deep reinforcement learning algorithm, determine the direction and size of the action force of the mechanical separation device based on the optimal mechanical separation action sequence, and perform a separation action.

[0051] A fourth unit is configured to collect a strain signal of the mechanical separation device during the execution of the separation action, perform time-frequency analysis on the strain signal by using a wavelet transform to obtain a separation characteristic frequency band, calculate a separation state evaluation value by using a support vector regression algorithm, predict separation trend data based on the separation state evaluation value, and obtain an optimal control strategy by calculating a microwave energy compensation value and a mechanical action force compensation value in real time based on the separation state evaluation value and then executing the control strategy.

[0052] In a third aspect of the embodiment of the present application, an electronic device is provided, which includes:

[0053] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0054] In the present application, the hierarchical structure and bonding characteristics of the aluminum-plastic composite packaging material are accurately identified by using a convolutional neural network, the accurate identification and parameter extraction of the interlayer characteristics of the material are realized, accurate data basis is provided for subsequent microwave energy transmission and mechanical separation, the pertinence and accuracy of the separation process are significantly improved, the microwave radiation technology and the mechanical separation method are combined, the particle swarm algorithm is used to optimize the microwave energy distribution, directional energy focusing is formed at the interlayer interface of the material, the deep reinforcement learning algorithm is used to solve the optimal mechanical separation action sequence, the efficient separation of the aluminum-plastic composite material is realized, the separation efficiency is greatly improved and the energy consumption is reduced, the separation state is evaluated in real time by using the wavelet transform and the support vector regression algorithm, the microwave energy and the mechanical action force parameters are dynamically adjusted, the closed-loop control strategy is formed, the problems of excessive processing or insufficient separation in the traditional separation method are effectively avoided, and the material recycling quality and the resource utilization rate are improved. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 FIG. 1 is a flowchart of an aluminum-plastic composite packaging material separation control method according to an embodiment of the present application;

[0056] Figure 2The flow chart of the cooperative sub-group optimization process of the separation control method of the aluminum-plastic composite packaging material of the embodiment of the present application. DETAILED DESCRIPTION

[0057] To make the purpose, technical solutions and advantages of the embodiment of the present application clearer, the technical solutions in the embodiment of the present application will be described clearly and completely below in combination with the drawings in the embodiment of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0058] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0059] Figure 1 The flow chart of the separation control method of the aluminum-plastic composite packaging material of the embodiment of the present application is shown in Figure 1 The method comprises:

[0060] Collecting image information of the aluminum-plastic composite packaging material, extracting and identifying features of the image information through a convolutional neural network to obtain a first feature map and extract interlayer joint parameters;

[0061] Constructing a microwave energy transfer equation based on the interlayer joint parameters, iteratively solving the microwave energy transfer equation through a particle swarm optimization algorithm to obtain optimal distribution data of the microwave energy, controlling the microwave radiation antenna array to form directional energy focusing at the interlayer interface according to the optimal distribution data to obtain an initial peeling area;

[0062] Collecting temperature distribution data of the initial peeling area to extract a second feature map and heat distribution law, calculating material stress distribution trend according to the heat distribution law, solving an optimal mechanical separation action sequence combining a deep reinforcement learning algorithm, determining the direction and size of the action force of the mechanical separation device based on the optimal mechanical separation action sequence and executing the separation action;

[0063] Collecting strain signals of the mechanical separation device during the execution of the separation action, performing time-frequency analysis on the strain signals through wavelet transform to obtain a separation characteristic frequency band and calculate a separation state evaluation value combining a support vector regression algorithm, predicting separation trend data based on the separation state evaluation value, and obtaining an optimal control strategy by real-time calculating microwave energy compensation values and mechanical action force compensation values combining the separation state evaluation value and executing the separation action.

[0064] In an alternative embodiment,

[0065] Collecting image information of the aluminum-plastic composite packaging material, extracting and identifying features of the image information through a convolutional neural network to obtain a first feature map and extracting interlayer bonding parameters including:

[0066] Collecting image information of the aluminum-plastic composite packaging material, the image information including surface topography features and internal structure features of the material;

[0067] Inputting the image information into a pre-trained convolutional neural network, extracting and identifying features of the image information through multi-layer convolutional operations and pooling operations of the convolutional neural network to obtain a first feature map containing interlayer structure features of the material;

[0068] Analyzing features of the first feature map and extracting interlayer bonding parameters of the aluminum-plastic composite packaging material, the interlayer bonding parameters including interlayer bonding strength, interlayer bonding area, and interlayer bonding defect distribution.

[0069] Collecting image information of the aluminum-plastic composite packaging material, using a high-resolution industrial camera in the collecting process, the pixel resolution of the camera being 4096x3072, the focal length range being 8-50mm, and the camera being adjustable according to the size of the material. In the light source setting, a combination of a ring-shaped LED light source and a lateral LED light source is used, the ring-shaped light source having a power of 30W and a color temperature of 5500K, and being used to obtain surface topography features of the material; the lateral light source having a power of 20W and an irradiation angle of 45°, and being used to enhance the display effect of the surface texture features of the material. When collecting the surface topography features of the material, the sample is placed on a flat black background plate, and the camera is vertically downward to shoot at a shooting distance of 30cm. For the collection of the internal structure features of the material, an industrial CT scanning device is used, the tube voltage is set to 120kV, the tube current is 100mA, the scanning layer thickness is 0.05mm, and the three-dimensional reconstruction accuracy is 0.01mm, the tomographic images of the material are obtained, and exemplarily, a typical aluminum-plastic composite packaging material sample has a size of 100mmx100mm, a thickness of 0.2mm, and contains a three-layer structure of polyethylene (PE), aluminum foil and polyester (PET).

[0070] After the image collection is completed, the obtained image information is input into a pre-trained convolutional neural network, the convolutional neural network adopting a VGG architecture and containing 5 convolutional blocks, each convolutional block containing 2-3 convolutional layers and one max-pooling layer. The first convolutional block uses 64 3x3 convolutional kernels, the second convolutional block uses 128 3x3 convolutional kernels, the third convolutional block uses 256 3x3 convolutional kernels, and the fourth and fifth convolutional blocks both use 512 3x3 convolutional kernels. The step length of the convolutional operation is 1, and the padding mode is SAME; the kernel size of the max-pooling layer is 2x2, and the step length is 2. ReLU is used as the activation function in the network to improve the nonlinear expression ability and calculation efficiency of the network.

[0071] The network was trained on 5000 pre-labeled A-PET images, 4000 of which were used for training and 1000 for validation. The batch size was set to 32, the initial learning rate was set to 0.001, the Adam optimizer was used, and the training was performed for 100 epochs. To prevent overfitting, a Dropout layer was added to the network with a dropout rate of 0.5.

[0072] The convolutional neural network performs multiple layers of convolution and pooling operations on the input image information. The first convolutional block extracts basic edge and texture features, such as fine cracks and texture changes on the material surface. The second convolutional block identifies simple shape features, such as the boundary lines and joint areas between material layers. The third convolutional block extracts more complex structural features, such as the integrity and continuity of the interlayer joint. The fourth and fifth convolutional blocks identify high-level semantic features, such as the overall interlayer structure distribution and defect patterns of the material. After a series of processing, the first feature map containing the interlayer structure features of the material is generated, with a size of 7x7x512.

[0073] The first feature map is analyzed for feature extraction, and the interlayer joint parameters of the A-PET packaging material are extracted. The feature analysis uses a fully connected network structure, which includes three fully connected layers with node numbers of 1024, 512, and 256, respectively. The first fully connected layer maps the flattened feature vector of the first feature map to a 1024-dimensional space, the second fully connected layer further compresses the features to 512 dimensions, and the third fully connected layer outputs a 256-dimensional feature vector containing the interlayer joint parameter information of the material. ReLU activation functions are used between the fully connected layers, and a linear activation function is used in the last layer.

[0074] The interlayer bonding strength, interlayer bonding area, and interlayer bonding defect distribution are extracted from the feature vector. The interlayer bonding strength is represented by the first 85 elements of the feature vector, with a range of 2.5-7.8 N / cm 2 in actual testing, and an accuracy of 0.1 N / cm 2 . For example, for a typical aluminum foil-PE layer bonding strength, the model predicts a value of 5.6 N / cm 2 , with an error of less than 5% compared to the actual measured value of 5.8 N / cm 2 . The interlayer bonding area is represented by elements 86-170 of the feature vector, given in percentage form, with a range of 75%-99% in actual applications, and a model detection accuracy of ±2%. For example, the predicted value of the aluminum foil-PET layer bonding area is 92.5%, and the actual measured value is 93.1%. The interlayer bonding defect distribution is represented by elements 171-256 of the feature vector, presented in the form of a heat map, with a defect size accuracy of 0.5 mm 2 and a position accuracy of 1 mm.

[0075] In this embodiment, by collecting the image information of the aluminum-plastic composite packaging material, the surface topography characteristics and internal structure characteristics of the material can be obtained at the same time, thereby providing comprehensive data support for subsequent analysis. Through the multi-layer convolution operation and pooling operation of the pre-trained convolutional neural network on the image information, the deep feature extraction and hierarchical recognition of the material interlayer structure can be automatically completed, avoiding the problems of low efficiency and insufficient precision of traditional manual detection methods. By analyzing the obtained first feature spectrum, the key parameters such as interlayer bonding strength, bonding area and defect distribution can be accurately extracted, thereby realizing the accurate evaluation of the interlayer bonding quality of the aluminum-plastic composite packaging material. This not only improves the automation and intelligence level of detection, but also significantly improves the reliability and accuracy of the detection results, providing strong support for subsequent process optimization and material quality control.

[0076] In an alternative embodiment,

[0077] A microwave energy transfer equation is constructed based on the interlayer bonding parameters, and the optimal distribution data of the microwave energy is obtained by iteratively solving the microwave energy transfer equation through a particle swarm optimization algorithm. The optimal distribution data is used to control the microwave radiation antenna array to form directional energy focusing at the interlayer interface, and an initial peeling area is obtained.

[0078] A microwave energy transfer equation is constructed based on the interlayer bonding parameters, and the initial microwave energy distribution data is obtained by establishing a transfer relationship combining the energy diffusion coefficient, electrical conductivity and electric field strength. The particle swarm in the particle swarm optimization algorithm is initialized based on the initial energy distribution data.

[0079] The particle swarm is divided into competitive subgroups and cooperative subgroups based on the initial microwave energy distribution data, and the optimization objective function corresponding to each subgroup is set. The optimal solutions of each subgroup are determined, the optimal solutions of each subgroup are constructed into a target vector, and the global optimal solution set is obtained by executing cooperative subgroup optimization. The target conflict degree is calculated based on the optimal solutions of each subgroup and the global optimal solution set, and the comprehensive optimization result is obtained by weight coefficient fusion.

[0080] Based on the comprehensive optimization result, the antenna element spacing, main lobe direction angle and phase compensation value are calculated, the phase control parameter and power distribution parameter are obtained, and the microwave radiation antenna array is controlled to form a directional energy focusing area at the interlayer interface by using the phase control parameter and power distribution parameter. The energy distribution characteristics and the corresponding initial peeling area are obtained.

[0081] Based on the obtained interlayer bonding parameters, a microwave energy transfer equation is constructed. The microwave energy transfer equation takes into account the differences in physical properties between the layers of the aluminum-plastic composite material. The energy diffusion coefficient D value is introduced in the energy transfer equation. This coefficient is positively correlated with the dielectric constant and loss angle of the material. In the aluminum-plastic composite material, the D value of the polyethylene layer is about 0.35x10^-3 m 2D value of the aluminum foil layer is about 1.25 x 10-3m 2 D value of the polyester layer is about 0.42 x 10-3m 2 Meanwhile, considering the conductivity of the materials, the conductivity of the aluminum foil layer is 3.8 x 107S / m, the conductivity of the polyethylene layer is 10-16S / m, and the conductivity of the polyester layer is 10-14S / m. The electric field intensity E and the material properties jointly affect the distribution of microwave energy in the material. Through experiments, it is verified that when the microwave frequency is 2.45 GHz and the input power is 1000 W, the electric field intensity at the PE-aluminum foil interface is about 2.3 x 104V / m, and the electric field intensity at the aluminum foil-PET interface is about 1.8 x 104V / m.

[0082] Based on the above parameters, the energy transfer relationship is established, and the finite difference method is used to numerically solve the transfer equation. The grid division precision is 0.2 mm, the time step is 0.01 s, and the iteration number is 1000 times. The initial microwave energy distribution data is obtained. The initial microwave energy distribution data is represented in the form of a three-dimensional matrix, with a matrix dimension of 50 x 50 x 3, corresponding to the spatial energy distribution of the aluminum-plastic composite material sample, and the data range is 0-1000 J / cm 3 At the PE-aluminum foil interface, the energy density peak value reaches 850 J / cm 3 At the aluminum foil-PET interface, the energy density peak value reaches 780 J / cm 3 .

[0083] According to the initial microwave energy distribution data, the particle swarm in the particle swarm optimization algorithm is initialized, and the total number of particles is set to 100. Each particle contains 20 dimensions, respectively corresponding to the position coordinates, power distribution coefficients and phase shift values of the microwave antenna array. The initial position of the particle is generated based on the initial energy distribution data, and the position range is limited to the [-10, 10] interval, and the initial speed is randomly generated in the [-1, 1] interval.

[0084] The particle swarm is divided into a competitive sub-swarm and a cooperative sub-swarm, the competitive sub-swarm contains 60 particles, and the cooperative sub-swarm contains 40 particles. The optimization objective function of the competitive sub-swarm is set to maximize the energy density value at the target interface to promote the separation of the material at the specific interface; the optimization objective function of the cooperative sub-swarm is set to minimize the energy diffusion in the non-target area to reduce the thermal damage to the non-target area. For the competitive sub-swarm, the fitness function is positively correlated with the energy density value at the interface, and when the interface energy density exceeds the critical value 900 J / cm 3 , the fitness value reaches the maximum; for the cooperative sub-swarm, the fitness function is negatively correlated with the standard deviation of the energy density in the non-target area, and when the standard deviation is less than 50 J / cm 3 , the fitness value reaches the maximum.

[0085] The optimization process adopts the inertia weight decreasing strategy, the initial inertia weight is 0.9, the terminal inertia weight is 0.4, the learning factors c1 and c2 are both set to 2.0, and the maximum iteration number is 200. In the iteration process, after every 10 iterations, the optimal solution of each sub-population is evaluated to determine the optimal particle position of each sub-population. The optimal solution of the competitive sub-population reaches stability at the 150th iteration, and the optimal solution corresponds to an interfacial energy density of 935 J / cm3; the optimal solution of the cooperative sub-population reaches stability at the 130th iteration, and the optimal solution corresponds to a non-target region energy density standard deviation of 42 J / cm3.

[0086] The optimal solution of each sub-population is constructed as a target vector, and the cooperative sub-population optimization is performed by convex combination. In the cooperative optimization process, an adaptive weight adjustment mechanism is introduced, and the weight coefficient is dynamically adjusted according to the current iteration round. In the initial stage, the weight of the competitive sub-population is 0.7, and the weight of the cooperative sub-population is 0.3; as the iteration proceeds, the weight of the cooperative sub-population gradually increases, and finally the weight of the competitive sub-population is 0.5 and the weight of the cooperative sub-population is 0.5. Through 30 rounds of cooperative optimization iteration, a global optimal solution set is obtained, which contains 10 non-dominated solutions, each corresponding to a different energy distribution scheme.

[0087] The target conflict degree between each sub-population optimal solution and the global optimal solution set is calculated, and the conflict measure adopts cosine similarity, with a value range of [-1, 1], and the closer the value is to 1, the lower the conflict degree. The average conflict degree of the competitive sub-population optimal solution and the global optimal solution set is 0.35, and the average conflict degree of the cooperative sub-population optimal solution and the global optimal solution set is 0.28. Based on the conflict degree result, the comprehensive optimization result is obtained by weight coefficient fusion, and the weight distribution is: the weight of the competitive sub-population is 0.6, and the weight of the cooperative sub-population is 0.4. The integrated solution after fusion has an energy density of 920 J / cm 3 at the PE-aluminum foil interface, and a non-target region energy density standard deviation of 55 J / cm 3 .

[0088] Based on the comprehensive optimization result, the microwave antenna array parameters are calculated, and the antenna array adopts a 4x4 planar array structure with a working frequency of 2.45 GHz. The calculated antenna element spacing is 62 mm, which is about half of the working wavelength; the main lobe direction angle azimuth is 35°, and the elevation angle is 42°; the phase compensation value range is 0-330°, with a step of 30°. The phase control parameter adopts 8-bit digital control, with an accuracy of 1.4°; the power distribution parameter is realized by a power distribution network, with a distribution ratio range of 0.1-1.0 and a step of 0.1, and the actual power range is 100-1000 W.

[0089] The phase control parameters and the power distribution parameters are applied to the microwave radiation antenna array to control the formation of a directional energy focusing region of microwaves at the PE-aluminum foil interlayer interface. The energy focusing region is elliptical, with a long axis of about 20 mm and a short axis of about 15 mm, and the central value of the energy density reaches 950 J / cm 3 , and the edge value is about 750 J / cm 3 . The energy distribution characteristics exhibit a Gaussian distribution with a high center and a low edge, and the attenuation gradient is 45 J / cm 3 / mm. Based on the energy distribution characteristics, an initial peeling region is formed on the surface of the material, with a diameter of about 18 mm and a depth of about 0.1 mm, and the peeling edge is neat without obvious thermal damage.

[0090] In this embodiment, the microwave energy transfer equation is established based on the interlayer bonding parameters, and the energy transfer relationship is constructed by combining the energy diffusion coefficient, the electrical conductivity and the electric field strength, which can accurately describe the distribution state of microwave energy in the aluminum plastic composite material, so as to obtain reasonable initial energy distribution data. Through the setting of different optimization objective functions and the cooperative updating of optimal solutions, the local search ability and the global exploration ability can be effectively balanced in the global range, the optimization convergence speed and the diversity of solutions are improved, and the comprehensive optimization result obtained by the weight coefficient fusion not only solves the conflict problem between different objectives, but also ensures the overall balance and stability of the optimization result. The antenna element spacing, the main lobe direction angle and the phase compensation value calculated by using the comprehensive optimization result can realize the accurate control of phase and power, so that the microwave radiation antenna array forms a directional energy focusing region at the interlayer interface, thereby obtaining a stable and controllable energy distribution characteristic and a corresponding initial peeling region. The efficiency and reliability of the aluminum plastic composite material peeling process are significantly improved.

[0091] In an alternative embodiment,

[0092] The optimal solutions of each sub-population are constructed into a target vector and a global optimal solution set is obtained by performing cooperative sub-population optimization, the conflict degree between each sub-population optimal solution and the global optimal solution set is calculated, and a comprehensive optimization result is obtained by weight coefficient fusion, including:

[0093] The component values of the target vector are calculated based on the optimal solutions of each sub-population and the initial weight coefficients corresponding to each objective function, differential evolution optimization is performed based on the target vector, an exponential decay function is constructed according to the ratio of the current iteration number to the maximum iteration number, an adaptive scaling factor is constructed based on the exponential decay function and the pre-set initial scaling factor, and the optimal search step is calculated.

[0094] The correlation coefficient between the objective functions in the target vector is calculated based on the optimal search step, an adaptive crossover operator is constructed based on the correlation coefficient, and a difference vector generated by a random walk strategy is combined with the adaptive crossover operator and the optimal solution of each sub-population to obtain a mutation vector;

[0095] Each sub-population is expanded based on the mutation vector to generate an expanded sub-population and is subjected to non-dominated sorting to obtain individual ranks. The crowding distance value of individuals with the same individual rank is calculated to determine high-quality individuals, and the high-quality individuals in each expanded sub-population are obtained to obtain a global optimal solution set.

[0096] The function value of the global optimal solution set on each objective function is calculated, the degree of conflict between the objectives is calculated in combination with the optimal solution of each sub-population, the weight coefficients of each objective function are updated in combination with the exponential decay function, and a comprehensive optimization result is calculated based on the updated weight coefficients and the degree of conflict between the objectives.

[0097] The component values of the target vector are calculated based on the optimal solution of each sub-population and the initial weight coefficients corresponding to each objective function. In the separation control of aluminum-plastic composite packaging materials, two main optimization objectives are set: one is to maximize the energy density at the interlayer interface, and the other is to minimize the thermal damage in the non-target area. The energy density function value corresponding to the optimal solution of the competitive sub-population is 935 J / cm3, and the thermal damage control function value corresponding to the optimal solution of the cooperative sub-population is 42 J / cm 3 In the initial stage, the weight coefficient of the energy density objective is set to 0.65, and the weight coefficient of the thermal damage control objective is set to 0.35. The component values of the target vector are calculated by weighted summation. The dimension of the target vector is 20, corresponding to the spatial position coordinates (x, y, z) of the microwave array, the power distribution coefficient, and the phase adjustment value. The initial value range of each component of the target vector is [-10, 10], the average value is 2.5, and the standard deviation is 3.2.

[0098] Differential evolution optimization is performed based on the target vector. The differential evolution optimization adopts an adaptive parameter control strategy. An exponential decay function is constructed according to the ratio of the current iteration number to the maximum iteration number. The maximum iteration number of the differential evolution algorithm is set to 500 times. When the iteration ratio is 0.2, the exponential decay function value is 0.82; when the iteration ratio is 0.5, the function value is 0.61; and when the iteration ratio is 0.8, the function value is 0.45. The initial scaling factor is set to 0.8, and an adaptive scaling factor is constructed in combination with the exponential decay function. In the early stage of iteration, the adaptive scaling factor value is about 0.76, which promotes global search; in the middle stage of iteration, the scaling factor decreases to about 0.65, which balances global and local search; and in the later stage of iteration, the scaling factor further decreases to below 0.5, which enhances the local search capability. The optimal search step is calculated based on the adaptive scaling factor, and the search step gradually decreases with iteration.

[0099] The correlation coefficient between the energy density objective and the thermal damage control objective is about-0.72, indicating that the two objectives have a strong negative correlation, i.e., increasing the interface energy density tends to increase the risk of thermal damage. Based on the correlation coefficient, an adaptive crossover operator is constructed, and the initial value of the crossover probability is set to 0.6, which increases with the absolute value of the correlation coefficient. When the absolute value of the correlation coefficient reaches 0.8, the crossover probability increases to 0.75, enhancing the exploration ability of the algorithm. A difference vector is generated through a random walk strategy, and the random walk step is set to 0.5. The walk dimension is randomly selected from 5 to 10 dimensions of the target vector. A mutation vector is calculated based on the adaptive crossover operator and the optimal solution of each sub-population. The generation of the mutation vector adopts the "optimal individual guidance" strategy, i.e., one base vector of the difference vector is selected as the optimal individual in the current population.

[0100] Based on the mutation vector, the sub-populations are expanded to generate expanded sub-populations, and the expansion coefficient is set to 1.5, i.e., the size of the expanded sub-population is 1.5 times that of the original sub-population. The competitive sub-population is expanded from 60 to 90, and the cooperative sub-population is expanded from 40 to 60. The expanded sub-populations are non-dominantly sorted, and the fast non-dominant sorting algorithm is used to calculate the dominated solution set and the dominated count of each individual. In the actual sorting, the number of non-dominant solutions in the first grade is 22, the second grade is 35, the third grade is 53, and the fourth grade is 40. The crowding distance value is calculated for individuals with the same individual grade. The crowding distance calculation considers the Euclidean distance between individuals in the objective space. The larger the distance, the higher the crowding distance value. In the first grade of non-dominant solutions, the individual with the highest crowding distance value corresponds to an energy density of 915 J / cm 3 and a thermal damage control value of 48 J / cm 3 . The individual with the lowest crowding distance value corresponds to an energy density of 908 J / cm 3 and a thermal damage control value of 52 J / cm 3 . Based on the grade and crowding distance, high-quality individuals are determined, and the selection criteria are: low grade first, and high crowding distance first in the same grade. The top 40% of individuals in the expanded sub-population are selected as high-quality individuals, and the high-quality individuals in each expanded sub-population form the global optimal solution set, with a size of 60.

[0101] The function values of the global optimal solution set on each objective function are calculated. The function value range of the energy density objective is 890-940 J / cm 3 , with an average value of 912 J / cm 3 . The function value range of the thermal damage control objective is 40-65 J / cm 3 , with an average value of 53 J / cm 3The conflict degree between the energy density target and the thermal damage control target in the current iteration is 0.42, which is lower than 0.72 in the initial stage, indicating that the optimization process effectively alleviates the target conflict. The weight coefficients of each target function are updated in combination with the exponential decay function. In the current iteration stage (iteration ratio 0.6), the weight coefficient of the energy density target is updated to 0.58, and the weight coefficient of the thermal damage control target is updated to 0.42. The weight adjustment reflects the gradual emphasis on thermal damage control as the optimization progresses.

[0102] The comprehensive optimization result is calculated based on the updated weight coefficients and the conflict degree between the targets. The linear weighting method is used to fuse multiple targets, and a conflict compensation factor is introduced. The compensation coefficient is set to 0.15. The comprehensive optimization result forms an energy focusing area with a diameter of 22mm at the PE-aluminum foil interface, with a center energy density of 925J / cm 3 , an edge energy density of 780J / cm 3 , an energy gradient of 40J / cm 3 / mm; the maximum energy density of the non-target area is 350J / cm 3 , the average energy density is 180J / cm 3 , and the standard deviation is 45J / cm 3 . The corresponding microwave array parameters are: element spacing 64mm, main lobe direction angle azimuth 32° and elevation 40°, power distribution ratio center element 1.0, edge element 0.65, and phase compensation value range 15°-330°.

[0103] In this embodiment, the exponential decay function and the adaptive scaling factor are introduced, so that the search step can be dynamically adjusted with the iteration process, improving the convergence speed and accuracy of the optimization. The adaptive crossover operator is constructed based on the correlation coefficient between the target functions, and the difference vector is generated by combining the random walk strategy to perform mutation, which expands the diversity of the search space and reduces the probability of falling into local optimum. Through non-dominated sorting and congestion calculation of the individuals after the expansion of the sub-population, the distribution balance of the solution set is improved while ensuring the convergence, so that a higher quality global optimal solution set is obtained. The conflict degree between the global optimal solution set and the sub-population optimal solution is calculated, and the weight coefficients are dynamically updated in combination with the exponential decay function, realizing the balance and coordination between different target functions, and improving the balance and stability of the multi-objective optimization result.

[0104] Figure 2 The collaborative sub-population optimization flowchart of the aluminum-plastic composite packaging material separation control method of the embodiment of the application.

[0105] In an alternative embodiment,

[0106] The temperature distribution data of the initial peeling area is collected to extract a second feature map and a heat distribution rule, and the stress distribution trend of the material is calculated according to the heat distribution rule, including:

[0107] The initial peeling area is dynamically scanned by a multi-band infrared thermal imaging array to obtain temperature distribution data, the temperature distribution data is enhanced to obtain a second feature map, and the heat distribution rule is obtained by combining a multi-head self-attention mechanism to capture long-range heat conduction dependence.

[0108] Based on the heat distribution rule and the pre-acquired energy distribution characteristics, a capsule network is trained, a coupling coefficient is calculated through a dynamic routing mechanism, a multi-level stress field is constructed according to the coupling coefficient, and the multi-level stress field is input into a variational autoencoder to obtain the material stress distribution trend through feature dimension reduction.

[0109] The initial peeling area is dynamically scanned by a multi-band infrared thermal imaging array to obtain temperature distribution data, the infrared thermal imaging array includes five-band thermal imaging sensors with wavelength ranges of 3-5 μm, 8-12 μm, 1-3 μm, 5-8 μm and 12-14 μm, a resolution of 640×480 pixels, a temperature measurement accuracy of ±0.05℃, and a sampling frequency of 60Hz. The array arrangement adopts a 3×2 matrix layout, covering an area of 100mm×80mm, which meets the full coverage scanning of the initial peeling area of the aluminum-plastic composite material. During the scanning process, the thermal imaging array moves along the material surface at a speed of 10mm / s, and a frame of thermal image is collected every 0.5s, and 60 frames of thermal images are continuously collected to form a dynamic temperature field sequence. Exemplarily, in a typical aluminum-plastic composite material PE-aluminum foil interface peeling area, the center temperature is about 175℃, the edge temperature is about 120℃, and the background area temperature is about 30℃.

[0110] The acquired temperature distribution data is enhanced to obtain a second feature map, and the multi-scale thermal gradient analysis method is used for feature enhancement, including spatial gradient calculation and time gradient calculation. The spatial gradient adopts a Sobel operator with a kernel size of 5×5 to enhance the extraction ability of edge features; the time gradient is calculated by difference between adjacent frames, and a sliding window size of 5 frames is used. In addition, a thermal conductivity estimation module is introduced to estimate the local thermal conductivity according to the material characteristics, the thermal conductivity of the polyethylene layer is about 0.35W / (m·K), the thermal conductivity of the aluminum foil layer is about 237W / (m·K), and the thermal conductivity of the polyester layer is about 0.15W / (m·K). The size of the second feature map after feature enhancement is 640×480×12, and the 12 channels correspond to the original temperature data of the 5 bands, the spatial gradient data of the 5 bands, the time gradient data and the thermal conductivity estimation data, respectively.

[0111] The second feature map is input into a spatio-temporal graph neural network for processing. The spatio-temporal graph neural network uses a graph structure to represent the temperature field distribution, and divides the temperature field into 1024 nodes, each representing the temperature characteristics of a specific spatial location. The node feature dimension is 16, including temperature value, gradient value, thermal conductivity, etc. The edge connection in the network represents the heat conduction relationship between adjacent regions, and a total of 3072 edges are constructed. The node feature matrix is initialized using the feature vector at the corresponding position in the second feature map, and the edge feature is initialized based on the spatial distance and thermal conduction difference between nodes.

[0112] The node feature update equation of the spatio-temporal graph neural network is used to extract the heat conduction spatio-temporal characteristics. The node feature update uses graph convolution operation, including 3 layers of graph convolution layers with hidden layer dimensions of 64, 128 and 64 respectively. Each layer of graph convolution is followed by a ReLU activation function and a batch normalization layer. The edge feature update uses edge convolution operation, and the update process considers the feature difference between adjacent nodes and the initial feature of the edge. In actual calculation, for the nodes at the PE-aluminum foil interface, the feature value increases by about 25% after the first layer of graph convolution; for the nodes in the homogeneous region, the feature value changes by about 5%.

[0113] The multi-head self-attention mechanism is combined to capture long-range heat conduction dependencies. The self-attention mechanism uses 8 attention heads, each with an output dimension of 32, and a total output dimension of 256. The scaling factor in attention calculation is set to 8, and the dropout rate is 0.1. Through the self-attention mechanism, it can identify regions in the temperature field that are far apart in space but similar in heat conduction characteristics, and has high sensitivity to the heat conduction fracture phenomenon at the interlayer interface in the aluminum plastic composite material. The output of the self-attention layer is passed through two fully connected layers with hidden layer dimensions of 128 and output dimensions of 64 to obtain a compact representation of the heat distribution law.

[0114] The capsule network is trained based on the heat distribution law and the pre-acquired energy distribution characteristics. The capsule network includes a primary capsule layer and two advanced capsule layers. The primary capsule layer includes 32 capsules, each outputting an 8-dimensional vector; the first advanced capsule layer includes 16 capsules, each outputting a 16-dimensional vector; and the second advanced capsule layer includes 8 capsules, each outputting a 32-dimensional vector. The primary capsule layer adopts a convolution operation, and the convolution kernel size is 9x9 with a step of 2; the advanced capsule layers are connected through a dynamic routing mechanism, and the coupling coefficients are calculated through the dynamic routing mechanism. The number of iterations of the dynamic routing is set to 3, and the initial routing logic is set to a zero vector. In each iteration, the coupling coefficients are updated based on the similarity between the current capsule output and the target capsule, and the similarity is calculated through vector dot product. For the interfacial region of the aluminum plastic material, the average coupling coefficient from the primary capsule to the first advanced capsule is about 0.65, which is significantly higher than 0.35 of the non-interfacial region; the coupling coefficient difference from the first advanced capsule to the second advanced capsule is further expanded, reaching 0.78 for the interfacial region and about 0.22 for the non-interfacial region.

[0115] A multi-level stress field is constructed according to the coupling coefficients. The stress field construction is divided into three levels: micro stress field, meso stress field and macro stress field. The micro stress field is constructed based on the output of the primary capsule, with a size of 320x240x8; the meso stress field is constructed based on the output of the first advanced capsule, with a size of 160x120x16; and the macro stress field is constructed based on the output of the second advanced capsule, with a size of 80x60x32. Each level of stress field is represented by a unified multi-level stress field through upsampling and feature fusion, with a size of 320x240x32. At the PE-aluminum foil interface, the micro stress field shows high-frequency oscillation characteristics, with a stress value fluctuation range of ±15MPa; the meso stress field shows stress concentration zone, with a maximum stress value of 25MPa; and the macro stress field shows stress gradient distribution, with a gradient value of about 2MPa / mm.

[0116] The multi-level stress field is input into a variational autoencoder for feature dimension reduction. The variational autoencoder is composed of an encoder and a decoder, the encoder includes 4 convolutional layers with convolution kernel sizes of 5x5, 5x5, 3x3 and 3x3 respectively, channel numbers of 64, 128, 256 and 512 respectively, and a step of 2, and the decoder includes 4 transposed convolutional layers with structures symmetrical to the encoder. The hidden space dimension is set to 128, which is divided into 64 dimensions of mean vector and 64 dimensions of logarithmic variance vector. In the training process, the weighted sum of the reconstruction loss and the KL divergence loss is used as the optimization objective, the weight ratio is 10:1, the learning rate is 0.0001, and the training rounds are 200 rounds.

[0117] The material stress distribution trend is obtained by dimension reduction through the variational autoencoder. In the hidden space, the stress distribution trend shows a manifold structure along the principal component direction, and the first 10 principal components explain about 92% of the variance. According to the hidden space representation, the deformation and separation behavior of the aluminum-plastic composite material under external force can be predicted.

[0118] In this embodiment, the initial peeling area is dynamically scanned by a multi-band infrared thermal imaging array, and a feature enhancement process is combined to improve the clarity and detail integrity of the temperature distribution data, and to improve the accuracy of heat conduction feature extraction. The long-range heat conduction dependency is captured through the multi-head self-attention mechanism, which effectively improves the analysis accuracy of complex heat distribution rules. The capsule network is trained based on the heat distribution law and energy distribution characteristics, and the coupling coefficient calculated by the dynamic routing mechanism is used to construct a multi-level stress field, realizing the layering and correlation enhancement of stress distribution representation. The multi-level stress field is input into the variational autoencoder for feature dimension reduction, reducing redundant information and improving the stability and generalization ability of stress distribution modeling. The stress distribution trend of the material is accurately obtained.

[0119] In an alternative embodiment,

[0120] The optimal mechanical separation action sequence is solved by combining a deep reinforcement learning algorithm, and the action force direction and action force size of the mechanical separation device are determined based on the optimal mechanical separation action sequence and the separation action is executed, including:

[0121] Topological features are extracted from the material stress distribution trend using a dual-channel spatial transformation network and input into a fractional order neural network. A state transition equation is constructed by combining a non-integer order differential operator. Non-uniform sampling of continuous state space is performed by a fractional order dynamic programming method. The nonlinear state sampled is mapped to the Fourier domain for fast solving by combining the Mellin transform, and the separation action strategy is obtained.

[0122] The optimal mechanical separation action sequence is generated according to the separation action strategy, and the action force direction and action force size of the mechanical separation device are determined based on the pre-set action evaluation index and the separation action is executed.

[0123] The topological features are extracted from the material stress distribution trends using a dual-channel spatial transformation network. The dual-channel spatial transformation network includes two parallel channels, which process macro-structure features and micro-texture features, respectively. The macro-channel adopts a three-layer spatial transformation module. The positioning network consists of three convolution layers with kernel sizes of 7x7, 5x5, and 3x3, and channel numbers of 32, 64, and 128, respectively. Two fully connected layers are connected at the end to output six transformation parameters. The grid generator uses bilinear interpolation to generate the sampling grid. The sampler uses bicubic interpolation for feature sampling. The micro-channel adopts a similar structure, but with smaller convolution kernel sizes of 5x5, 3x3, and 3x3, and the same channel numbers to capture more subtle texture changes. The features of the two channels are combined through a feature fusion module. The feature fusion module uses a channel attention mechanism to assign weights to the features of different channels. The initial weight of the macro-channel is 0.65, and the initial weight of the micro-channel is 0.35. For the PE-aluminum foil interface region of the aluminum-plastic composite material, the topological features extracted by the spatial transformation network show obvious layered structure and interface discontinuity, with a feature dimension of 128x128x64.

[0124] The extracted topological features are input into a fractional-order neural network for processing. The fractional-order neural network uses fractional-order derivatives instead of integer-order derivatives in traditional neural networks to improve the modeling capability of long-range dependencies. The network contains five layers of fractional-order convolution layers with orders of 0.7, 0.8, 0.9, 0.8, and 0.7, respectively. The convolution kernel sizes are 5x5, 5x5, 3x3, 3x3, and 3x3, and the channel numbers are 64, 128, 256, 128, and 64, respectively. Each layer of fractional-order convolution is followed by a fractional-order ReLU activation function with the same order as the corresponding convolution layer. The implementation of fractional-order convolution uses the discrete form of the Grünwald-Letnikov definition with a memory length of 20 and a truncation error less than 10^-6. The network training uses the fractional-order Adam optimizer with a learning rate of 0.0005, 1000 training samples, and 150 training rounds.

[0125] The state transition equation is constructed using non-integer order differential operators. The state vector has a dimension of 128 and contains key parameters such as stress, strain, and temperature of the material. The order of the non-integer order differential operator is set to 0.85, which is determined through experiments and has the best effect in describing the viscoelastic behavior of aluminum-plastic composite materials. The discrete form of the state transition equation is implemented using the short memory principle with a memory length of 15 and a step size of 0.1s. For the state evolution at the PE-aluminum foil interface, when the initial stress is 5.5MPa, the change amplitude of the state vector within 0.5s is about 20%, showing obvious nonlinear characteristics. When the stress increases to 7.0MPa, the state change amplitude increases to 35%, and an inflection point appears after 1.2s, indicating that the interface is about to separate.

[0126] The continuous state space is non-uniformly sampled by the fractional dynamic programming method. The state space dimension is 128, and the action space dimension is 6, corresponding to the displacement and force of the mechanical separation device in three directions. The number of sampling points is set to 500, and the sampling density is proportional to the state gradient. In the area with large stress gradient (such as the PE-aluminum foil interface), the sampling interval is 0.05 MPa; in the area with small stress gradient, the sampling interval increases to 0.5 MPa. The value function is initialized with a zero function, and the discount factor is set to 0.95. The dynamic programming iteration is 200 times, and the convergence threshold is 10^-4.

[0127] The nonlinear state sampled is mapped to the Fourier domain for fast solving by combining the Mellin transform, and the integral interval of the Mellin transform is set to [0.1, 10], the discrete sampling point number is 128, and the sampling method is logarithmic uniform sampling. In the Fourier domain, the low-frequency component mainly represents the overall deformation trend of the material, the medium-frequency component corresponds to the interface separation process, and the high-frequency component reflects the microstructure change. Analysis of the PE-aluminum foil interface shows that when the frequency range is 0.5-2 Hz, the signal energy proportion is the highest, reaching 65%, indicating that this frequency range is the key frequency band of material separation. The Fourier domain solving adopts the fast Fourier transform algorithm, and the computational complexity is reduced from O(n^2) to O(nlogn). The solving result is converted back to the time domain by inverse Mellin transform to obtain the separation action strategy, which is represented as a series of state-action pairs, containing 30 key points.

[0128] The optimal mechanical separation action sequence is generated according to the separation action strategy. The action sequence includes multiple stages: preloading stage, initial peeling stage, stable separation stage, and complete separation stage. Each stage contains multiple discrete action points, and the action point interval is 0.2s. The force direction of the preloading stage is perpendicular to the interface, and the force size increases from 0 to 3.5N; in the initial peeling stage, the force direction is gradually adjusted to 30° with the interface, and the force size increases to 5.8N; in the stable separation stage, the force direction remains in the range of 35°±5°, and the force size maintains at 6.0-6.5N; in the complete separation stage, the force direction is adjusted to 45°, and the force size gradually decreases to 2.0N. For a 100mm×100mm aluminum-plastic composite sample, the complete separation process takes about 12 seconds, and the separation speed is controlled at 8mm / s.

[0129] The force direction and force size of the mechanical separation device are determined by combining the pre-set action evaluation index. The evaluation index includes three aspects: separation efficiency, interface integrity, and energy consumption. The separation efficiency index requires that the separation speed be not less than 5mm / s; the interface integrity index requires that the surface integrity of the PE layer and the aluminum foil layer after separation be not less than 95%; and the energy consumption index requires that the unit area separation power consumption be not more than 0.02J / cm 2Based on these indicators, the generated action sequence is evaluated and fine-tuned. For the PE-aluminum foil interface, the optimal force direction is 35°±3° deviated from the normal direction of the interface plane, and the optimal force size is 6.2N±0.3N. Experimental verification shows that, using this parameter combination, the separation speed can reach 8.5mm / s, the surface integrity of the PE layer and the aluminum foil layer is 98.2% and 97.5% respectively, and the unit area separation power consumption is 0.018J / cm 2 , which meets the preset evaluation index requirements.

[0130] In this embodiment, the double-channel spatial transformation network is used to extract topological features from the material stress distribution trend, improving the completeness and spatial correlation of the stress distribution feature representation. The fractional order dynamic programming method is used for non-uniform sampling of the continuous state space, effectively reducing redundant state points and improving computational efficiency. The nonlinear state is mapped to the Fourier domain for fast solving through the Mellin transform, accelerating the optimization process and improving the solving stability. The optimal mechanical separation action sequence generated based on the separation action strategy, combined with the action evaluation index, determines the direction and size of the mechanical separation device's force, enabling precise control of the separation process.

[0131] In an alternative embodiment,

[0132] The strain signal of the mechanical separation device during the execution of the separation action is collected, and the strain signal is analyzed in time and frequency through wavelet transform to obtain the separation characteristic frequency band and calculate the separation state evaluation value based on the support vector regression algorithm, including:

[0133] The strain signal of the mechanical separation device during the execution of the separation action is collected;

[0134] The strain signal is segmented according to a preset scale interval using the box dimension calculation formula, and the minimum number of boxes required to cover the strain signal is calculated within each scale segment. Based on the rate of change of the minimum number of boxes with respect to the scale, the self-similarity feature is obtained. The delay sequence is constructed according to a preset time delay interval to obtain the state trajectory, and the rate of change of the distance between adjacent sampling times of the state trajectory is calculated to obtain the maximum Lyapunov exponent.

[0135] The strain signal is transformed by wavelet transform to obtain a time-frequency distribution. According to the self-similarity feature, a first threshold value is set, and the region in the time-frequency distribution that exceeds the first threshold value is determined as a characteristic scale interval. According to the maximum Lyapunov exponent, a second threshold value is set, and the region in the time-frequency distribution that exceeds the second threshold value is determined as a dynamically unstable region. The overlapping part of the characteristic scale interval and the dynamically unstable region is extracted to obtain the separation characteristic frequency band.

[0136] The separation feature frequency band is input into a support vector regression algorithm to calculate a separation state evaluation value.

[0137] The strain signal during the separation action of the mechanical separation device is collected. A high-precision strain sensor is used, with a sensitivity of 2.1 and a measurement range of ±5000με, and a sampling frequency of 1000 Hz. The strain sensor is arranged in a four-arm Wheatstone bridge connection mode, and can simultaneously measure the strain values in the X, Y and Z directions, with a resolution of 0.5με and a signal-to-noise ratio greater than 60 dB. During the separation process of the aluminum-plastic composite material PE-aluminum foil interface, the strain signal usually shows a characteristic of slowly rising first and then rapidly falling after reaching a peak value, with a peak strain of about 3500με, an upward stage lasting about 4 seconds, and a downward stage lasting about 0.5 seconds. The strain signal is collected through a 16-bit analog-to-digital converter, and after filtering to remove high-frequency noise above 50 Hz, a continuous and smooth strain time series data is obtained.

[0138] The strain signal is segmented according to a preset scale interval using the box dimension calculation formula. The scale interval is set in a logarithmic uniform manner, with a minimum scale of 0.01 s and a maximum scale of 2 s, and a total of 20 scale points. The minimum number of boxes required to cover the strain signal is calculated in each scale segment. The box size decreases with the decrease of the scale, and the minimum box height is 1 / 500 of the strain value range. For a typical PE-aluminum foil interface separation process, the number of boxes required to cover is about 850 at a scale of 0.01 s, reduces to 105 at a scale of 0.1 s, and further reduces to 15 at a scale of 1 s. Based on the change rate of the minimum box number with the scale, the self-similarity feature is obtained, and the logarithmic values of the box number and the scale are linearly fitted. The negative value of the slope is the box dimension, and the typical value is about 1.65, indicating that the strain signal has obvious fractal characteristics.

[0139] The state trajectory is constructed by the delay sequence of the strain signal with preset time delay interval. The time delay interval is determined by the mutual information method, and the optimal delay value of the separation process of the aluminum plastic composite material is about 0.05 s. The embedding dimension is determined by the false nearest neighbor method, and the typical value is 4, indicating that the system can be fully described by a four-dimensional phase space. The distance between adjacent sampling time points of the state trajectory is calculated by the Euclidean distance. The distance change rate is small in the initial stage, about 0.02; with the separation process, the distance change rate increases rapidly when the interface starts to peel off, and the maximum can reach 0.35; after the separation is completed, the distance change rate decreases to about 0.05. The maximum Lyapunov exponent is calculated by calculating the distance change rate between adjacent sampling time points of the state trajectory. The maximum Lyapunov exponent represents the degree of chaos and sensitivity to initial conditions of the system. For the separation process of the aluminum plastic composite material, the maximum Lyapunov exponent is about 0.15 in the stable separation stage, indicating that the system has a certain predictability; when the interface suddenly breaks, the exponent value increases to 0.65 instantaneously, indicating that the system enters an unstable state; after the separation is completed, the exponent value decreases to about 0.08, and the system returns to stability.

[0140] The time-frequency distribution of the strain signal is obtained by wavelet transform. The wavelet transform uses the continuous wavelet transform method, the mother wavelet is selected as Morlet wavelet, the scale range is from 1 to 128, and 64 scale points are set, corresponding to the frequency range of 1 Hz to 500 Hz. The transformation result forms a time-frequency energy distribution diagram with a resolution of 1000x64, representing the energy distribution of 1000 time points and 64 frequency points. In the separation process of the PE-aluminum foil interface, the low frequency band (1-10 Hz) energy is mainly distributed in the whole separation process, the medium frequency band (10-100 Hz) energy is concentrated in the initial separation and the separation moment, and the high frequency band (100-500 Hz) energy mainly appears in the interface fracture moment.

[0141] According to the self-similarity feature, a first threshold value is set, and a characteristic scale interval is determined in the time-frequency distribution which exceeds the first threshold value. The first threshold value is calculated based on the statistical distribution of the self-similarity feature value and the time-frequency energy, which is the energy mean value plus the self-similarity feature value multiplied by the energy standard deviation, and the typical value is about 2.65 times the energy mean value. For the separation of the PE-aluminum foil interface, the characteristic scale interval is mainly distributed in the 15-85 Hz frequency band, and is concentrated in the 3.5-5.5 s interval of the separation process. According to the maximum Lyapunov exponent, a second threshold value is set, and a dynamic unstable region is determined in the time-frequency distribution which exceeds the second threshold value. The second threshold value is calculated based on the statistical distribution of the maximum Lyapunov exponent and the time-frequency energy, which is the energy mean value plus the maximum Lyapunov exponent multiplied by the energy standard deviation, and the typical value is about 1.85 times the energy mean value. The dynamic unstable region is mainly distributed in the 25-120 Hz frequency band, and is concentrated in the 4.0-5.0 s interval corresponding to the period when the interface starts to separate rapidly.

[0142] The separated characteristic frequency band is obtained by extracting the overlapping part of the characteristic scale interval and the dynamic instability region. For PE-aluminum foil interface separation, the separated characteristic frequency band is 25-85 Hz, and the time interval is 4.0-5.0 seconds. In this frequency band, the energy distribution presents a clear peak, and the peak frequency is about 55 Hz, which corresponds to the moment of complete fracture of the interface. The energy integral value of the separated characteristic frequency band is significantly positively correlated with the separation quality, and the correlation coefficient reaches 0.92, which can be used as an important indicator to evaluate the separation state.

[0143] The separated characteristic frequency band is input into the support vector regression algorithm to calculate the separation state evaluation value. The support vector regression algorithm uses a radial basis function kernel, the kernel parameter γ is 0.05, the penalty factor C is 10, and the ε value is 0.01. The characteristic input includes the energy integral value, the energy center frequency, the frequency bandwidth, the time duration length and the energy maximum value of the separated characteristic frequency band, a total of 5 characteristic quantities. The training sample is 200 groups of pre-labeled separation process data, each group of data contains characteristic input and artificial evaluation of separation quality score (0-100 points). Training uses 5-fold cross-validation, the mean square error is 3.25, and the correlation coefficient is 0.95.

[0144] In this embodiment, the state trajectory is constructed based on the delay sequence and the maximum Lyapunov exponent is calculated, which improves the accuracy of the characterization of the nonlinear dynamics of the strain signal. The strain signal is subjected to wavelet transform to obtain time-frequency distribution, and threshold values are set in combination with self-similarity features and maximum Lyapunov exponent, so that the characteristic scale interval and the dynamic instability region can be effectively distinguished, and the accuracy and robustness of the separation feature extraction are significantly improved. The separated characteristic frequency band is input into the support vector regression algorithm to obtain a quantitative separation state evaluation value, and the precise evaluation of the mechanical separation process state is realized.

[0145] In an alternative embodiment,

[0146] Based on the separation state evaluation value, the separation trend data is predicted, and the microwave energy compensation value and the mechanical force compensation value are calculated in real time to obtain an optimal control strategy and execute, which includes:

[0147] The separation state evaluation value sequence is input into the long short-term memory network, and the separation trend data is predicted by the long short-term memory network;

[0148] The deviation value between the separation trend data and the target separation state is calculated, and the microwave energy compensation value and the mechanical force compensation value are calculated according to the deviation value and the change rate of the deviation value;

[0149] The microwave energy compensation value is added to a preset microwave energy reference value to obtain a microwave energy control instruction, and the mechanical action force compensation value is added to a preset mechanical action force reference value to obtain a mechanical action force control instruction.

[0150] Separation control is performed according to the microwave energy control instruction and the mechanical action force control instruction.

[0151] The separation state evaluation value sequence is input into a long short-term memory network to obtain separation trend data through prediction. The separation state evaluation value sequence is organized in a time sequence form, with a sampling interval of 0.1 seconds and a sequence length of 50, covering a separation process of 5 seconds. For the PE-aluminum foil interface separation process, the evaluation value sequence has a value of 40-55 in the early separation stage, a value of 60-75 in the middle separation stage, and a value of 80-90 in the late separation stage. The long short-term memory network structure includes an input layer, two LSTM hidden layers, and an output layer. The input layer receives an evaluation value sequence with a window size of 10; the first LSTM hidden layer includes 64 neurons, and the second LSTM hidden layer includes 32 neurons, both of which use a tanh activation function, and the forgetting gate bias is set to 1.0 to enhance long-term memory capability; and the output layer is a fully connected layer including 10 neurons corresponding to the predicted values of the next 10 time steps.

[0152] The network training uses an Adam optimizer with a learning rate of 0.001, a batch size of 32, and 200 training rounds. The training data set includes 500 separation process records, of which 400 are used for training and 100 are used for verification. The training loss function uses mean square error, and the early stopping mechanism is introduced during the training process. The training is stopped when the validation set loss does not improve for 10 consecutive rounds. The average prediction error on the validation set is 4.5, and the prediction accuracy reaches 92%. For the PE-aluminum foil interface separation, when the current evaluation value sequence is [65, 68, 70, 73, 75, 78, 80, 82, 84, 85], the network predicts the evaluation values of the next 10 time steps as [86, 87, 88, 89, 90, 91, 91, 92, 92, 93], indicating that the separation process will continue to develop towards a good state.

[0153] The deviation value between the separation trend data and the target separation state is calculated. The target separation state is set to a separation quality score of 90 points or more, corresponding to a high-quality separation result. Taking the above prediction result as an example, the deviation values of the first 5 time steps are [4, 3, 2, 1, 0], and the last 5 time steps have reached the target state with a deviation value of 0. The change rate of the deviation value is calculated by the difference between adjacent time step deviation values. In this example, the change rate is [-1, -1, -1, -1, 0, 0, 0, 0, 0], indicating that the deviation is decreasing at a rate of 1 point per time step, and the separation process is converging steadily towards the target state.

[0154] The microwave energy compensation value and the mechanical force compensation value are calculated according to the deviation value and the change rate of the deviation value, the compensation value calculation adopts an adaptive fuzzy control strategy, the input is the deviation value and the change rate of the deviation value, and the output is the two compensation values. The fuzzy rule base contains 25 rules, covering various combinations of the deviation value and the change rate. The deviation value is divided into five fuzzy sets: large negative (DN), small negative (SN), zero (Z), small positive (SP), and large positive (DP); the change rate is also divided into five fuzzy sets. The fuzzy reasoning adopts the Mamdani method, the membership function adopts a triangular function, and the defuzzification adopts the center average method. For the case that the deviation value is 4 and the change rate is -1 in the separation process of the PE-aluminum foil interface, the calculated microwave energy compensation value is +50 W, and the mechanical force compensation value is +0.5 N; when the deviation value is -3 and the change rate is +1, the microwave energy compensation value is -70 W, and the mechanical force compensation value is -0.8 N.

[0155] The microwave energy compensation value and the mechanical force compensation value are adjusted through gain coefficients, and the gain coefficients are pre-set according to the material characteristics. For the PE-aluminum foil-PET three-layer structure aluminum-plastic composite material with a thickness of 0.2 mm, the microwave energy gain coefficient is 1.2, and the mechanical force gain coefficient is 0.9. After adjustment, the aforementioned microwave energy compensation value is +60 W, and the mechanical force compensation value is +0.45 N.

[0156] The microwave energy compensation value is added to the pre-set microwave energy reference value to obtain a microwave energy control instruction, and the microwave energy reference value is pre-set according to the material type and the interface characteristics. For the PE-aluminum foil interface, the reference value is set to 800 W. The microwave energy control instruction after adding the compensation value is 860 W. The microwave energy control instruction adopts a 16-bit binary format, with a resolution of 1 W and a control range of 0-1000 W. The control instruction is transmitted to the microwave generator through an industrial control bus, and the update frequency is 10 Hz.

[0157] The mechanical force compensation value is added to the pre-set mechanical force reference value to obtain a mechanical force control instruction, and the mechanical force reference value is pre-set according to the material bonding strength. For the PE-aluminum foil interface, the reference value is 5.5 N in the X direction, 0 N in the Y direction, and 3.2 N in the Z direction. The mechanical force control instruction after adding the compensation value is 5.95 N in the X direction, 0 N in the Y direction, and 3.2 N in the Z direction. The mechanical force control instruction also adopts a 16-bit binary format, with a resolution of 0.01 N and a control range of 0-10 N. The control instruction is transmitted to the servo controller through an industrial control bus, and the update frequency is 20 Hz.

[0158] According to the microwave energy control instruction and the mechanical action force control instruction, separation control is performed, the microwave energy control is realized by adjusting the output power of the microwave generator, the power adjustment accuracy is ±5W, and the response time is less than 0.1 seconds. The microwave energy is transmitted to the target interface area of the aluminum-plastic composite material through the directional antenna array, the energy distribution is in a Gaussian distribution, the central energy density is 950J / cm 3 , and the energy attenuation in the circular area with a radius of 10mm is not more than 15%. The mechanical action force control is realized by the mechanical arm driven by the precision servo motor, the force control accuracy is ±0.05N, the position control accuracy is ±0.01mm, and the response time is less than 0.05 seconds. The mechanical arm is provided with a special separation tool at the end, the contact area is 50mm², and the surface hardness is HRC55, so that the material surface is not damaged in the separation process.

[0159] In the embodiment, the separation state evaluation value sequence is input into the long short-term memory network, the accuracy and stability of the separation trend prediction are improved, the microwave energy compensation value and the mechanical action force compensation value are calculated based on the deviation value and the change rate between the separation trend data and the target separation state, the deviation in the separation process is dynamically corrected, the compensation value is combined with the preset reference value to generate corresponding control instructions, the regulation and control of the microwave energy and the mechanical action force in the separation process are more fine, the accumulation of the deviation in the separation process is effectively reduced by executing the separation control based on the compensation, the real-time performance and the stability of the separation control are improved, and the accuracy and the reliability of the separation operation are improved.

[0160] In a second aspect of the embodiment of the present application, an aluminum-plastic composite packaging material separation control system is provided, comprising:

[0161] A first unit is configured to collect image information of the aluminum-plastic composite packaging material, extract features and perform hierarchical recognition on the image information through a convolutional neural network to obtain a first feature map and extract interlayer bonding parameters;

[0162] A second unit is configured to construct a microwave energy transmission equation based on the interlayer bonding parameters, iteratively solve the microwave energy transmission equation to obtain optimal distribution data of the microwave energy through a particle swarm optimization algorithm, control the microwave radiation antenna array to form directional energy focusing at the interlayer interface according to the optimal distribution data, and obtain an initial peeling area;

[0163] A third unit is configured to collect temperature distribution data of the initial peeling area, extract a second feature map and heat distribution law, calculate a material stress distribution trend according to the heat distribution law, solve an optimal mechanical separation action sequence through a deep reinforcement learning algorithm, determine an action force direction and an action force size of a mechanical separation device based on the optimal mechanical separation action sequence, and perform a separation action;

[0164] The fourth unit is used for collecting strain signals of the mechanical separation device during the execution of the separation action, performing time-frequency analysis on the strain signals through wavelet transform, obtaining a separation characteristic frequency band, calculating a separation state evaluation value in combination with a support vector regression algorithm, predicting separation trend data based on the separation state evaluation value, calculating a microwave energy compensation value and a mechanical action force compensation value in real time in combination with the separation state evaluation value, obtaining an optimal control strategy, and executing the optimal control strategy.

[0165] In a third aspect, the present application provides an electronic device, comprising:

[0166] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0167] The present application can be a method, an apparatus, a system, and / or a computer program product. The computer program product can include a computer-readable storage medium having computer-readable program instructions loaded thereon, which instructions, when executed by a computer, cause the computer to perform various aspects of the present application.

[0168] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for controlling the separation of aluminum-plastic composite packaging materials, characterized in that, include: Image information of aluminum-plastic composite packaging materials is collected. Feature extraction and hierarchical recognition are performed on the image information using a convolutional neural network to obtain the first feature map and extract inter-layer coupling parameters, including: Collect image information of aluminum-plastic composite packaging materials, including the surface morphology features and internal structural features of the materials; The image information is input into a pre-trained convolutional neural network. The image information is then subjected to feature extraction and hierarchical recognition through multi-layer convolution and pooling operations of the convolutional neural network to obtain a first feature map containing the interlayer structural features of the material. The first feature map is analyzed to extract the interlayer bonding parameters of the aluminum-plastic composite packaging material. The interlayer bonding parameters include interlayer bonding strength, interlayer bonding area and interlayer bonding defect distribution. The microwave energy transfer equation is constructed based on the interlayer interface parameters. The optimal distribution data of microwave energy is obtained by iteratively solving the microwave energy transfer equation through the particle swarm optimization algorithm. Based on the optimal distribution data, the microwave radiating antenna array is controlled to form directional energy focusing at the interlayer interface to obtain the initial stripping region. Temperature distribution data of the initial stripping area is collected to extract the second feature map and heat distribution pattern. The stress distribution trend of the material is calculated based on the heat distribution pattern. The optimal mechanical separation action sequence is solved by combining deep reinforcement learning algorithm. Based on the optimal mechanical separation action sequence, the force direction and magnitude of the mechanical separation device are determined and the separation action is executed. The strain signal of the mechanical separation device during the separation process is collected. The time-frequency analysis of the strain signal is performed by wavelet transform to obtain the separation characteristic frequency band. The separation state evaluation value is calculated by combining the support vector regression algorithm. Based on the separation state evaluation value, the separation trend data is predicted. The microwave energy compensation value and mechanical force compensation value are calculated in real time by combining the separation state evaluation value to obtain the optimal control strategy and execute it. Strain signals from the mechanical separation device during the separation process are collected. Wavelet transform is used to perform time-frequency analysis on the strain signals to obtain the separation characteristic frequency bands. The separation state evaluation value is then calculated using a support vector regression algorithm, including: Collect strain signals during the separation process performed by the mechanical separation device; The strain signal is segmented according to a preset scale interval using the box dimension calculation formula. The minimum number of boxes required to cover the strain signal is calculated in each scale segment. Self-similarity features are obtained based on the rate of change of the minimum number of boxes with scale. The strain signal is constructed into a delay sequence according to a preset time delay interval to obtain a state trajectory. The distance change rate of the state trajectory between adjacent sampling times is calculated to obtain the maximum Lyapunov exponent. Wavelet transform is performed on the strain signal to obtain the time-frequency distribution. A first threshold is set according to the self-similarity feature. The region in the time-frequency distribution that exceeds the first threshold is determined as the feature scale interval. A second threshold is set according to the maximum Lyapunov exponent. The region in the time-frequency distribution that exceeds the second threshold is determined as the dynamic unstable region. The overlapping part of the feature scale interval and the dynamic unstable region is extracted to obtain the separated feature frequency band. The separated characteristic frequency bands are input into the support vector regression algorithm to calculate the separation state evaluation value.

2. The method according to claim 1, characterized in that, A microwave energy transfer equation is constructed based on interlayer interface parameters. The optimal distribution data of microwave energy is obtained by iteratively solving the microwave energy transfer equation using a particle swarm optimization algorithm. Based on the optimal distribution data, a microwave radiating antenna array is controlled to form directional energy focusing at the interlayer interface, resulting in the initial stripping region including: Microwave energy transfer equations are constructed based on interlayer parameters. Initial microwave energy distribution data are obtained by combining energy diffusion coefficient, conductivity and electric field strength to establish a transfer relationship. The particle swarm in the particle swarm optimization algorithm is initialized based on the initial energy distribution data. Based on the initial microwave energy distribution data, the particle swarm is divided into competing subgroups and cooperative subgroups, and an optimization objective function is set for each subgroup. The optimal solution for each subgroup is determined, and the optimal solutions of each subgroup are constructed into an objective vector. Cooperative subgroup optimization is performed to obtain the global optimal solution set. The degree of conflict between objectives is calculated based on the optimal solutions of each subgroup and the global optimal solution set. The comprehensive optimization result is obtained by fusion through weight coefficients. Based on the comprehensive optimization results, the antenna element spacing, main lobe azimuth angle, and phase compensation value are calculated. Phase control parameters and power allocation parameters are obtained. The phase control parameters and power allocation parameters are used to control the microwave radiating antenna array to form a directional energy focusing region at the interlayer interface, thereby obtaining the energy distribution characteristics and the corresponding initial stripping region.

3. The method according to claim 2, characterized in that, The optimal solutions of each subgroup are constructed into an objective vector, and collaborative subgroup optimization is performed to obtain a global optimal solution set. The degree of conflict between objectives is calculated based on the optimal solutions of each subgroup and the global optimal solution set. The comprehensive optimization result is obtained by fusing the solutions with weight coefficients, including: The component values ​​of the target vector are calculated based on the optimal solutions of each subgroup and the initial weight coefficients corresponding to each objective function. Differential evolution optimization is performed based on the target vector. An exponential decay function is constructed based on the ratio of the current iteration number to the maximum iteration number. An adaptive scaling factor is constructed based on the exponential decay function and the pre-set initial scaling factor, and the optimal search step size is calculated. The correlation coefficient between objective functions in the target vector is calculated based on the optimal search step size. An adaptive crossover operator is constructed based on the correlation coefficient and a difference vector is generated by a random walk strategy. The mutation vector is calculated by combining the adaptive crossover operator and the optimal solution of each subgroup. Based on the mutation vector, each subgroup is expanded to generate extended subgroups and then sorted non-dominated to obtain individual levels. For individuals with the same individual level, the crowding value is calculated and the high-quality individuals are determined. The high-quality individuals in each extended subgroup are obtained to obtain the global optimal solution set. By calculating the function values ​​of the global optimal solution set on each objective function, combining the optimal solutions of each subgroup to calculate the degree of conflict between objectives, updating the weight coefficients of each objective function with the exponential decay function, and calculating the comprehensive optimization result based on the updated weight coefficients and the degree of conflict between objectives.

4. The method according to claim 1, characterized in that, Temperature distribution data from the initial peeling region was collected to extract the second feature map and heat distribution pattern. Based on the heat distribution pattern, the material stress distribution trend was calculated, including: A multi-band infrared thermal imaging array is used to dynamically scan the initial stripped area to obtain temperature distribution data. The temperature distribution data is then enhanced to obtain a second feature map, which is input into a spatiotemporal graph neural network. The spatiotemporal characteristics of heat conduction are extracted using the node feature matrix and edge feature update equation of the spatiotemporal graph neural network. The long-range heat conduction dependence is captured by combining a multi-head self-attention mechanism to obtain the heat distribution law. Based on the heat distribution pattern and the pre-acquired energy distribution features, a capsule network is trained. The coupling coefficient is calculated through a dynamic routing mechanism. A multi-level stress field is constructed based on the coupling coefficient. The multi-level stress field is then input into a variational autoencoder for feature dimensionality reduction to obtain the material stress distribution trend.

5. The method according to claim 1, characterized in that, The optimal mechanical separation action sequence is obtained by combining deep reinforcement learning algorithms. Based on the optimal mechanical separation action sequence, the direction and magnitude of the force applied by the mechanical separation device are determined, and the separation action is executed, including: A dual-channel spatial transformation network is used to extract topological features from the stress distribution trend of the material and input them into a fractional neural network. A state transition equation is constructed by combining a non-integer differential operator. The continuous state space is non-uniformly sampled by a fractional dynamic programming method. The nonlinear state obtained by sampling is mapped to the Fourier domain by Merlin transform for fast solution, thus obtaining the separation action strategy. The optimal mechanical separation action sequence is generated based on the separation action strategy. The direction and magnitude of the force of the mechanical separation device are determined by combining the pre-set action evaluation index, and the separation action is executed.

6. The method according to claim 1, characterized in that, Based on the separation state assessment value, the separation trend data is predicted. The optimal control strategy is then obtained and executed by combining the separation state assessment value with real-time calculations of microwave energy compensation and mechanical force compensation values, including: The sequence of separation state evaluation values ​​is input into a long short-term memory network, and separation trend data is predicted by the long short-term memory network. Calculate the deviation between the separation trend data and the target separation state, and calculate the microwave energy compensation value and the mechanical force compensation value based on the deviation value and the rate of change of the deviation value; The microwave energy compensation value is added to the preset microwave energy reference value to obtain the microwave energy control command, and the mechanical force compensation value is added to the preset mechanical force reference value to obtain the mechanical force control command; Separation control is executed according to the microwave energy control command and the mechanical force control command.

7. An aluminum-plastic composite packaging material separation control system, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to collect image information of aluminum-plastic composite packaging materials, and to extract features and perform hierarchical recognition of the image information through a convolutional neural network to obtain the first feature map and extract interlayer combination parameters. The second unit is used to construct the microwave energy transfer equation based on the interlayer combination parameters. The optimal distribution data of microwave energy is obtained by iteratively solving the microwave energy transfer equation through the particle swarm optimization algorithm. Based on the optimal distribution data, the microwave radiating antenna array is controlled to form directional energy focusing at the interlayer interface to obtain the initial stripping region. The third unit is used to collect temperature distribution data of the initial stripping area, extract the second feature map and heat distribution law, calculate the material stress distribution trend based on the heat distribution law, solve the optimal mechanical separation action sequence by combining deep reinforcement learning algorithm, determine the force direction and magnitude of the mechanical separation device based on the optimal mechanical separation action sequence, and execute the separation action. The fourth unit is used to collect strain signals from the mechanical separation device during the separation process. It performs time-frequency analysis on the strain signals through wavelet transform to obtain the separation characteristic frequency band and calculates the separation state evaluation value using the support vector regression algorithm. Based on the separation state evaluation value, it predicts the separation trend data and calculates the microwave energy compensation value and mechanical force compensation value in real time using the separation state evaluation value to obtain and execute the optimal control strategy.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

Images