Blast furnace tuyere inspection robot control method and system

Abstract

The invention belongs to the technical field of industrial automatic control, and particularly relates to a blast furnace tuyere inspection robot control method and system, a navigation positioning module in the system performs fusion processing on environment point cloud data acquired by a laser radar and image data acquired by a visual sensor, and positioning, map construction and path planning of an inspection robot are realized; in the positioning and map building process, the three-dimensional motion vector of a point in a scene is calculated in combination with optical flow information and depth information, and a dynamic object area is separated and removed through motion residual clustering; the data acquisition module controls a sensor group mounted on the inspection robot and detects the blast-furnace tuyere; the state diagnosis module identifies the tuyere burning loss defect and diagnoses the abnormal state of the equipment by using the multi-source data; the predictive maintenance module predicts the remaining service life or the fault occurrence probability of the key component through a predictive maintenance model based on the historical operation data and the abnormal state data of the equipment.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial automatic control, and particularly relates to a blast furnace tuyere inspection robot control method and system. BACKGROUND

[0002] The statements in this section merely refer to the background art of the present application and do not necessarily constitute the prior art.

[0003] As a core equipment for delivering hot air into the furnace, the blast furnace tuyere has long been subjected to extremely harsh working conditions of high temperature, high pressure, high dust, thermal fatigue and coal dust erosion, and is extremely prone to burning loss. Once the tuyere is damaged, the cooling water will leak into the furnace, causing the furnace temperature to fluctuate sharply and the production index to deteriorate. In severe cases, contact with high-temperature molten iron may cause water vapor explosion, causing disastrous accidents such as molten iron overflow, posing a great threat to personnel safety and equipment stability. At present, the tuyere damage is usually detected early by inspection and other methods.

[0004] The traditional blast furnace tuyere inspection work mainly relies on manual methods. The inspection personnel need to regularly go to the tuyere platform with high temperature, high noise and high dust, and judge the tuyere running state through sensory experience such as "seeing, listening, touching and smelling", such as observing whether the tuyere is red, listening to whether the tuyere airflow sound is abnormal, and sensing the temperature of the peephole. The inspection quality of the traditional method is poor, and is greatly affected by the quality of the inspection personnel, and there is a phenomenon that the inspection is not in place.

[0005] If the inspection robot is used instead of manual work, although the inspection efficiency is improved, the current inspection robot is usually a mobile inspection platform based on track or magnetic guidance (AGV). The blast furnace tuyere belongs to existing facilities, and large-scale civil construction cannot be carried out during production to lay the track of the robot. The magnetic navigation AGV needs to pre-embed magnetic strips or magnetic nails, which has high requirements for site modification and is difficult to implement. At the same time, the fixed track inspection method has insufficient flexibility and is difficult to cover all tuyeres, and the track is easy to deform and accumulate dust in the high temperature and high dust environment, affecting the operation reliability. SUMMARY

[0006] The present application provides a blast furnace tuyere inspection robot control method and system, which solves the problems of poor quality and low safety of traditional manual inspection of blast furnaces, and insufficient navigation adaptability of existing automation solutions, realizes accurate, automatic and intelligent monitoring and early warning of the state of tuyere equipment, ensures safe production of blast furnaces, and promotes the change of maintenance mode from passive repair to active prediction.

[0007] In order to achieve the above purpose, the present application adopts the following technical scheme: The first aspect of the present application discloses a blast furnace tuyere inspection robot control system, comprising: The navigation positioning module is configured to fuse the environment point cloud data acquired by the laser radar and the image data acquired by the visual sensor, and realize positioning, map construction and path planning of the inspection robot. The dynamic object processing module is configured to calculate the three-dimensional motion vector of the points in the scene in combination with the optical flow information and the depth information during the positioning and map construction, and separate and remove the dynamic object region through motion residual clustering. The data acquisition module is configured to control the sensor group installed on the inspection robot to detect the blast furnace tuyere, and the sensor group at least includes one or more of an infrared thermal imager, a CO concentration measuring instrument and a visible light camera. The state diagnosis module is configured to receive the multi-source data acquired by the data acquisition module, identify the tuyere burning loss defect based on the deep learning model, and determine the equipment abnormal state based on the anomaly detection model. The predictive maintenance module is configured to predict the remaining useful life or failure probability of the key components based on the historical operation data and abnormal state data of the equipment through the predictive maintenance model.

[0008] Further, in the navigation positioning module, the fusion of the environment point cloud data and the image data is realized based on the improved unscented Kalman filter, the improved unscented Kalman filter uses a spherical simplex Sigma point set to approximate the mean and covariance, and an adaptive factor is introduced when updating the state estimation covariance, and the scaling of the covariance is dynamically adjusted according to the information size.

[0009] Further, the deep learning model in the state diagnosis module is a multi-model fusion architecture, including a convolutional neural network for image classification, an image segmentation network for positioning the burning loss area, and a convolutional neural network combined with a long short-term memory network when the input is a video sequence.

[0010] Further, the predictive maintenance model includes: A multi-scale feature editor uses multiple one-dimensional convolutional neural networks to process historical data of different time scales to extract short-term, medium-term and long-term features. A cross-scale attention fusion module dynamically fuses features of different time scales from the multi-scale feature editor using a multi-head self-attention mechanism. A diffusion probability prediction model takes the context vector output by the cross-scale attention fusion module as a condition, performs a reverse generation process through a conditional U-Net structure, and predicts the future complete data distribution.

[0011] Further, during the identification of the tuyere burning loss defect by the state diagnosis module, at least one of the following determination logics is executed: When the thermal imaging data shows that a point has high temperature and the CO concentration measurement data also rises at the same time, it is determined that there is a coal gas leakage. When the thermal imaging data shows local high temperature in the front end of the tuyere and the visible light image observes a burning gap, it is determined that the tuyere is burned out; When the thermal imaging data shows local high temperature in the direct blowing pipe but no CO concentration rise, it is determined that the lining is damaged.

[0012] The second aspect of the present application discloses a blast furnace tuyere inspection robot control method, comprising the following steps: Obtain image data and laser point cloud data of the blast furnace tuyere and preprocess, perform data fusion through improved unscented Kalman filtering, realize synchronous positioning and map construction of the robot, complete autonomous navigation and path planning; In the positioning and map construction process, the three-dimensional motion vector of the points in the scene is calculated combining the optical flow and depth information, and the dynamic objects are separated and removed through motion residual clustering; The sensor group carried by the inspection robot is controlled to detect the blast furnace tuyere, and multi-source data including at least infrared thermal imaging, CO concentration and visible light image are obtained; The collected multi-source data are identified for tuyere burning defects using a deep learning model, and the equipment abnormal state is diagnosed based on an anomaly detection model; Based on the historical operation data and abnormal state data of the equipment, the residual service life or failure probability of the key components is predicted through a predictive maintenance model.

[0013] Further, the improved unscented Kalman filtering adopts a spherical simplex Sigma point set, and an adaptive factor is introduced in the state estimation covariance update, and the scaling of the covariance is dynamically adjusted according to the information size.

[0014] Further, the automatic identification of tuyere burning defects based on the deep learning model includes: Using a convolutional neural network for image classification; Using an image segmentation network to locate the burning area, and classifying according to the size and degree of the burning area; If the input is a video sequence, a convolutional neural network and a long short-term memory network are combined for time series modeling.

[0015] Further, the prediction is performed through a predictive maintenance model, including: Through a multi-scale feature editor, a plurality of one-dimensional convolutional neural networks are used to extract features of the input time series at different time scales; Through a cross-scale attention fusion module, a multi-head self-attention mechanism is used to dynamically fuse features at different time scales; Taking the fused context features as conditions, the reverse generation process of the diffusion probability prediction model is used to generate the future data distribution for prediction.

[0016] Furthermore, the collected multi-source data utilizes a deep learning model to identify air vent burn-out defects, and an anomaly detection model is used to diagnose abnormal equipment conditions, specifically: When thermal imaging data shows a high temperature at a certain point and the CO concentration data also increases at the same time, it is determined to be a gas leak; When thermal imaging data shows localized high temperatures at the front end of the air vent and visible light images reveal burn marks, it is determined that the air vent is burnt. When thermal imaging data shows localized high temperatures in the direct-blowing pipe but no increase in CO concentration, it is determined that the lining is damaged.

[0017] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects: 1. By integrating lidar with vision, the robot can achieve stable and accurate positioning and navigation on complex blast furnace platforms with high dust levels, strong light, and human interference. This not only completely avoids the safety risks of manual inspection but also overcomes the problem of large-scale platform modifications required by traditional magnetic navigation AGVs, achieving non-contact, highly adaptable autonomous operation.

[0018] 2. By fusing data from multiple sensors, including infrared thermal imaging, CO concentration monitoring, and visible light vision, cross-validation and accurate diagnosis of various faults such as vent burnout, gas leakage, and lining damage were achieved. This overcomes the shortcomings of single sensors, which have high false alarm and false negative rates, transforming fault identification from qualitative judgment based on experience to intelligent and quantitative diagnosis based on the correlation of multi-source information, greatly improving the accuracy and reliability of condition monitoring.

[0019] 3. By introducing a predictive maintenance algorithm that integrates multi-scale feature extraction and diffusion probability models, it is possible to deeply explore the degradation patterns of equipment performance and achieve probabilistic prediction of the remaining service life of key components. This allows equipment maintenance strategies to be upgraded from traditional "planned maintenance" or "emergency repair" to data-driven "predictive maintenance," enabling the early detection of potential faults, the rational scheduling of maintenance windows, the reduction of unplanned downtime, and ensuring the long-term smooth operation and safe production of blast furnaces. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a schematic diagram of the control process of a blast furnace tuyere inspection robot provided in one or more embodiments of the present invention; Figure 2 A schematic diagram of a blast furnace tuyere inspection robot control system provided for one or more embodiments of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] As described in the background section, to overcome the shortcomings of manual inspection, the industry has gradually introduced automated inspection technologies, such as track-mounted inspection robots or magnetically guided (AGV)-based mobile inspection platforms. However, these existing technological solutions still face significant challenges in the specific scenario of blast furnace tuyeres: The blast furnace tuyere platform is an existing facility, and large-scale civil construction is strictly prohibited during production. Traditional magnetic navigation AGVs require pre-embedded magnetic strips or nails, which places high demands on on-site modifications and is difficult to implement. At the same time, the fixed track inspection method lacks flexibility, cannot cover all tuyeres, and the track is prone to deformation and dust accumulation in high-temperature and high-dust environments, affecting operational reliability.

[0025] The blast furnace tuyeres platform has a complex environment, with varying lighting, large amounts of steam and dust, and occasional movement of personnel and vehicles. Existing inspection robot positioning systems generally suffer from positioning drift and susceptibility to interference in such dynamic and unstructured environments, preventing the robot from accurately stopping at the designated inspection position at each tuyer, directly impacting the consistency and comparability of data collection.

[0026] Existing automated monitoring methods are mostly limited to threshold alarms based on single sensor data, such as relying solely on infrared temperature measurement to determine anomalies. However, early signs of air vent malfunctions (such as burn-out, gas leaks, or lining damage) are often subtle manifestations of multi-source information coupling. The lack of effective multi-sensor information fusion and intelligent analysis algorithms leads to high false alarm and false negative rates, making it impossible to achieve deep perception of equipment health status and accurate location of fault roots.

[0027] Most existing systems remain at the level of "post-event alarm" or "periodic monitoring," failing to fully utilize historical and real-time data to build equipment performance degradation models. They cannot effectively predict the remaining service life of critical components (such as tuyeres), resulting in maintenance strategies still primarily based on planned maintenance or reactive emergency repairs. This fails to achieve the transition from "preventative" to "predictive" maintenance, hindering the realization of the goal of long-term stable operation of blast furnaces.

[0028] Therefore, this solution provides a control method and system for a blast furnace tuyere inspection robot. It adopts SLAM navigation that combines laser and vision to overcome harsh working conditions and achieve precise positioning. Through multi-sensor (infrared, CO, visible light) data fusion and AI algorithms (deep learning, anomaly detection, diffusion model), it achieves accurate fault identification and life prediction, and builds a complete intelligent inspection and predictive maintenance system.

[0029] like Figure 1 As shown, a blast furnace tuyere inspection robot control system includes three parts: an airborne control system, a relay data station, and a central control system.

[0030] The airborne control system is equipped with detection and positioning devices such as lidar, position sensors, and cameras. The control algorithm runs on the controller to realize on-site path planning and motion posture control of the inspection robot, thereby completing the inspection operation of key locations.

[0031] The relay data station is deployed at the work site to be responsible for data transfer and signal enhancement, ensuring stable communication between the robot and the central control system.

[0032] The central control system receives inspection data and combines it with data analysis algorithms to achieve equipment status assessment and anomaly warning. At the same time, based on the results of the equipment status assessment, it formulates personalized and professional inspection plans.

[0033] Through this control system and method, the inspection robot can operate stably under complex working conditions, achieve high-precision positioning and intelligent navigation, improve inspection efficiency and accuracy, and provide strong protection for the safe operation of blast furnace tuyeres.

[0034] Most blast furnace tuyere platforms are completed during the initial construction phase. Traditional AGV navigation methods require the installation of magnetic strips, which involves significant modifications to the tuyere platform and is not permitted during production. This solution proposes a navigation and positioning method based on the fusion of laser and vision. It acquires environmental point cloud data through LiDAR and combines it with visual recognition algorithms to match feature markers, enabling the robot to achieve autonomous localization and path planning.

[0035] This method requires no large-scale modification to existing air vent platforms, is highly adaptable, and can maintain stable operation under complex lighting and temperature conditions. Simultaneously, the fusion and correction of laser and visual data improves positioning accuracy and anti-interference capabilities, ensuring that the inspection robot can still complete high-precision navigation tasks in high-temperature and high-dust environments, providing solid support for the efficient operation of the entire inspection system.

[0036] The fusion of laser and visual information fully utilizes the high-precision distance measurement information of lidar to construct accurate geometric maps, and fully combines the rich texture information of visual images to perform semantic understanding of the scene, thereby improving the integrity and accuracy of map construction and laying a solid foundation for high-precision inspection.

[0037] During information fusion, an improved unscented Kalman filter is used to fuse laser and camera data, primarily focusing on pose estimation and loop closure detection. In the information fusion process, image information is converted into point cloud information, and the LiDAR point cloud data is combined and fused using the improved unscented Kalman filter.

[0038] Standard unscented Kalman filtering uses 2n+1 Sigma points (n being the state dimension). When n is large, the computational cost (especially the calculation of the covariance matrix and weighted summation) increases significantly. Using a spherical simplex Sigma point set, fewer Sigma points (at least n+2 points) are used to approximate the mean and covariance, thereby reducing the computational cost and improving the overall computational efficiency.

[0039] The fused data is used to estimate the positions between adjacent frames. The loop closure detection section uses the ORB feature descriptor, which is rotation-invariant and size-invariant, to extract features from the image. This allows for stable feature point extraction under different operating conditions. A matching algorithm is then used to detect similarity and determine loop closures.

[0040] Meanwhile, operations are performed directly on the square root of the covariance matrix, avoiding explicit calculation and storage of the complete covariance matrix and ensuring numerical stability. An adaptive factor is introduced during the state estimation covariance update to dynamically adjust the scaling of the covariance based on the information magnitude, suppressing the influence of outliers or model errors and enhancing robustness.

[0041] In simultaneous localization and mapping (SLAM) systems, the presence of dynamic objects can severely interfere with the accuracy of location estimation and map building. Blast furnace tuyere platforms cannot be completely unmanned; traditional algorithms typically assume a static environment. Therefore, dynamic object removal is a key technology for improving robustness. This system uses motion segmentation based on depth information. Without requiring prior semantics, it addresses the depth discontinuity or motion inconsistency between dynamic objects and the background. Through scene flow analysis, the scene flow of static points should be consistent with camera motion, while that of dynamic points deviates significantly. Therefore, it combines optical flow and depth information to calculate 3D motion vectors, compares them with camera motion, and clusters the motion residuals to separate dynamic object regions.

[0042] The system is equipped with infrared thermal imaging, CO measuring instruments, cameras, and other detection instruments. Through multi-sensor data fusion, it achieves automated, intelligent, and data-driven monitoring of equipment status in windy areas. The massive amounts of data collected are transformed into valuable equipment health information and risk warnings. Through advanced data analysis technologies (especially multi-source data fusion and AI), it enables early detection, precise location, and risk quantification of potential equipment failures.

[0043] In terms of multi-sensor fusion, data from a single sensor may contain false alarms or incomplete information, while fusion analysis can significantly improve diagnostic accuracy.

[0044] Thermal imaging showed high temperature and elevated CO concentration at a certain point, strongly indicating a gas leak; Thermal imaging showed localized high temperatures at the front end of the air vent, and visible light images revealed burn marks, confirming that the air vent was burned. Thermal imaging showed localized high temperatures in the direct-blowing pipe without CO leakage, which may indicate that the lining was damaged but had not yet leaked, or that the leak point had not been detected.

[0045] By fusing real-time data from multiple sensors, accurate judgments are achieved, greatly reducing the probability of missed or false alarms and significantly improving the reliability of intelligent inspection.

[0046] In data analysis and processing, a deep learning-based CV algorithm automatically identifies defects such as vent burn-out. An anomaly detection model is proposed, trained using historical data to automatically identify abnormal heat maps, CO concentration patterns, or sound spectra. A predictive maintenance model is also proposed, combining equipment operating time, operating parameters (air temperature, air pressure, coal ratio, etc.) and historical data from multiple sensors (temperature trends, slow rise in CO concentration baseline, etc.) to predict the remaining service life or probability of failure of key components. Based on this, a preventative maintenance strategy is established, transforming reactive emergency repairs into proactive prevention and planned maintenance into precise predictive maintenance.

[0047] like Figure 2 As shown, a control method for a blast furnace tuyere inspection robot includes the following steps: Image data and laser point cloud data of blast furnace tuyere are acquired and preprocessed. Data fusion is performed through improved unscented Kalman filtering to achieve synchronous localization and map building of the robot, and complete autonomous navigation and path planning. During the localization and map building process, the three-dimensional motion vectors of points in the scene are calculated by combining optical flow and depth information, and dynamic objects are separated and removed by motion residual clustering. The sensor array carried by the inspection robot detects the blast furnace tuyeres and acquires multi-source data, including at least infrared thermal imaging, CO concentration, and visible light images. The collected multi-source data is used to identify air outlet burn-out defects using a deep learning model, and the abnormal state of the equipment is diagnosed based on an anomaly detection model. Based on historical operating data and abnormal status data of equipment, predictive maintenance models are used to predict the remaining service life or probability of failure of key components.

[0048] The steps for automatically identifying air vent burn-out using a deep learning-based CV algorithm are as follows.

[0049] (1) Data collection and labeling: Collect images of normal air outlets and burned air outlets and label them (divide the burn level according to the degree of burn in a finer granular manner, and divide the burn level into 0 to 9 ten levels according to the actual situation on site, with 0 being intact and 9 being damaged).

[0050] (2) Image preprocessing: including image denoising, contrast enhancement, color correction, etc., to improve image quality.

[0051] (3) Feature extraction: Use CNN to extract deep features, and combine them with traditional image features (color histogram, texture features, etc.).

[0052] (4) Model Construction: A multi-model fusion approach is adopted. The specific implementation is as follows: a. Using CNN for image classification; b. Use image segmentation to locate the burned areas, and then classify them according to the size and extent of the burned areas; c. If video sequences are used, CNN and LSTM can be combined for temporal modeling.

[0053] (5) Model training and validation: Train the model using labeled data and validate the model performance.

[0054] (6) Deployment: Deploy the trained model to the monitoring system to achieve real-time detection.

[0055] By combining multi-scale burn area detection with global appearance and local texture features, dynamic burn process modeling based on video sequence burn evolution analysis, and multi-modal feature fusion integrating infrared thermal imaging and visible light visual information, automatic identification of vent burn was achieved.

[0056] The predictive maintenance model integrates deep learning and uncertainty quantification to not only capture patterns across multiple time scales but also predict the complete future distribution through probabilistic generation. The model architecture is as follows: 1. Multi-scale Feature Editor. This module aims to extract features at different time scales from the input time series data. It uses three one-dimensional convolutional neural networks (Conv1D) to process short-term, medium-term, and long-term historical data respectively, as shown in the following equation: ; ; ; Among them, L S L m L I These correspond to short-term (e.g., 7 days), medium-term (e.g., 30 days), and long-term (e.g., 1 year) observation windows, respectively. The input time series is Y. t-L :t, where L is the length of the look-back window.

[0057] 2. Cross-scale attention fusion module. The goal of this module is to dynamically fuse features from different time scales, using a multi-head self-attention mechanism to learn the importance of features at different scales.

[0058] When using a multi-head attention mechanism to dynamically fuse multi-scale features, the weight allocation is time-dependent; the more recent the time period, the larger the weight, as shown in the following formula: ; ; ; ; Among them, X S X m S I Historical data for the short, medium and long term, respectively, W Q To query the weight matrix, W K W is the key weight matrix. V The value weight matrix.

[0059] 3. Diffusion Probability Prediction Model. The diffusion model consists of two processes: forward diffusion and backward generation. The specific processes are as follows.

[0060] Forward diffusion process (training phase): ; ; Where q is the forward diffusion process distribution, defining a Markov process from clean data to noise; Y0 is the original true value of the future sequence, with dimensions (batch_size, forecast_horizon, feature_dim), and is the true future sequence label used during training; Y 1：T The noise sequence in the diffusion process is the noise-added sequence from time step t=1 to t=T; T is the total number of diffusion steps, usually taken as 1000 steps, which is the total number of steps from clean data to pure noise.

[0061] Among them, Y t Y is the noisy version of the sequence at step t after being diffused for t steps; t-1 Let be the sequence at step t-1, and be the sequence state at the previous step; N is a Gaussian distribution, and the diffusion process follows a Gaussian distribution; β is the signal retention coefficient, which controls how much of the signal from the previous moment is retained. t This is a noise scheduling parameter, with values ​​ranging from β. t =10 -4 To β t =0.02 increases linearly to control the amount of noise added at each step; I is the identity matrix and covariance matrix, assuming that each dimension is independent and identically distributed.

[0062] Reverse generation process (prediction stage): ; in, This is a parameterized reverse generation process, a denoising process learned by a neural network; Y 0:T For the complete path from noise to clean data, all intermediate states during the generation process; p(Y T ) represents the initial noise distribution, generated starting from a standard Gaussian distribution; H is the conditional feature vector with dimensions (batch_size, context_dim), derived from the fusion features of the second module, guiding the generation process.

[0063] ; in, This is the mean value predicted by the neural network. The variance of the neural network prediction, the degree of uncertainty in the prediction, is usually simplified to a fixed value or a time-dependent parameter of the learning process; For neural network parameters, all trainable weights in U-Net.

[0064] The reverse process is a denoising network that uses a conditional U-Net structure, which takes a noise sequence and time-step embedding as input and is conditional on the context vector H output by the attention fusion module.

[0065] The structure of U-Net is as follows: Input noise sequence Y t The embedding vector at time step t, and the condition vector H; First, the time step t is embedded into a vector, and the dimensions are adjusted through a fully connected layer; Then, the noise sequence Y tDownsampling (encoder) is performed through several one-dimensional convolutional layers, while time step embeddings and conditional vectors H are injected in each downsampling layer by addition or concatenation. Next, upsampling (decoder) is performed through a one-dimensional transposed convolutional layer, and skip connections are made with the corresponding layers of the encoder.

[0066] The output of U-Net is a noise prediction with the same shape as the input noise sequence.

[0067] This solution integrates LiDAR and vision, employing improved unscented Kalman filtering and dynamic object removal technology. This enables the robot to achieve stable and accurate positioning and navigation on complex blast furnace platforms with high dust levels, strong lighting, and human interference. It not only completely eliminates the safety risks of manual inspection but also overcomes the challenge of large-scale platform modifications required by traditional magnetic navigation AGVs, achieving non-contact, highly adaptable autonomous operation.

[0068] By fusing data from multiple sensors, including infrared thermal imaging, CO concentration monitoring, and visible light vision, cross-validation and accurate diagnosis of various faults such as vent burnout, gas leakage, and lining damage were achieved. This effectively overcomes the shortcomings of single sensors, which have high false alarm and false negative rates. It transforms fault identification from qualitative judgment based on experience to intelligent and quantitative diagnosis based on the correlation of multi-source information, greatly improving the accuracy and reliability of condition monitoring.

[0069] By introducing a predictive maintenance algorithm that integrates multi-scale feature extraction and diffusion probability models, it is possible to deeply explore the degradation patterns of equipment performance and achieve probabilistic predictions of the remaining service life of key components. This allows equipment maintenance strategies to be upgraded from traditional "planned maintenance" or "emergency repair" to data-driven "predictive maintenance," which can identify potential faults in advance, rationally schedule maintenance windows, significantly reduce unplanned downtime, ensure the long-term smooth operation and safe production of blast furnaces, and effectively reduce operation and maintenance costs.

[0070] The massive amounts of on-site data collected by the inspection robot are transformed into equipment health information and risk warnings with clear guiding significance through condition diagnosis and predictive maintenance models, providing strong data support for blast furnace operation and maintenance decisions.

[0071] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A blast furnace tuyere inspection robot control system, characterized in that, include: The navigation and positioning module is configured to fuse environmental point cloud data acquired by lidar and image data acquired by visual sensors to realize the positioning, map building and path planning of the inspection robot. The dynamic object processing module is configured to: calculate the three-dimensional motion vectors of points in the scene by combining optical flow information and depth information during the localization and map building process, and separate and remove dynamic object regions by motion residual clustering; The data acquisition module is configured to control a sensor group installed on the inspection robot to detect the blast furnace tuyeres. The sensor group includes at least one or more of an infrared thermal imager, a CO concentration meter, and a visible light camera. The status diagnosis module is configured to: receive multi-source data acquired by the data acquisition module, identify air outlet burn-out defects based on a deep learning model, and determine abnormal equipment status based on an anomaly detection model; The predictive maintenance module is configured to predict the remaining service life or probability of failure of key components based on historical operating data and abnormal status data of the equipment through a predictive maintenance model.

2. The blast furnace tuyere inspection robot control system as described in claim 1, characterized in that, In the navigation and positioning module, the fusion of environmental point cloud data and image data is achieved based on the improved unscented Kalman filter. The improved unscented Kalman filter uses a spherical simplex Sigma point set to approximate the mean and covariance, and introduces an adaptive factor when updating the state estimation covariance to dynamically adjust the scaling of the covariance according to the information size.

3. The blast furnace tuyere inspection robot control system as described in claim 1, characterized in that, The deep learning model in the state diagnosis module is a multi-model fusion architecture, including a convolutional neural network for image classification, an image segmentation network for locating burned areas, and a temporal model combining a convolutional neural network and a long short-term memory network when the input is a video sequence.

4. The blast furnace tuyere inspection robot control system as described in claim 1, characterized in that, The predictive maintenance model includes: A multi-scale feature editor utilizes multiple one-dimensional convolutional neural networks to process historical data at different time scales to extract short-term, medium-term, and long-term features. The cross-scale attention fusion module utilizes a multi-head self-attention mechanism to dynamically fuse features from different time scales from a multi-scale feature editor. The diffusion probability prediction model uses the context vector output by the cross-scale attention fusion module as a condition and performs a reverse generation process through a conditional U-Net structure to predict the future complete data distribution.

5. The blast furnace tuyere inspection robot control system as described in claim 1, characterized in that, During the identification of air vent burn-out defects, the status diagnosis module executes at least one of the following judgment logic: When thermal imaging data shows a high temperature at a certain point and CO concentration measurement data rises at the same time, it is determined to be a gas leak; When thermal imaging data shows localized high temperatures at the front end of the air vent and visible light images reveal burn marks, it is determined that the air vent is burnt. When thermal imaging data shows localized high temperatures in the direct-blowing pipe but no increase in CO concentration, it is determined that the lining is damaged.

6. A control method for a blast furnace tuyere inspection robot, implemented based on the blast furnace tuyere inspection robot control system as described in claim 1, characterized in that, Includes the following steps: Image data and laser point cloud data of blast furnace tuyere are acquired and preprocessed. Data fusion is performed through improved unscented Kalman filtering to achieve synchronous localization and map building of the robot, and complete autonomous navigation and path planning. During the localization and map building process, the three-dimensional motion vectors of points in the scene are calculated by combining optical flow and depth information, and dynamic objects are separated and removed by motion residual clustering. The sensor array carried by the inspection robot detects the blast furnace tuyeres and acquires multi-source data, including at least infrared thermal imaging, CO concentration, and visible light images. The collected multi-source data is used to identify air outlet burn-out defects using a deep learning model, and the abnormal state of the equipment is diagnosed based on an anomaly detection model. Based on historical operating data and abnormal status data of equipment, predictive maintenance models are used to predict the remaining service life or probability of failure of key components.

7. The control method for a blast furnace tuyere inspection robot as described in claim 6, characterized in that, The improved unscented Kalman filter uses a spherical simplex Sigma point set and introduces an adaptive factor when updating the state estimation covariance to dynamically adjust the scaling of the covariance according to the information size.

8. The control method for a blast furnace tuyere inspection robot as described in claim 6, characterized in that, Automatic identification of air vent burn-out defects based on deep learning models includes: Image classification using convolutional neural networks; An image segmentation network is used to locate the burned areas and classify them according to the size and extent of the burned areas; If the input is a video sequence, then a temporal model is performed by combining a convolutional neural network and a long short-term memory network.

9. The control method for a blast furnace tuyere inspection robot as described in claim 6, characterized in that, Predictions are made through predictive maintenance models, including: The multi-scale feature editor utilizes multiple one-dimensional convolutional neural networks to extract features of the input time series at different time scales. By using a cross-scale attention fusion module, features from different time scales are dynamically fused using a multi-head self-attention mechanism. Using the fused contextual features as conditions, the future data distribution is generated for prediction through the reverse generation process of the diffusion probability prediction model.

10. The control method for a blast furnace tuyere inspection robot as described in claim 6, characterized in that, The collected multi-source data is used to identify air vent burn-out defects using a deep learning model, and the abnormal state of the equipment is diagnosed based on an anomaly detection model, specifically: When thermal imaging data shows a high temperature at a certain point and the CO concentration data also increases at the same time, it is determined to be a gas leak; When thermal imaging data shows localized high temperatures at the front end of the air vent and visible light images reveal burn marks, it is determined that the air vent is burnt. When thermal imaging data shows localized high temperatures in the direct-blowing pipe but no increase in CO concentration, it is determined that the lining is damaged.

Images