An industrial cleaning system based on multimodal perception and intelligent control

Abstract

This invention belongs to the field of industrial automation and relates to an industrial cleaning system based on multimodal perception and intelligent control. It achieves accurate stain identification by integrating multiple sensors and utilizing multimodal perception and spectral analysis; intelligently plans cleaning paths using topology optimization and reinforcement learning; intelligently controls cleaning parameters using adaptive control and model prediction; constructs a multi-robot collaborative cleaning system to handle large-scale complex objects; and establishes an intelligent monitoring and management platform. The aim is to solve problems such as inaccurate stain identification, unreasonable path planning, and unintelligent parameter control in industrial cleaning, thereby improving the efficiency, quality, and adaptability of industrial cleaning and meeting the needs of modern industrial production.

Description

Technical Field

[0001] This invention relates to the field of industrial robot automation, and more particularly to an industrial cleaning system based on multimodal perception and intelligent control. Background Technology

[0002] Industrial cleaning is a crucial step in many stages of industrial production, ensuring product quality and extending equipment lifespan. Industries such as machinery manufacturing, electronics processing, and food and beverage all have stringent requirements for industrial cleaning. With the development of industrial automation, industrial robots are increasingly being used in industrial cleaning tasks to replace traditional manual cleaning methods. However, industrial robots still face numerous challenges in industrial cleaning, severely hindering their widespread application and further development in this field.

[0003] Industrial production involves cleaning a wide variety of objects, with significant differences in the type, distribution, and adhesion of surface stains. Traditional industrial robot cleaning systems rely primarily on simple visual recognition technology, which can only roughly determine the presence of stains and cannot accurately identify their specific type, composition, and distribution. This prevents the robot from selecting appropriate cleaning processes and parameters based on the actual condition of the stains, resulting in poor cleaning effects and potentially causing unnecessary damage to the object being cleaned.

[0004] Different objects requiring cleaning have different shapes and structures, necessitating different cleaning paths to ensure thorough and uniform cleaning. Existing industrial robot cleaning path planning methods are often based on fixed patterns and simple geometric models, failing to adequately consider the complex shapes, surface features, and dirt distribution of the objects being cleaned. This leads to unreasonable cleaning path planning, resulting in blind spots and repeated cleaning, wasting cleaning resources and reducing cleaning efficiency.

[0005] In industrial cleaning processes, cleaning parameters such as the flow rate, pressure, and temperature of the cleaning fluid, as well as the speed and force of the cleaning tools, have a significant impact on cleaning effectiveness and cost. Different cleaning objects and types of stains require different combinations of cleaning parameters, and these parameters need to be adjusted in real time as stains are removed and the surface condition of the object changes. However, most current industrial robot cleaning systems use preset, fixed cleaning parameters, which cannot be intelligently adjusted according to actual cleaning conditions, making it difficult to achieve efficient and energy-saving cleaning operations. Summary of the Invention

[0006] This invention provides an industrial cleaning system based on multimodal sensing and intelligent control, comprising:

[0007] The stain identification algorithm module based on multimodal perception and spectral analysis integrates various types of sensors, such as high-resolution vision sensors, infrared sensors, laser displacement sensors, and spectral analyzers, into an industrial robot to acquire multimodal information such as images, temperature distribution, surface morphology, and spectral characteristics of the object being cleaned in real time. It then uses multimodal perception technology to fuse the data and combines spectral analysis technology with deep learning algorithms to identify detailed information such as the type, composition, distribution, and adhesion degree of stains.

[0008] The intelligent cleaning path planning algorithm module based on topology optimization and reinforcement learning utilizes 3D modeling technology and topology optimization methods to construct a topological model of the cleaning object based on its shape, structure, and stain distribution. This transforms the cleaning path planning problem into an optimal path search problem on this model. Reinforcement learning algorithms, such as Deep Q-Network (DQN) or Proximal Policy Optimization (PPO), are then used to define the state space, action space, and reward function, enabling the robot to learn the optimal cleaning path planning strategy.

[0009] The intelligent control algorithm module for cleaning parameters based on adaptive control and model prediction installs pressure sensors, flow sensors, temperature sensors, force sensors, etc. in the cleaning system to monitor parameters such as cleaning fluid flow rate, pressure, temperature, and the interaction force between cleaning tools and objects in real time. The adaptive control algorithm automatically adjusts the cleaning parameters based on real-time parameters and stain identification results. Combined with the dynamic model of the cleaning process, the model predictive control algorithm predicts the future cleaning state and effect and adjusts the parameters in advance.

[0010] The large-scale complex object cleaning system module based on multi-robot collaboration can rationally allocate cleaning tasks to multiple industrial robots to complete collaboratively according to the size, shape and complexity of the object to be cleaned. It establishes a communication and collaborative control mechanism between multiple robots based on a communication network, realizing information sharing and collaborative operation among robots.

[0011] The intelligent monitoring and management platform module for industrial cleaning processes interacts with the industrial robot cleaning system to collect and store data such as stain identification results, cleaning paths, cleaning parameters, cleaning effects, and equipment status in real time. It uses big data analysis and visualization technology to analyze and display the collected data, enabling intelligent monitoring and management of the cleaning process, and providing intelligent early warning and fault diagnosis functions.

[0012] Furthermore, in the stain accurate identification algorithm module based on multimodal perception and spectral analysis, the multimodal data fusion adopts a fusion network based on an attention mechanism.

[0013] Furthermore, in the intelligent planning algorithm module for cleaning paths based on topology optimization and reinforcement learning, the state space includes information such as the current state of the object being cleaned, the distribution of stains, and the position and posture of the robot; the action space includes various motion actions of the robot and cleaning operations; and the reward function is designed based on factors such as cleaning effect, cleaning efficiency, and resource consumption.

[0014] Furthermore, in the intelligent control algorithm module for cleaning parameters based on adaptive control and model prediction, the established dynamic model of the cleaning process is optimized through the analysis and experimental verification of a large amount of cleaning data to improve the accuracy and reliability of the model.

[0015] Furthermore, the large-scale complex object cleaning system module based on multi-robot collaboration considers factors such as the robot's working capacity, working range, and current task status when allocating cleaning tasks.

[0016] Furthermore, the intelligent monitoring and management platform module for the industrial cleaning process can monitor data in real time during the cleaning process by setting thresholds and analysis models, thereby achieving intelligent early warning and fault diagnosis functions.

[0017] Furthermore, in the stain identification algorithm module based on multimodal perception and spectral analysis, the deep learning algorithm uses a convolutional neural network (CNN) or a recurrent neural network (RNN) to train the stain sample data and establish a correlation model between stain features and cleaning process.

[0018] Furthermore, in the intelligent planning algorithm module for cleaning paths based on topology optimization and reinforcement learning, 3D modeling technology is used to obtain the precise shape and structural information of the cleaning object, and then topology modeling is performed.

[0019] Furthermore, in the intelligent control algorithm module for cleaning parameters based on adaptive control and model prediction, the adaptive control algorithm automatically adjusts parameters such as the flow rate, pressure, and temperature of the cleaning fluid, as well as the movement speed and force of the cleaning tool, according to real-time monitoring parameters and stain identification results.

[0020] Furthermore, in the large-scale complex object cleaning system module based on multi-robot collaboration, multiple robots coordinate their movements and operations through a collaborative control algorithm to ensure that different parts of the object are cleaned simultaneously.

[0021] Beneficial effects:

[0022] Algorithms based on multimodal perception and spectral analysis accurately identify stains, increasing the accuracy rate from 60% to over 90%, improving cleaning effectiveness by approximately 35%, and avoiding cleaning damage. Path planning algorithms based on topology optimization and reinforcement learning reduce the cleaning blind spot rate to below 5%, control the re-cleaning rate to within 5%, increase cleaning efficiency by approximately 40%, and save cleaning resources. Parameter control algorithms based on adaptive control and model prediction reduce the defect rate from 15% to below 5%, lower cleaning costs by approximately 30%, and stabilize cleaning quality. The multi-robot collaborative cleaning system is suitable for large-scale complex objects, significantly improving cleaning efficiency and expanding the application scope of industrial robots. The intelligent monitoring and management platform enables real-time monitoring, intelligent early warning, and fault diagnosis, ensuring smooth cleaning processes, improving production management, and comprehensively promoting the upgrading of industrial cleaning technology. Attached Figure Description

[0023] Figure 1 Overall process flow diagram. Detailed Implementation

[0024] Example of a stain identification algorithm based on multimodal sensing and spectral analysis

[0025] Sensor Deployment and Data Acquisition: A high-resolution vision sensor with a resolution of 2000×1500 is installed at a 45° angle at the front end of the industrial robot's arm to ensure comprehensive imaging of the surface to be cleaned; an infrared sensor is attached closely to the vision sensor to acquire surface temperature distribution; a laser displacement sensor is installed vertically downwards, achieving a measurement accuracy of up to [missing information].

[0026] ±0.05mm is used to acquire surface morphology; a spectrometer connected via fiber optic cable enables rapid analysis of the stain spectrum. Each sensor acquires data in real time at a frequency of 50Hz and transmits it to the data processing layer via high-speed Ethernet.

[0027] Multimodal data fusion and stain feature extraction: The data processing layer employs an attention-based fusion network to extract texture features from visual images through convolution operations; it performs gridding processing on infrared temperature data; it converts laser displacement data into height maps; and it normalizes spectral data. The fusion network learns the weights of each modality to construct a comprehensive information model. The stain feature extraction module uses spectral analysis techniques, combined with a chemical feature library, to determine the chemical composition and type of stains.

[0028] Stain Recognition Model Training and Application: A dataset containing 1000 different stain samples was collected, covering common stains such as oil, rust, and dust. The model was trained using a Convolutional Neural Network (CNN), which consists of 7 convolutional layers and 3 fully connected layers. The convolutional layers used 5×5 convolutional kernels to extract features, and ReLU activation function was used to increase non-linearity. In actual cleaning, real-time multimodal data was input into the trained model, and the model output detailed stain information.

[0029] Algorithm modeling and solution process: Taking CNN training as an example, the loss function is cross-entropy loss, the Adam optimizer is used, the learning rate is set to 0.001, and the training is iterated 500 times. The network parameters are continuously adjusted through backpropagation, so that the model's accuracy in stain recognition is continuously improved.

[0030] Enhancement Description: Traditional simple visual recognition can only roughly determine the presence or absence of stains, with an accuracy rate of about 60%. This algorithm improves the accuracy rate to over 90%, accurately determining the type and composition of stains, allowing for more precise selection of cleaning processes, improving cleaning effectiveness by about 35%, and effectively reducing cleaning damage.

[0031] Example of an intelligent path planning algorithm for cleaning based on topology optimization and reinforcement learning

[0032] Topological modeling of the cleaning object: A precise 3D model of the cleaning object is obtained by scanning it using 3D laser scanning technology. Topology optimization software is then used to simplify the model into a topological structure based on the stain distribution and the object's structure, removing irrelevant details and retaining key connections and boundaries.

[0033] Reinforcement learning model training: A state space is defined, containing information such as the distribution of dirt on the surface of the object to be cleaned, the robot's position and orientation, and the cleaned area. The action space includes operations such as linear movement, rotation, and opening and closing of the cleaning tools. The reward function uses cleaning coverage, cleaning efficiency, and resource consumption as indicators; a 10% increase in cleaning coverage is rewarded with +5 points, while exceeding the resource consumption limit results in a -10 penalty. The model is trained using the Proximal Policy Optimization (PPO) algorithm, with a discount factor of 0.95 and an update step size of 20.

[0034] Intelligent cleaning path planning and execution: During actual cleaning, the intelligent cleaning path planning module generates cleaning path instructions based on the topology model and the trained reinforcement learning model, and the robot executes the cleaning task according to the instructions.

[0035] Algorithm modeling and solution process: The PPO algorithm samples the trajectory under the current policy, calculates the advantage function, updates the policy network and value network using importance sampling, and continuously iterates and optimizes the policy so that the robot learns the optimal cleaning path.

[0036] Enhancement Description: Traditional fixed-mode cleaning path planning results in a cleaning blind spot rate of 20% and a repeated cleaning rate of 15%. This algorithm reduces the cleaning blind spot rate to below 5%, controls the repeated cleaning rate to within 5%, improves cleaning efficiency by approximately 40%, and significantly saves cleaning resources.

[0037] Example of an intelligent control algorithm for cleaning parameters based on adaptive control and model prediction

[0038] Sensor Installation and Data Acquisition: A flow sensor with an accuracy of ±0.1 L / min is installed in the cleaning fluid pipeline; a pressure sensor with an accuracy of ±0.05 MPa is installed at the nozzle; a temperature sensor with an accuracy of ±1℃ is installed on the cleaning tool; and a force sensor with an accuracy of ±0.05 N is installed at the connection between the cleaning tool and the robotic arm. The sensors acquire data in real time at a frequency of 10 Hz and transmit it to the data processing layer.

[0039] A dynamic model of the cleaning process was established: based on the jet dynamics of the cleaning fluid and the interaction principle between the cleaning tool and the stain, a dynamic model of the cleaning process was established, considering the effects of cleaning fluid flow rate, pressure, and temperature on the stain removal rate, as well as the relationship between the force of the cleaning tool and the cleaning effect. The model parameters were fitted with a large amount of experimental data to ensure that the model accurately reflects the cleaning process.

[0040] Intelligent control and execution of cleaning parameters: The adaptive control module adjusts cleaning parameters using an adaptive PID control algorithm based on real-time sensor data and stain identification results. The model predictive control module predicts future cleaning conditions based on a dynamic model and adjusts parameters in advance. For example, if it predicts that stains in a certain area will be difficult to remove, it increases the cleaning fluid flow rate and pressure in advance.

[0041] Algorithm modeling and solution process: Adaptive PID control adjusts the proportional, integral, and derivative coefficients in real time according to the deviation; Model predictive control predicts the future cleaning state and calculates the parameter adjustment amount by solving the dynamic model.

[0042] Enhancement Description: Traditional fixed-parameter cleaning results in inconsistent cleaning quality and a defect rate of approximately 15%. This algorithm stabilizes cleaning quality, reduces the defect rate to below 5%, and lowers cleaning costs by approximately 30%, effectively improving both cleaning efficiency and quality.

Claims

1. An industrial robot cleaning system based on multimodal perception and intelligent control, characterized in that, include: The stain identification algorithm module based on multimodal perception and spectral analysis integrates high-resolution vision sensors, infrared sensors, laser displacement sensors, and spectral analyzers on industrial robots to acquire images, temperature distribution, surface morphology, and spectral characteristics of the surface to be cleaned in real time. It uses multimodal perception technology to fuse data and combines spectral analysis technology with deep learning algorithms to identify the type, composition, distribution, and adhesion degree of stains. The intelligent cleaning path planning algorithm module based on topology optimization and reinforcement learning utilizes 3D modeling technology and topology optimization methods to construct a topological model of the cleaning object based on its shape, structure, and stain distribution. This transforms the cleaning path planning problem into an optimal path search problem on this model. By using deep Q-networks or proximal policy optimization algorithms, the state space, action space, and reward function are defined, enabling the robot to learn the optimal cleaning path planning strategy. The intelligent control algorithm module for cleaning parameters based on adaptive control and model prediction installs pressure sensors, flow sensors, temperature sensors, and force sensors in the cleaning system to monitor the cleaning fluid flow, pressure, temperature, and the force parameters between the cleaning tools and the cleaning objects in real time. The adaptive control algorithm automatically adjusts the cleaning parameters based on the real-time parameters and stain identification results. Combined with the dynamic model of the cleaning process, the model prediction control algorithm predicts the future cleaning state and effect and adjusts the parameters in advance. The large-scale complex object cleaning system module based on multi-industrial robot collaboration can rationally allocate cleaning tasks to multiple industrial robots to complete collaboratively according to the size, shape and complexity of the object to be cleaned. It establishes a communication and collaborative control mechanism between multiple industrial robots based on a communication network, realizing information sharing and collaborative operation among robots. The intelligent monitoring and management platform module for industrial cleaning processes interacts with the industrial robot cleaning system to collect and store real-time data on stain identification results, cleaning paths, cleaning parameters, cleaning effects, and equipment status during the cleaning process. Utilizing big data analytics and visualization technology, it analyzes and displays the collected data, enabling intelligent monitoring and management of the cleaning process, and providing intelligent early warning and fault diagnosis functions.

2. The industrial robot cleaning system based on multimodal perception and intelligent control according to claim 1, characterized in that, In the stain identification algorithm module based on multimodal perception and spectral analysis, the multimodal data fusion adopts a fusion network based on an attention mechanism.

3. The industrial robot cleaning system based on multimodal perception and intelligent control according to claim 1, characterized in that, In the intelligent cleaning path planning algorithm module based on topology optimization and reinforcement learning, the state space includes the current state of the cleaning object, the distribution of stains, and the position and posture information of the robot; the action space includes various motion actions of the robot and cleaning operations; and the reward function is designed based on cleaning effect, cleaning efficiency, and resource consumption factors.

4. The industrial robot cleaning system based on multimodal perception and intelligent control according to claim 1, characterized in that, In the intelligent control algorithm module for cleaning parameters based on adaptive control and model prediction, the established dynamic model of the cleaning process is optimized through the analysis and experimental verification of a large amount of cleaning data to improve the accuracy and reliability of the model.

5. The industrial robot cleaning system based on multimodal perception and intelligent control according to claim 1, characterized in that, The large-scale complex object cleaning system module based on multi-industrial robot collaboration considers the robot's working capacity, working range, and current task status when assigning cleaning tasks.

6. The industrial robot cleaning system based on multimodal perception and intelligent control according to claim 1, characterized in that, The intelligent monitoring and management platform module for the industrial cleaning process monitors data in real time during the cleaning process by setting thresholds and analysis models, thereby enabling intelligent early warning and fault diagnosis functions.

7. The industrial robot cleaning system based on multimodal perception and intelligent control according to claim 1, characterized in that, In the stain identification algorithm module based on multimodal perception and spectral analysis, the deep learning algorithm uses a convolutional neural network (CNN) or a recurrent neural network (RNN) to train the stain sample data and establish a correlation model between stain features and cleaning process.

8. The industrial robot cleaning system based on multimodal perception and intelligent control according to claim 1, characterized in that, In the intelligent planning algorithm module for cleaning paths based on topology optimization and reinforcement learning, 3D modeling technology is used to obtain the precise shape and structural information of the cleaning object, and then topology modeling is performed.

9. The industrial robot cleaning system based on multimodal perception and intelligent control according to claim 1, characterized in that, In the intelligent control algorithm module for cleaning parameters based on adaptive control and model prediction, the adaptive control algorithm automatically adjusts the flow rate, pressure, and temperature of the cleaning fluid, as well as the movement speed and force parameters of the cleaning tool, according to real-time monitoring parameters and stain identification results.

10. The industrial robot cleaning system based on multimodal perception and intelligent control according to claim 1, characterized in that, In the large-scale complex object cleaning system module based on multi-industrial robot collaboration, multiple industrial robots coordinate their movements and operations through a collaborative control algorithm to ensure that different parts of the object are cleaned simultaneously.

Images