Multi-system cooperation-based redundant material self-adaptive removal device and multi-system cooperation-based redundant material self-adaptive removal method

Abstract

The invention relates to a multi-system cooperation-based redundancy self-adaptive removal device and method.The self-adaptive removal device comprises a visual recognition module, a machine learning module, a central control module and a redundancy removal module, the visual recognition module collects an image of a to-be-welded area, obtains information such as features, areas and positions of redundancy and sends the information to the central control module; the machine learning module adopts an A3C machine learning algorithm model to establish an optimal strategy that multiple parallel threads carry out redundancy removal at the same time, the central control module cooperates with the redundancy removal module to adjust the redundancy removal strategy in a self-adaptive mode based on the redundancy removal strategy, and on the premise that a metal layer on the surface of a battery is not damaged, the redundancy removal module carries out redundancy removal on the surface of the battery. And various redundancies in the to-be-welded area are efficiently and accurately removed. According to the method, the accuracy of removing the redundancy can be remarkably improved, meanwhile, the labor cost is reduced, the reworking number of welding of the solar cell and the interconnection piece is greatly reduced, and the bonding strength of a welding interface is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of pre-welding treatment, and particularly relates to a redundant material adaptive removal device and method under multi-system cooperation. Background Art

[0002] During the series resistance welding process of triple-junction gallium arsenide solar cells and interconnection strips, redundant materials (such as overflow glue, metal debris, dust, fingerprints, etc. left over from the previous process) existing in the area to be welded will form a physical barrier during the welding process, absorb the heat between interfaces, making it difficult for the heat transmitted by the electrode head to reach the surface of the object to be welded, thus greatly affecting the interface connection strength and causing the phenomenon of solar cell string detachment during the in-orbit service of satellites. Traditional methods for removing redundant materials have many limitations. For example, the efficiency of manual inspection is low, and it is difficult to ensure accuracy and batch consistency. General cleaning means cannot effectively handle different types of redundant materials. With the rapid development of automation and intelligence, there is an urgent need for a device and method for adaptive removal of redundant materials. Summary of the Invention

[0003] In order to overcome the deficiencies in the prior art, the present invention provides a redundant material adaptive removal device and method under multi-system cooperation to solve problems such as low automation degree of redundant material removal and poor bonding of the welding interface. The present invention can achieve high automation and precision in redundant material removal, significantly increase the strength of the welding interface, and is simple to operate and convenient to implement.

[0004] The above object of the present invention is mainly achieved through the following technical solutions:

[0005] A redundant material adaptive removal device under multi-system cooperation includes a visual recognition module, a machine learning module, a central control module, and a redundant material removal module; wherein:

[0006] The visual recognition module is used to locate and identify redundant materials in the area to be welded, and send the recognition information to the machine learning module; and monitor the remaining redundant materials during the redundant material removal process, and send the monitoring information to the central control module;

[0007] The machine learning module is used to receive the redundant material recognition information from the visual recognition module, analyze features through a deep neural network, generate a strategy instruction and send it to the central control module; receive the monitoring information sent by the central control module during the redundant material removal process, adjust the strategy instruction according to the monitoring information, and send it to the central control module; and is also used for offline training to improve the model convergence efficiency;

[0008] The central control module is used to receive the strategy instruction sent by the machine learning module and send it to the redundant material removal module, and receive the monitoring information sent by the visual recognition module and send it to the machine learning module;

[0009] The foreign matter removal module is used to receive the policy instructions from the central control module and perform foreign matter removal operations.

[0010] The method for the visual recognition module to locate and identify foreign matters in the area to be welded includes: collecting the surface image of the solar cell to be welded through a camera, extracting the feature information of the edges of foreign matters in the image, and obtaining the features of foreign matters, the geometric dimensions of the covered area, and the position information.

[0011] The machine learning module includes multiple policy networks, value networks, experience replay buffers, and a global parameter server; among them:

[0012] The policy network receives the recognition information and monitoring information sent from the visual recognition module and the central control module, analyzes them through a deep neural network, combines the historical removal effect data, calculates the action probability distribution, generates policy instructions, sends the action probability distribution and policy instructions to the value network, and sends the action selection result, environmental state, and next state information to the experience replay buffer; receives the data sent from the experience replay buffer for offline training;

[0013] The value network receives the recognition information and monitoring information sent from the visual recognition module and the central control module, calculates the state value function through a neural network, evaluates the long-term expected return of the current state, calculates the advantage function according to the output value of the state value function and the action probability distribution and policy instructions sent by multiple policy networks, selects the optimal policy instruction, and sends the optimal policy instruction to the global parameter server; receives the data sent from the experience replay buffer for offline training;

[0014] The experience replay buffer receives the action selection result, environmental state, and next state information sent from the policy network, forms a state-action-reward-next state tuple, eliminates the temporal correlation by randomly sampling historical data, and distributes the batch data to the policy network and the value network for offline training to improve the model convergence efficiency;

[0015] The global parameter server receives the optimal policy instruction sent from the value network, sends the optimal policy instruction to the central control module, adjusts the global network parameters using an asynchronous update algorithm, sends the updated parameters to all policy networks, and outputs the uniformly optimized model weights and distributes them to all policy networks to maintain the collaborative learning process.

[0016] The method for generating and adjusting the policy instructions of the machine learning module is as follows:

[0017] (1). Define the state space, action space, and incentive function; the state space is used to describe the categories of foreign objects, the action space is used to store all foreign object removal actions, and the incentive function uses a discounted reward function, with the specific formula as follows:

[0018]

[0019] Among them, G t represents the reward starting from time step t; r t+k represents the immediate reward at time step t + k; γ represents the discount factor (0 ≤ γ ≤ 1), which is used to control the importance of future rewards;

[0020] (2). Establish an A3C model, which is used to represent the policy instructions and value function of each agent, execute the policy instructions, adjust the neuron connection weights and incentive function according to the feedback of the foreign object removal effect, optimize the policy instructions, and form the optimal policy instructions, with the specific formula as follows:

[0021]

[0022] Among them, y represents the output removal strategy; x i represents the output foreign object eigenvalue; w i represents the weight corresponding to the eigenvalue; b i represents the bias term; f represents the incentive function; n represents the number of input eigenvalues.

[0023] The foreign object removal module includes a gas blowing device and a liquid spraying device; the gas blowing device includes a gas storage tank, a pressure regulating valve, a gas delivery pipeline, and a jet nozzle; the gas storage tank is used to store gas, and the gas is ejected from the jet nozzle through the gas delivery pipeline to blow the foreign objects outside the area to be welded, while preventing the oxidation of the metal layer in the area to be welded; the pressure regulating valve is installed between the gas storage tank and the gas delivery pipeline and is connected to the central control module through a control motor to adjust the output pressure of the gas in real time according to the policy instructions; the jet nozzle is located on one side of the electrode head and moves through a circular track to adjust the jet angle in real time according to the policy instructions; the liquid spraying device includes a liquid storage tank, a liquid delivery pump, a flow regulating valve, a spraying pipeline, and a spray head; the liquid storage tank stores the liquid used to remove foreign objects; the liquid delivery pump is connected to the liquid storage tank to pump out and deliver the liquid to the spraying pipeline; the flow regulating valve is located between the gas delivery pipeline and the spraying pipeline to adjust the liquid flow rate and speed according to the policy instructions; the spray head is located on the other side of the electrode head and moves through a circular track to adjust the spraying angle in real time according to the policy instructions.

[0024] The gas pressure of the gas blowing device is 0.1 to 1.5 MPa, the flow rate is 0 to 120 L / min, and the air flow direction is 0 to 360°; the liquid injection volume of the liquid spraying device is 0 to 60 ml / s, the spraying angle is 0 to 360°, and the spraying speed is 0 to 15 m / s.

[0025] The cooperation logic between the gas blowing device and the liquid spraying device includes:

[0026] (1). Priority determination: Different removal strategies are adopted for different foreign matters. For loose impurities, the gas blowing device is started to blow them with gas; for adhesive impurities, the liquid spraying device is started to dissolve them with chemical solvents and then assisted by gas blowing for cleaning; for mixed impurities, a gas-liquid alternating spraying strategy is adopted to avoid the influence of liquid residue on the welding quality.

[0027] (2). Parameter linkage control: When the gas blowing device is started, the nozzle angle of the liquid spraying device is automatically adjusted to the avoidance mode, and the avoidance mode is that the nozzle angle is 90° to the air flow direction to prevent liquid splashing from interfering with the gas blowing effect; when the liquid spraying device is started, the pressure regulating valve of the gas blowing device synchronously reduces the output pressure to avoid air flow disturbing the liquid spraying trajectory; for foreign matters distributed over a large area, the gas blowing device and the liquid spraying device operate in a coordinated manner in different regions. The gas blowing device removes loose impurities at the edge, and the liquid spraying device removes the adhered area.

[0028] (3). Safety protection mechanism: The spraying volume of the liquid spraying device is dynamically adjusted according to the removal situation of foreign matters, and the maximum flow rate is limited by the flow regulating valve to avoid liquid penetration from damaging the metal layer on the surface of the solar cell; the jet angle of the gas blowing device is synchronously calibrated with the movement trajectory of the electrode head to prevent the air flow impact force of the gas blowing device from causing the battery to break.

[0029] The gas stored in the gas blowing device is an inert gas.

[0030] A method for adaptive removal of foreign matters under the cooperation of multiple systems includes the following steps:

[0031] (1). Before welding, the visual recognition module collects image data of the area to be welded, extracts the characteristic information of foreign matters, and transmits it to the machine learning module;

[0032] (2). The machine learning module further analyzes the foreign matter characteristic data provided by the visual recognition module, judges the type of foreign matters, generates removal strategies, calculates different removal strategies through multiple parallel threads, and generates a strategy instruction after the analysis is completed and transmits it to the central control module;

[0033] (3). The central control module transmits the strategy instruction to the foreign matter removal module to remove foreign matters.

[0034] (4) Monitor the removal progress of foreign matters through the visual recognition module and send it to the central control module. The central control module sends it to the machine learning module, and the machine learning module adjusts the policy instruction.

[0035] (5) Execute the policy instruction in a loop until the removal of foreign matters is completed.

[0036] The calculation method of the policy instruction of the machine learning module is as follows:

[0037] (1) Define the state space, action space, and incentive function. The state space is used to describe the categories of foreign matters. The action space is used to store all foreign matter removal actions. The incentive function adopts the discounted return function, and the specific formula is as follows:

[0038]

[0039] Among them, G t represents the return starting from time step t; r t+k represents the immediate reward at time step t + k; γ represents the discount factor (0 ≤ γ ≤ 1), which is used to control the importance of future rewards.

[0040] (2) Establish an A3C model, which is used to represent the policy instruction and value function of each agent, execute the policy instruction, adjust the neuron connection weights and incentive function according to the feedback of the foreign matter removal effect, optimize the policy instruction, and form an optimal policy instruction. The specific formula is as follows:

[0041]

[0042] Among them, y represents the output removal policy; x i represents the output foreign matter eigenvalue; w i represents the weight corresponding to the eigenvalue; b i represents the bias term; f represents the incentive function; n represents the number of input eigenvalues.

[0043] The present invention has at least the following beneficial effects compared with the prior art:

[0044] (1) By precisely removing foreign matters in the area to be welded, the present invention avoids foreign matters from absorbing heat during the welding process, ensures that the interface temperature can reach the melting point of the metal, so that the diffusion interface is combined more firmly, significantly improves the welding quality, eliminates the physical barrier between the solar cell and the interconnection strip, effectively enhances the connection effect between the two, and thus greatly improves the stability of the photoelectric conversion efficiency of the series-connected solar cell panel.

[0045] (2) The present invention extracts the features, area, and position information of the foreign objects through the visual recognition module, the machine learning module determines the types and sizes of the foreign objects, and based on the environmental feedback, explores multiple foreign object removal strategies simultaneously through multiple parallel threads, adjusts and optimizes in real time, and forms an optimal execution strategy for different foreign object features, improving the accuracy and automation of foreign object removal and realizing an efficient and targeted foreign object removal method. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is the overall schematic diagram of a foreign object adaptive removal device under the coordination of multiple systems according to the present invention;

[0047] Figure 2 It is the partial enlarged schematic diagram of the position of the electrode head in a foreign object adaptive removal device under the coordination of multiple systems according to the present invention;

[0048] Figure 3 It is the spatial position schematic diagram of the electrode head, the air jet nozzle, and the spray head in a foreign object adaptive removal device under the coordination of multiple systems according to the present invention;

[0049] Figure 4 It is the algorithm model training flow chart of a foreign object adaptive removal device under the coordination of multiple systems according to the present invention;

[0050] Figure 5 It is the operation flow chart of a foreign object adaptive removal device and method under the coordination of multiple systems according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The present invention will be further described in detail below with reference to the drawings and specific embodiments:

[0052] As Figure 1 shown, a foreign object adaptive removal device under the coordination of multiple systems includes a visual recognition module 10, a machine learning module 20, a central control module 30, and a foreign object removal module 40; wherein:

[0053] The visual recognition module 10 is used for positioning and identifying foreign objects in the area to be welded, and sending the recognition information to the machine learning module 20; and monitoring the remaining foreign objects during the foreign object removal process, and sending the monitoring information to the central control module 30;

[0054] The machine learning module 20 is used for receiving the foreign object recognition information from the visual recognition module 10, analyzing the features through a deep neural network, generating a strategy instruction and sending it to the central control module 30; receiving the monitoring information sent by the central control module 30 during the foreign object removal process, adjusting the strategy instruction according to the monitoring information, and sending it to the central control module 30; and also used for offline training to improve the model convergence efficiency;

[0055] The central control module 30 is configured to receive the policy instructions sent by the machine learning module 20 and send them to the foreign object removal module 40, and receive the monitoring information sent by the visual recognition module 10 and send it to the machine learning module 20;

[0056] The foreign object removal module 40 is configured to receive the policy instructions from the central control module 30 and perform foreign object removal operations.

[0057] The visual recognition module 10 collects the surface image of the solar cell to be welded through a camera, extracts the feature of the foreign object edge information in the image, obtains information such as the feature of the foreign object, the geometric size and position of the covered area, and sends the above information to the machine learning module 20.

[0058] The machine learning module 20 includes multiple policy networks, value networks, experience replay buffers, and a global parameter server; among them:

[0059] The policy network receives the recognition information and monitoring information sent from the visual recognition module 10 and the central control module 30, analyzes them through a deep neural network, combines the historical removal effect data, calculates the action probability distribution, generates policy instructions, sends the action probability distribution and policy instructions to the value network, and sends the action selection result, environmental state, and next state information to the experience replay buffer; receives the data sent by the experience replay buffer for offline training;

[0060] The value network receives the recognition information and monitoring information sent from the visual recognition module 10 and the central control module 30, calculates the state value function through a neural network, evaluates the long-term expected return of the current state, calculates the advantage function according to the output value of the state value function and the action probability distribution and policy instructions sent by multiple policy networks, selects the optimal policy instruction, and sends the optimal policy instruction to the global parameter server; receives the data sent by the experience replay buffer for offline training;

[0061] The experience replay buffer receives the action selection result, environmental state, and next state information sent by the policy network, forms a state-action-reward-next state tuple, and after eliminating the temporal correlation by randomly sampling historical data, distributes the batch data to the policy network and the value network for offline training to improve the model convergence efficiency;

[0062] The global parameter server receives the optimal policy instruction sent by the value network, sends the optimal policy instruction to the central control module 30, adjusts the global network parameters using an asynchronous update algorithm, sends the updated parameters to all policy networks, and outputs the uniformly optimized model weights and distributes them to all policy networks to maintain the collaborative learning process.

[0063] Such as Figure 2 、Figure 3 As shown in the figure, the excess material removal module 40 includes a gas blowing device and a liquid spraying device; the gas blowing device includes a gas storage tank 41, a pressure regulating valve 42, a gas delivery pipeline 43, and a jet nozzle 44; the gas storage tank 41 is used to store inert gas, and the inert gas is ejected from the jet nozzle 44 through the gas delivery pipeline 43 to blow the excess material outside the welding area, while preventing the oxidation of the metal layer in the welding area; the pressure regulating valve 42 is installed between the gas storage tank 41 and the gas delivery pipeline 43 and is connected to the central control module 30 through a control motor. The central control module 30 adjusts the output pressure of the gas in real time through the pressure regulating valve 42; the gas delivery pipeline 43 is made of a material that is resistant to high temperature, high pressure, and has a smooth inner wall to reduce pressure loss and impurity generation during gas transmission; the jet nozzle 44 is located on the side of the electrode head 50, moves through an annular track, and adjusts the jet angle in real time based on the machine learning module 20; the excess material liquid spraying device includes a liquid storage tank 45, a liquid delivery pump 46, a flow regulating valve 47, a spraying pipeline 48, and a spray head 49; the liquid storage tank 45 stores the liquid used to remove excess material; the liquid delivery pump 46 is connected to the liquid storage tank 45 to pump out and deliver the liquid to the spraying pipeline 48; the flow regulating valve 47 is located between the gas delivery pipeline 43 and the spraying pipeline 48 to control the liquid flow rate and speed based on the machine learning module 20; the spraying pipeline 48 is made of a corrosion-resistant material that does not react with the liquid; the spray head 49 is located on the side of the electrode head 50, moves through an annular track, and adjusts the spraying angle in real time based on the machine learning module 20 to ensure that the liquid can efficiently remove excess material without damaging the battery.

[0064] The gas pressure regulation range of the excess material gas blowing device is 0.1 - 1.5 MPa, the flow regulation range is 0 - 120 L / min, and the air flow direction regulation range is 0 - 360°; the liquid spraying volume regulation range of the excess material liquid spraying device is 0 - 60 ml / s, the spraying angle regulation range is 0 - 360°, and the spraying speed regulation range is 0 - 15 m / s.

[0065] As Figure 4 shown, the strategy instruction calculation method of the machine learning module 20 is as follows:

[0066] (1). Define the state space, action space, and incentive function; the state space is used to describe the category of excess material, the action space is used to store all excess material removal actions, and the incentive function uses a discounted return function, and the specific formula is as follows:

[0067]

[0068] where, G t represents the return starting from time step t; r t+kdenotes the immediate reward at time step \(t + k\); \(\gamma\) represents the discount factor (\(0\leqslant\gamma\leqslant1\)), which is used to control the importance of future rewards;

[0069] (2) Establish an A3C model, which is used to represent the policy instructions and value function of each agent, execute the policy instructions, adjust the neuron connection weights and activation functions according to the feedback of the foreign matter removal effect, optimize the policy instructions, and form the optimal policy instructions. The specific formula is as follows:

[0070]

[0071] where \(y\) represents the output removal strategy; \(x\) i represents the output foreign matter eigenvalue; \(w\) i represents the weight corresponding to the eigenvalue; \(b\) i represents the bias term; \(f\) represents the activation function; \(n\) represents the number of input eigenvalues.

[0072] During the process of removing foreign matter, the state monitoring module collects the geometric size changes and distribution information of the remaining foreign matter in real time, and feeds it back to the machine learning module 20 through the central control module 30; according to the dynamic feedback data, the A3C model updates the global policy parameters through an asynchronous thread, and adjusts the air blowing pressure, liquid spraying flow rate, spraying angle, etc. in real time to form a closed-loop control.

[0073] The cooperation logic between the air blowing device and the liquid spraying device includes the following levels:

[0074] (1) Priority determination: For loose impurities (such as metal powders, particulate matters, etc.), the air blowing device is preferentially started and purged with high-pressure inert gas; for adhesive impurities (such as colloids, oil stains, etc.), the liquid spraying device is preferentially started, dissolved with chemical solvents and then assisted by air blowing for cleaning; for mixed impurities, an air-liquid alternating spraying strategy is adopted to avoid the influence of liquid residue on the welding quality;

[0075] (2) Parameter linkage control: When the air blowing device is started, the nozzle angle of the liquid spraying device is automatically adjusted to the avoidance mode (at \(90^{\circ}\) to the air flow direction) to prevent liquid splashing from interfering with the air blowing effect; when the liquid spraying device is started, the pressure regulating valve of the air blowing device synchronously reduces the output pressure to avoid air flow disturbing the liquid spraying trajectory; for foreign matter distributed over a large area, the air blowing and liquid spraying devices cooperate in different regions. The air blowing device is responsible for removing loose impurities at the edge, and the liquid spraying device is responsible for removing the adhered area;

[0076] (3) Safety protection mechanism: The spraying volume of the liquid spraying device is dynamically adjusted according to the removal of excess materials. The maximum flow rate is restricted by a flow regulating valve to prevent the liquid from penetrating and damaging the palladium, gold, and silver metal layers on the surface of the solar cell. The jet angle of the air blowing device is synchronously calibrated with the movement trajectory of the electrode head to prevent the air flow impact of the air blowing device from causing the battery to break.

[0077] The gas stored in the air blowing device is an inert gas, which will not cause an oxidation reaction on the palladium, gold, and silver metal layers on the surface of the solar cell during the process of removing excess materials. The liquid stored in the liquid spraying device has good volatility and the ability to carry excess materials, and will not damage the palladium, gold, and silver metal layers on the surface of the solar cell.

[0078] An adaptive excess material removal method under the coordination of multiple systems includes the following steps:

[0079] (1) Before welding, the visual recognition module 10 collects image data of the area to be welded, extracts characteristic information of the excess materials, and transmits it to the machine learning module 20.

[0080] (2) The machine learning module 20 further analyzes the excess material characteristic data provided by the visual recognition module 10, determines the type of excess materials, generates removal strategies, calculates different removal strategies through multiple parallel threads, and generates a strategy instruction and transmits it to the central control module 30 after the analysis is completed.

[0081] (3) The central control module 30 transmits the strategy instruction to the excess material removal module 40 to remove the excess materials.

[0082] (4) The visual recognition module 10 monitors the progress of the removal of excess materials and sends it to the central control module 30. The central control module 30 sends it to the machine learning module 20, and the machine learning module 20 adjusts the strategy instruction.

[0083] (5) The execution of the strategy instruction is looped until the excess materials are completely removed.

[0084] The gas stored in the air blowing device is an inert gas, which will not cause an oxidation reaction on the palladium, gold, and silver metal layers on the surface of the solar cell during the process of removing excess materials. The liquid stored in the liquid spraying device has good volatility and the ability to carry excess materials, and will not damage the palladium, gold, and silver metal layers on the surface of the solar cell.

[0085] The solar cell is a triple-junction gallium arsenide solar cell.

[0086] The machine learning module 20 adopts the A3C machine learning algorithm model.

[0087] As Figure 5As shown below, taking a redundant object adaptive removal device and method under multi-system cooperation as an example, the complete process of the assembly and use of the present invention will be described:

[0088] (1) Preparation before welding: Place the triple-junction gallium arsenide solar cell and the interconnection piece to be welded into the tooling platform, and ensure that each module and device is in the on state.

[0089] (2) Image acquisition and feature extraction: The vision recognition module 10 acquires the image of the area to be welded of the triple-junction gallium arsenide solar cell to be welded, extracts the redundant object information in the image, obtains information such as the features, area, and position of the redundant object, and transmits the above information to the machine learning module 20.

[0090] (3) Analysis and strategy generation by the machine learning module 20: The machine learning module 20 adopts the A3C machine learning algorithm model. Based on the redundant object information data provided by the vision recognition module 10, it identifies and judges the types of redundant objects (such as overflow glue, metal debris, dust, fingerprints, etc. left over from the previous process), their areas, positions, etc. It uses a policy network, a value network, an experience replay buffer, and a global parameter server, and uses multiple parallel threads to simultaneously explore the redundant object removal strategy. At the same time, according to the environmental feedback, it adjusts the removal strategy in real time, performs adaptive parallel thread strategy optimization, forms the optimal execution strategy, and transmits it to the central control module 30.

[0091] (4) Instruction transmission and monitoring adjustment by the central control module 30: The instruction transmission module of the central control module 30 transmits the redundant object removal strategy generated by the machine learning module 20 to the redundant object removal module 40 to achieve real-time control of the redundant object removal process. During the process, the status monitoring module monitors the remaining redundant object information, adjusts the redundant object removal strategy in real time according to the monitoring results, and loops to execute the strategy. The redundant object air blowing device and the redundant object liquid spraying device adjust the air blowing and liquid spraying flow rates and angles in real time according to the instructions, and cooperate to remove the redundant objects until the central control module 30 determines that the redundant objects in the area to be welded are cleared to a weldable state.

[0092] (5) Formal welding: Perform resistance welding of the triple-junction gallium arsenide solar cell and the interconnection piece.

[0093] The above is only the best specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

[0094] The content not detailedly described in the specification of the present invention belongs to the well-known technology of those skilled in the art.

Claims

1. A redundant object adaptive removal device under multi-system coordination, characterized in that: The system comprises a visual recognition module (10), a machine learning module (20), a central control module (30) and a redundant object removal module (40); wherein: A visual recognition module (10) is used to locate and identify the excess material in the area to be welded, and to send the identification information to the machine learning module (20); and to monitor the remaining excess material during the excess material removal process, and to send the monitoring information to the central control module (30); The machine learning module (20) is used to receive the redundant object recognition information from the visual recognition module (10), analyze the features through a deep neural network, generate a strategy instruction, and send it to the central control module (30); receive the monitoring information sent by the central control module (30) during the redundant object removal process, adjust the strategy instruction according to the monitoring information, and send it to the central control module (30); and perform offline training to improve the model convergence efficiency; A central control module (30) is used to receive the strategy instructions sent by the machine learning module (20) and send them to the redundant object removal module (40), and receive the monitoring information sent by the visual recognition module (10) and send it to the machine learning module (20); The redundant object removal module (40) is used to receive the strategy instruction of the central control module (30) and execute the redundant object removal operation.

2. The redundant object adaptive removal device under multi-system coordination according to claim 1 is characterized in that: The method for the visual recognition module (10) to locate and identify the redundant objects in the area to be welded comprises: collecting the surface image of the solar cell to be welded by a camera, extracting the features of the edge information of the redundant objects in the image, and obtaining the features of the redundant objects, the geometric size of the covered area and the position information.

3. The redundant object adaptive removal device under multi-system coordination according to claim 1 is characterized in that: The machine learning module (20) includes a plurality of policy networks, a value network, an experience replay buffer and a global parameter server; wherein: The strategy network receives the recognition information and monitoring information sent from the visual recognition module (10) and the central control module (30), analyzes them through a deep neural network, combines the historical removal effect data, calculates the action probability distribution, generates a strategy instruction, sends the action probability distribution and the strategy instruction to the value network, and sends the action selection result, the environment state and the next state information to the experience replay buffer; receives the data sent by the experience replay buffer for offline training; The value network receives the recognition information and monitoring information sent from the visual recognition module (10) and the central control module (30), calculates the state value function through the neural network, evaluates the long-term expected return of the current state, calculates the advantage function according to the output value of the state value function and the action probability distribution and strategy instructions sent by the multiple strategy networks, selects the optimal strategy instruction, and sends the optimal strategy instruction to the global parameter server; receives the data sent by the experience playback buffer for offline training; The experience replay buffer receives the action selection results, environment state, and next state information sent by the policy network, forms a state-action-reward-next state tuple, and distributes batch data to the policy network and value network for offline training by randomly sampling historical data and eliminating time series correlation, thereby improving the model convergence efficiency. The global parameter server receives the optimal strategy instruction sent by the value network, sends the optimal strategy instruction to the central control module (30), uses an asynchronous update algorithm to adjust the global network parameters, sends the updated parameters to all strategy networks, outputs a uniformly optimized model weight and distributes it to all strategy networks to maintain the collaborative learning process.

4. The redundant object adaptive removal device under multi-system coordination according to claim 1 is characterized in that: The method for generating and adjusting the strategy instructions of the machine learning module (20) is as follows: (1) Define the state space, action space and activation function; the state space is used to describe the category of redundant objects, the action space is used to store all redundant object removal actions, and the activation function adopts a discounted reward function. The specific formula is as follows: Among them, G t represents the return starting from time step t; r t+k represents the immediate reward at time step t+k; γ represents the discount factor (0≤γ≤1), which is used to control the importance of future rewards; (2) Establish an A3C model to represent the policy instructions and value functions of each agent, execute the policy instructions, adjust the neuron connection weights and the incentive function based on the feedback of the redundant object removal effect, optimize the policy instructions, and form the optimal policy instructions. The specific formula is as follows: Among them, y represents the output removal strategy; x i represents the redundant feature value of the output; w i represents the weight of the corresponding eigenvalue; b i represents the bias term; f represents the activation function; n represents the number of input eigenvalues.

5. The redundant object adaptive removal device under multi-system coordination according to claim 1, characterized in that: The excess material removal module (40) includes an air blowing device and a liquid spraying device; the air blowing device includes a gas storage tank (41), a pressure regulating valve (42), a gas delivery pipeline (43) and an air nozzle (44); the gas storage tank (41) is used to store gas, and the gas is sprayed out from the air nozzle (44) through the gas delivery pipeline (43), so as to blow the excess material out of the area to be welded and prevent oxidation of the metal layer in the area to be welded; the pressure regulating valve (42) is installed between the gas storage tank (41) and the gas delivery pipeline (43), and is connected to the central control module (30) through a control motor, so as to adjust the output pressure of the gas in real time according to the strategy instruction; the air nozzle (44) is located on one side of the electrode head (50), and is connected to the central control module (30) through a control motor; the air nozzle (44) is located on one side of the electrode head (50), and is connected to the central control module (30) through a control motor; the air nozzle (44) is located on one side of the electrode head (50), and is connected to the central control module (30) through a control motor; the air nozzle (44) is connected to the central control module (30) through a control motor; the air nozzle (44) is connected to the central control module (30) through a control motor, and ... The liquid spraying device moves along a circular track and adjusts the spray angle in real time according to the strategic instructions; the liquid spraying device comprises a liquid storage tank (45), a liquid delivery pump (46), a flow regulating valve (47), a spray pipe (48) and a nozzle (49); the liquid storage tank (45) stores liquid for removing excess; the liquid delivery pump (46) is connected to the liquid storage tank (45) to extract the liquid and deliver it to the spray pipe (48); the flow regulating valve (47) is located between the gas delivery pipe (43) and the spray pipe (48) to adjust the liquid flow and speed according to the strategic instructions; the nozzle (49) is located on the other side of the electrode head (50), moves along a circular track, and adjusts the spray angle in real time according to the strategic instructions.

6. The redundant object adaptive removal device under multi-system coordination according to claim 5, characterized in that: The gas pressure of the air blowing device is 0.1-1.5MPa, the flow rate is 0-120L / min, and the air flow direction is 0-360°; the liquid injection amount of the liquid injection device is 0-60ml / s, the injection angle is 0-360°, and the injection speed is 0-15m / s.

7. The redundant object adaptive removal device under multi-system coordination according to claim 5, characterized in that: The coordination logic between the air blowing device and the liquid spraying device includes: (1) Priority determination: Different removal strategies are used for different excess materials. For loose impurities, start the air blowing equipment to purge with gas; for adhesive impurities, start the liquid spraying equipment to dissolve with chemical solvents and then use air blowing to assist in cleaning; for mixed impurities, adopt the gas-liquid alternating spraying strategy to avoid liquid residue affecting the welding quality; (2) Parameter linkage control: when the air blowing device is started, the nozzle angle of the liquid spraying device is automatically adjusted to an avoidance mode, wherein the nozzle angle is 90° to the airflow direction to prevent liquid splashing from interfering with the air blowing effect; when the liquid spraying device is started, the pressure regulating valve (42) of the air blowing device simultaneously reduces the output pressure to prevent the airflow from disturbing the liquid spraying trajectory; for the redundant objects distributed over a large area, the air blowing device and the liquid spraying device work together in different areas, the air blowing device removes loose impurities on the edge, and the liquid spraying device removes the adhesion area; (3) Safety protection mechanism: The spray volume of the liquid spray device is dynamically adjusted according to the removal of excess materials, and the maximum flow rate is limited by the flow control valve (47) to prevent liquid from penetrating and damaging the metal layer on the surface of the solar cell; the spray angle of the air blowing device is synchronously calibrated with the movement trajectory of the electrode head (50) to prevent the air flow impact of the air blowing device from causing the battery to break.

8. The redundant object adaptive removal device under multi-system coordination according to claim 5, characterized in that: The gas stored in the gas blowing equipment is an inert gas.

9. A method for adaptively removing redundant objects under multi-system collaboration, characterized in that: The following steps are involved: (1) Before welding, the visual recognition module (10) collects image data of the area to be welded, extracts feature information of the redundant objects, and transmits the feature information to the machine learning module (20); (2) The machine learning module (20) further analyzes the redundant object feature data provided by the visual recognition module (10), determines the category of the redundant objects, generates a removal strategy, calculates different removal strategies through multiple parallel threads, generates a strategy instruction after the analysis is completed, and transmits it to the central control module (30); (3) The central control module (30) transmits the strategy instruction to the redundant object removal module (40) to remove the redundant objects; (4) monitoring the progress of removing the excess material through the visual recognition module (10) and sending it to the central control module (30), which sends it to the machine learning module (20), and the machine learning module (20) adjusts the strategy instructions; (5) The strategy instructions are executed repeatedly until the redundant objects are removed.

10. The method for adaptively removing redundant objects under multi-system collaboration according to claim 9, characterized in that: The method for calculating the strategy instruction of the machine learning module (20) is as follows: (1) Define the state space, action space and activation function; the state space is used to describe the category of redundant objects, the action space is used to store all redundant object removal actions, and the activation function adopts a discounted reward function. The specific formula is as follows: Among them, G t represents the return starting from time step t; r t+k represents the immediate reward at time step t+k; γ represents the discount factor (0≤γ≤1), which is used to control the importance of future rewards; (2) Establish an A3C model to represent the policy instructions and value functions of each agent, execute the policy instructions, adjust the neuron connection weights and the incentive function based on the feedback of the redundant object removal effect, optimize the policy instructions, and form the optimal policy instructions. The specific formula is as follows: Among them, y represents the output removal strategy; x i represents the redundant feature value of the output; w i represents the weight of the corresponding eigenvalue; b i represents the bias term; f represents the activation function; n represents the number of input eigenvalues.

Images