Artificial Intelligence-Based Status Monitoring Method and System

Abstract

This application discloses an artificial intelligence-based condition monitoring method and system, relating to the field of intelligent monitoring technology for industrial equipment. The disclosed artificial intelligence-based condition monitoring method and system are applied to high-temperature industrial equipment. By fusing visible light and infrared thermal image data, combined with thermal inertia prediction models and dynamic threshold analysis technology, it effectively distinguishes between real equipment faults and transient external interference, thereby improving the accuracy and reliability of condition monitoring for high-temperature industrial equipment and reducing false alarm and missed alarm rates.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring technology for industrial equipment, and in particular to a status monitoring method and system based on artificial intelligence. Background Technology

[0002] In the field of infrared status monitoring of high-temperature industrial equipment (such as metallurgical furnaces, reaction vessels, and power generation boilers), artificial intelligence-based thermal imaging data analysis has become an important means of fault early warning. However, existing technologies generally face the problem of temperature data distortion caused by transient occlusion (such as workers passing by, maintenance tools blocking the view, and steam / fume interference) and dynamic thermal interference (such as splashing slag, condensate dripping, and environmental reflection). Current solutions generally rely on image classification and semantic segmentation techniques using convolutional neural networks (CNNs) or employing sliding window temporal statistical methods. These methods lack modeling of the equipment's thermophysical characteristics (such as thermal inertia and heat conduction paths) at the model level, leading to inaccurate judgments of temperature anomalies in occluded areas. Furthermore, these methods can only passively respond to transient interference and cannot distinguish the spatial pattern differences between actual equipment faults and brief external disturbances. In addition, the use of simple data discarding or interpolation compensation when occlusion occurs disrupts the continuity of temperature field evolution. These problems result in a high false alarm rate, causing not only a large amount of ineffective maintenance but also the potential for missed detection of real faults, threatening industrial safety. Especially in high-temperature industrial environments, the dynamic changes in the temperature field of equipment surfaces exhibit significant thermal inertia. Existing methods fail to effectively utilize this physical characteristic for anomaly detection, resulting in insufficient sensitivity in identifying early equipment failures. Furthermore, existing systems often employ a one-size-fits-all approach to anomaly detection in obstructed areas, failing to differentiate the varying impacts of different obstruction types on the temperature field, further reducing the reliability of the monitoring system.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main purpose of this application is to provide a condition monitoring method and system based on artificial intelligence, which aims to improve the accuracy of condition monitoring of high-temperature industrial equipment and reduce false alarm rate and missed alarm rate.

[0005] To achieve the above objectives, this application proposes an artificial intelligence-based condition monitoring method for use in high-temperature industrial equipment. The method includes:

[0006] Acquire visible light image data and infrared thermal image sequence data of the target device;

[0007] The infrared thermal image sequence data is input into a pre-trained thermal inertia prediction model for processing to generate interference-free thermal image data and thermal inertia parameter data.

[0008] The current infrared thermal image data from the visible light image data and the infrared thermal image sequence data is input into the pre-trained occlusion detection model for processing to generate actual occlusion mask data and predicted occlusion mask data.

[0009] Based on the interference-free thermal image data and the infrared thermal image data at the current moment, residual thermal image data is calculated and generated;

[0010] Based on the predicted occlusion mask data, the residual heatmap data is segmented into residual data of the occluded area and residual data of the unoccluded area;

[0011] The residual data of the occluded area is combined with the actual occlusion mask data to perform interference analysis and generate interference discrimination flag data.

[0012] The residual data of the unobstructed area is combined with the thermal inertia parameter data to perform dynamic threshold analysis and generate abnormal pixel coordinate data.

[0013] By fusing the interference discrimination flag data and the abnormal pixel coordinate data, device status alarm command data is generated.

[0014] In one embodiment, the method further includes:

[0015] Acquire a sequence of continuous, unobstructed infrared thermal images of the target device under normal operating conditions to form an infrared thermal image training sequence set;

[0016] A spatiotemporal joint feature extraction network containing convolutional layers and LSTM layers is constructed. The infrared thermal image training sequence set is input into the spatiotemporal joint feature extraction network, and the spatiotemporal feature tensor is output.

[0017] Based on the specific heat capacity data, thermal conductivity data and three-dimensional structural model of the target equipment's metallic material, a partial differential equation for heat conduction on the equipment surface is established.

[0018] The theoretical surface temperature gradient tensor at the location corresponding to the spatiotemporal characteristic tensor is calculated using the partial differential equation of heat conduction on the surface of the device.

[0019] The mean square error between the temperature gradient predicted by the hidden layer of the spatiotemporal joint feature extraction network and the theoretical surface temperature gradient tensor is calculated to generate a physical regularization loss value.

[0020] The sum of absolute errors between the predicted heatmap and the actual heatmap pixels in the final output of the spatiotemporal joint feature extraction network is calculated to generate pixel-level reconstruction loss values.

[0021] The weighted fusion of physical regularization loss and pixel-level reconstruction loss is used to update the weights of the spatiotemporal joint feature extraction network through backpropagation until the loss converges, thereby generating the pre-trained thermal inertia prediction model.

[0022] In one embodiment, the step of establishing the partial differential equation for heat conduction on the surface of the device based on the specific heat capacity data, thermal conductivity data, and three-dimensional structural model of the target device's metallic material includes:

[0023] Obtain the data set of thermal conductivity coefficients of the materials of each component in the preset 3D model of the target device;

[0024] Construct heat conduction topology network data based on the material thermal conductivity dataset and the component connection relationships of the target device;

[0025] Based on the heat conduction topology network data, the surface temperature distribution equation is derived, and simplified heat conduction equation data is generated.

[0026] The simplified heat conduction equation data is discretized into a partial differential equation for heat conduction on the surface of the device.

[0027] In one embodiment, the step of inputting the current-time infrared thermal image data from the visible light image data and the infrared thermal image sequence data into a pre-trained occlusion detection model for processing to generate actual occlusion mask data and predicted occlusion mask data includes:

[0028] Target detection is performed on the visible light image data to generate bounding box data for moving objects;

[0029] Based on the motion trajectory of the bounding box data, the predicted occlusion mask data is generated by extrapolation.

[0030] Analyze the regions with temperatures below a threshold in the infrared thermal image data to generate low-temperature region contour data;

[0031] By fusing the bounding box data of the moving object with the contour data of the low-temperature region, calibrated actual occlusion data is generated.

[0032] In one embodiment, the step of extrapolating the predicted occlusion mask data based on the motion trajectory of the bounding box data includes:

[0033] Motion vector data is established based on the bounding box data;

[0034] Based on the motion vector data, a trajectory for a future preset duration is simulated, and corresponding trajectory point set data is generated;

[0035] The trajectory point set data is expanded by convex hull to generate the predicted occlusion mask data.

[0036] In one embodiment, the interference discrimination flag data includes first interference discrimination flag data and second interference discrimination flag data; the step of combining the residual data of the occluded area with the actual occlusion mask data to perform interference analysis and generate interference discrimination flag data includes:

[0037] When the actual occlusion mask data covers the occluded area, a low confidence flag data is set and used as the first interference discrimination flag data.

[0038] When the actual occlusion mask data does not cover the occluded area, the maximum residual value of the residual data of the occluded area is calculated, and the corresponding tolerance threshold data is determined based on the thermal inertia parameter data. When the maximum residual value of the area exceeds the tolerance threshold data, a high anomaly risk flag is set and used as the second interference discrimination flag data.

[0039] In one embodiment, the step of setting high anomaly risk flag data includes:

[0040] Extract the spatial distribution feature data of the residual data in the occluded area;

[0041] Compare the spatial distribution feature data with the structural features of the preset target device 3D model;

[0042] When spatial distribution characteristic data matches flange structure or bearing structure, the risk level is increased to generate structural associated risk indicator data.

[0043] When the spatial distribution characteristic data is in a random and discrete form, transient interference marker data is generated;

[0044] The structurally correlated risk indicator data and the transient interference indicator data are combined to form the high anomaly risk indicator data.

[0045] In one embodiment, the step of combining the residual data of the unobstructed region with the thermal inertia parameter data to perform dynamic threshold analysis and generate abnormal pixel coordinate data includes:

[0046] For each pixel in the residual data of the unobstructed area, the corresponding thermal inertia parameter data is matched according to its spatial location coordinate data;

[0047] Based on the unobstructed residual sample dataset of the device under historical normal operating conditions, a probability distribution model bound to spatial location is constructed.

[0048] The parameters of the probability distribution model are adjusted based on the thermal inertia parameter data to generate dynamic adaptive threshold data;

[0049] Calculate the confidence probability data of the current pixel residual value in the dynamic adaptive threshold data;

[0050] When the confidence probability data is lower than the preset risk threshold, the pixel is marked as an abnormal candidate point data;

[0051] The abnormal pixel coordinate data is generated by summarizing all abnormal candidate point data.

[0052] In one embodiment, the step of fusing the interference discrimination flag data and the abnormal pixel coordinate data to generate device status alarm command data includes:

[0053] The abnormal pixel coordinate data is input into a clustering algorithm to generate abnormal region block coordinate data.

[0054] Based on the interference discrimination flag data, mark the corresponding abnormal region blocks to generate a risk labeling dataset;

[0055] Initial alarm command data is generated by matching the risk labeling dataset with a preset alarm rule base.

[0056] The initial alarm command data is corrected by combining the equipment operating status data, and the final equipment status alarm command data is output.

[0057] Furthermore, to achieve the above objectives, this application also proposes an artificial intelligence-based condition monitoring system for use in high-temperature industrial equipment. The artificial intelligence-based condition monitoring system includes: a memory, a processor, and an artificial intelligence-based condition monitoring program stored in the memory and executable on the processor. The artificial intelligence-based condition monitoring program is configured to implement the steps of the artificial intelligence-based condition monitoring method.

[0058] The artificial intelligence-based condition monitoring method and system proposed in this application are applied to high-temperature industrial equipment. By fusing visible light and infrared thermal image data, and combining thermal inertia prediction models and dynamic threshold analysis technology, it can effectively distinguish between real equipment faults and transient external interference, thereby improving the accuracy and reliability of condition monitoring of high-temperature industrial equipment and reducing false alarm and missed alarm rates. Attached Figure Description

[0059] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0060] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating an embodiment of the artificial intelligence-based state monitoring method of this application.

[0062] Figure 2 This is a schematic diagram of a structural embodiment of the artificial intelligence-based status monitoring system of this application.

[0063] Explanation of icon numbers:

[0064] 10. Memory; 20. Processor.

[0065] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0066] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0067] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0068] In existing technologies, infrared status monitoring of high-temperature industrial equipment often results in distorted temperature data due to momentary obstruction and dynamic thermal interference. Current solutions rely on convolutional neural networks or sliding window statistical methods, which lack modeling of the equipment's thermophysical characteristics. This makes it difficult to accurately identify temperature anomalies in obstructed areas, distinguish between spatial patterns of genuine faults and transient disturbances, and disrupt the continuity of temperature field evolution through data discarding or interpolation compensation, leading to high false alarm and false negative rates, threatening industrial safety. For example, during the operation of a metallurgical furnace, splashing molten slag may create momentary high-temperature interference, while maintenance tools may obscure true temperature anomalies; existing methods cannot effectively identify such complex scenarios.

[0069] To address the aforementioned issues, a method capable of simultaneously handling occlusion interference and dynamic thermal anomalies is needed. Considering the impact of equipment thermal inertia on temperature field evolution, establishing a thermophysical model to predict temperature distribution under undisturbed conditions becomes a key breakthrough. Simultaneously, it is necessary to combine multimodal data for dynamic analysis of occlusion areas to distinguish between genuine anomalies and external interference. The core approach to solving this problem lies in combining thermal inertia prediction with occlusion detection to achieve accurate anomaly localization while preserving the continuity of the temperature field.

[0070] Based on this, embodiments of this application provide an artificial intelligence-based state monitoring method, referring to... Figure 1 The artificial intelligence-based state monitoring method includes steps S100 to S800, wherein:

[0071] Step S100: Acquire visible light image data and infrared thermal image sequence data of the target device;

[0072] Step S200: Input the infrared thermal image sequence data into the pre-trained thermal inertia prediction model for processing to generate interference-free thermal image data and thermal inertia parameter data;

[0073] Step S300: Input the current infrared thermal image data from the visible light image data and the infrared thermal image sequence data into the pre-trained occlusion detection model for processing to generate actual occlusion mask data and predicted occlusion mask data.

[0074] Step S400: Based on the interference-free thermal image data and the infrared thermal image data at the current moment, calculate and generate residual thermal image data;

[0075] Step S500: Based on the predicted occlusion mask data, the residual heatmap data is divided into occluded region residual data and unoccluded region residual data.

[0076] Step S600: Combine the residual data of the occluded area with the actual occlusion mask data to perform interference analysis and generate interference discrimination flag data;

[0077] Step S700: Combine the residual data of the unobstructed area with the thermal inertia parameter data to perform dynamic threshold analysis and generate abnormal pixel coordinate data.

[0078] Step S800: The interference discrimination flag data and the abnormal pixel coordinate data are fused to generate device status alarm command data.

[0079] In this embodiment, visible light image data refers to visual information of the equipment surface acquired by an optical camera. Specifically, an industrial-grade explosion-proof camera can be used to capture images in the visible light band to detect obstructed areas formed by moving objects. Infrared thermal image sequence data refers to the temperature distribution data of the equipment surface continuously acquired by an infrared thermal imager. Specifically, an uncooled infrared detector with a frame rate of not less than 10Hz can be used to record the dynamic changes in the temperature field. The thermal inertia prediction model refers to a temperature field evolution prediction model established based on the thermophysical characteristics of the equipment. Specifically, a spatiotemporal joint feature extraction network containing convolutional layers and LSTM layers can be used to simulate the heat conduction process by fusing material thermal conductivity and three-dimensional structural data. Residual thermal image data refers to the difference distribution between the actual observed thermal image and the predicted interference-free thermal image. Specifically, it can be generated by pixel-by-pixel difference operations to characterize abnormal fluctuations in the temperature field. Dynamic threshold analysis refers to adjusting the sensitivity of anomaly detection based on thermal inertia parameters. Specifically, a probability distribution model bound to spatial location can be used to determine abnormal pixels through confidence probability calculation.

[0080] In this embodiment, visible light images are used to identify moving objects forming occlusion areas, and the actual occlusion range is calibrated by combining the low-temperature region contour from infrared thermal images. A pre-trained thermal inertia model is used to predict the theoretical temperature distribution under undisturbed conditions, generating a residual thermal map reflecting the deviation between the actual and theoretical values. Based on the predicted occlusion mask, the residual data is segmented into occluded and unoccluded regions, and interference analysis and anomaly detection are performed separately for each. For occluded regions, the actual occlusion mask is used to determine whether the anomaly is caused by occlusion, and different risk level discrimination flags are set. For unoccluded regions, the anomaly threshold is dynamically adjusted based on thermal inertia parameters to identify temperature anomalies exceeding the normal fluctuation range. Finally, the two types of analysis results are fused to generate accurate alarm commands, avoiding misjudgments caused by instantaneous interference or occlusion.

[0081] In this embodiment, the proposed solution achieves dynamic analysis of shading areas while preserving the continuity of the temperature field through multimodal data fusion and thermophysical model prediction. Furthermore, this solution establishes a dynamic threshold mechanism based on thermal conduction characteristics to improve the accuracy of anomaly detection. Simultaneously, through residual segmentation and interference analysis, this solution maintains monitoring continuity even when shading occurs. Thus, this application effectively distinguishes between actual equipment faults and instantaneous external interference, maintaining the continuity of temperature field analysis during shading and reducing the risk of false alarms and missed alarms. Dynamically adjusting the detection threshold through thermal inertia parameters improves the accuracy of anomaly identification in different areas. Multimodal data fusion and physical model embedding enhance the system's adaptability to complex industrial scenarios, providing reliable assurance for the safe operation of high-temperature equipment.

[0082] In one feasible implementation, the method further includes: acquiring a sequence of continuous, unobstructed infrared thermal images of the target device under normal operating conditions to form an infrared thermal image training sequence set; constructing a spatiotemporal joint feature extraction network including convolutional layers and LSTM layers, inputting the infrared thermal image training sequence set into the spatiotemporal joint feature extraction network, and outputting a spatiotemporal feature tensor; establishing a partial differential equation for heat conduction on the device surface based on the specific heat capacity data, thermal conductivity data, and three-dimensional structural model of the target device's metallic material; calculating the theoretical surface temperature gradient tensor at the corresponding position of the spatiotemporal feature tensor using the partial differential equation for heat conduction on the device surface; calculating the mean square error between the temperature gradient predicted by the hidden layer of the spatiotemporal joint feature extraction network and the theoretical surface temperature gradient tensor to generate a physical regularization loss value; calculating the sum of absolute errors between the predicted thermal image and the actual thermal image pixels finally output by the spatiotemporal joint feature extraction network to generate a pixel-level reconstruction loss value; weightedly fusing the physical regularization loss value and the pixel-level reconstruction loss value, updating the weights of the spatiotemporal joint feature extraction network through a backpropagation algorithm until the loss converges, and generating the pre-trained thermal inertia prediction model.

[0083] In this embodiment, the spatiotemporal joint feature extraction network refers to a composite structure composed of a convolutional neural network and a long short-term memory network. Specifically, it can be implemented by using three-dimensional convolutional layers to extract spatial features and bidirectional LSTM layers to extract temporal dependencies, used to capture the spatiotemporal evolution patterns in the heat map sequence. The partial differential equation of heat conduction refers to a differential equation established based on the material's thermal property parameters. Specifically, it can be implemented by using the finite volume method to discretize the three-dimensional model of the device into a mesh, and then solving the discretized equation to obtain the theoretical temperature gradient distribution. The physical regularization loss value is a measure of the difference between the model's predicted temperature gradient and the theoretical calculated value. Specifically, it can be calculated using the mean square error function to ensure that the neural network learning process conforms to the physical laws of heat conduction. The pixel-level reconstruction loss value is a measure of the difference between the predicted heat map and the actual heat map at the pixel level. Specifically, it can be calculated by comparing the differences pixel by pixel using the absolute value error function to ensure the accuracy of heat map reconstruction.

[0084] In this embodiment, in the high-temperature industrial equipment monitoring scenario, an unobstructed infrared thermal image sequence during normal equipment operation is first collected as training data. A spatiotemporal joint feature extraction network extracts spatial features from a single frame of the thermal image through a 3D convolutional layer, and then captures the temporal variation patterns between multiple frames of thermal images through an LSTM layer, outputting a feature tensor that fuses spatiotemporal information. Simultaneously, based on the thermal conductivity, specific heat capacity parameters, and 3D structural model of the equipment material, a heat conduction equation describing the surface temperature field changes is established, and theoretical temperature gradient data is generated through numerical calculation methods. During model training, not only is the pixel difference between the predicted thermal image and the actual thermal image calculated, but the difference between the predicted temperature gradient and the theoretical gradient is also introduced as a physical constraint. The network parameters are updated by weighted fusion of two types of loss functions, ensuring that the model can accurately reconstruct the thermal image while conforming to the physical laws of heat conduction.

[0085] In this embodiment, the proposed solution generates a theoretical temperature gradient by establishing a heat conduction equation, and uses it as a regularization term to constrain the training of the neural network. This ensures that the model learning process simultaneously satisfies data fitting and physical laws, improving the physical rationality of thermal inertia prediction. It solves the problem of temperature prediction distortion caused by ignoring the thermophysical characteristics of equipment in the prior art. By integrating physical equation constraints with deep learning models, the generated interference-free heat map data is more consistent with the actual heat conduction laws, providing a reliable temperature field benchmark for subsequent residual analysis, thereby reducing the risk of false alarms caused by thermal inertia modeling errors.

[0086] In one feasible implementation, the step of establishing a partial differential equation for heat conduction on the surface of the target device based on the specific heat capacity data, thermal conductivity data, and three-dimensional structural model of the target device's metallic material includes: obtaining a dataset of the thermal conductivity data of each component in a preset three-dimensional model of the target device; constructing a heat conduction topology network based on the thermal conductivity data and the component connection relationships of the target device; deriving a surface temperature distribution equation based on the heat conduction topology network data to generate simplified heat conduction equation data; and discretizing the simplified heat conduction equation data into the partial differential equation for heat conduction on the surface of the device.

[0087] In this embodiment, the material thermal conductivity dataset refers to the set of thermal conductivity performance parameters of the corresponding metallic materials for each component of the target device. Specifically, this can be achieved by matching the material property database with the component labels of the 3D model, used to determine the differences in thermal conductivity between different metallic components. The thermal conduction topology network data refers to graph-structured data reflecting the thermal conduction path relationships between device components. Specifically, it can be generated using a 3D model adjacency relationship combined with a weld point detection algorithm, used to construct a physical path model for heat flow transfer. The simplified thermal conduction equation data refers to the mathematical expression of the temperature field after ignoring secondary influencing factors. Specifically, it can be derived using Fourier's law combined with boundary condition constraints, used to reduce computational complexity while retaining the core thermal conduction laws. The partial differential equation for thermal conduction on the device surface refers to the discretized temperature gradient calculation model. Specifically, the finite volume method can be used to mesh the continuous equations, facilitating numerical computer solutions.

[0088] In this embodiment, during the heat conduction modeling of high-temperature equipment, the material identification information of each metal component is first extracted from the 3D model, and the thermal conductivity set of the corresponding materials is obtained by querying a material database. A heat conduction path topology is constructed based on the physical connection relationships between components; for example, the weld surfaces of flanges and pipes are marked as heat conduction boundary nodes. Based on the connection methods of each node in the topology network, a temperature gradient equation is established using Fourier's law of heat conduction, simplifying the equation by neglecting secondary factors such as environmental radiation. Finally, the continuous equation is discretized into a set of difference equations suitable for iterative computer solving using the finite volume method, forming a partial differential equation for heat conduction on the equipment surface that can be directly used for training a thermal inertia prediction model.

[0089] In this embodiment, the proposed solution constructs a heat conduction topology network to accurately describe the heat conduction path relationships between various components of the equipment. Combined with discretization processing to preserve structural details, this significantly improves the physical accuracy of the heat conduction model and solves the problem of heat conduction modeling errors caused by neglecting equipment structural details in existing technologies. By establishing a topology network model based on the actual three-dimensional structure, the differences in heat conduction characteristics at key locations such as flange connections and weld areas can be accurately captured, providing more precise physical constraints for subsequent thermal inertia prediction models. In the test of predicting the temperature of the flange sealing surface of a high-temperature reactor, this method reduces the calculation error of temperature gradients at key locations and improves the reliability of anomaly detection.

[0090] In one feasible implementation, the step of inputting the current-time infrared thermal image data from the visible light image data and the infrared thermal image sequence data into a pre-trained occlusion detection model for processing to generate actual occlusion mask data and predicted occlusion mask data includes: performing target detection on the visible light image data to generate moving object bounding box data; extrapolating the predicted occlusion mask data based on the motion trajectory of the bounding box data; analyzing regions with temperatures below a threshold in the infrared thermal image data to generate low-temperature region contour data; and fusing the moving object bounding box data and the low-temperature region contour data to generate calibrated actual occlusion data.

[0091] In this embodiment, the moving object bounding box data refers to the smallest bounding rectangle of a dynamic object in a visible light image, identified by a target detection algorithm. Specifically, it can be implemented using YOLO or Faster R-CNN algorithms to capture moving objects that may cause occlusion in the device monitoring scene. The predicted occlusion mask data refers to a binary mask predicting the occlusion area at future times based on the object's motion trajectory. Specifically, it can be generated by extrapolating the trajectory point set using Kalman filtering or multinomial fitting to create a convex hull expansion region, used to predict the occlusion range in advance. The low-temperature region contour data refers to the set of boundary coordinates of low-temperature anomaly regions extracted based on the temperature distribution of infrared thermograms. Specifically, it can be implemented using a region growing algorithm combined with temperature gradient threshold segmentation, used to identify local temperature anomalies caused by heat absorption or reflection from occluding objects.

[0092] In this embodiment, moving objects in the visible light image are located in real time using a target detection algorithm. Their motion trajectory is modeled using historical frame data to simulate their movement path within a preset time period, thus forming a predicted occlusion region. Simultaneously, areas in the infrared thermal image with temperatures significantly lower than the surrounding environment are extracted as low-temperature contours and spatially superimposed and fused with the bounding box from the visible light detection, eliminating misjudged areas caused by thermal reflection or vapor interference. The resulting calibrated actual occlusion data includes both the real-time position of the dynamic object and the correction of the occlusion boundary based on thermal imaging characteristics, ensuring the physical consistency of the occlusion mask.

[0093] Understandably, existing technologies generally rely on single sensor data to determine occlusion areas, such as identifying occlusion solely through temperature abrupt changes in infrared thermal images, which is susceptible to misjudgment due to environmental thermal noise interference. This solution, however, integrates visible light dynamic object detection with infrared low-temperature region analysis. It utilizes the high resolution of visible light to capture the contours of moving objects while combining thermal imaging data to eliminate non-occlusion low-temperature interference, achieving dual-modal cross-validation of occlusion areas. Thus, this application effectively solves the misjudgment problem caused by instantaneous occlusion and thermal reflection interference, improving the spatiotemporal consistency of occlusion detection. The collaborative processing of visible light and infrared data can distinguish between real moving objects and thermal environmental noise, avoiding misidentification of low-temperature areas caused by vapor diffusion or metal reflection, ensuring that the actual occlusion mask accurately reflects the real physically occluded area on the device surface, providing a reliable data foundation for subsequent residual analysis.

[0094] In one feasible implementation, the step of extrapolating the predicted occlusion mask data based on the motion trajectory of the bounding box data includes: establishing motion vector data based on the bounding box data; simulating a trajectory for a preset duration based on the motion vector data to generate corresponding trajectory point set data; and performing convex hull expansion on the trajectory point set data to generate the predicted occlusion mask data.

[0095] In this embodiment, motion vector data refers to the motion direction and velocity parameters calculated from the position changes of the bounding box in a continuous time series. Specifically, it can be implemented using optical flow or Kalman filtering algorithms to capture the dynamic change characteristics of moving objects. Trajectory point set data refers to the predicted set of positions of the bounding box center points within a future time period, calculated based on the motion vectors. Specifically, it can be implemented using linear extrapolation or polynomial fitting methods to simulate the motion trend of objects. Convex hull expansion refers to the operation of expanding the smallest convex polygon of the trajectory point set data outward to form a covering region. Specifically, it can be achieved by increasing the buffer radius using computational geometry algorithms to cover the potential spatial range of the object's motion trajectory.

[0096] In this embodiment, the continuous motion trajectory of the bounding box of a moving object is modeled as a motion vector, and an extrapolation algorithm is used to predict the set of possible future locations. The spatial distribution formed by the trajectory point set is expanded into a continuous region using a convex hull algorithm, generating a predictive mask that covers the potential occlusion range in the future. For example, when a maintenance tool approaches a device at a constant speed, its motion vector is extracted and used to generate trajectory points for the next five seconds. After convex hull expansion, a predictive occlusion region covering the tool's movement path is formed.

[0097] In this embodiment, by extrapolating the trajectory point set and expanding the convex hull, the occlusion prediction within the future time window is generated in advance, which solves the problem of incomplete mask coverage caused by changes in the object's moving speed and avoids misjudgment of temperature data due to prediction deviation. This allows the present application to accurately predict the occlusion range of moving objects in the future time period, mark potentially affected areas in advance, avoid temperature data distortion and misjudgment due to instantaneous occlusion, and provide accurate region segmentation basis for subsequent residual analysis, thereby improving the anti-interference capability and alarm accuracy of the high-temperature equipment status monitoring system.

[0098] In one feasible implementation, the interference discrimination flag data includes first interference discrimination flag data and second interference discrimination flag data; the step of combining the residual data of the occluded area with the actual occlusion mask data to perform interference analysis and generate interference discrimination flag data includes: when the actual occlusion mask data covers the occluded area, setting low confidence flag data and using it as the first interference discrimination flag data; when the actual occlusion mask data does not cover the occluded area, calculating the maximum residual value data of the occluded area residual data, and determining the corresponding tolerance threshold data based on the thermal inertia parameter data; when the maximum residual value data of the area exceeds the tolerance threshold data, setting high anomaly risk flag data and using it as the second interference discrimination flag data.

[0099] In this embodiment, the actual occlusion mask data refers to the binarized marker data of the occlusion area generated by analyzing visible light images and infrared thermal images. Specifically, it can be generated by fusing target detection algorithms and temperature threshold segmentation algorithms to accurately identify the physical location of the device surface being occluded by external objects at the current moment. The occluded area residual data refers to the local temperature deviation data obtained by calculating the difference between the interference-free thermal image and the current infrared thermal image. Specifically, it can be generated by pixel-level difference operations to characterize the degree of deviation between the actual temperature of the occluded area and the theoretically predicted value. The tolerance threshold data refers to the residual allowable range data that is dynamically adjusted based on the thermal inertia parameters of the device material. Specifically, it can be generated by weighting the parameters of the heat conduction equation with historical residual statistics to distinguish between normal thermal fluctuations and abnormal temperature abrupt changes.

[0100] In this embodiment, when the actual obstruction mask completely covers the obstructed area, the system determines that the temperature data of that area is affected by the continuous obstruction from external objects. At this time, a low confidence flag is set to exclude the abnormal judgment of that area. When the actual obstruction mask does not completely cover the area, the system extracts the maximum residual value of that area and dynamically calculates the tolerance threshold in combination with the current thermal inertia parameters of the equipment. For example, in a high-temperature pipeline scenario, if the residual of a certain area suddenly increases but the thermal inertia parameters indicate that the part should be in a state of rapid heat dissipation, the tolerance threshold is automatically increased to avoid misjudgment. If the residual distribution pattern matches the flange sealing surface structure, a structural association risk flag is triggered. This achieves refined classification and discrimination of abnormal states in obstructed areas.

[0101] In this embodiment, the proposed solution dynamically adjusts the threshold using thermal inertia parameters and combines this with spatial distribution feature matching to effectively distinguish between actual equipment faults and transient interference such as steam and smoke. For example, in a scenario where condensate droplets slide off, existing technologies may misjudge the residual temperature of the droplets as a leak point. However, this solution can accurately identify the actual leak signal by analyzing the spatial matching degree between the residual distribution pattern and the flange structure. Through the above technical solution, this application can effectively improve the accuracy of anomaly detection under operating conditions with dynamic obstruction interference. Specifically, it automatically suppresses false alarms in areas where obstruction is completely covered, avoiding unnecessary maintenance operations; dynamically adjusts the judgment threshold using thermal inertia parameters in partially obstructed areas, reducing the false alarm rate caused by environmental reflection; and accurately identifies actual faults in key areas such as flange connections through matching analysis of residual distribution patterns and equipment structural features.

[0102] In one feasible implementation, the step of setting high anomaly risk indicator data includes: extracting spatial distribution feature data of the residual data of the obscured area; comparing the spatial distribution feature data with the structural features of a preset target equipment 3D model; when the spatial distribution feature data matches the flange structure or bearing structure, increasing the risk level to generate structural association risk indicator data; when the spatial distribution feature data is in a random discrete form, generating transient interference indicator data; and merging the structural association risk indicator data and the transient interference indicator data to form the high anomaly risk indicator data.

[0103] In this embodiment, spatial distribution feature data refers to the two-dimensional distribution pattern of residual values ​​in the obscured area. Specifically, it can be achieved by using morphological analysis algorithms to extract the geometric shape and connected component distribution features of the residual area, used to distinguish between anomalies related to equipment structure and random external interference. The structural features of the target equipment's three-dimensional model refer to the three-dimensional geometric position information of key components such as flanges and bearings on the equipment surface. Specifically, this can be obtained through analysis of a computer-aided design model, used to match the spatial correspondence between the residual distribution and potentially faulty components. Flange or bearing structures refer to key connecting components in high-temperature equipment that are prone to thermal fatigue or mechanical wear. Specifically, the structural features of high-frequency fault areas can be screened based on equipment maintenance records, used to prioritize the identification of abnormal signals in high-risk areas. Random discrete morphology refers to the spatial distribution of residual areas that are multi-point dispersed and without a fixed geometric pattern. Specifically, it can be quantified by calculating the area-to-perimeter ratio and the standard deviation of the centroid spacing of the residual area, used to identify noise signals generated by transient external interference.

[0104] In this embodiment, the residual data of the obscured area is extracted based on its spatial distribution features, and its morphological characteristics are matched with the structural positions of key components such as flanges and bearings in the 3D model of the equipment. If the spatial distribution of the residual area overlaps with the geometric position of the flange or bearing more than a preset threshold, the anomaly is determined to be related to a potential failure of a key component of the equipment, triggering a structural association risk flag. If the residual area is scattered across multiple points without a fixed geometric pattern, it is determined to be transient interference such as steam smoke or splashing molten slag, triggering a transient interference flag. The two flag data are merged to form a comprehensive risk level, providing a more discriminative basis for subsequent alarm decisions.

[0105] In this embodiment, the proposed solution effectively improves the accuracy of identifying faults in critical components by introducing a three-dimensional structural feature matching mechanism for the equipment and combining it with residual spatial distribution morphology analysis. Through the above technical solution, this application can accurately distinguish between actual structural faults in the equipment and residual signals generated by external transient interference, avoiding misjudging abnormal thermal expansion at flange connections as steam interference, and preventing random residuals caused by condensate dripping from being falsely reported as bearing overheating faults. This solution can reduce the number of invalid maintenance work orders due to misjudgments and improve the detection rate of early-stage faults in critical components.

[0106] In one feasible implementation, the step of performing dynamic threshold analysis by combining the unobstructed area residual data with the thermal inertia parameter data to generate abnormal pixel coordinate data includes: for each pixel in the unobstructed area residual data, matching its spatial location coordinate data with corresponding thermal inertia parameter data; constructing a probability distribution model bound to spatial location based on the unobstructed residual sample dataset under historical normal operating conditions of the device; adjusting the parameters of the probability distribution model according to the thermal inertia parameter data to generate dynamic adaptive threshold data; calculating the confidence probability data of the current pixel residual value in the dynamic adaptive threshold data; when the confidence probability data is lower than a preset risk threshold, marking the pixel as an abnormal candidate point data; and summarizing all abnormal candidate point data to generate the abnormal pixel coordinate data.

[0107] In this embodiment, thermal inertia parameter data refers to physical parameters reflecting the thermal response delay characteristics of equipment materials. Specifically, it can be implemented using the thermal conductivity and specific heat capacity data of metal components, used to characterize the hysteresis effect of temperature changes in different regions. The probability distribution model refers to a statistical distribution model constructed based on historical residual data. Specifically, it can be implemented using a Gaussian mixture model or kernel density estimation method, used to describe the spatial distribution law of residual data under normal operating conditions. Dynamic adaptive threshold data refers to the anomaly judgment boundary value dynamically adjusted according to the thermal inertia parameters. Specifically, it can be implemented by adjusting the standard deviation multiple or quantile parameters of the probability distribution model, used to adapt to the differences in thermal conduction characteristics in different regions. Confidence probability data refers to the probability of the current residual value appearing in the probability distribution model. Specifically, it can be implemented by calculating the cumulative probability density function value of the residual value in the distribution model, used to quantify the anomaly risk level.

[0108] In this embodiment, during the operation of high-temperature equipment, the temperature changes in unobstructed areas exhibit spatial variability due to the thermal inertia of the equipment materials. By associating the residual data of each pixel with its corresponding thermal inertia parameter, the temperature fluctuation range of that area under normal operating conditions can be accurately reflected. The probability distribution model constructed based on historical residual samples can capture the statistical regularity of residual data at different locations, while the introduction of the thermal inertia parameter allows the model parameters to be dynamically adjusted according to the thermal response characteristics of the material. For example, in areas with high thermal inertia, the temperature change rate is slower, and the dynamic threshold is appropriately relaxed to reduce false alarms; in areas with low thermal inertia, the temperature change rate is faster, and the dynamic threshold is tightened to improve detection sensitivity. When the confidence probability of the real-time residual value in the adjusted probability distribution is lower than the preset threshold, it indicates that the possibility of the pixel's temperature deviating from the normal pattern exceeds the allowable range, and it needs to be marked as an abnormal candidate point. The finally generated abnormal pixel coordinate data provides accurate spatial positioning information for subsequent alarm decisions.

[0109] Compared to existing technologies, traditional methods typically employ fixed thresholds or sliding window thresholds based on global statistics, which cannot adapt to the differences in thermal conductivity characteristics across different areas of the equipment surface. For example, at flange connections, due to the high thermal resistance of metal contact and the slow rate of temperature change, using a uniform threshold may lead to missed anomalies; while in thin-walled areas, due to low thermal inertia and large temperature fluctuations, fixed thresholds are prone to false alarms. This method effectively solves the problem of decreased detection accuracy caused by spatial heterogeneity by dynamically adjusting local thresholds in conjunction with thermal inertia parameters. Simultaneously, it utilizes a probability distribution model to quantify anomaly risks, avoiding the risk of misjudgment caused by simple threshold determination. Through the above technical solution, this application can dynamically optimize the anomaly detection threshold based on the differences in thermal conductivity characteristics across different areas of high-temperature equipment, solving the problems of false alarms and missed alarms caused by neglecting the influence of thermal inertia in traditional methods. By combining residual data with a probability distribution model, normal temperature fluctuations can be accurately distinguished from true anomaly signals, improving the spatial positioning accuracy of anomaly detection and providing a reliable basis for subsequent alarm decisions.

[0110] In one feasible implementation, the step of fusing the interference discrimination flag data and the abnormal pixel coordinate data to generate device status alarm command data includes: inputting the abnormal pixel coordinate data into a clustering algorithm to generate abnormal region block coordinate data; marking the corresponding abnormal region blocks according to the interference discrimination flag data to generate a risk labeling dataset; matching the risk labeling dataset with a preset alarm rule base to generate initial alarm command data; correcting the initial alarm command data by combining it with device operating status data, and outputting the final device status alarm command data.

[0111] In this embodiment, the clustering algorithm refers to the technique of grouping discretely distributed abnormal pixels according to spatial proximity. Specifically, it can be implemented using density clustering or hierarchical clustering algorithms to aggregate scattered abnormal points into spatially continuous regions. The risk labeling dataset refers to the data set with interference discrimination labels attached to each abnormal region block. This can be achieved through coordinate mapping and label binding, used to distinguish abnormal regions from different risk sources. The alarm rule base refers to the set of preset alarm conditions and response strategies. This can be stored and matched using a rule engine or decision tree model, used to convert risk labels into executable alarm commands. The equipment operating status data refers to the set of current equipment operating parameters, specifically including load rate, runtime, and pressure value, used to adjust alarm commands for environmental adaptability.

[0112] In this embodiment, abnormal pixel coordinate data is clustered to form spatially continuous abnormal region blocks. Interference discrimination flag data is associated with the corresponding region blocks to form risk labeling information. The risk labeling dataset is matched with the alarm rule base to generate a preliminary alarm command. This command is further combined with real-time equipment operating parameters for logical verification and priority adjustment. For example, when the equipment is operating under high load, the alarm threshold for abnormal temperature regions is dynamically lowered to improve sensitivity. The resulting final alarm command includes both the spatial location information of the abnormal region and integrates the interference discrimination results with the equipment's real-time operating parameters.

[0113] In some specific implementations, the clustering algorithm can employ the DBSCAN algorithm based on Euclidean distance, setting the neighborhood radius to 5 pixels and the minimum number of points to 3. The alarm rule base can be configured to immediately trigger a level-one alarm when a structural association risk indicator exists, and to perform a secondary verification after a 3-second delay when a transient interference indicator exists. Equipment operating status data can be obtained in real time from the PLC control system via the OPC protocol, used to adjust the response level of alarm commands.

[0114] Understandably, existing technologies either discard abnormal data directly or use linear interpolation compensation when occlusion occurs, leading to the masking of true fault signals. This solution preserves the spatial correlation of abnormal areas through cluster analysis and performs multi-dimensional verification by combining interference indicators and equipment status, avoiding missed fault reports due to data discarding. For example, traditional methods directly filter transient high-temperature areas caused by molten slag splashing, while this solution distinguishes between molten slag interference and actual overheating through risk labeling, triggering an alarm only when the interference indicator is absent. In this way, this application can accurately distinguish between actual equipment faults and abnormal areas caused by transient external interference, dynamically adjusting the alarm strategy based on the real-time operating status of the equipment, effectively reducing the false alarm rate caused by transient occlusion or environmental reflection, while avoiding the risk of missed reports of actual faults due to data compensation processing, thus improving the alarm accuracy and reliability of the high-temperature industrial equipment status monitoring system.

[0115] In this embodiment, an artificial intelligence-based condition monitoring method is applied to high-temperature industrial equipment. By fusing visible light and infrared thermal image data, and combining thermal inertia prediction models and dynamic threshold analysis technology, it can effectively distinguish between real equipment faults and transient external interference, thereby improving the accuracy and reliability of condition monitoring of high-temperature industrial equipment and reducing false alarm and missed alarm rates.

[0116] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the state monitoring method based on artificial intelligence in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0117] This application also provides an artificial intelligence-based condition monitoring system for use in high-temperature industrial equipment, see reference. Figure 2 The artificial intelligence-based state monitoring system includes: a memory 10, a processor 20, and an artificial intelligence-based state monitoring program stored on the memory 10 and executable on the processor 20. The artificial intelligence-based state monitoring program is configured to implement the steps of the artificial intelligence-based state monitoring method.

[0118] The AI-based condition monitoring system provided in this application, employing the AI-based condition monitoring method described in the above embodiments, can improve the accuracy of condition monitoring for high-temperature industrial equipment and reduce false alarm and missed alarm rates. Compared with the prior art, the beneficial effects of the AI-based condition monitoring system provided in this application are the same as those of the AI-based condition monitoring method provided in the above embodiments, and other technical features of the AI-based condition monitoring system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0119] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0120] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. All equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. A condition monitoring method based on artificial intelligence, applied to high-temperature industrial equipment, characterized in that, The method includes: Acquire visible light image data and infrared thermal image sequence data of the target device; The infrared thermal image sequence data is input into a pre-trained thermal inertia prediction model for processing to generate interference-free thermal image data and thermal inertia parameter data. The current infrared thermal image data from the visible light image data and the infrared thermal image sequence data is input into the pre-trained occlusion detection model for processing to generate actual occlusion mask data and predicted occlusion mask data. Based on the interference-free thermal image data and the infrared thermal image data at the current moment, residual thermal image data is calculated and generated; Based on the predicted occlusion mask data, the residual heatmap data is segmented into residual data of the occluded area and residual data of the unoccluded area; The residual data of the occluded area is combined with the actual occlusion mask data to perform interference analysis and generate interference discrimination flag data. The residual data of the unobstructed area is combined with the thermal inertia parameter data to perform dynamic threshold analysis and generate abnormal pixel coordinate data. By fusing the interference discrimination flag data and the abnormal pixel coordinate data, device status alarm command data is generated.

2. The artificial intelligence-based state monitoring method as described in claim 1, characterized in that, The method further includes: Acquire a sequence of continuous, unobstructed infrared thermal images of the target device under normal operating conditions to form an infrared thermal image training sequence set; A spatiotemporal joint feature extraction network containing convolutional layers and LSTM layers is constructed. The infrared thermal image training sequence set is input into the spatiotemporal joint feature extraction network, and the spatiotemporal feature tensor is output. Based on the specific heat capacity data, thermal conductivity data and three-dimensional structural model of the target equipment's metallic material, a partial differential equation for heat conduction on the equipment surface is established. The theoretical surface temperature gradient tensor at the location corresponding to the spatiotemporal characteristic tensor is calculated using the partial differential equation of heat conduction on the surface of the device. The mean square error between the temperature gradient predicted by the hidden layer of the spatiotemporal joint feature extraction network and the theoretical surface temperature gradient tensor is calculated to generate a physical regularization loss value. The sum of absolute errors between the predicted heatmap and the actual heatmap pixels in the final output of the spatiotemporal joint feature extraction network is calculated to generate pixel-level reconstruction loss values. The weighted fusion of physical regularization loss and pixel-level reconstruction loss is used to update the weights of the spatiotemporal joint feature extraction network through backpropagation until the loss converges, thereby generating the pre-trained thermal inertia prediction model.

3. The artificial intelligence-based state monitoring method as described in claim 2, characterized in that, Based on the specific heat capacity data, thermal conductivity data, and three-dimensional structural model of the target equipment's metallic material, the steps for establishing the partial differential equation for heat conduction on the equipment surface include: Obtain the data set of thermal conductivity coefficients of the materials of each component in the preset 3D model of the target device; Construct heat conduction topology network data based on the material thermal conductivity dataset and the component connection relationships of the target device; Based on the heat conduction topology network data, the surface temperature distribution equation is derived, and simplified heat conduction equation data is generated. The simplified heat conduction equation data is discretized into a partial differential equation for heat conduction on the surface of the device.

4. The artificial intelligence-based state monitoring method as described in claim 1, characterized in that, The steps of inputting the current-time infrared thermal image data from the visible light image data and infrared thermal image sequence data into a pre-trained occlusion detection model for processing to generate actual occlusion mask data and predicted occlusion mask data include: Target detection is performed on the visible light image data to generate bounding box data for moving objects; Based on the motion trajectory of the bounding box data, the predicted occlusion mask data is generated by extrapolation. Analyze the regions with temperatures below a threshold in the infrared thermal image data to generate low-temperature region contour data; By fusing the bounding box data of the moving object with the contour data of the low-temperature region, calibrated actual occlusion data is generated.

5. The artificial intelligence-based state monitoring method as described in claim 4, characterized in that, The step of extrapolating the predicted occlusion mask data based on the motion trajectory of the bounding box data includes: Motion vector data is established based on the bounding box data; Based on the motion vector data, a trajectory for a future preset duration is simulated, and corresponding trajectory point set data is generated; The trajectory point set data is expanded by convex hull to generate the predicted occlusion mask data.

6. The artificial intelligence-based state monitoring method as described in claim 1, characterized in that, The interference discrimination flag data includes first interference discrimination flag data and second interference discrimination flag data; The steps of combining the residual data of the occluded area with the actual occlusion mask data to perform interference analysis and generate interference discrimination flag data include: When the actual occlusion mask data covers the occluded area, a low confidence flag data is set and used as the first interference discrimination flag data. When the actual occlusion mask data does not cover the occluded area, the maximum residual value of the residual data of the occluded area is calculated, and the corresponding tolerance threshold data is determined based on the thermal inertia parameter data. When the maximum residual value of the area exceeds the tolerance threshold data, a high anomaly risk flag is set and used as the second interference discrimination flag data.

7. The artificial intelligence-based state monitoring method as described in claim 6, characterized in that, The steps for setting high anomaly risk flags include: Extract the spatial distribution feature data of the residual data in the occluded area; Compare the spatial distribution feature data with the structural features of the preset target device 3D model; When spatial distribution characteristic data matches flange structure or bearing structure, the risk level is increased to generate structural associated risk indicator data. When the spatial distribution characteristic data is in a random and discrete form, transient interference marker data is generated; The structurally correlated risk indicator data and the transient interference indicator data are combined to form the high anomaly risk indicator data.

8. The artificial intelligence-based state monitoring method as described in claim 1, characterized in that, The steps for generating abnormal pixel coordinate data by combining the residual data of the unobstructed region with the thermal inertia parameter data for dynamic threshold analysis include: For each pixel in the residual data of the unobstructed area, the corresponding thermal inertia parameter data is matched according to its spatial location coordinate data; Based on the unobstructed residual sample dataset of the device under historical normal operating conditions, a probability distribution model bound to spatial location is constructed. The parameters of the probability distribution model are adjusted based on the thermal inertia parameter data to generate dynamic adaptive threshold data; Calculate the confidence probability data of the current pixel residual value in the dynamic adaptive threshold data; When the confidence probability data is lower than the preset risk threshold, the pixel is marked as an abnormal candidate point data; The abnormal pixel coordinate data is generated by summarizing all abnormal candidate point data.

9. The artificial intelligence-based state monitoring method as described in claim 1, characterized in that, The step of integrating the interference discrimination flag data and the abnormal pixel coordinate data to generate device status alarm command data includes: The abnormal pixel coordinate data is input into a clustering algorithm to generate abnormal region block coordinate data. Based on the interference discrimination flag data, mark the corresponding abnormal region blocks to generate a risk labeling dataset; Initial alarm command data is generated by matching the risk labeling dataset with a preset alarm rule base. The initial alarm command data is corrected by combining the equipment operating status data, and the final equipment status alarm command data is output.

10. An artificial intelligence-based condition monitoring system, applied to high-temperature industrial equipment, characterized in that, The AI-based state monitoring system includes: a memory, a processor, and an AI-based state monitoring program stored in the memory and executable on the processor, wherein the AI-based state monitoring program is configured to implement the steps of the AI-based state monitoring method as described in any one of claims 1 to 9.

Images