Submerged arc furnace intelligent inspection system based on cloud side end cooperation

Abstract

A submerged arc furnace intelligent inspection system based on cloud side-end cooperation belongs to the technical field of submerged arc furnace intelligent robot inspection, and comprises an end side device, an edge side platform, a cloud platform and a cross-layer cooperative processing module, the end side device is in communication connection with the edge side platform, and the edge side platform is in communication connection with the cloud platform; and the cross-layer cooperative processing module is used for intelligently deploying global resources of the end-side equipment, the edge-side platform and the cloud platform, predicting a future load trend by adopting a dynamic sparse LSTM model algorithm, inputting a prediction result into a security constraint PPO reinforcement learning algorithm to adaptively adjust a task and a computing power allocation proportion, and carrying out dynamic resource allocation according to an output result. According to the invention, the cross-layer cooperative processing module interacts and communicates with the cloud side, the edge side and the end side, so that a dynamic allocation strategy of resources is realized, and the problems of low resource utilization rate and poor real-time performance caused by a rigid resource allocation strategy during routing inspection are avoided.

Description

Technical Field

[0001] The present invention relates to an intelligent inspection system for an ore-bearing furnace based on cloud-edge-end collaboration, belonging to the technical field of intelligent robot inspection of ore-bearing furnaces. Background Art

[0002] As an important equipment in the metallurgical industry, the operating environment of the submerged arc furnace is extremely harsh, with multiple dangerous factors such as high temperature, high pressure, dust, noise, and large network fluctuations. The traditional manual inspection method has great safety risks and is difficult to achieve all-weather, high-frequency monitoring. In recent years, intelligent inspection robots have gradually become popular, but when applied to extreme environments such as submerged arc furnaces, there are still some problems: First, the traditional submerged arc furnace inspection system adopts a fixed resource allocation strategy. The computing resources of the edge platform and the end-side equipment (i.e., the inspection robot) are statically divided and cannot be dynamically adjusted according to the real-time load. When multiple acquisition devices collect images at the same time, the edge server lacks computing resources due to the large amount of data transmission, resulting in image processing delays, while the cloud server has idle computing power that is not utilized. In addition, when sudden abnormal events need to be handled as a priority, the traditional system cannot respond quickly due to the rigid task scheduling, resulting in a backlog of critical tasks and response time exceeding the safety threshold. This problem of low resource utilization not only increases the risk of equipment overload, but also restricts the real-time and scalability of the inspection system. Second, the existing inspection system The system often concentrates computing tasks on a single computing node. In the harsh environment of submerged arc furnaces, it is difficult to make adaptive adjustments when network fluctuations or single-point failures occur, and operational stability cannot be guaranteed. At the same time, data transmission and model updates between different layers also face security and timeliness issues, which restrict the practical application of intelligent inspection systems. Third, submerged arc furnace inspections involve multimodal data such as infrared thermal imaging, visible light images, vibration signals, gas composition, and correlation analysis of process parameters such as current, voltage, and furnace pressure. However, the inspection system usually only focuses on a single data item for alarm. When a certain indicator such as pressure or temperature exceeds the threshold, the system will issue an alarm. This alarm rule can only identify known fault modes, not potential faults or unknown fault modes, and is easily affected by environmental factors, resulting in a high false alarm rate. Data formats are not unified among equipment manufacturers and are stored in independent databases or local files, forming "data islands" that make it difficult to achieve spatiotemporal alignment analysis. Therefore, due to the lack of data integration capabilities, traditional submerged arc furnace intelligent inspection systems cannot quickly locate the root cause of the fault, delaying the opportunity for emergency repairs. Summary of the Invention

[0003] In order to solve the technical problems that the resource allocation strategy of the existing intelligent inspection system of electric arc furnace is rigid, resulting in insufficient cloud-edge detection utilization and poor real-time performance, poor reliability of the intelligent inspection system in harsh environments, and insufficient multi-source data fusion resulting in poor fault identification accuracy, the present invention proposes an intelligent inspection system for electric arc furnace based on cloud-edge-end collaboration, which includes an end-side device, an edge-side platform, a cloud platform and a cross-layer collaborative processing module, the end-side device is communicatively connected to the edge-side platform, and the edge-side platform is communicatively connected to the cloud platform; the cross-layer collaborative processing module is used for global intelligent allocation of resources of the end-side device, the edge-side platform and the cloud platform, and adopts a dynamic sparse LSTM model algorithm to predict future load trends, inputs the prediction results into the safety-constrained PPO reinforcement learning algorithm to adaptively adjust the task and computing power allocation ratio, and performs dynamic resource allocation according to the output results;

[0004] The end-side device is used to complete the data collection task of target detection, obtain data images, trigger the preset hierarchical control strategy through autonomous decision-making to determine the detection priority, run the LSTM predictor in minimalist mode, pre-process the collected data, and execute the detection data processing task based on the dynamic resource allocation results of the cross-layer collaborative processing module;

[0005] The edge-side platform is used to deploy and adjust inspection strategies and detailed tasks for end-side devices, receive pre-processed data sent by end-side devices and data from connected multi-port acquisition devices, run a lightweight LSTM predictor, and perform data collation and analysis tasks based on the dynamic resource allocation results of the cross-layer collaborative processing module;

[0006] The cloud platform is used to receive and organize data uploaded by terminal devices and edge-side platforms, conduct data mining to identify potential risks, perform deep learning analysis based on the dynamic resource allocation results of the cross-layer collaborative processing module, and return the processing results.

[0007] The cross-layer collaborative processing module includes a dynamic sparse LSTM load prediction unit, which is used to predict the future load trends of the end-side devices, edge-side platforms, and cloud platforms, and adaptively adjust the consumed computing power in combination with the gating mechanism. Specifically:

[0008] Preprocess the input data. The input time series data includes:

[0009] Resource utilization, including CPU, GPU, and memory: ;

[0010] Network bandwidth: ;

[0011] Task queue length: ;

[0012] Normalization is performed on the input time series data to form an input vector:

[0013] ;

[0014] in, They are the average utilization rates of CPU, GPU and memory resources respectively. They are CPU, GPU, and memory standard deviations, is the maximum queue length;

[0015] Define device type-dependent activation ratios based on gating mechanisms Among them, cloud refers to the cloud, edge refers to the edge side, and device refers to the device side:

[0016] ;

[0017] The hidden state update formula of each LSTM unit is:

[0018]

[0019] in, They are forget gate, input gate, and output gate respectively. is the candidate cell state, are the current cell state and the cell state at the previous moment, respectively. is the weight matrix corresponding to each gate and candidate state, is the previous hidden state and the current hidden state, is the current input, is the bias term for each gate and candidate state, is the Sigmoid function, is the hyperbolic tangent function, ☉ is the Hadamard product;

[0020] Perform dynamic sparsification:

[0021] ;

[0022] in, is the LSTM hidden state after sparse gating, To retain the neuron outputs of the first k maximum values, the rest are set to zero. is the total number of neurons in the LSTM hidden layer;

[0023] Use sparse gating to extract features, independently predict multiple quantiles, and generate asymmetric confidence intervals;

[0024] Predict the load value in the next T minutes:

[0025] ;

[0026] in, is the load value predicted for the next T minutes, is the hidden state update formula of the LSTM unit, is the input value before time t, are the parameters of the LSTM model, including the weight matrix and bias term;

[0027] Confidence intervals are calculated using quantile regression:

[0028] ;

[0029] in, is the predicted value at the quantile q in the future T time step, is the quantile regression function, which outputs the predicted value corresponding to the quantile q;

[0030] right The predicted values are combined and output as a prediction matrix with confidence intervals, which optimizes LSTM in a dynamic sparse manner:

[0031] .

[0032] Furthermore, the cross-layer collaborative processing module also includes a safety constraint PPO decision unit, which performs resource allocation scheduling according to the predicted load value output by the dynamic sparse LSTM load prediction unit, specifically:

[0033] For the state space S of the algorithm t , that is, the input of the algorithm is defined as:

[0034] ;

[0035] in, is the predicted value obtained by the LSTM algorithm based on dynamic sparseness, The real-time resource status indicates the real-time resource availability status of the current cloud / edge / device devices.

[0036] in, The task characteristics include the following three key attributes of the task:

[0037] Computational intensity c: The amount of computing resources required by the task per unit of data, expressed in FLOPS / Byte. Computational intensity c is obtained through offline calibration, which first benchmarks various tasks and calculates their computational effort. The amount of input data , directly assign the pre-calibrated c value according to the task type label and normalize the c value:

[0038] ;

[0039] in, is the amount of statistical calculation, is the amount of input data.

[0040] The delay sensitivity l, that is, the sensitivity of the task to the response time, reflects the urgency. Normalization is performed and l is scaled to the interval [0,1]:

[0041] ;

[0042] ;

[0043] in, is the maximum delay time allowed for the task, l max is the maximum value of delay sensitivity, l min is the minimum value of delay sensitivity, l norm is the normalized feature parameter;

[0044] The data volume d is the total amount of data that needs to be transmitted or processed by the task. The value is normalized according to the maximum transmission capacity of the device. To zoom:

[0045] ;

[0046] According to the above three parameters describing the task characteristics, the final task feature vector is integrated:

[0047] ;

[0048] Furthermore, the action space A of the algorithm, that is, the output of the algorithm, is defined as follows: the action vector α t :

[0049] ;

[0050] in, Task assignment encoding uses one-hot encoding with a length of 3, corresponding to three types of devices: cloud, edge, and device. The resource allocation ratio indicates the proportion of resources allocated to cloud, edge, and device devices, and the total ratio must be ≤ 1.

[0051] Furthermore, safety constraints are modeled, including hard and soft constraints;

[0052] The hard constraints are mandatory constraints:

[0053] ;

[0054] in, For the current device The resource status under the action;

[0055] The soft constraints are non-mandatory constraints and can be violated probabilistically:

[0056] ;

[0057] in, is the probability that the task completion time is greater than the deadline;

[0058] Furthermore, the specific default probability is estimated based on Monte Carlo sampling:

[0059] ;

[0060] in, To estimate the probability, is the number of sampling times, is an indicator function, which takes 1 when the condition is met and 0 otherwise;

[0061] Furthermore, the constraints are incorporated into the objective function of reinforcement learning. Based on the Lagrange dual method, the policy network is updated and the objective is optimized in combination with the constraints:

[0062] ;

[0063] Among them, E t is the expected operator, which represents the average performance of the objective function under the policy action distribution;

[0064] is the action probability of the current policy network parameter θ;

[0065] is the action probability ratio of the new and old strategies, which measures the change in action preference of the current strategy relative to the old strategy;

[0066] is the advantage function, which evaluates the quality of the action relative to the average level;

[0067] The default amounts for hard and soft constraints are:

[0068] ;

[0069] It is an adaptive Lagrange multiplier that dynamically adjusts the constraint penalty strength according to the degree of constraint violation:

[0070] ;

[0071] If the cloud resource utilization rate exceeds the standard ( ),but , Increase to strengthen the punishment; among them, is the learning rate, which controls the adjustment amplitude of the multiplier;

[0072] According to the updated policy network formula, the degree of violation of the safety constraint is determined based on the factual information data collected by the system, and the adaptive Lagrange multiplier is adjusted to obtain a new policy network, and the new policy network is updated to the system;

[0073] The policy network is, in a given state Select Action The rules are expressed as probability distributions. In deep reinforcement learning, the strategy is represented by a neural network with parameters θ. ; The current system status is When the system selects the one with the highest probability , through the action vector α t Get the task assignment code and resource allocation ratio , that is, the PPO algorithm based on safety constraints realizes the dynamic allocation of tasks and resources.

[0074] The terminal device includes the following modules:

[0075] The data acquisition module is used by the inspection robot to perform data acquisition tasks and obtain real-time environmental information and data images of the submerged arc furnace and its surroundings;

[0076] The autonomous decision-making module is used to control the inspection robot to complete the scheduled data collection tasks. Based on behavior tree task deployment, it automatically modifies and adjusts the inspection route based on image detection results. It communicates with the edge platform in real time to obtain overall information about the submerged arc furnace. It autonomously completes data collection and detection tasks under network anomalies and uploads detection results back to the edge platform after the network is restored. It autonomously switches to disaster recovery mode under extreme working conditions. Based on historical data from dynamic sparse LSTM load prediction, it triggers a preset hierarchical control strategy to determine detection priority, autonomously reduces the inspection robot's energy consumption, and suspends non-urgent tasks when the load exceeds the safety threshold, ensuring the core control of the robotic arm and the collection of key data.

[0077] The data processing module performs local computational analysis and preprocessing on the end-side tasks of the data acquisition module, the dynamic resource allocation strategy results of the cross-layer collaborative processing module, and the detection tasks of the autonomous decision-making module.

[0078] The end-side collaboration module is used to schedule the end-side execution of the dynamic resource allocation strategy results of the cross-layer collaborative processing module, run the minimalist LSTM predictor of the cross-layer collaborative processing module, and send the processing results of the data processing module to the autonomous decision-making module for decision adjustment or upload to the edge platform.

[0079] The edge-side platform includes the following modules:

[0080] The device access module is used to receive pre-processed data sent by the end-side device and receive data from the connected multi-port acquisition device to achieve unified fusion of multi-modal data;

[0081] The task management module is used to formulate inspection strategies for end-side devices, deploy detailed tasks, and adjust strategies based on demand;

[0082] The data storage module is used to store data collected by multi-port devices. It adopts a storage mode that combines relational database and time series database to store raw data and intermediate detection results to support real-time analysis.

[0083] The intelligent detection module performs intelligent detection and accelerated reasoning based on deep learning on data from end-side devices and connected multi-port acquisition devices;

[0084] The data analysis module is used for in-depth analysis of multi-source fusion data. It uses efficient deep learning algorithms to analyze multi-source data streams in real time and quickly locate operational anomalies and fault points.

[0085] The edge collaboration module is used to execute and schedule the edge-side platform based on the dynamic resource allocation strategy results of the cross-layer collaborative processing module. It rationally allocates and schedules detection tasks according to the resource load, runs the lightweight LSTM predictor of the cross-layer collaborative processing module, and achieves resource load prediction in seconds. When the edge-side platform load is too high, the detection task is promptly dispatched to the cloud platform for calculation according to the scheduling of the cross-layer collaborative processing module.

[0086] The cloud platform includes the following modules:

[0087] The data warehouse module is used to organize and store test result data uploaded by end-side devices and edge-side platforms, store uploaded historical resource load data, and store structured data assets for a long time through a partition index mechanism;

[0088] The data mining module is used to extract key historical data from the data warehouse module and deeply explore potential risks in the data through machine learning algorithms, statistical analysis and pattern recognition technology;

[0089] The intelligent computing core module is used to handle computationally intensive inspection tasks, collect historical inspection result data from device-side devices and edge-side platforms, and regularly train and optimize inspection models.

[0090] Remote alarm module, used to set alarm mechanisms, remotely push alarm information to terminals, and conduct in-depth analysis of potential dangers;

[0091] The cloud-side collaboration module is used to deploy and execute cloud-side platform policies for the dynamic resource allocation strategy results of cross-layer collaborative processing modules. It receives data processing requests from end-side devices and edge-side platforms and returns the processing results, optimizes the global collaborative scheduling strategy, and updates the LSTM model.

[0092] The beneficial effects of the present invention are:

[0093] The interaction and communication between the terminal equipment, edge platform, cloud platform and cross-layer collaborative processing module of the electric arc furnace inspection system are utilized, and the dynamic sparse LSTM load of the cross-layer collaborative processing module is used to predict the future load trend of the terminal equipment, edge platform and cloud platform. The gating mechanism is combined to adaptively adjust the computing power allocation ratio, so that the cloud platform is fully activated, the edge platform is partially activated, and the terminal equipment is in a minimalist mode to achieve dynamic sparsity; the historical load, network status and other time series data are used to predict the confidence interval of the resource load in the future period, while ensuring accuracy and reducing the computing overhead. The output results enter the decision layer, and the safety constraint PPO task dynamic resource allocation reinforcement learning algorithm is used to integrate the load prediction results, real-time task characteristics and equipment status into a multi-dimensional state space. The task allocation and resource quota decisions are generated through the deep policy network. At the same time, the hard constraint of the equipment load threshold and the soft constraint of the task default probability are introduced to ensure the stability of the system, and a dynamic reward function is designed. Balance real-time performance, energy efficiency, and resource utilization; the edge-side platform deploys and adjusts inspection strategies and detailed tasks for end-side equipment, performs lightweight prediction and rapid pre-scheduling, and can respond to sudden tasks; the cloud platform performs global strategy optimization and model iteration, and implements dynamic tuning through parameter distribution; the end-side equipment is used to complete data collection tasks for target detection and obtain data images. In the event of network anomalies, the end-side operates based on minimalist LSTM autonomous degradation, triggers preset hierarchical control strategies through autonomous decision-making to determine detection priorities, and ensures the continuous execution of critical tasks; the system achieves a 65% increase in load balancing under complex working conditions, and a significant effect of reducing the response delay of high-priority tasks to 200ms, becoming the core technical support for the highly reliable and adaptive operation of intelligent inspection of electric arc furnaces, realizing a dynamic resource allocation strategy, and avoiding the problems of insufficient computing resources, image data processing delays, low resource utilization, and poor real-time performance caused by rigid resource allocation strategies during inspections.

[0094] By utilizing the autonomous decision-making module of the end-side device, the inspection robot can make autonomous behavioral decisions based on behavior tree task deployment. In the event of network interruption, it no longer relies on instructions sent by the edge-side platform and the cloud platform, but can still independently complete data collection and detection tasks, and upload the detection results back to the edge-side platform after the network is restored. This mechanism enables the end-side device to still ensure the completion rate of key tasks in a network-disconnected environment, and its reliability is greatly improved compared to traditional solutions. It can autonomously switch to disaster recovery mode under extreme working conditions, trigger the preset hierarchical control strategy to determine the detection priority based on the historical data of the dynamic sparse LSTM load forecast, and autonomously reduce the energy consumption of the inspection robot. When the load is low, the data cache cycle is extended and basic inspection functions are maintained. When the load climbs to a medium level, computing degradation is initiated. When the load exceeds the safety threshold, the core control of the robotic arm and key data collection are prioritized, and non-emergency tasks are suspended to minimize the impact of accidents such as abnormal working conditions and network failures.

[0095] The edge-side platform device access module is used to uniformly access the data of multi-port acquisition devices. The end-side equipment cannot guarantee continuous inspection. Multiple acquisition devices are deployed around the submersible arc furnace and connected to the edge side. Data from different sensors and different dimensions are effectively integrated, breaking down data silos and achieving seamless docking and integration of multimodal data. It can achieve all-weather and full-coverage detection of the submersible arc furnace, and realize comprehensive analysis of multi-source data. Using various big data analysis methods such as time series analysis, association rule mining, and cluster analysis, we deeply explore the trend characteristics of the equipment operation process and the regular patterns of fault occurrence, realize efficient, reliable and flexible inspection and monitoring, effectively improve the operating efficiency and safety of the submersible arc furnace, and provide strong technical support for the intelligent management of the operation and maintenance of the submersible arc furnace. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 It is a structural diagram of the intelligent inspection system for submerged arc furnaces of the present invention.

[0097] Figure 2 It is a schematic diagram of the cross-layer collaborative processing module.

[0098] Figure 3 yes Figure 2 Schematic diagram of the algorithm flow.

[0099] Figure 4 yes Figure 1 Schematic diagram of the structure of mid-end equipment.

[0100] Figure 5 yes Figure 1 Schematic diagram of the structure of the mid-edge side platform.

[0101] Figure 6 yes Figure 1 Schematic diagram of the structure of the cloud platform. DETAILED DESCRIPTION

[0102] In order to make the objectives, technical solutions and beneficial effects of the present invention more clear, embodiments of the present invention are described in detail below with reference to the accompanying drawings. The embodiments are exemplary and intended to be used to explain the present invention, but should not be understood as limiting the present invention.

[0103] See also Figure 1 , an intelligent inspection system for electric arc furnaces based on cloud-edge collaboration, which includes end-side devices, edge-side platforms, cloud platforms and cross-layer collaborative processing modules. The end-side devices are communicated with the edge-side platform, and the edge-side platform is communicated with the cloud platform; the cross-layer collaborative processing module is used for global resource intelligent allocation of end-side devices, edge-side platforms and cloud platforms, adopts dynamic sparse LSTM model algorithm to predict future load trends, combines the gating mechanism to adaptively adjust the computing power allocation ratio, and performs dynamic resource allocation based on the output results using the security constraint PPO reinforcement learning algorithm; the end-side device is used to complete the data acquisition task of target detection, obtain data images, trigger the preset hierarchical control strategy through autonomous decision-making to judge the detection priority, and run extremely The simplified LSTM predictor performs preprocessing operations on the collected data and executes the detection data processing task according to the dynamic resource allocation results of the cross-layer collaborative processing module; the edge-side platform is used to deploy and adjust the inspection strategy and detailed tasks of the end-side equipment, receive the preprocessed data sent by the end-side equipment and receive the data from the connected multi-port acquisition equipment, run the lightweight LSTM predictor, and execute the data sorting and analysis tasks according to the dynamic resource allocation results of the cross-layer collaborative processing module; the cloud platform is used to receive and sort the data uploaded by the end-side equipment and the edge-side platform, perform data mining to identify potential dangers, perform deep learning analysis according to the dynamic resource allocation results of the cross-layer collaborative processing module, and return the processing results.

[0104] Furthermore, in order to realize the dynamic allocation of computing resources, a real-time monitoring system is established to monitor the operating status, network status, and computing load on the edge and cloud sides of the electric arc furnace in real time; based on the monitoring results, the system uses a dynamic resource allocation algorithm to dynamically allocate computing tasks between the cloud, edge, and end; innovatively integrates dynamic sparse LSTM load prediction and security-constrained PPO task scheduling technology to build a cloud-edge-end collaborative intelligent resource allocation system, the core of which is to achieve dynamic resource optimization through the "prediction-decision-making" closed loop.

[0105] Specifically, the prediction layer is based on a dynamic sparse LSTM model. To address the differences in computing power between heterogeneous devices in the cloud, edge, and end, it dynamically adjusts the neuron activation ratio through gating, with full activation on the cloud, partial activation on the edge, and a minimalist mode on the end side to achieve dynamic sparsity. It uses time series data such as historical load and network status to predict the confidence interval of resource load in the next 5 minutes, reducing computing overhead while ensuring accuracy. The decision layer adopts a security-constrained PPO reinforcement learning algorithm to integrate load prediction results, real-time task characteristics (delay sensitivity, data scale), and device status into a multi-dimensional state space. Task allocation and resource quota decisions are generated through a deep policy network. At the same time, hard constraints (device load threshold) and soft constraints (task default probability) are introduced to ensure system stability, and a dynamic reward function is designed to balance real-time performance, energy efficiency, and resource utilization. The algorithm achieves efficient operation through cross-layer collaborative processing of the cloud, edge, and end.

[0106] Further, see Figure 2 、 Figure 3 The cross-layer collaborative processing module includes a dynamic sparse LSTM load prediction unit, which is used to predict the future load trends of the terminal side equipment, edge side platform and cloud platform, and adaptively adjust the computing power allocation ratio in combination with the gating mechanism;

[0107] Specifically:

[0108] Preprocess the input data. The input time series data includes:

[0109] Resource utilization, including CPU, GPU, and memory: ;

[0110] Network bandwidth: ;

[0111] Task queue length: ;

[0112] Normalization is performed on the input time series data to form an input vector:

[0113] ;

[0114] in, They are the average utilization rates of CPU, GPU and memory resources respectively. They are CPU, GPU, and memory standard deviations, is the maximum queue length;

[0115] Define device type-dependent activation ratios based on gating mechanisms Based on actual usage and results, the cloud side is set to 95%, the edge side is 70%, and the device side is 30%.

[0116] ;

[0117] The hidden state update formula of each LSTM unit is:

[0118] ;

[0119] in, They are forget gate, input gate, and output gate respectively. is the candidate cell state, are the current cell state and the cell state at the previous moment, respectively. is the weight matrix corresponding to each gate and candidate state, is the previous hidden state and the current hidden state, is the current input, is the bias term for each gate and candidate state, is the Sigmoid function, is the hyperbolic tangent function, ☉ is the Hadamard product;

[0120] Perform dynamic sparsification and impose activation rate constraints on the hidden layer output:

[0121] ;

[0122] in, is the LSTM hidden state after sparse gating, To retain the neuron outputs of the first k maximum values, the rest are set to zero. is the total number of neurons in the LSTM hidden layer;

[0123] Using sparse gating to extract features, multiple quantiles are independently predicted, and asymmetric confidence intervals are generated. This system can both predict the most likely value (median, quantile 0.5) and quantify risk (confidence interval, quantile 0.05-0.95), making it ideal for resource scheduling scenarios requiring safety redundancy.

[0124] Predict the load value for the next T=5 minutes:

[0125] ;

[0126] in, is the load value predicted for the next T minutes, is the hidden state update formula of the LSTM unit, is the input value before time t, are the parameters of the LSTM model, including the weight matrix and bias term;

[0127] Confidence intervals are calculated using quantile regression:

[0128] ;

[0129] in, is the predicted value at the quantile q in the future T time step, is the quantile regression function, which outputs the predicted value corresponding to the quantile q;

[0130] right The predicted values are combined and output as a prediction matrix with confidence intervals, which optimizes LSTM in a dynamic sparse manner:

[0131] ;

[0132] The predicted values at different quantiles for the next 1 to 5 minutes are combined into a prediction matrix with confidence intervals, which is used as part of the input for the subsequent scheduling strategy based on the safety constraint PPO algorithm. This takes into account the load prediction under different situations and different probability distributions, which facilitates the accuracy of the subsequent scheduling strategy generation.

[0133] The LSTM is optimized through dynamic sparseness to meet the load prediction needs of different devices on the cloud edge; load prediction is achieved through the LSTM algorithm based on dynamic sparseness, tasks and resources are allocated and scheduled according to the predicted load values and other information, and the PPO algorithm based on safety constraints is used to generate scheduling strategies.

[0134] Furthermore, the cross-layer collaborative processing module also includes a security constraint PPO decision unit, and the security constraint PPO decision unit resources are allocated and scheduled according to the predicted load value output by the dynamic sparse LSTM load prediction unit, specifically:

[0135] For the state space S of the algorithm t , that is, the input of the algorithm is defined as:

[0136] ;

[0137] in, is the predicted value obtained by the LSTM algorithm based on dynamic sparseness, The real-time resource status indicates the real-time resource availability status of the current cloud / edge / device devices.

[0138] in, The task characteristics include the following three key attributes of the task:

[0139] Computational intensity c: The amount of computing resources required by the task per unit of data volume, expressed in FLOPS / Byte. Computational intensity c is obtained through offline calibration, which first benchmarks each type of task (such as image recognition and temperature analysis) and calculates its computational effort. (FLOPs) and the amount of input data (Bytes), and then directly assign the pre-calibrated c value according to the task type label, and normalize the c value:

[0140] ;

[0141] in, FLOPs is the number of floating-point operations used to measure the complexity of an algorithm or model. This number is related to the number of model parameters. Specific programs can be used to calculate the number of floating-point operations based on the weight file. The amount of input data is calculated based on the image resolution and number of channels;

[0142] The delay sensitivity l, that is, the sensitivity of the task to the response time, reflects the urgency. Normalization is performed and l is scaled to the interval [0,1]:

[0143] ;

[0144] ;

[0145] in, is the maximum delay time allowed for the task, l max is the maximum value of delay sensitivity, l min is the minimum value of delay sensitivity, l norm is the normalized characteristic parameter; for example, assuming that the maximum delay time allowed for a task is between 10 seconds and 60 seconds, if the maximum delay time allowed for a task is 25 seconds, then its delay sensitivity is ;

[0146] The data volume d is the total amount of data that needs to be transmitted or processed by the task. The value is normalized according to the maximum transmission capacity of the device. To zoom:

[0147] ;

[0148] According to the above three parameters describing the task characteristics, the final task feature vector is integrated:

[0149] ;

[0150] This vector is input as part of the PPO scheduler's state space to guide resource allocation decisions; if the current edge node has 30% remaining CPU and the task requires high computational intensity, the state vector will contain the information needed for these scheduling decisions.

[0151] Furthermore, the action space A of the algorithm, that is, the output of the algorithm, is defined as follows: the action vector αt :

[0152] ;

[0153] in, For the task assignment code, use one-hot encoding (One-Hot) with a length of 3, corresponding to the three types of devices: cloud, edge, and terminal. Preferably, [1,0,0] indicates that the task is assigned to the cloud, [0,1,0] to the edge, and [0,0,1] to the terminal.

[0154] is the resource allocation ratio, which indicates the proportion of resources allocated to the cloud, edge, and device devices, and the sum is ≤ 1. Preferably, [0.6, 0.3, 0.1] means that 60% of computing resources are allocated to the cloud, 30% to the edge, and 10% to the device.

[0155] Furthermore, safety constraints ensure that the agent's behavior does not violate certain safety conditions during training or execution, such as avoiding collisions, staying within a safe zone, or limiting the frequency of certain dangerous actions. This is very important in real-world applications such as autonomous driving and industrial robotics, where safety is the primary consideration. Safety constraints are modeled, including hard and soft constraints.

[0156] The hard constraints are mandatory and cannot be violated:

[0157] ;

[0158] in, For the current device The resource status under the action is to ensure that the cloud resource usage is less than 95%, the edge side is less than 85%, and the device side is less than 70%;

[0159] The soft constraints are non-mandatory constraints and can be violated probabilistically:

[0160] ;

[0161] in, is the probability that the task completion time is greater than the deadline;

[0162] Furthermore, the specific default probability is estimated based on Monte Carlo sampling:

[0163] ;

[0164] in, To estimate the probability, is the number of sampling times, is an indicator function, which takes 1 when the condition is met and 0 otherwise;

[0165] Furthermore, the constraints are incorporated into the objective function of reinforcement learning to ensure that both performance and safety are taken into account when optimizing the strategy. Based on the Lagrange duality method, the objective is optimized in combination with the constraints:

[0166] ;

[0167] Among them, E t is the expected operator, which represents the average performance of the objective function under the policy action distribution;

[0168] is the action probability of the current policy network parameter θ;

[0169] is the action probability ratio of the new and old strategies, which measures the change in action preference of the current strategy relative to the old strategy;

[0170] is the advantage function, which evaluates the quality of the action relative to the average level;

[0171] The default amounts for hard and soft constraints are:

[0172] ;

[0173] It is an adaptive Lagrange multiplier that dynamically adjusts the constraint penalty strength according to the degree of constraint violation:

[0174] ;

[0175] If the cloud resource utilization rate exceeds the standard ( ),but , Increase to intensify punishment; is the learning rate, which controls the adjustment amplitude of the multiplier;

[0176] According to the updated policy network formula, the degree of violation of the safety constraint is determined based on the factual information data collected by the system, and the adaptive Lagrange multiplier is adjusted to obtain a new policy network, and the new policy network is updated to the system;

[0177] The policy network is, in a given state Select Action The rules are expressed as probability distributions. In deep reinforcement learning, the strategy is represented by a neural network with parameters θ. ; The current system status is When the system selects the one with the highest probability , through the action vector α t Get the task assignment code and resource allocation ratio ;Task assignment coding Indicates which layer the task is assigned to, for example [1,0,0] means the task is assigned to the cloud; resource allocation ratio It represents the proportion of resources allocated to the cloud, edge, and end devices, and the total is ≤ 1. In this way, we achieve dynamic allocation of tasks and resources through the PPO algorithm based on security constraints.

[0178] Furthermore, the terminal side device includes the following modules:

[0179] The data acquisition module is used by the inspection robot to perform data acquisition tasks and obtain real-time environmental information and image data of the submerged arc furnace and its surroundings; including the acquisition of high-definition visible light images, infrared thermal imaging, and various environmental information data;

[0180] The autonomous decision-making module is used to control the inspection robot to complete the scheduled data collection tasks. Based on behavior tree task deployment, it automatically modifies and adjusts the inspection route based on image detection results. It communicates with the edge platform in real time to obtain overall information about the submerged arc furnace. It autonomously completes data collection and detection tasks under network anomalies and uploads detection results back to the edge platform after the network is restored. It autonomously switches to disaster recovery mode under extreme working conditions. Based on historical data from dynamic sparse LSTM load prediction, it triggers a preset hierarchical control strategy to determine detection priority, autonomously reduces the inspection robot's energy consumption, and suspends non-urgent tasks when the load exceeds the safety threshold, ensuring the core control of the robotic arm and the collection of key data.

[0181] The data processing module is responsible for local analysis and computation, as well as image data preprocessing, of image data collected by the data acquisition module, device-side tasks assigned by the cross-layer collaborative processing module, and detection tasks by the autonomous decision-making module. Computations are performed locally on the device, ensuring real-time performance and reliability.

[0182] The end-side collaborative module is used for end-side execution scheduling of the dynamic resource allocation strategy results of the cross-layer collaborative processing module, runs the minimalist mode LSTM predictor of the cross-layer collaborative processing module, and sends the processing results of the data processing module to the autonomous decision-making module for decision adjustment or upload to the edge-side platform; according to the dynamic resource allocation strategy of the cross-layer collaborative processing module, the detection tasks are classified, packaged and compressed. Some of them are calculated by the local data processing module, and the results are uploaded to the edge-side platform or sent to the autonomous decision-making module to assist decision-making. The others are packaged into a standard data format, compressed in blocks, and uploaded to the edge-side platform for processing.

[0183] Furthermore, the edge platform is deployed near the submerged arc furnace and receives pre-processed data from the edge devices. The edge platform is primarily responsible for real-time data analysis and processing, quickly extracting features and conducting preliminary fault diagnosis, and running lightweight deep learning models to meet real-time requirements. The edge platform includes the following modules:

[0184] The device access module is used to receive pre-processed data sent by end-side devices (single or multiple inspection robots) and receive data from connected multi-port acquisition devices (such as pan-tilt cameras, temperature and vibration sensors, etc.), realizing the unified fusion of multimodal data;

[0185] The task management module is used to develop inspection strategies for end-side devices and deploy detailed tasks and adjust them as needed. It allows users to deploy and adjust inspection tasks based on actual conditions to adapt to changing work requirements. If a task is no longer required, the module also provides a deletion function to easily remove irrelevant tasks. In addition to deploying, modifying, and deleting tasks, the task management module manages task schedules, allowing users to schedule inspection robots for periodic, routine, and special inspections, enabling adaptive adjustments to submerged arc furnace inspections.

[0186] The data storage module is used to store data collected by multi-port devices. It uses a storage model that combines relational databases and time-series databases to store raw data and intermediate detection results to support real-time analysis. Business logic data uses a relational database, while time-series data such as sensor data uses a time-series database.

[0187] The intelligent detection module applies deep learning technology to data collected by end-side devices and other devices connected to the system, performing intelligent detection such as target detection and semantic segmentation, and implementing accelerated inference functions to improve the real-time and stability of detection.

[0188] The data analysis module is used for in-depth analysis of multi-source data. It uses efficient deep learning algorithms to analyze multi-source data streams in real time and quickly locate operational anomalies and fault points. This module can promptly detect current equipment faults and predict potential future problems, thereby achieving predictive maintenance, reducing equipment downtime, and improving production efficiency and equipment reliability.

[0189] The edge collaboration module is used to execute scheduling on the edge platform based on the dynamic resource allocation strategy results of the cross-layer collaborative processing module, reasonably allocate and schedule detection tasks according to the resource load situation, and run the lightweight LSTM predictor of the cross-layer collaborative processing module to achieve second-level prediction of resource load. When the edge platform load is too high, the detection task will be promptly dispatched to the cloud platform for calculation according to the scheduling of the cross-layer collaborative processing module; according to the dynamic resource allocation strategy of the cross-layer collaborative processing module, the detection task uploaded by the receiving end side will be sorted, key data and alarm information will be uploaded to the cloud after screening and filtering, and the detection task will be reasonably allocated and scheduled according to the resource load situation. When the edge platform load is too high, the module will promptly package and compress the detection task and upload it to the cloud platform for calculation, effectively avoiding the risks brought by the accumulation of detection tasks.

[0190] Furthermore, the cloud platform has powerful computing and storage capabilities, responsible for in-depth analysis and model training of large amounts of inspection data. The cloud platform runs complex deep learning models to improve the accuracy of fault diagnosis. It is also responsible for updating and managing edge and device software. The cloud platform includes the following modules:

[0191] The data warehouse module is used to organize and store test result data uploaded by end-side devices and edge-side platforms, store uploaded historical resource load data, and store structured data assets for a long time through a partition index mechanism;

[0192] The data mining module extracts key historical data from the data warehouse module and uses machine learning algorithms, statistical analysis, and pattern recognition techniques to deeply explore potential data risks. This module supports the cleaning and feature engineering of massive amounts of structured and semi-structured data of different types and sources, and performs predictive maintenance based on business scenarios. Its functions include cluster analysis, association rule mining, time series data analysis, and visual insight generation.

[0193] The intelligent computing core module is used to handle computationally intensive inspection tasks, collect historical inspection result data from device-side devices and edge-side platforms, and regularly train and optimize inspection models.

[0194] The remote alarm module is used to set alarm mechanisms, remotely push alarm information to terminals, and conduct in-depth mining and analysis of potential dangers. When the detection results meet the preset alarm rules, or the data mining module analyzes the existence of potential dangers, the alarm information is pushed to the user through mini-programs and other means.

[0195] The cloud-side collaboration module is used for cloud-side platform strategy deployment and execution scheduling of the dynamic resource allocation strategy results of the cross-layer collaborative processing module, receives data processing requests from terminal devices and edge-side platforms, and returns the processing results; receives data processing requests uploaded by edge-side and terminal devices according to the dynamic resource allocation strategy of the cross-layer collaborative processing module, performs complex calculations or model inference in the cloud, and then returns the results to the device, summarizes the equipment operation data of the edge-side platform, optimizes the global scheduling strategy in the cloud and updates the LSTM model, and continuously improves the efficiency of the electric arc furnace inspection system.

Claims

1. An intelligent inspection system for submerged arc furnaces based on cloud-edge-end collaboration, comprising end-side devices, an edge-side platform, and a cloud-side platform, characterized by: It also includes a cross-layer collaborative processing module, wherein the end-side device is connected to the edge-side platform in communication, and the edge-side platform is connected to the cloud platform in communication; the cross-layer collaborative processing module is used for global intelligent resource allocation of the end-side device, edge-side platform and cloud platform, and uses a dynamic sparse LSTM model algorithm to predict future load trends, inputs the prediction results into the security-constrained PPO reinforcement learning algorithm to adaptively adjust the task and computing power allocation ratio, and dynamically allocates resources based on the output results; The end-side device is used to complete the data collection task of target detection, obtain data images, trigger the preset hierarchical control strategy through autonomous decision-making to determine the detection priority, run the LSTM predictor in minimalist mode, pre-process the collected data, and execute the detection data processing task based on the dynamic resource allocation results of the cross-layer collaborative processing module; The edge-side platform is used to deploy and adjust inspection strategies and detailed tasks for end-side devices, receive pre-processed data sent by end-side devices and data from connected multi-port acquisition devices, run a lightweight LSTM predictor, and perform data collation and analysis tasks based on the dynamic resource allocation results of the cross-layer collaborative processing module; The cloud platform is used to receive and organize data uploaded by terminal devices and edge-side platforms, conduct data mining to identify potential risks, perform deep learning analysis based on the dynamic resource allocation results of the cross-layer collaborative processing module, and return the processing results.

2. The intelligent inspection system for submerged arc furnaces based on cloud-edge-end collaboration according to claim 1 is characterized by: The cross-layer collaborative processing module includes a dynamic sparse LSTM load prediction unit, which is used to predict the future load trends of the end-side devices, edge-side platforms, and cloud platforms, and adaptively adjust the consumed computing power in combination with the gating mechanism. Specifically: Preprocess the input data. The input time series data includes: Resource utilization, including CPU, GPU, and memory: ; Network bandwidth: ; Task queue length: ; Normalization is performed on the input time series data to form an input vector: ； in, They are the average utilization rates of CPU, GPU and memory resources respectively. They are CPU, GPU, and memory standard deviations, is the maximum queue length; Define device type-dependent activation ratios based on gating mechanisms Among them, cloud refers to the cloud, edge refers to the edge side, and device refers to the device side: ； The hidden state update formula of each LSTM unit is: ； in, They are forget gate, input gate, and output gate respectively. is the candidate cell state, are the current cell state and the cell state at the previous moment, respectively. is the weight matrix corresponding to each gate and candidate state, is the previous hidden state and the current hidden state, is the current input, is the bias term for each gate and candidate state, is the Sigmoid function, is the hyperbolic tangent function, ☉ is the Hadamard product; Perform dynamic sparsification: ； in, is the LSTM hidden state after sparse gating, To retain the neuron outputs of the first k maximum values, the rest are set to zero. is the total number of neurons in the LSTM hidden layer; Use sparse gating to extract features, independently predict multiple quantiles, and generate asymmetric confidence intervals; Predict the load value in the next T minutes: ； in, is the load value predicted for the next T minutes, is the hidden state update formula of the LSTM unit, is the input value before time t, are the parameters of the LSTM model, including the weight matrix and bias term; Confidence intervals are calculated using quantile regression: ； in, is the predicted value at the quantile q in the future T time step, is the quantile regression function, which outputs the predicted value corresponding to the quantile q; right The predicted values are combined and output as a prediction matrix with confidence intervals, which optimizes LSTM in a dynamic sparse manner: 。 3. The intelligent inspection system for submerged arc furnaces based on cloud-edge-end collaboration according to claim 2 is characterized by: The cross-layer collaborative processing module also includes a safety constraint PPO decision unit, which performs resource allocation scheduling according to the predicted load value output by the dynamic sparse LSTM load prediction unit, specifically: For the state space S of the algorithm t , that is, the input of the algorithm is defined as: ； in, is the predicted value obtained by the LSTM algorithm based on dynamic sparseness, The real-time resource status indicates the real-time resource availability status of the current cloud / edge / device devices. in, The task characteristics include the following three key attributes of the task: Computational intensity c: The amount of computing resources required by the task per unit of data, expressed in FLOPS / Byte. Computational intensity c is obtained through offline calibration, which first benchmarks various tasks and calculates their computational effort. The amount of input data , directly assign the pre-calibrated c value according to the task type label and normalize the c value: ； in, is the amount of statistical calculation, is the amount of input data; The delay sensitivity l, that is, the sensitivity of the task to the response time, reflects the urgency. Normalization is performed and l is scaled to the interval [0,1]: ； ； in, is the maximum delay time allowed for the task, l max is the maximum value of delay sensitivity, l min is the minimum value of delay sensitivity, l norm is the normalized feature parameter; The data volume d is the total amount of data that needs to be transmitted or processed by the task. The value is normalized according to the maximum transmission capacity of the device. To zoom: ； According to the above three parameters describing the task characteristics, the final task feature vector is integrated: ； Furthermore, the action space A of the algorithm, that is, the output of the algorithm, is defined as follows: the action vector α t : ； in, Task assignment encoding uses one-hot encoding with a length of 3, corresponding to three types of devices: cloud, edge, and device. The resource allocation ratio indicates the proportion of resources allocated to cloud, edge, and device devices, and the total ratio must be ≤ 1. Furthermore, safety constraints are modeled, including hard and soft constraints; The hard constraints are mandatory constraints: ； in, For the current device The resource status under the action; The soft constraints are non-mandatory constraints and can be violated probabilistically: ； in, is the probability that the task completion time is greater than the deadline; Furthermore, the specific default probability is estimated based on Monte Carlo sampling: ； in, To estimate the probability, is the number of sampling times, is an indicator function, which takes 1 when the condition is met and 0 otherwise; Furthermore, the constraints are incorporated into the objective function of reinforcement learning. Based on the Lagrange dual method, the policy network is updated and the objective is optimized in combination with the constraints: ； Among them, E t is the expected operator, which represents the average performance of the objective function under the policy action distribution; is the action probability of the current policy network parameter θ; is the action probability ratio of the new and old strategies, which measures the change in action preference of the current strategy relative to the old strategy; is the advantage function, which evaluates the quality of the action relative to the average level; The default amounts for hard and soft constraints are: ； It is an adaptive Lagrange multiplier that dynamically adjusts the constraint penalty strength according to the degree of constraint violation: ； in, is the learning rate, which controls the adjustment amplitude of the multiplier; According to the updated policy network formula, the degree of violation of the safety constraint is determined based on the factual information data collected by the system, and the adaptive Lagrange multiplier is adjusted to obtain a new policy network, and the new policy network is updated to the system; The policy network is, in a given state Select Action The rules are expressed as probability distributions. In deep reinforcement learning, the strategy is represented by a neural network with parameters θ. ; The current system status is When the system selects the one with the highest probability , through the action vector α t Get the task assignment code and resource allocation ratio , that is, the PPO algorithm based on safety constraints realizes the dynamic allocation of tasks and resources.

4. The intelligent inspection system for submerged arc furnaces based on cloud-edge-end collaboration according to claim 1 is characterized by: The terminal device includes the following modules: The data acquisition module is used by the inspection robot to perform data acquisition tasks and obtain real-time environmental information and data images of the submerged arc furnace and its surroundings; The autonomous decision-making module is used to control the inspection robot to complete the scheduled data collection tasks. Based on behavior tree task deployment, it automatically modifies and adjusts the inspection route based on image detection results. It communicates with the edge platform in real time to obtain overall information about the submerged arc furnace. It autonomously completes data collection and detection tasks under network anomalies and uploads detection results back to the edge platform after the network is restored. It autonomously switches to disaster recovery mode under extreme working conditions. Based on historical data from dynamic sparse LSTM load prediction, it triggers a preset hierarchical control strategy to determine detection priority, autonomously reduces the inspection robot's energy consumption, and suspends non-urgent tasks when the load exceeds the safety threshold, ensuring the core control of the robotic arm and the collection of key data. The data processing module performs local computational analysis and preprocessing on the end-side tasks of the data acquisition module, the dynamic resource allocation strategy results of the cross-layer collaborative processing module, and the detection tasks of the autonomous decision-making module. The end-side collaboration module is used to schedule the end-side execution of the dynamic resource allocation strategy results of the cross-layer collaborative processing module, run the minimalist LSTM predictor of the cross-layer collaborative processing module, and send the processing results of the data processing module to the autonomous decision-making module for decision adjustment or upload to the edge platform.

5. The intelligent inspection system for submerged arc furnaces based on cloud-edge-end collaboration according to claim 1 is characterized by: The edge-side platform includes the following modules: The device access module is used to receive pre-processed data sent by the end-side device and receive data from the connected multi-port acquisition device to achieve unified fusion of multi-modal data; The task management module is used to formulate inspection strategies for end-side devices, deploy detailed tasks, and adjust strategies based on demand; The data storage module is used to store data collected by multi-port devices. It adopts a storage mode that combines relational database and time series database to store raw data and intermediate detection results to support real-time analysis. The intelligent detection module performs intelligent detection and accelerated reasoning based on deep learning on data from end-side devices and connected multi-port acquisition devices; The data analysis module is used for in-depth analysis of multi-source fusion data. It uses efficient deep learning algorithms to analyze multi-source data streams in real time and quickly locate operational anomalies and fault points. The edge collaboration module is used to execute and schedule the edge-side platform based on the dynamic resource allocation strategy results of the cross-layer collaborative processing module. It rationally allocates and schedules detection tasks according to the resource load, runs the lightweight LSTM predictor of the cross-layer collaborative processing module, and achieves resource load prediction in seconds. When the edge-side platform load is too high, the detection task is promptly dispatched to the cloud platform for calculation according to the scheduling of the cross-layer collaborative processing module.

6. The intelligent inspection system for submerged arc furnaces based on cloud-edge-end collaboration according to claim 1 is characterized by: The cloud platform includes the following modules: The data warehouse module is used to organize and store test result data uploaded by end-side devices and edge-side platforms, store uploaded historical resource load data, and store structured data assets for a long time through a partition index mechanism; The data mining module is used to extract key historical data from the data warehouse module and deeply explore potential risks in the data through machine learning algorithms, statistical analysis and pattern recognition technology; The intelligent computing core module is used to handle computationally intensive inspection tasks, collect historical inspection result data from device-side devices and edge-side platforms, and regularly train and optimize inspection models. Remote alarm module, used to set alarm mechanisms, remotely push alarm information to terminals, and conduct in-depth analysis of potential dangers; The cloud-side collaboration module is used to deploy and execute cloud-side platform policies for the dynamic resource allocation strategy results of cross-layer collaborative processing modules. It receives data processing requests from end-side devices and edge-side platforms and returns the processing results, optimizes the global collaborative scheduling strategy, and updates the LSTM model.

Images