Robot inspection method, device and system for polycrystalline silicon production park

Abstract

The invention discloses a robot inspection method, device and system for a polycrystalline silicon production park. The method comprises the steps that firstly, a robot inspection path is planned according to a dynamic risk map, a global planning path is obtained, and the dynamic risk map is constructed based on multi-dimensional environment data; according to the global planning path, controlling the robot to inspect in the polycrystalline silicon production park; performing local adjustment on the global planning path based on the regional risk in the robot inspection process to obtain an adjusted global planning path; and finally, according to the adjusted global planning path, controlling the robot of the polycrystalline silicon production park to continue to inspect. According to the routing inspection method, the dynamic risk map is combined, and real-time dynamic optimization of the routing inspection path can be realized through cooperation of global planning and local adjustment, so that emergencies can be quickly responded.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot inspection, and in particular relates to a robot inspection method, device and system for a polysilicon production park. Background Art

[0002] In polysilicon production parks, especially during the production process, real-time monitoring of the production environment and equipment inspections are crucial for ensuring production safety and improving efficiency. However, traditional manual inspections have numerous shortcomings, such as low efficiency, slow response, and the tendency to miss critical issues. These issues are particularly prominent in polysilicon production, as the environment is often characterized by complex factors such as high corrosion, high temperatures, and large amounts of dust. This greatly increases the difficulty and risk of manual inspections, severely impacting their reliability and safety.

[0003] With the development of automated industrial inspection technology, it has begun to be gradually applied and promoted in polysilicon production parks. However, most current automated industrial inspection systems have certain limitations. They typically rely on fixed inspection routes or simple obstacle avoidance strategies, lacking the effective integration of global planning and local adjustments. This means that they cannot dynamically adjust inspection routes based on real-time risk conditions. For example, when an emergency occurs, such as a sudden leak in a piece of equipment, existing inspection systems often still follow a predetermined route, making it difficult to respond in a timely manner. This delay not only misses the optimal response time but also potentially leads to greater safety hazards. Summary of the Invention

[0004] The present invention addresses the aforementioned shortcomings of the existing technology by proposing a robotic inspection method, device, and system for polysilicon production parks. This method controls robotic inspections based on a dynamic risk map, combining global planning with local adjustments. By incorporating a dynamic risk map and synergizing global planning with local adjustments, this inspection method achieves real-time dynamic optimization of inspection paths, enabling rapid response to emergencies.

[0005] In a first aspect, the present invention provides a robot inspection method for a polysilicon production park, the method comprising the following steps:

[0006] Plan the robot inspection path based on the dynamic risk map to obtain the global planning path;

[0007] According to the global planning path, the robot is controlled to conduct inspections in the polysilicon production park;

[0008] During the robot inspection process, the global planning path is locally adjusted based on regional risks to obtain the adjusted global planning path;

[0009] According to the adjusted global planning path, the robots in the polysilicon production park are controlled to continue their inspections.

[0010] Furthermore, before planning the robot inspection path according to the dynamic risk map and obtaining the global planned path, the method further includes:

[0011] Based on the multi-dimensional environmental data of the polysilicon production park, a dynamic risk map of the polysilicon production park is constructed.

[0012] Furthermore, the multi-dimensional environmental data of the polysilicon production park includes equipment operating parameters, environmental parameters, and visual data of the polysilicon production park;

[0013] Based on multi-dimensional environmental data, a dynamic risk map of the polysilicon production park is constructed, which includes the following steps:

[0014] Step A1: Divide the polysilicon production park into multiple sub-areas; and normalize the multi-dimensional environmental data of each sub-area to obtain normalized measurement values ​​of the environmental data of each sub-area;

[0015] Step A2: Based on the historical fault correlation and the weights corresponding to the normalized measurement values ​​and multi-dimensional environmental data, the discontinuity risk value of each area is obtained;

[0016] Step A3: Generate a risk distribution surface based on the discontinuous risk value;

[0017] Step A4: superimpose the risk distribution surface with the polysilicon production park building information model to form a three-dimensional dynamic risk map, thereby obtaining a dynamic risk map of the polysilicon production park.

[0018] Furthermore, step A3 specifically includes the following steps:

[0019] Based on the discontinuous risk value, the unmonitored risk value is estimated using the Kriging interpolation method;

[0020] Summarize discontinuous risk values ​​and unmonitored risk values ​​to obtain a risk value grid;

[0021] According to the preset resolution, the risk value grid is refined using the inverse distance weighted method to obtain the risk distribution surface.

[0022] Furthermore, after step A4, the method further includes:

[0023] Step A5: Visualize the risk level in the three-dimensional dynamic risk map through color gradient;

[0024] If the risk level is greater than or equal to the first risk threshold, the risk level is set as a high-risk area and marked with a first color; if the risk level is greater than or equal to the second risk threshold and less than the first risk threshold, the risk level is set as a medium-risk area and marked with a second color; if the risk level is less than the second risk threshold, the risk level is set as a low-risk area and marked with a third color;

[0025] A pulse warning box is generated in the high-risk area of ​​the three-dimensional dynamic risk map, and a real-time data floating window is displayed on the AR terminal.

[0026] Furthermore, the robot inspection path is planned according to the dynamic risk map to obtain a global planning path, which specifically includes the following steps:

[0027] Based on the dynamic risk map, the graph structure of the graph neural network model is constructed;

[0028] The graph nodes in the graph structure of the graph neural network model represent the center point of a key device or a high-risk area, and the edges between the graph nodes in the graph structure of the graph neural network model represent the paths that the robot can pass through;

[0029] The graph neural network model is used to learn the graph structure and generate a global planning path.

[0030] Furthermore, during the robot inspection process, the global planning path is locally adjusted based on regional risks to obtain the adjusted global planning path, which specifically includes the following steps:

[0031] Dividing the global planning path into a plurality of continuous path nodes; and controlling the robot to perform inspections according to each path node of the global planning path;

[0032] Collect the current state of the robot at the current path node; the current state includes the robot state, environment state and task state;

[0033] Based on the current state, the policy network of the proximal policy optimization model is used to generate new actions for the robot; the new actions include movement actions and emergency actions;

[0034] Update the current path according to the new action, generate new path points or adjust existing path points to obtain new path points;

[0035] Calculate the reward value of the new path point;

[0036] Update the policy network of the proximal policy optimization model according to the reward value and optimize the path adjustment strategy;

[0037] According to the path adjustment strategy, the global planning path is adjusted to obtain the adjusted global planning path.

[0038] In a second aspect, the present invention provides a robot inspection device for a polysilicon production park, the device comprising:

[0039] The planning unit is used to plan the robot inspection path according to the dynamic risk map to obtain the global planning path;

[0040] A first control unit is connected to the planning unit and is used to control the robot to perform inspections in the polysilicon production park according to the global planning path;

[0041] an adjustment unit connected to the first control unit, configured to locally adjust the global planning path based on regional risks during the robot inspection process to obtain an adjusted global planning path;

[0042] The second control unit is connected to the adjustment unit and is used to control the robot in the polysilicon production park to continue inspection according to the adjusted global planning path.

[0043] In a third aspect, the present invention provides a robot inspection system for a polysilicon production park, the system comprising:

[0044] Inspection robots, used to inspect polysilicon production parks;

[0045] Inspection server, communicating with the inspection robot;

[0046] The inspection server includes a processor configured to execute the robot inspection method for a polysilicon production park according to the first aspect.

[0047] Furthermore, the number of inspection robots is N, where N is a natural number greater than or equal to 1;

[0048] Among them, the inspection robot's body is covered with polytetrafluoroethylene coating and is equipped with a multispectral camera and a hydrogen fluoride concentration sensor.

[0049] This invention uses multi-dimensional environmental data to construct a dynamic risk map, and combines global planning with local adjustments during the inspection process to achieve real-time dynamic optimization of inspection paths, thereby quickly responding to emergencies. The specific effective effects are as follows:

[0050] 1. The present invention optimizes inspection routes and avoids safety failures while improving inspection efficiency through the construction of dynamic risk maps and local adjustment of paths.

[0051] 2. The present invention combines global route planning with local adjustment strategies to effectively enhance the flexibility and adaptability of inspections, and can cope with complex and changing production environments, thereby improving maintenance warning and emergency response capabilities.

[0052] 3. The present invention adopts an automated control method to reduce manual intervention, improve the stability and reliability of inspections, and ensure that key equipment and important areas can be inspected in a timely and comprehensive manner.

[0053] 4. The present invention can meet the needs of the complex environment of polysilicon production parks and provide reliable technical support for ensuring production safety and improving production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of a robot inspection method for a polysilicon production park in an embodiment of the present invention;

[0055] Figure 2 Schematic diagram of a robot inspection framework for a polysilicon production park in an embodiment of the present invention;

[0056] Figure 3 Schematic diagram of a robot inspection device for a polysilicon production park in an embodiment of the present invention;

[0057] Figure 4 Schematic diagram of the framework of the robot inspection system for a polysilicon production park in an embodiment of the present invention.

[0058] Reference numerals: 10, acquisition unit, 20, construction unit, 30, planning unit, 40, adjustment unit, 50, control unit, 100, inspection server, 200, inspection robot. DETAILED DESCRIPTION

[0059] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0060] It should be understood that the specific embodiments and drawings described herein are only used to explain the present invention rather than to limit the present invention.

[0061] It is understood that, in the absence of conflict, the various embodiments of the present invention and the various features in the embodiments may be combined with each other.

[0062] It can be understood that, for the convenience of description, the drawings of the present invention only show parts related to the present invention, while parts unrelated to the present invention are not shown in the drawings.

[0063] It can be understood that each unit and module involved in the embodiments of the present invention may correspond to only one physical structure, or may be composed of multiple physical structures, or multiple units and modules may be integrated into one physical structure.

[0064] It will be understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of the present invention may occur in an order different from that marked in the drawings.

[0065] It is understood that the flowcharts and block diagrams of the present invention illustrate the possible architectures, functions, and operations of the systems, devices, equipment, and methods according to various embodiments of the present invention. Each box in the flowchart or block diagram may represent a unit, module, program segment, or code, which contains executable instructions for implementing the specified functions. Moreover, each box or combination of boxes in the block diagram and flowchart may be implemented using a hardware-based system that implements the specified functions, or may be implemented using a combination of hardware and computer instructions.

[0066] It can be understood that the units and modules involved in the embodiments of the present invention can be implemented by software or hardware. For example, the units and modules can be located in a processor.

[0067] Example 1:

[0068] This embodiment provides a robot inspection method for a polysilicon production park. This method is mainly used in daily inspection and safety monitoring scenarios of polysilicon production enterprises, and is particularly suitable for production environments with complex environments, dense equipment, and high risks. In terms of equipment operation status monitoring, real-time collection of environmental parameters, and rapid identification of potential risks, this method can automatically perform key equipment detection, gas leak warning, dust concentration monitoring, and visual anomaly identification to ensure equipment safety and environmental hygiene during the production process. At the same time, when responding to emergencies such as equipment failures, leakage accidents, or environmental anomalies, the system can dynamically adjust the inspection path, give priority to detecting dangerous areas, and improve the early detection and response efficiency of accidents. This method is not only suitable for routine inspections, but can also play an important role in various scenarios such as emergency response to emergencies, maintenance plan optimization, and production site risk assessment, significantly improving the safety and operational efficiency of polysilicon production parks.

[0069] like Figure 1 As shown, the inspection method includes the following steps:

[0070] Step S1: Plan the robot inspection path according to the dynamic risk map to obtain the global planning path.

[0071] Specifically, before planning the robot inspection path according to the dynamic risk map and obtaining the global planned path, the inspection method further includes:

[0072] Based on the multi-dimensional environmental data of the polysilicon production park, a dynamic risk map of the polysilicon production park is constructed.

[0073] The multi-dimensional environmental data of the polysilicon production park includes the equipment operating parameters, environmental parameters, and visual data of the polysilicon production park.

[0074] Based on multi-dimensional environmental data, a dynamic risk map of the polysilicon production park is constructed, which includes the following steps:

[0075] Step A1: Divide the polysilicon production park into multiple sub-areas; and normalize the multi-dimensional environmental data of each sub-area to obtain normalized measurement values ​​of the environmental data of each sub-area.

[0076] The calculation formula of the normalized measurement value is as follows:

[0077]

[0078] in,

[0079] S k,i represents the measured value of the kth type of environmental data in area i;

[0080] Represents the normalized value of the measured value of the kth type of environmental data in area i;

[0081] U k,max Indicates the upper limit of the parameter safety threshold of the kth category of environmental data;

[0082] L k,min Indicates the lower limit of the parameter safety threshold of the kth category of environmental data.

[0083] Step A2: Based on the historical fault correlation and the weights corresponding to the normalized measurement value and the multi-dimensional environmental data, the discontinuity risk value of each area is obtained.

[0084] Without considering the correlation of historical faults, the calculation formula of the discontinuous risk value of each area is as follows:

[0085]

[0086] w k Represents the weight of the k-th category of environmental data.

[0087] Considering the historical fault correlation, the calculation formula of the discontinuity risk value of each area is as follows:

[0088] R i-total =R i +λ×H i,t (3);

[0090] Among them, R i-total represents the total risk value of region i; H i,t Indicates the historical fault correlation;

[0091] λ represents the weight coefficient of historical fault correlation;

[0092] Historical fault correlation Hi,t The calculation formula is as follows:

[0093]

[0094] Where D represents the total number of historical days;

[0095] d represents the historical days index;

[0096] τ represents the time decay constant;

[0097] I i,t-d is the fault indicator function, which is 1 if a fault occurs in area i within td days, and 0 otherwise.

[0098] Step A3: Generate a risk distribution surface based on the discontinuous risk values.

[0099] Step A3 specifically includes the following steps:

[0100] Based on the discontinuous risk value, the unmonitored risk value is estimated using the Kriging interpolation method;

[0101] Summarize discontinuous risk values ​​and unmonitored risk values ​​to obtain a risk value grid;

[0102] According to the preset resolution, the risk value grid is refined using the inverse distance weighted method to obtain the risk distribution surface.

[0103] Specifically, the Kriging interpolation method is used to calculate the risk value of the unmonitored area for discrete sensor points; the inverse distance weighted (IDW) method is used around key equipment to increase the spatial resolution to 0.1m×0.1m.

[0104] Inverse distance weighting (IDW) is a distance-based spatial interpolation method. Its detailed process is as follows: First, the location of the point to be interpolated is selected, and the distances from all known points to that point are calculated. Next, a weight is calculated based on the distance from each known point to the point to be interpolated. This weight is typically calculated by dividing 1 by the distance raised to the power of p, where p is a set exponent used to adjust the effect of farther points on the result. The value of each known point is then multiplied by the corresponding weight to obtain the sum of the weighted values ​​of all known points. Finally, this sum is divided by the sum of all weights to obtain the value of the point to be interpolated. This way, closer known points have a greater influence on the interpolation result, achieving smoothing and refinement of spatial information.

[0105] Step A4: superimpose the risk distribution surface with the polysilicon production park building information model to form a three-dimensional dynamic risk map, thereby obtaining a dynamic risk map of the polysilicon production park.

[0106] After step A4, the method further comprises:

[0107] Step A5: Visualize the risk level in the three-dimensional dynamic risk map through color gradient;

[0108] If the risk level is greater than or equal to the first risk threshold, the risk level is set as a high-risk area and marked with a first color; if the risk level is greater than or equal to the second risk threshold and less than the first risk threshold, the risk level is set as a medium-risk area and marked with a second color; if the risk level is less than the second risk threshold, the risk level is set as a low-risk area and marked with a third color;

[0109] Generate a pulse warning box in the high-risk area of ​​the 3D dynamic risk map, and display a real-time data floating window on the AR terminal.

[0110] In specific implementation, the risk level is visualized through a color gradient, with red indicating Ri ≥ 0.8, yellow indicating 0.5 ≤ Ri < 0.8, and green indicating Ri < 0.5.

[0111] The robot inspection path is planned based on the dynamic risk map to obtain the global planning path. The specific steps include:

[0112] Based on the dynamic risk map, the graph structure of the graph neural network model is constructed;

[0113] The graph nodes in the graph structure of the graph neural network model represent the center point of a key device or a high-risk area, and the edges between the graph nodes in the graph structure of the graph neural network model represent the paths that the robot can pass through;

[0114] The graph neural network model is used to learn the graph structure and generate a global planning path.

[0115] As a specific implementation method, the global planning layer adopts the GNN model, which specifically includes the following input feature construction, network structure design, and network training.

[0116] 1) Input feature construction: In terms of feature construction, the node features of the topology map include static attributes such as equipment type, installation coordinates and maintenance records, as well as dynamic risk indicators.

[0117] Topology map node features:

[0118] Static attributes: equipment type, installation coordinates, maintenance records.

[0119] Dynamic risk: current Ri value, risk gradient Historical failure frequency.

[0120] Edge features: Edge features are achieved by updating edge weights according to formula (5).

[0121]

[0122] Among them, W ij is the edge weight from node i to node j; R i is the real-time risk value of node i; R j The real-time risk value of node j; ε is the minimum connectivity weight, which is 0.01.

[0123] A graph neural network (GNN) is a deep learning model specifically designed for processing graph-structured data. Its basic process is as follows: First, each node in the graph is initialized as a feature vector. Then, during training, each node communicates and integrates information with its neighbors through a message passing mechanism. At each layer, a node's features are updated based on the features of its neighbors. The new features are a weighted sum or other fusion of the neighboring node features, combined with the node's own features, and ultimately undergo a nonlinear transformation to generate a new node representation. This process can be repeated multiple times, allowing the node representation to continuously incorporate information from more distant neighbors. Formulaically, this can be described as follows: the node's updated features are equal to the weighted sum of the features of its neighbors (each neighbor's feature multiplied by its corresponding weight), plus the node's previous features. The new features are then passed through an activation function, such as ReLU, to obtain the new node features. After multiple rounds of this message passing, the network can learn complex representations of the graph, both global and local, for tasks such as node classification, edge prediction, and graph classification.

[0124] 2) Network structure, including input layer, graph convolution layer and output layer.

[0125] Input layer:

[0126] Receive the node feature matrix X∈R of the polysilicon production park topology map N×d (N is the number of nodes, d=20 is the feature dimension), including static attributes and dynamic risk indicators of the equipment;

[0127] Edge feature matrix E∈R N×N×3 , including path length, travel difficulty and real-time weight Wij.

[0128] Graph convolution layer (3 layers in total):

[0129] For the lth layer of graph convolution (3 layers in total), the feature update of node i is The formula is as follows:

[0130]

[0131] Where σ is the activation function; N is the number of nodes in the graph; A ij is the weight of the connecting edge, when A ij = 0, the characteristics of device j do not participate in the update of device i. ij = 1, the feature of device j affects i according to the normalized weight; is the degree of node i (including self-connections), indicating the number of neighbors directly connected to node i (determined by the physical connection relationship of the device); d j represents the degree of node j; is the feature vector of node j in layer l; W (l) is the trainable weight matrix of the lth graph convolutional layer, d in is the input feature dimension (output dimension of the previous layer), d out Output feature dimension for this layer.

[0132] Output layer:

[0133] Generate node access probability distribution F∈R N×1 , represents the relative priority of each device node (N in total) that should be accessed in the next inspection cycle (1 hour), sorts the i-th access probability Fi, and generates an inspection path sequence (from high to low access). Through softmax normalization, its calculation formula is as follows (7):

[0134]

[0135] Among them, h i represents the original output feature value of node i, h j Represents the original output feature value of node j.

[0136] 3) Network Training: The network training process primarily measures the model's predictive performance by defining a loss function (such as mean squared error or cross entropy). Optimizers (such as Adam and SGD) are then used to adjust model parameters based on the gradient of the loss function to minimize the loss and improve the model's predictive accuracy. During training, the model is continuously iterated, using the backpropagation algorithm to calculate gradients and gradually optimize parameters, ultimately resulting in a model with superior performance.

[0137] The loss function consists of two parts: the supervision loss term and the graph regularization term. The first part represents minimizing the mean square error between the node access frequency predicted by the model and the optimal frequency manually annotated. The second part represents constraining the access frequency of adjacent nodes to change smoothly to avoid overfitting. The calculation formula is as follows:

[0138]

[0139] Among them, K is the number of training samples; F k is the predicted access frequency vector of the kth sample; F k *is the corresponding true label vector; λ is the regularization coefficient; L = DA, where D is the degree matrix and A is the adjacency matrix; F represents the predicted access frequency matrix; F T represents the device matrix of F; tr represents the trace of the matrix.

[0140] Adam optimizer is used, with an initial learning rate of 3×10 -4 , decaying by 10% every 20 rounds.

[0141] The Adam (Adaptive Moment Estimation) optimizer is an optimization algorithm that combines momentum and adaptive learning rates. Its detailed process is as follows: For each parameter update, it first calculates the first-order moment estimate of the gradient, which is the exponentially decaying average of the gradient. This is similar to the concept of momentum and is expressed as follows: the first moment estimate is the current gradient multiplied by a decay coefficient (such as beta1) plus the decay of the previous first moment. Next, it calculates the second-order moment estimate of the gradient, which is the exponentially decaying average of the square of the gradient. Similarly, this is multiplied by another decay coefficient (such as beta2) plus the decay of the previous second moment. To prevent division by zero or numerical instability, these two moments are bias-corrected to obtain the bias-corrected first and second moments. Finally, the parameter update is the corrected first moment divided by the square root of the corrected second moment, multiplied by the learning rate, and subtracted from the corresponding gradient correction value. This entire process enables Adam to adaptively adjust the learning rate of each parameter during training, accelerating convergence and providing good robustness, making it particularly suitable for processing large-scale and sparse data.

[0142] Step S2: According to the global planned path, the robot is controlled to perform inspections in the polysilicon production park.

[0143] Step S3: During the robot inspection process, the global planning path is locally adjusted based on the regional risk to obtain the adjusted global planning path.

[0144] Specifically, this step includes:

[0145] The specific steps include:

[0146] Dividing the global planning path into a plurality of continuous path nodes; and controlling the robot to perform inspections according to each path node of the global planning path;

[0147] Collect the current state of the robot at the current path node; the current state includes the robot state, environment state and task state;

[0148] Based on the current state, the policy network of the proximal policy optimization model is used to generate new actions for the robot; the new actions include movement actions and emergency actions;

[0149] Update the current path according to the new action, generate new path points or adjust existing path points to obtain new path points;

[0150] Calculate the reward value of the new path point;

[0151] Update the policy network of the proximal policy optimization model according to the reward value and optimize the path adjustment strategy;

[0152] According to the path adjustment strategy, the global planning path is adjusted to obtain the adjusted global planning path.

[0153] The local adjustment layer uses the Proximal Policy Optimization (PPO) model. PPO (Proximal Policy Optimization) is a deep reinforcement learning algorithm used in reinforcement learning, designed to achieve stable and efficient policy optimization when training agents. It introduces a clever objective function to constrain policy updates to a certain range, avoiding excessive policy adjustments and reducing training instability. The PPO model specifically covers core aspects such as state space, action space, real-time decision-making, as well as training and optimization. By defining a set of states and actions, it supports real-time decision-making in complex environments and improves policy performance and adaptability through continuous training and optimization.

[0154] 1) State space:

[0155] Robot status: current position (x, y, z), remaining battery power, motor temperature, communication delay;

[0156] Environmental status: current regional risk value Wi, risk spatial gradient Coordinates of adjacent high-risk areas;

[0157] Task status: distance to the next target node, proportion of inspected nodes.

[0158] 2) Action Space:

[0159] Movement: 8-direction (0°, 45°, ..., 315°) displacement, step length 0.3-1.2m;

[0160] Emergency actions: emergency braking, charging request, local scanning.

[0161] 3) Real-time decision making:

[0162] When R i When the value is greater than 0.7, the path replanning is completed within 450ms, and a smooth Bézier curve path is output. The constraints include: maximum curvature radius ≥ 1.5m; path risk integral is minimized;

[0163] During execution, the risk status is updated every 200ms. i When >0.9, emergency stop is forced.

[0164] 4) Training and optimization:

[0165] Offline training: 10,000 sets of historical emergency event samples;

[0166] Online learning: Daily incremental updates (taking less than 5 minutes), using the EWC algorithm to prevent forgetting.

[0167] The EWC (Elastic Weight Consolidation) algorithm is a method used to prevent catastrophic forgetting in neural networks during continuous learning. Its detailed process is as follows: First, before learning a new task, the trained model is used to calculate the optimal values ​​of the parameters and the importance of these parameters in the previous task. This is usually measured by calculating the Fisher information matrix of the parameters. The diagonal elements of the Fisher information matrix reflect the importance of each parameter. Then, when training a new task, a regularization term is added to limit the deviation of the parameters from the important parameter values ​​of the previous task. This regularization term is the weighted sum of the squares of the difference between the current values ​​of all parameters and the optimal values ​​saved in the previous task, where the weight of each parameter is determined by the value of the corresponding Fisher information. The specific formula is: A term is added to the error function, which is equal to the Fisher information multiplied by the sum of the squares of the parameter deviation from the previous parameter value. In this way, the model can maintain its performance on the old task while learning the new task, achieving the effect of preventing forgetting in continuous learning.

[0168] Step S4: According to the adjusted global planning path, the robot in the polysilicon production park is controlled to continue inspection.

[0169] During specific control, the AR remote collaborative interface supports managers to mark virtual no-entry areas, and the marked data is automatically updated to the environmental map; and the movement speed is dynamically adjusted according to the path risk level (0.3-0.5m / s in high-risk areas, 1.0-1.2m / s in low-risk areas).

[0170] Figure 2The automated inspection system architecture for a polysilicon production park is presented, divided into four layers. First, the data acquisition layer utilizes multispectral cameras, temperature sensors, vibration sensors, and hydrofluoric acid (HF) concentration detectors to collect real-time environmental parameters and equipment operating status information. Next, the edge computing layer, equipped with servers equipped with field-programmable gate arrays (FPGAs) and graphics processing units (GPUs), performs rapid risk assessments. Global path planning is implemented using a graph neural network (GNN), and combined with the proximal policy optimization (PPO) model from deep reinforcement learning, it dynamically adjusts inspection paths to adapt to environmental changes. The digital twin layer uses a three-dimensional factory model to render a dynamic risk map in real time, color-coding the risk level and providing pulsed warning boxes to indicate potential safety hazards. Finally, the execution control layer is implemented by a corrosion-resistant inspection robot equipped with a multi-source positioning system and an emergency control module to ensure precise positioning, obstacle avoidance, and emergency stop, ensuring a safe and efficient inspection process.

[0171] Let me illustrate this with a specific example:

[0172] Step 1: System deployment scenario

[0173] This embodiment is applied to a polysilicon production base (with an area of ​​50,000 square meters, including 12 reduction furnaces and 1 cold hydrogenation fluidized bed), and three corrosion-resistant inspection robots are deployed to cover key areas such as production workshops, pipeline corridors, and storage areas.

[0174] Step 2: Dynamic Risk Map Construction: Regional risk values ​​are calculated through multi-source data collection, including equipment temperature, vibration, harmonic distortion, particulate matter concentration, and visual defects. In the risk assessment, temperature, vibration, dust, and visual defects each have a certain weight, resulting in a real-time regional risk index of 0.82. Combined with historical fault correlation, the final risk value reaches 0.86, issuing a red alert. Kriging interpolation is used to construct a three-dimensional risk map, showing a risk gradient of 0.15 / m for the pipeline corridor and a real-time pulse warning box displayed around the reduction furnace. The specific process of dynamic risk map construction is as follows:

[0175] (1) Data collection

[0176] Equipment operating parameters:

[0177] The temperature of reduction furnace #3 suddenly increased to 865℃ (current safety threshold ≤850℃), and the vibration amplitude reached 7.2mm / s (threshold 5.0mm / s).

[0178] Current harmonic distortion rate THD = 8.5% (normal ≤ 5%).

[0179] Environmental parameters:

[0180] Particle concentration = 65 mg / m3 (Triggers a high risk threshold).

[0181] Visual data:

[0182] The multispectral camera detected 12% of defective pixels in the pipeline weld (threshold 5%).

[0183] (2) Calculation of risk value

[0184] Real-time risk value of area A (5m around reduction furnace #3):

[0185] R A =0.6×(865-850) / (900-850)+0.2×(7.2-5.0) / (10.0-5.0)+0.1×8.5 / 5+0.1×65 / 50=0.82

[0186] Among them, the weight coefficients are temperature w1 = 0.6, vibration w2 = 0.2, THD w3 = 0.1, and dust w4 = 0.1.

[0187] Historical fault correlation H A (3 failures in the past 7 days):

[0188] H A =(0.8^0×1+0.8^1×0+0.8^2×1+...) / (1+0.8+0.8^2+...)=0.42

[0189] Final risk value R A :

[0190] R A =0.82+0.1×0.42=0.86 (red alert)

[0191] (3) Generation of three-dimensional risk maps

[0192] Kriging interpolation shows the risk gradient of pipeline corridors (High-risk diffusion trend); the AR terminal displays a pulse warning box around reduction furnace #3 in real time.

[0193] Step 3: Dual-model path planning execution

[0194] (1) Global planning layer (GNN model): The global planning layer uses a graph neural network model to update the edge weights between nodes by inputting features such as the node's equipment type, coordinates, maintenance records, and risk assessment value. The edge weight calculation takes into account factors such as risk value and distance, and obtains W = 0.232. The model outputs the access priority of each node, with reduction furnace #3 having the highest priority, followed by the cold hydrogenation fluidized bed, and finally generates the inspection path for the next hour. Graph Neural Network (GNN) is a type of deep learning model specifically designed for processing graph-structured data. It learns the potential features of nodes and edges by transferring and aggregating information between nodes and their neighbors, thereby modeling complex network relationships. GNN is widely used in social network analysis, recommendation systems, path planning, equipment maintenance, and other fields, and has the ability to capture local and global structural features. In the inspection system, GNN can help evaluate the relationship and risk between each device, optimize the inspection path, and improve maintenance efficiency.

[0195] The global planning layer specifically includes:

[0196] Input features:

[0197] Node feature matrix X (dimension 20): contains equipment type, coordinates (120, 45), maintenance records (last maintenance 7 days ago), current R A =0.86.

[0198] Edge weight update:

[0199] W{A→B}=0.5×(1-0.86)+0.3×(1-0.45)+0.2×0.01=0.232

[0200] Output:

[0201] Node access priority: reduction furnace #3 (F=0.92)>cold hydrogenation fluidized bed (F=0.67).

[0202] Generate a 1-hour cycle path.

[0203] (2) Local adjustment layer (PPO model): The local adjustment layer uses the PPO model to dynamically optimize the path of the inspection robot. After detecting that the danger index RA is higher than the threshold, the robot will trigger replanning. The system defines the robot's current position, speed, and high-risk area and other status information. The candidate actions include 26 different directions, which are divided into basic directions, diagonal directions, adjacent directions, and adjacent diagonal directions. Each action is represented by a three-dimensional displacement vector and a step length. The optimal action is selected by evaluating the path risk integral and length penalty. The path is parameterized by a cubic Bézier curve. The starting point is the current position of the robot and the end point is the target position. The control point adjusts the path shape to meet the curvature constraint. The risk integral calculation uses discrete sampling points on the path to evaluate the potential danger of the path. The obstacle avoidance path and end point are finally output to ensure the safety of the path curvature. The path takes 420 milliseconds. After optimization, the risk level of the new path is reduced by 46% compared with the original path. The specific process of the local adjustment layer is as follows:

[0204] ① Emergency response:

[0205] Robot #2 detects R A =0.86>0.7, triggering replanning.

[0206] ②Input state definition:

[0207] The robot's current state S t :

[0208]

[0209] Among them, the three-dimensional position coordinate p:

[0210]

[0211] Three-dimensional velocity vector V:

[0212]

[0213] High-risk area H:

[0214] Η={p|||p-[120,45,25] T ||2≤2}

[0215] ③Action generation:

[0216] Candidate action set A (26 directions):

[0217]

[0218] Where Δp i is a candidate displacement action (three-dimensional vector); r i is the step length (meters); u iis the unit direction vector, S 2 It represents a sphere in three-dimensional space where the distance from all points to the origin is 1.

[0219] 26 different directions, which can be divided into the following categories:

[0220] Four basic directions: front (F), back (B), left (L), right (R).

[0221] Four diagonal directions: front left (FL), front right (FR), back left (BL), and back right (BR).

[0222] Eight adjacent directions: adjacent directions for each cardinal direction and diagonal direction, for example, front left front (FLL), front left (FL), front left back (FLB), etc.

[0223] Eight diagonally adjacent directions: each diagonal direction and its adjacent direction, for example, front left front (FLL), front left diagonal (FL), front right diagonal (FR), etc.

[0224] Optimal action selection:

[0225]

[0226] Where Δp* represents the optimal action change, ∫R(p)dp is the risk integral along the path; ||Δp|| 2 is the path length penalty term; λ is the weight coefficient, and A represents the candidate action set.

[0227] ④Bézier path parameterization

[0228] Control Points:

[0229]

[0230] Cubic Bézier curve:

[0231] B(t)=(1-t) 3 P0+3t(1-t) 2 P1+3t 2 (1-t)P2+t 3 P3,t∈[0,1]

[0232] Among them, P0 is the starting point (the current position of the robot [118, 43, 23]); P3 is the end point (such as [117.43, 42.43, 23]); P1 and P2 are control points (control the shape of the curve and must meet the curvature constraint); t is a parameterized variable, p t is the position at time t.

[0233] ⑤ Risk score calculation

[0234]

[0235] Wherein, N=50 is the number of discrete sampling points of the path; is the risk value of the kth sampling point; ||B(k / N)-B((k-1) / N)||2 is the Euclidean distance between adjacent points.

[0236] ⑥ Output results

[0237] Avoidance path: B(t), t∈[0,1].

[0238] End point: [117.43, 42.43, 23]T.

[0239] Curvature radius: ρ ≥ 1.8m (the minimum curvature radius ρ is 1.8m).

[0240] Time taken: 420ms.

[0241] Original path risk: 1.24; new path risk: 0.67 (reduced by 46%).

[0242] Step 4: Emergency Operations for the Anti-Corrosion Robot: The robot automatically stopped when the RA suddenly rose to 0.91, 2.3 meters from the leak point. Multispectral scanning confirmed that the silane gas concentration reached 9 ppm, exceeding the threshold. Simultaneously, a scraper removed silicon powder from the lens to ensure clear imaging in preparation for an unexpected leak. The emergency operations for the anti-corrosion robot included:

[0243] Emergency braking: Robot #1 automatically stopped when RA suddenly increased to 0.91, 2.3 m away from the leak point.

[0244] Multispectral scanning: 17-band imaging (400-1700nm) was performed on the leak point, confirming the silane gas concentration was 9ppm (threshold 5ppm).

[0245] Cleaning mechanism: The scraper removes the silicon powder (thickness 0.3mm) attached to the lens to ensure image clarity.

[0246] Step 5: Typical exception handling process

[0247] Scenario: HF leakage caused by seal failure of cold hydrogenation fluidized bed (R B=0.88). HF leaks refer to the leakage of hydrogen fluoride (HF) during production or use. HF is highly corrosive and toxic, posing a significant threat to both humans and the environment. Leaks can be caused by ruptured storage containers, operational errors, or equipment failures. Once a leak occurs, it can cause severe chemical burns, respiratory irritation, or poisoning, and can even threaten life. Therefore, industrial sites typically implement strict protective measures and develop emergency response plans to address HF leaks, ensuring both personnel safety and environmental protection.

[0248] Risk map update: HF sensor detected 6.8ppm, dynamic weighting adjusted to w HF =0.4.

[0249] Path replanning:

[0250] The PPO model generates a detour path to avoid areas with concentrations > 5 ppm (path length increases by 2.1 m and time increases by 8 s).

[0251] Co-processing:

[0252] The AR interface pushes the three-dimensional coordinates of the leak point to the maintenance personnel's helmet display.

[0253] Robot #3 continuously monitors the downwind concentration and updates the risk value every 10 seconds.

[0254] The inspection process in this embodiment can combine multi-source sensors to collect equipment and environmental data in real time and establish a dynamic risk map. Using GNN and Kriging interpolation, regional risks are assessed, three-dimensional risk visualizations are generated, and high-risk areas are identified. A global inspection path is developed through a graph neural network, and local path optimization is performed in conjunction with the PPO model to achieve path clarity, dynamic adjustment, and risk control. In emergency situations, the robot automatically shuts down, brakes, performs visual inspection, and cleans the lens to ensure safety. This allows for rapid path replanning, avoids dangerous areas, and provides real-time monitoring and early warning, thereby improving production safety and efficiency.

[0255] Example 2:

[0256] like Figure 3 As shown, this embodiment provides a robot inspection device for a polysilicon production park, the device comprising:

[0257] The planning unit 10 is used to plan the robot inspection path according to the dynamic risk map to obtain a global planning path;

[0258] The first control unit 20 is connected to the planning unit 10 and is used to control the robot to perform inspections in the polysilicon production park according to the global planning path;

[0259] The adjustment unit 30 is connected to the first control unit 20 and is used to make local adjustments to the global planning path based on regional risks during the robot inspection process to obtain an adjusted global planning path;

[0260] The second control unit 40 is connected to the adjustment unit 30 and is used to control the robot in the polysilicon production park to continue inspection according to the adjusted global planning path.

[0261] This embodiment provides a robotic inspection device for a polysilicon production park. The device consists of multiple control units, including a planning unit 10 for developing a global inspection path based on a dynamic risk map; a first control unit 20, connected to the planning unit, for guiding the robot during inspections according to the planned path; an adjustment unit 30, connected to the first control unit, for making local adjustments to the global path based on regional risk conditions in real time during the inspection process, thereby generating a more reasonable inspection route; and a second control unit 40, connected to the adjustment unit, for controlling the robot to continuously perform inspection tasks according to the adjusted path. This system combines global planning with local adjustments to achieve dynamic, efficient, and accurate inspection path management, enabling timely response to unexpected risks and ensuring production safety and inspection efficiency.

[0262] The device in this embodiment can execute the method in embodiment 1.

[0263] Example 3:

[0264] like Figure 4 As shown, this embodiment provides a robot inspection system for a polysilicon production park, the system comprising:

[0265] Inspection robot 200, used to inspect polysilicon production parks;

[0266] The inspection server 100 is in communication with the inspection robot 200;

[0267] The inspection server 100 includes a processor configured to execute the robot inspection method for a polysilicon production park according to the first embodiment.

[0268] As a specific implementation, the number of the inspection robots 200 is N, where N is a natural number greater than or equal to 1;

[0269] Among them, the body of the inspection robot 200 is covered with a polytetrafluoroethylene coating and is equipped with a multispectral camera and a hydrogen fluoride concentration sensor.

[0270] Specifically, the inspection robot 200 is a corrosion-resistant inspection robot. Its body is covered with a polytetrafluoroethylene coating with a thickness of not less than 2 mm, which effectively enhances its corrosion resistance and adapts to multi-corrosive environments. At the same time, the robot is equipped with a multi-spectral camera (its working band is 400-1700nm), which can achieve multi-dimensional environmental perception and monitoring; and is equipped with an HF concentration sensor with a measurement range of 0 to 20ppm, which is used to detect the concentration of hydrofluoric acid in the gas to ensure production safety. In addition, the corrosion-resistant inspection robot also includes a hub motor equipped with a current harmonic detection module with a sampling frequency of not less than 1kHz, which is used to monitor the current changes of the drive system to ensure the stability and safety of operation. In order to keep the lens clean, the robot is also equipped with a front scraper-type cleaning mechanism, which is specifically used to remove silicon powder and other dirt attached to the lens surface to ensure the normal operation of the sensor and camera equipment, thereby achieving efficient and stable inspection tasks.

[0271] It will be understood that the above embodiments are merely exemplary embodiments for illustrating the principles of the present invention, and the present invention is not limited thereto. Those skilled in the art will appreciate that various modifications and improvements can be made without departing from the spirit and substance of the present invention, and such modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A robot inspection method for a polysilicon production park, characterized in that: The method comprises the following steps: Plan the robot inspection path based on the dynamic risk map to obtain the global planning path; According to the global planned path, the robot is controlled to perform inspections in the polysilicon production park; During the robot inspection process, the global planning path is locally adjusted based on regional risks to obtain an adjusted global planning path; According to the adjusted global planning path, the robot in the polysilicon production park is controlled to continue inspection.

2. The robot inspection method for a polysilicon production park according to claim 1, characterized in that: Before planning the robot inspection path according to the dynamic risk map to obtain the global planning path, the method further includes: Based on the multi-dimensional environmental data of the polysilicon production park, a dynamic risk map of the polysilicon production park is constructed.

3. The robot inspection method for a polysilicon production park according to claim 2, characterized in that: The multi-dimensional environmental data of the polysilicon production park includes equipment operating parameters, environmental parameters, and visual data of the polysilicon production park; The step of constructing a dynamic risk map of a polysilicon production park based on the multi-dimensional environmental data specifically includes the following steps: Step A1: Divide the polysilicon production park into multiple sub-areas; and normalize the multi-dimensional environmental data of each sub-area to obtain a normalized measurement value of the environmental data of each sub-area; Step A2: Based on the historical fault correlation and the weights corresponding to the normalized measurement value and the multi-dimensional environmental data, a discontinuity risk value of each area is obtained; Step A3: generating a risk distribution surface according to the discontinuous risk value; Step A4: superimposing the risk distribution surface with the polysilicon production park building information model to form a three-dimensional dynamic risk map, thereby obtaining a dynamic risk map of the polysilicon production park.

4. The robot inspection method for a polysilicon production park according to claim 3, characterized in that: The step A3 specifically includes the following steps: Based on the discontinuous risk values, the unmonitored risk values ​​are estimated using the Kriging interpolation method; Summarizing the discontinuous risk values ​​and the unmonitored risk values ​​to obtain a risk value grid; According to the preset resolution, the risk value grid is refined using the inverse distance weighted method to obtain the risk distribution surface.

5. The robot inspection method for a polysilicon production park according to claim 3, characterized in that: After step A4, the method further includes: Step A5: Visualize the risk level in the three-dimensional dynamic risk map through color gradient; If the risk level is greater than or equal to the first risk threshold, the risk level is set as a high-risk area and marked with a first color; if the risk level is greater than or equal to the second risk threshold and less than the first risk threshold, the risk level is set as a medium-risk area and marked with a second color; if the risk level is less than the second risk threshold, the risk level is set as a low-risk area and marked with a third color; A pulse warning box is generated in the high-risk area in the three-dimensional dynamic risk map, and a real-time data floating window is displayed on the AR terminal.

6. The robot inspection method for a polysilicon production park according to claim 1, characterized in that: Planning the robot inspection path according to the dynamic risk map to obtain a global planning path specifically includes the following steps: Constructing a graph structure of a graph neural network model based on the dynamic risk map; The graph nodes in the graph structure of the graph neural network model represent the center point of a key device or a high-risk area, and the edges between the graph nodes in the graph structure of the graph neural network model represent paths that the robot can pass through; The graph structure is learned using a graph neural network model to generate a global planning path.

7. The robot inspection method for a polysilicon production park according to any one of claims 1 to 6, characterized in that: During the robot inspection process, locally adjusting the global planning path based on regional risks to obtain an adjusted global planning path specifically includes the following steps: Dividing the global planning path into a plurality of continuous path nodes; and controlling the robot to perform inspections according to each path node of the global planning path; Collecting the current state of the robot at the current path node; wherein the current state includes the robot state, the environment state and the task state; According to the current state, a policy network of the proximal policy optimization model is used to generate a new action of the robot; the new action includes a moving action and an emergency action; Update the current path according to the new action, generate new path points or adjust existing path points to obtain new path points; Calculate the reward value of the new path point; Update the policy network of the proximal policy optimization model according to the reward value and optimize the path adjustment strategy; According to the path adjustment strategy, the global planning path is adjusted to obtain an adjusted global planning path.

8. A robot inspection device for a polysilicon production park, characterized in that: include: The planning unit is used to plan the robot inspection path according to the dynamic risk map to obtain the global planning path; a first control unit, connected to the planning unit, for controlling the robot to perform inspections in the polysilicon production park according to the global planning path; an adjustment unit connected to the first control unit, configured to locally adjust the global planning path based on regional risks during the robot inspection process to obtain an adjusted global planning path; The second control unit is connected to the adjustment unit and is used to control the robot in the polysilicon production park to continue inspection according to the adjusted global planning path.

9. A robot inspection system for a polysilicon production park, characterized in that: The system comprises: Inspection robots, used to inspect polysilicon production parks; An inspection server, communicatively connected to the inspection robot; The inspection server includes a processor configured to execute the robot inspection method for a polysilicon production park according to any one of claims 1 to 7.

10. The robot inspection system for a polysilicon production park according to claim 9, characterized in that: The number of the inspection robots is N, where N is a natural number greater than or equal to 1; The body of the inspection robot is covered with a polytetrafluoroethylene coating and is equipped with a multispectral camera and a hydrogen fluoride concentration sensor.

Images