Bearing operation state detection method and system based on infrared thermal imaging and temperature field modeling

Abstract

The application relates to the technical field of intelligent nondestructive testing and fault diagnosis, in particular to a bearing operation state detection method and system based on infrared thermal imaging and temperature field modeling; the method comprises the following steps: acquiring an infrared thermal image sequence containing space-time temperature distribution information; constructing a generative model with a built-in thermodynamic conduction law as a prior constraint; inputting the image sequence into the model to generate an initial temperature field, and calculating physical consistency entropy representing the conflict degree between visual features and physical facts; when the entropy value is less than a threshold value, a high-definition reconstruction mode is executed to generate an internal temperature field inversion image with pixel-level precision; when the entropy value is greater than or equal to the threshold value, a heat flow topology deduction mode is executed to suppress texture generation and generate a heat risk topology graph based on the heat flow conservation principle; and the generated image or topology graph is used to locate internal thermal anomaly singular points; the application realizes physical constraint and dynamic mode switching, significantly reduces the false alarm rate caused by environmental thermal shock or shielding, and improves the reliability of fault detection results.

Description

Technical Field

[0001] This invention relates to the field of intelligent non-destructive testing and fault diagnosis technology, specifically to a bearing operating status detection method and system based on infrared thermal imaging and temperature field modeling. Background Technology

[0002] In the current high-speed train mechanical component operation status monitoring environment, the system typically deploys high-frame-rate infrared thermal imagers at the trackside or undercarriage to continuously collect infrared thermal image sequences of key components such as bearings, serving as thermal radiation analysis data with spatiotemporal dimensions. To detect anomalies in this data, existing solutions generally employ traditional computer vision algorithms or purely data-driven deep learning models, directly classifying the status based on surface radiation temperature values ​​or texture features. Although this solution has a certain detection efficiency under steady-state conditions, when the train encounters thermal shocks from unsteady environments such as rain or heavy snow, or when the component surface is obscured by oil sludge contaminants, an asymmetric thermal state of simultaneous surface rapid cooling and internal heating, as well as nonlinear distortion of local emissivity, will occur. Because existing visual models rely solely on image appearance features and lack the laws of thermodynamic conduction as physical prior constraints, they cannot dynamically identify such infrared visual deception phenomena. As a result, when the visual signal is distorted due to environmental interference, the model is prone to hallucinations or being unable to see through surface cold illusions, leading to serious missed detections of internal overheating faults.

[0003] Therefore, how to introduce physical laws to quantitatively evaluate the credibility of visual signals, and accurately invert the internal thermal risk distribution through the principle of heat flow conservation under extreme conditions of visual failure, thereby improving the robustness and accuracy of the detection system, has become an urgent technical problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a bearing operating status detection method based on infrared thermal imaging and temperature field modeling. This method can effectively avoid missed fault detection when infrared visual signals are distorted due to thermal shock or surface contaminant obstruction in unsteady environments. Furthermore, by introducing physical consistency entropy, it achieves adaptive switching of detection modes, thereby improving the system's adaptability and robustness in all-weather operating environments. Specifically, the technical solution of this invention is as follows:

[0005] A bearing operating condition detection method based on infrared thermal imaging and temperature field modeling includes:

[0006] Acquire an infrared thermal image sequence of the mechanical component under test, the infrared thermal image sequence containing spatiotemporal temperature distribution information of the component surface;

[0007] Based on the infrared thermal image sequence, a generative model constrained by physical information is constructed, and the generative model is embedded with the thermodynamic conduction law as a priori constraint.

[0008] The temperature field inversion and state determination process includes:

[0009] Step 1: Input the infrared thermal image sequence into the generative model, extract the spatiotemporal residual features of the image, and generate initial temperature field distribution data;

[0010] Step 2: Based on the degree of deviation between the initial temperature field distribution data and the thermodynamic conduction law, calculate the physical consistency entropy, which is used to characterize the degree of conflict between visual image features and thermodynamic physical facts;

[0011] Step 3: In response to the physical consistency entropy being less than a preset confidence threshold, execute a high-definition reconstruction mode to generate a pixel-level precision internal temperature field inversion image;

[0012] Step 4: In response to the physical consistency entropy being greater than or equal to the confidence threshold, execute the heat flow topology deduction mode, suppress the generation of image texture details, and generate a thermal risk topology map based on the principle of heat flow conservation.

[0013] Step 5: Based on the internal temperature field inversion image or the thermal risk topology map, locate the internal thermal anomaly singularity and output the operating status detection result of the tested mechanical component.

[0014] Preferably, step 2, calculating the physical consistency entropy, includes:

[0015] Extract the surface temperature gradient features from the initial temperature field distribution data;

[0016] Based on the aforementioned thermodynamic conduction law, calculate the theoretical heat flux density corresponding to the surface temperature gradient characteristics;

[0017] Calculate the residual vector between the theoretical heat flux density and the actual observed rate of change of thermal radiation in the infrared thermogram sequence;

[0018] The physical consistency entropy is generated based on the norm of the residual vector; wherein the value of the physical consistency entropy increases monotonically as the residual vector increases.

[0019] Preferably, step 3 involves performing a high-definition reconstruction mode, including:

[0020] Activate the super-resolution reconstruction module in the generative model;

[0021] With minimizing visual perception loss as the optimization objective, the initial temperature field distribution data is subjected to texture enhancement processing;

[0022] The mapping generates an internal temperature field inversion image with high spatial resolution, which is used to characterize the fine temperature gradient distribution inside the component.

[0023] Preferably, step 4 involves executing a heat flux topology derivation mode, including:

[0024] Freeze the texture generation layer in the generative model and call the preset thermal flow topology energy function as a constraint condition; and activate the thermal flow topology constraint layer;

[0025] With the goal of maximizing the probability of thermodynamic existence, the initial temperature field distribution data is fuzzified to remove high-frequency visual noise that violates the thermodynamic law of conduction.

[0026] Extract and amplify the low-frequency signal that conforms to the internal heat source diffusion law in the spatiotemporal residual features;

[0027] The thermal risk topology map is reconstructed and generated. The thermal risk topology map is used to characterize the skeleton structure of heat flow transmission inside the component and the probability distribution of heat sources.

[0028] Preferably, physical consistency entropy is used to identify infrared visual deception states, which include:

[0029] An asymmetric thermal state caused by unsteady thermal shock, resulting in both rapid surface cooling and internal heating.

[0030] The local emissivity nonlinear distortion state caused by surface contaminant shielding;

[0031] In the infrared visual deception state, the physical consistency entropy increases significantly, causing the system to determine that the heat flow topology deduction mode should be executed.

[0032] Preferably, the generative model employs a physically based adversarial generative network architecture, including:

[0033] A generator network is used to generate a predicted temperature field based on the infrared thermal image sequence;

[0034] A discriminator network is used to determine the visual realism of the predicted temperature field;

[0035] The physical constraint module is used to calculate the residual of the predicted temperature field on the partial differential equation of heat conduction, and feed the residual as a penalty term into the loss function of the generator network.

[0036] Preferably, locating the internal thermal anomaly singularity in step 5 includes:

[0037] In the internal temperature field inversion image or the thermal risk topology map, identify spatial coordinate points where the temperature value exceeds a preset safety threshold and the duration exceeds a preset time window.

[0038] A bearing operating condition detection system based on infrared thermal imaging and temperature field modeling includes:

[0039] The image acquisition module is used to acquire a sequence of infrared thermal images of the mechanical parts under test.

[0040] The model loading module is used to load generative models with physical information constraints.

[0041] The status detection module includes:

[0042] The feature extraction unit is used to input the infrared thermal image sequence into the generative model, extract the spatiotemporal residual features of the image, and generate initial temperature field distribution data.

[0043] The entropy calculation unit is used to calculate the physical consistency entropy based on the degree of deviation between the initial temperature field distribution data and the thermodynamic conduction law;

[0044] The mode switching unit is used to trigger a high-definition reconstruction mode in response to the physical consistency entropy being less than a preset confidence threshold; and to trigger a thermal flow topology inference mode in response to the physical consistency entropy being greater than or equal to the confidence threshold.

[0045] The image generation unit is used to generate an internal temperature field inversion image in the high-definition reconstruction mode, or to generate a thermal risk topology map in the heat flow topology extrapolation mode.

[0046] The result output unit is used to locate internal thermal anomaly singularities based on the internal temperature field inversion image or the thermal risk topology map, and output the operating status detection results of the tested mechanical component.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. This invention introduces physical consistency entropy as a dynamic evaluation index for the reliability of visual signals. By calculating the residual between the actual observed rate of change of thermal radiation and the theoretical heat flux density derived from the thermodynamic conduction law, the degree of conflict between visual image features and physical facts is quantified. This enables the system to actively identify infrared visual deception caused by thermal shock in unsteady environment or surface contaminant occlusion. Thus, in cases where the surface temperature is masked but the internal temperature is actually overheated, it avoids missed faults caused by blindly trusting visual appearances, significantly improving the robustness of the system under complex all-weather conditions.

[0049] 2. Based on the feedback of physical consistency entropy, this invention dynamically switches between two modes: high-definition reconstruction and thermal flow topology deduction. When the visual signal is reliable, the super-resolution reconstruction module is activated to enhance texture details, achieving pixel-level accurate capture of early weak fault signals such as minor peeling of bearing raceways. When the visual signal is distorted by strong interference, texture generation is automatically frozen and thermal flow topology constraints are activated to suppress high-frequency visual noise, and the internal thermal flow skeleton structure is inverted based on the principle of thermal flow conservation. This mechanism ensures that the system can provide precise fault diagnosis and can also see through surface cold artifacts under extreme conditions of visual failure, locking in the risk of internal heat accumulation.

[0050] 3. This invention adopts a generative model architecture constrained by physical information, explicitly embedding the Fourier partial differential equation of heat conduction as a prior constraint into the loss function of the generator network, forcing the neural network to strictly abide by the law of energy conservation during training and inference. Compared with pure data-driven deep learning models, this method effectively limits the generation of illusions that violate physical common sense, ensuring that the generated temperature field data is reasonable and conserved at the physical level, while reducing the model's dependence on massive amounts of high-quality labeled data and improving the model's generalization ability under unknown conditions.

[0051] 4. When locating internal thermal anomaly singularities, this invention employs a judgment logic that combines spatial amplitude and temporal duration. It not only requires that the temperature value or thermal risk probability exceeds a safety threshold, but also requires that the duration of the abnormal state exceeds a preset time window. This design can effectively filter out instantaneous electronic noise or random environmental interference from the sensor, ensuring that the system only responds to real faults with a certain thermal energy accumulation effect. Thus, in high-dynamic background environments such as high-speed train operation, it significantly reduces the false alarm rate and improves the reliability of fault detection results. Attached Figure Description

[0052] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2 This is a structural diagram of the system of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0056] Example 1:

[0057] Please see Figure 1A bearing operating status detection method based on infrared thermal imaging and temperature field modeling includes: acquiring an infrared thermal image sequence of the tested mechanical component, the sequence containing spatiotemporal temperature distribution information of the component surface; constructing a generative model with physical information constraints based on the infrared thermal image sequence, the generative model embedding the thermodynamic conduction law as a priori constraint; and executing a temperature field inversion and state determination process, including:

[0058] Step 1: Input the infrared thermal image sequence into the generative model, extract the spatiotemporal residual features of the image, and generate initial temperature field distribution data;

[0059] Step 2: Based on the degree of deviation between the initial temperature field distribution data and the thermodynamic conduction law, calculate the physical consistency entropy. The physical consistency entropy is used to characterize the degree of conflict between visual image features and thermodynamic physical facts.

[0060] Step 3: In response to the physical consistency entropy being less than the preset confidence threshold, execute the high-definition reconstruction mode to generate a pixel-level precision internal temperature field inversion image;

[0061] Step 4: In response to the physical consistency entropy being greater than or equal to the confidence threshold, execute the heat flow topology deduction mode, suppress the generation of image texture details, and generate a thermal risk topology map based on the principle of heat flow conservation.

[0062] Step 5: Based on the internal temperature field inversion image or thermal risk topology map, locate the internal thermal anomaly singularity and output the operating status detection results of the tested mechanical component.

[0063] This embodiment provides a bearing operating status detection method based on infrared thermal imaging and temperature field modeling, which aims to solve the problem of missed detection caused by infrared visual signal distortion when traditional computer vision faces unsteady environment thermal shock or surface contaminant occlusion. The core logic of this method is to introduce thermodynamic laws as a real-time arbiter to dynamically evaluate the physical credibility of visual images.

[0064] The system acquires a sequence of infrared thermal images of the mechanical components under test by using a high frame rate infrared thermal imager installed beside the train track or under the bogie area. The infrared thermal image sequence is obtained by continuous acquisition by the infrared thermal imager and physically means thermal radiation image data containing both time and spatial dimensions, with units of pixel grayscale value or radiation temperature value.

[0065] The system constructs a generative model constrained by physical information. This model embeds the thermodynamic law of conduction, namely Fourier's law of heat conduction and its corresponding partial differential equation, as a priori constraints. This aims to limit the illusions of the generative AI model and ensure that the generated temperature field is physically conserved. The generative model is not limited to a specific network architecture. In addition to the adversarial generative network based on physical information preferred in this embodiment, a variational autoencoder constrained by physical information can also be used. The sampling of the latent space is constrained by adding the residual term of the heat conduction equation to the reconstruction loss of the decoder; or a physical-guided diffusion probability model is used to integrate the guiding term of the physical gradient into the inverse denoising process to generate temperature field samples that conform to the thermodynamic distribution.

[0066] These models all embed thermodynamic laws into the network optimization objective or inference process explicitly or implicitly. The system executes a temperature field inversion and state determination process, inputting infrared thermal image sequences into the generative model, extracting the spatiotemporal residual features of the images through convolutional neural network layers, and generating initial temperature field distribution data accordingly. Based on this, the system calculates the physical consistency entropy, which, as a dimensionless scalar, quantifies the degree of conflict between visual image features and thermodynamic physical facts. In response to the physical consistency entropy being less than a preset confidence threshold, such as 0.3, the system determines that the current visual signal is reliable, executes a high-definition reconstruction mode, and uses super-resolution technology to generate an internal temperature field inversion image with pixel-level precision.

[0067] Conversely, in response to a physical consistency entropy greater than or equal to the confidence threshold, the system determines the presence of strong interference such as rain, snow, rapid cooling, or oil sludge blockage, and forcibly switches to the thermal flow topology deduction mode to suppress the generation of image texture details and generate a thermal risk topology map based on the principle of thermal flow conservation; based on the generated image or topology map, it locates internal thermal anomaly singularities and outputs the detection results.

[0068] This embodiment introduces physical consistency entropy as a dynamic switch. When high-speed trains are running under extreme conditions such as heavy rain or snow, and external cooling masks internal overheating, causing visual deception, the system can automatically abandon the pursuit of high-definition images and instead provide a risk topology based on heat flow conservation. This effectively prevents the underreporting of catastrophic mechanical failures caused by visual signal distortion.

[0069] Example 2:

[0070] Step 2 calculates the physical consistency entropy, including: extracting surface temperature gradient features from the initial temperature field distribution data; calculating the theoretical heat flux density corresponding to the surface temperature gradient features based on the thermodynamic conduction law; calculating the residual vector between the theoretical heat flux density and the actual observed rate of change of thermal radiation in the infrared thermogram sequence; and quantizing the physical consistency entropy based on the norm of the residual vector. The value of the physical consistency entropy increases monotonically with the increase of the residual vector.

[0071] This embodiment details the specific calculation logic of physical consistency entropy, aiming to quantify the contradiction between visual observation and physical laws; the system generates initial temperature field distribution data. In this study, the Sobel operator or the Laplacian operator is used to extract the surface temperature gradient features in the spatial dimension. Based on the thermodynamic laws of conduction, i.e., Fourier's law, calculate the theoretical heat flux density corresponding to the surface temperature gradient characteristics. The calculation formula is as follows:

[0072] ;

[0073] in, The thermal conductivity of the material of the component being tested is given in units of 1000 kJ / m². ;

[0074] The system calculates the residual vector between the theoretical heat flux density and the actual observed rate of change of thermal radiation in the infrared thermogram sequence. Before this calculation, radiation calibration is required; the actual observed rate of change of thermal radiation is calculated using differential operations. ;

[0075] in, Specifically defined as the equivalent temperature time-varying rate, that is, the rate of change of the surface temperature of the measured component over time after radiation calibration. The unit is This step converts the original infrared radiation grayscale changes into a time-varying temperature field with clear thermodynamic significance.

[0076] residual vector The calculation formula is:

[0077] ;

[0078] The physical meaning refers to the vector difference between the theoretically derived heat flux and the actual observed equivalent heat flux; where the observed heat flux vector... Defined as:

[0079] ;

[0080] in, The equivalent heat flux density amplitude, This serves as a unit vector for the reference heat flow direction; it is to avoid introducing incorrect direction information when rapid cooling due to rain or snow causes a reversal of the surface temperature gradient. Instead of directly using the gradient direction at the current moment, it is defined as the normal vector of the geometry of the measured component, or the gradient direction at the previous stable moment is taken. Temperature gradient direction:

[0081] ;

[0082] in, This is the temperature gradient vector from the previous time step. The norm of the temperature gradient vector at the previous moment; The preset numerical stability constant has physical dimensions that are related to the temperature gradient. To maintain consistency and prevent the denominator from being zero, in this embodiment the value is taken as... It should be noted that when the system is at the initial startup moment, there is no data from the previous moment; at this time, directly set... This is equal to the surface normal vector of the pre-calibrated three-dimensional geometric model of the measured part under the current view, in order to complete the initialization calculation;

[0083] This setting ensures that even at the current moment... The calculated residual vector when the surface is rapidly cooled It can accurately reflect the significant vector difference between the theoretical external heat dissipation flow and the actual internal cooling flow, thus ensuring the sensitivity of physical consistency entropy; mapping function The formula is:

[0084] ;

[0085] in, The input temperature change rate variable corresponds to the local calculation. or global calculation ; For reference to the absolute temperature variable, corresponding to the local calculation or global calculation , Surface emissivity; It is the Stefan constant; The thermal response time constant of the system is expressed in seconds. This constant is obtained by calibrating the system through a step heat source response experiment, which records the time required for the infrared detector to reach 63.2% of the steady-state value in response to a sudden heat source. In this embodiment, the typical value is 0.1s to 0.5s.

[0086] This function is used to calculate the observed equivalent temperature time-varying rate according to the linearized Stefan-Boltzmann law. Mapped to the equivalent heat flux density amplitude, where, This is the system's thermal response time constant, used to express the instantaneous rate of temperature change. The integral is converted into the equivalent temperature difference within the effective response time. Thus making The calculated results have the correct dimensions of heat flux density, achieving consistency with the theoretical heat flux density. A direct comparison; Accurate acquisition depends on the aforementioned radiometric calibration and temporal difference processing; based on the norm of the residual vector Physical consistency entropy is generated by quantizing using the Sigmoid activation function. The specific calculation formula is as follows:

[0087] ;

[0088] in, This is a sensitivity coefficient used to adjust the steepness of the entropy change with the residual. Its physical dimension is the reciprocal of the heat flux density. In this embodiment, the preferred value range corresponds to the normalized value. to This is to ensure that the entropy value can respond quickly when the residual exceeds the threshold; This is the tolerance threshold, used to set the allowable physical error baseline; The specific values ​​are obtained through benchmark calibration: within a known steady-state and fault-free operating cycle of the equipment, the residual vector norm is statistically calculated. mean with standard deviation ,set up:

[0089] ;

[0090] in, This represents the average absolute temperature of the entire surface of the measured component at the current moment; this is achieved by calling the aforementioned general mapping function. Calculate the equivalent heat flux density threshold component corresponding to the background thermal noise of the entire image; This represents the absolute value of the average rate of temperature change across the entire graph. This is a dynamic adjustment coefficient, and its value range is typically [range missing]. to ;

[0091] This setting introduces adaptability to unsteady states under normal operating conditions, preventing the threshold from being falsely triggered due to normal global thermal fluctuations; It covers 99.7% of the system's inherent thermal noise, ensuring physical consistency. Entropy only responds to anomalies that significantly violate physical laws; the higher the entropy value, the better.

[0092] This embodiment establishes a mathematical mapping between visual features and physical laws, giving the AI ​​model self-awareness. This enables it to accurately identify physical conflicts when faced with contradictory situations where the surface temperature is extremely cold but the theoretically calculated internal heat flux density is extremely high, providing an objective mathematical basis for subsequent mode switching.

[0093] Example 3:

[0094] Step 3 executes a high-resolution reconstruction mode, including: activating the super-resolution reconstruction module in the generative model; performing texture enhancement processing on the initial temperature field distribution data with the optimization objective of minimizing visual perception loss; and mapping to generate an internal temperature field inversion image with high spatial resolution, which is used to characterize the fine temperature gradient distribution inside the component.

[0095] This embodiment details the processing flow when the physical consistency entropy is low, i.e. the visual signal is reliable; in response to the system's determination that the environmental interference is low, the pre-set super-resolution reconstruction module in the generative model is activated, which is usually composed of residual dense blocks or attention mechanism layers.

[0096] The system uses this module to perform texture enhancement processing on the initial temperature field distribution data. This processing is based on the parameters learned by the module during the training phase, and the optimization objective of its training is set to minimize visual perception loss. The source is the Euclidean distance between high-level feature maps extracted by pre-trained feature extraction networks such as VGG-19. The specific calculation formula is as follows:

[0097] ;

[0098] in, The first feature extraction network is represented by the first feature extraction network. Layer feature mapping; in this embodiment, the feature extraction network uses a pre-trained VGG-19 network. Specifically, it refers to the activation feature map before the fourth convolutional layer of the fifth convolutional block in the network, in order to capture the high-level semantic texture information of the image; These represent the number of channels, height, and width of the feature map, respectively. For the enhanced temperature field image, The training set consists of paired real-world benchmark images;

[0099] The physical meaning is the degree of difference of the image at the level of human visual perception, and the unit is the dimensionless loss value; the optimization objective is to make the generated image more in line with human visual perception, rather than simply minimizing the pixel-level mean square error.

[0100] The system maps and generates an internal temperature field inversion image with high spatial resolution. Under favorable environmental conditions, this embodiment fully utilizes the image enhancement capabilities of generative AI to recover clear texture details, thereby presenting the point-like temperature rise caused by tiny peeling on the bearing raceway surface. Its resolution is sufficient to distinguish two hot spots only a few millimeters apart, achieving accurate capture of early weak fault signals.

[0101] Example 4:

[0102] Step 4 executes the thermal flow topology derivation mode, including: freezing the texture generation layer in the generative model and calling the preset thermal flow topology energy function as a constraint condition; activating the thermal flow topology constraint layer; fuzzing the initial temperature field distribution data with the optimization objective of maximizing the thermodynamic existence probability, and removing high-frequency visual noise that violates the thermodynamic conduction law; extracting and amplifying low-frequency signals in the spatiotemporal residual features that conform to the internal heat source diffusion law; and reconstructing and generating a thermal risk topology map, which is used to characterize the skeleton structure of internal heat flow transmission and the probability distribution of heat sources in the component.

[0103] This embodiment details the processing flow when the physical consistency entropy is high, i.e. the visual signal is deceived; in response to the system's determination of strong interference, the texture generation layer in the generative model is automatically frozen, and the preset thermal flow topology energy function is called as a constraint condition; the generation of high-frequency details is stopped, and the thermal flow topology constraint layer is activated.

[0104] Specifically, maximizing the thermodynamic existence probability is equivalent to minimizing the topological energy function of heat flux. The function consists of a low-frequency heat conduction residual term and a skeleton continuity constraint term, and the specific calculation formula is as follows:

[0105] ;

[0106] in, The low-frequency temperature field component after low-pass filtering; and These are the weighting coefficients; the first term is based on the physical rationality of constraining the heat flow path according to the steady-state heat conduction equation, and the second term is based on Sobolev space. Norm constraints ensure the topological continuity of the heat flux skeleton over time; in practical calculations, this term expands into a joint constraint of the numerical value and its gradient.

[0107] ;

[0108] in, This is the temperature field distribution data from the previous moment. These are characteristic weighting coefficients with dimensions of length squared, used to balance the unit difference between the temperature and gradient terms; This represents the spatial gradient operator, where the constraint ensures the temporal smoothness of the temperature field in terms of both numerical amplitude and spatial morphology.

[0109] The system updates iteratively using gradient descent. To minimize This eliminates high-frequency visual noise that violates the laws of thermodynamic conduction. The source of high-frequency visual noise is edge abrupt changes caused by cold spots or sludge caused by raindrops. Physically, it is a sharp hot-cold abrupt change signal formed within milliseconds, with the unit being frequency amplitude.

[0110] The system extracts and amplifies low-frequency signals that conform to the internal heat source diffusion law in the spatiotemporal residual features. It utilizes the characteristic that the heat generated by the internal fault is conducted in the form of low-frequency diffused waves to see through the cold illusion on the surface. It reconstructs and generates a thermal risk topology map, which does not show specific temperature values, but shows the skeleton structure of heat flow transmission inside the component and the probability distribution of heat sources.

[0111] In the event of visual data failure, this embodiment abandons accurate but erroneous image reconstruction and instead provides a fuzzy but correct thermal flow risk assessment. This ensures that even when the surface is cooled or blocked, the internal thermal energy risk skeleton that is accumulating can still be identified, thus achieving a safety margin under extreme operating conditions.

[0112] Example 5:

[0113] Physical consistency entropy is used to identify infrared visual deception states, which include: an asymmetric thermal state caused by surface rapid cooling and internal heating due to unsteady environmental thermal shock; and a local emissivity nonlinear distortion state caused by surface contaminant shading. Among these, the physical consistency entropy increases significantly under infrared visual deception states, causing the system to determine that the thermal flow topology deduction mode should be executed.

[0114] This embodiment specifically defines the physical boundary conditions for trigger mode switching; the system monitors the asymmetric thermal state caused by thermal shock in an unsteady environment, such as when a train encounters a sudden rainstorm or heavy snow while running at high speed. At this time, the external rain and snow cause the surface of the axle box to cool rapidly, while the inside of the bearing is in an internal heating state due to high-speed operation. The surface temperature gradient is reversed with the theoretical heat flow direction generated by the internal heat source.

[0115] The system monitors the nonlinear distortion of local emissivity caused by surface contaminants, such as uneven sludge or oxide layer covering the axle box surface, which changes the infrared emissivity of the surface and results in huge grayscale differences in thermal images at the same temperature. Under such infrared visual deception, the physical consistency entropy will increase significantly, exceeding the preset confidence threshold. In response to this increase in entropy, the system immediately triggers a switch to thermal flow topology deduction mode.

[0116] This embodiment clearly defines the scenarios of physical conflicts, enabling the system to distinguish between sensor malfunctions and environmental interference. When faced with complex working conditions such as rain, snow, rapid cooling, or mud blockage, it can automatically adjust the detection strategy, significantly improving the system's adaptability and robustness in all-weather operating environments.

[0117] Example 6:

[0118] The generative model adopts a physically-based adversarial generative network architecture, which includes: a generator network for generating a predicted temperature field based on infrared thermal image sequences; a discriminator network for judging the visual realism of the predicted temperature field; and a physical constraint module for calculating the residual of the predicted temperature field with respect to the partial differential equation of heat conduction, and feeding the residual as a penalty term into the loss function of the generator network.

[0119] This embodiment details the network architecture of the generative model, namely, a physically-based generative adversarial network (GAN). The generator network, typically employing a U-Net structure, is constructed to receive low-resolution, noisy infrared image sequences and output prediction data containing two channels: a predicted temperature field. With potential heat source distribution field The generator network adopts the U-Net architecture, where the last two upsampled convolutional blocks of the decoder are defined as texture generation layers to recover high-frequency detail features of the image, and the rest are defined as structural feature layers.

[0120] Simultaneously, a discriminator network is constructed, and adversarial training forces the generator to produce images that closely approximate realistic infrared thermal images in terms of texture and distribution. Building upon this, the system integrates a physical constraint module, which contains no trainable parameters but instead embeds the partial differential equation of heat conduction. This physical constraint module calculates the point-to-point residual field of the predicted temperature field on the heat conduction equation. The specific calculation formula is as follows:

[0121] ;

[0122] in, For density, For specific heat capacity, Thermal conductivity; The predicted temperature field output by the generator; For the Laplace operator; The distribution field of the internal fault heat source to be inverted;

[0123] and residual field L2 norm and heat source The L1 sparse regularization terms are fed back as penalty terms into the loss function of the generator network; the total loss function for:

[0124] ;

[0125] in, To generate the adversarial loss term for the adversarial network, For physical constraint weights, For residual field; Adaptive adjustment using a gradient balancing strategy, the last term of the formula For heat source Regularization sparsity constraints, The sparsity penalty coefficient is used to force the model to generate non-zero heat sources only at necessary fault points, preventing heat source dispersion; Weights in the next iteration The calculation formula is:

[0126] ;

[0127] in, This indicates the current training iteration round of the neural network; Specifically refers to the set of trainable parameters for a generator network; This indicates the parameters of the generator in the current training batch. The average gradient magnitude;

[0128] This strategy ensures that the gradient contribution of the physical residual term and the gradient contribution of the adversarial loss term are on the same order of magnitude, preventing training collapse due to the difference in dimensions; the value is typically between 0.1 and 1.0.

[0129] This embodiment explicitly embeds physical constraints into the network structure, so that the neural network not only learns image features during training, but is also forced to learn physical laws. This ensures that the temperature field generated by the model during inference is naturally physically reasonable, reducing dependence on a large amount of labeled data while avoiding the non-physical illusions that may be generated by a purely data-driven model.

[0130] Example 7:

[0131] Step 5 involves locating internal thermal anomaly singularities, including identifying spatial coordinates of points in the internal temperature field inversion image or thermal risk topology map where the temperature value exceeds a preset safety threshold and the duration exceeds a preset time window.

[0132] This embodiment describes the specific logic for locating singularities of internal thermal anomalies. The system scans all pixels in the internal temperature field inversion image or thermal risk topology map to identify points whose temperature values ​​or thermal risk probability values ​​exceed a preset safety threshold. The thermal risk probability value comes from the probability distribution map output by the thermal flow topology deduction mode. This map is processed by the Sigmoid activation function, and the pixel value range is [0,1], which independently represents the confidence that a heat source exists at that location. For example, a threshold of 0.85 is set as the threshold for a temperature exceeding 90°C or a heat accumulation probability exceeding 85%.

[0133] The system introduces a time dimension constraint to determine whether the duration of the abnormal state exceeds a preset time window. The preset time window is derived from the number of consecutive frames set by the system, such as 10 frames or 0.2 seconds. The physical meaning is the time scale that distinguishes transient noise from actual heat accumulation, and the unit is seconds or frames. Only when the spatial coordinate point meets both the amplitude and time conditions is it marked as an internal thermal anomaly singularity.

[0134] This embodiment effectively filters out instantaneous electronic noise or random interference from the sensor by introducing a duration window as a judgment condition, ensuring that only real thermal anomalies with a certain energy accumulation effect are captured by the system, thereby significantly reducing the false alarm rate in high-speed dynamic environments.

[0135] Example 8:

[0136] Please see Figure 2 A bearing operating status detection system based on infrared thermal imaging and temperature field modeling is used to execute the above method, including: an image acquisition module for acquiring a sequence of infrared thermal images of the mechanical component under test;

[0137] The model loading module is used to load generative models with physical information constraints.

[0138] The state detection module includes: a feature extraction unit, which is used to input the infrared thermal image sequence into the generative model, extract the spatiotemporal residual features of the image, and generate initial temperature field distribution data;

[0139] The entropy calculation unit is used to calculate the physical consistency entropy based on the degree of deviation between the initial temperature field distribution data and the thermodynamic conduction law.

[0140] The mode switching unit is used to trigger the high-definition reconstruction mode in response to the physical consistency entropy being less than a preset confidence threshold; and to trigger the thermal flow topology inference mode in response to the physical consistency entropy being greater than or equal to the confidence threshold.

[0141] The image generation unit is used to generate an internal temperature field inversion image in high-definition reconstruction mode, or to generate a thermal risk topology map in thermal flow topology extrapolation mode.

[0142] The result output unit is used to locate internal thermal anomaly singularities based on the internal temperature field inversion image or thermal risk topology map, and output the detection results of the operating status of the tested mechanical component.

[0143] This embodiment provides a bearing operating status detection system based on infrared thermal imaging and temperature field modeling. This system realizes the hardware deployment of the above method. The image acquisition module uses an uncooled focal plane array infrared thermal imager, which is installed at key parts of the train to acquire infrared thermal image sequences of the tested mechanical parts in real time. The model loading module is based on an edge computing unit equipped with a high-performance GPU and pre-loads PI-GAN model weights trained based on the PyTorch or TensorFlow framework. The feature extraction unit in the status detection module inputs the image sequence into the model, extracts spatiotemporal residual features, and generates an initial temperature field.

[0144] During this period, the entropy calculation unit calculates the physical consistency entropy in real time and monitors the physical reliability of the model; the mode switching unit, as the logical core, responds to the change in physical consistency entropy and dynamically triggers the high-definition reconstruction mode or the thermal flow topology deduction mode; the image generation unit generates the corresponding inversion image or topology map according to the triggered mode; the result output unit locates the internal thermal anomaly singularity based on the image data and sends an alarm signal through the vehicle network.

[0145] This embodiment utilizes a modular design to deploy complex physical perception AI algorithms at the edge, achieving full-process automation from data acquisition, physical consistency verification, dynamic mode switching, and fault warning. It is particularly suitable for industrial scenarios with extremely high requirements for real-time performance and reliability, such as high-speed trains.

[0146] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A bearing operating condition detection method based on infrared thermal imaging and temperature field modeling, characterized in that, include: Acquire an infrared thermal image sequence of the mechanical component under test, the infrared thermal image sequence containing spatiotemporal temperature distribution information of the component surface; Based on the infrared thermal image sequence, a generative model constrained by physical information is constructed, and the generative model is embedded with the thermodynamic conduction law as a priori constraint. The temperature field inversion and state determination process includes: Step 1: Input the infrared thermal image sequence into the generative model, extract the spatiotemporal residual features of the image, and generate initial temperature field distribution data; Step 2: Based on the degree of deviation between the initial temperature field distribution data and the thermodynamic conduction law, calculate the physical consistency entropy, which is used to characterize the degree of conflict between visual image features and thermodynamic physical facts; Step 3: In response to the physical consistency entropy being less than a preset confidence threshold, execute a high-definition reconstruction mode to generate a pixel-level precision internal temperature field inversion image; Step 4: In response to the physical consistency entropy being greater than or equal to the confidence threshold, execute the heat flow topology deduction mode, suppress the generation of image texture details, and generate a thermal risk topology map based on the principle of heat flow conservation. Step 5: Based on the internal temperature field inversion image or the thermal risk topology map, locate the internal thermal anomaly singularity and output the operating status detection result of the tested mechanical component.

2. The bearing operating status detection method based on infrared thermal imaging and temperature field modeling according to claim 1, characterized in that, Step 2, calculating the physical consistency entropy, includes: Extract the surface temperature gradient features from the initial temperature field distribution data; Based on the aforementioned thermodynamic conduction law, calculate the theoretical heat flux density corresponding to the surface temperature gradient characteristics; Calculate the residual vector between the theoretical heat flux density and the actual observed rate of change of thermal radiation in the infrared thermogram sequence; The physical consistency entropy is generated based on the norm of the residual vector; wherein the value of the physical consistency entropy increases monotonically as the residual vector increases.

3. The bearing operating status detection method based on infrared thermal imaging and temperature field modeling according to claim 1, characterized in that, Step 3 involves performing a high-definition reconstruction, including: Activate the super-resolution reconstruction module in the generative model; With minimizing visual perception loss as the optimization objective, the initial temperature field distribution data is subjected to texture enhancement processing; The mapping generates an internal temperature field inversion image with high spatial resolution, which is used to characterize the fine temperature gradient distribution inside the component.

4. The bearing operating status detection method based on infrared thermal imaging and temperature field modeling according to claim 1, characterized in that, Step 4 involves executing the heat flow topology derivation mode, including: Freeze the texture generation layer in the generative model and call the preset thermal flow topology energy function as a constraint condition; and activate the thermal flow topology constraint layer; With the goal of maximizing the probability of thermodynamic existence, the initial temperature field distribution data is fuzzified to remove high-frequency visual noise that violates the thermodynamic law of conduction. Extract and amplify the low-frequency signal that conforms to the internal heat source diffusion law in the spatiotemporal residual features; The thermal risk topology map is reconstructed and generated. The thermal risk topology map is used to characterize the skeleton structure of heat flow transmission inside the component and the probability distribution of heat sources.

5. The bearing operating status detection method based on infrared thermal imaging and temperature field modeling according to claim 2, characterized in that, The physical consistency entropy is used to identify infrared visual deception states, which include: An asymmetric thermal state caused by unsteady thermal shock, resulting in both rapid surface cooling and internal heating. The local emissivity nonlinear distortion state caused by surface contaminant shielding; In the infrared visual deception state, the physical consistency entropy increases significantly, causing the system to determine that the heat flow topology deduction mode should be executed.

6. The bearing operating status detection method based on infrared thermal imaging and temperature field modeling according to claim 1, characterized in that, The generative model employs a physically based adversarial generative network architecture, including: A generator network is used to generate a predicted temperature field based on the infrared thermal image sequence; A discriminator network is used to determine the visual realism of the predicted temperature field; The physical constraint module is used to calculate the residual of the predicted temperature field on the partial differential equation of heat conduction, and feed the residual as a penalty term into the loss function of the generator network.

7. The bearing operating status detection method based on infrared thermal imaging and temperature field modeling according to claim 1, characterized in that, Step 5, locating the internal thermal anomaly singularity, includes: In the internal temperature field inversion image or the thermal risk topology map, identify spatial coordinate points where the temperature value exceeds a preset safety threshold and the duration exceeds a preset time window.

8. A bearing operating condition detection system based on infrared thermal imaging and temperature field modeling, used to execute the bearing operating condition detection method based on infrared thermal imaging and temperature field modeling as described in any one of claims 1-7, characterized in that, include: The image acquisition module is used to acquire a sequence of infrared thermal images of the mechanical parts under test. The model loading module is used to load generative models with physical information constraints. The status detection module includes: The feature extraction unit is used to input the infrared thermal image sequence into the generative model, extract the spatiotemporal residual features of the image, and generate initial temperature field distribution data. The entropy calculation unit is used to calculate the physical consistency entropy based on the degree of deviation between the initial temperature field distribution data and the thermodynamic conduction law; The mode switching unit is used to trigger a high-definition reconstruction mode in response to the physical consistency entropy being less than a preset confidence threshold; and to trigger a thermal flow topology inference mode in response to the physical consistency entropy being greater than or equal to the confidence threshold. The image generation unit is used to generate an internal temperature field inversion image in the high-definition reconstruction mode, or to generate a thermal risk topology map in the heat flow topology extrapolation mode. The result output unit is used to locate internal thermal anomaly singularities based on the internal temperature field inversion image or the thermal risk topology map, and output the operating status detection results of the tested mechanical component.

Images