A roller coaster track thermal deformation monitoring method based on infrared feature extraction

Abstract

The application discloses a roller coaster track thermal deformation monitoring method based on infrared feature extraction, and through the construction of each track section reference temperature field model, the track surface infrared image is collected in real time during train operation, and a thermal radiation distribution map is obtained by processing; macro temperature gradient features, local thermal spot topological features and thermal pulse response features are extracted from the thermal radiation distribution map, input into a temperature strain conversion model, and the real-time thermal deformation amount and direction vector of each node are inferred; an anomaly detector combined with an isolated forest algorithm and a long short-term memory network is used to identify thermal deformation anomalies and calculate a comprehensive risk index, and yellow, orange and red three-level early warnings are triggered according to an adaptive threshold. The application realizes full-section non-contact, real-time and high-precision thermal deformation monitoring and adaptive graded early warning of the roller coaster track, and significantly improves the operation safety level of the amusement facilities.

Description

Technical Field

[0001] This invention belongs to the field of intelligent track monitoring technology, specifically relating to a method for monitoring the thermal deformation of roller coaster tracks based on infrared feature extraction. Background Technology

[0002] Roller coaster tracks are subjected to the impact loads generated by high-speed trains passing by, the release of braking heat, and the cyclical thermal effects of day and night temperature differences. Irreversible thermal deformation accumulation is very likely to occur at bolted joints, welded areas, and curved sections, which can lead to track geometric deviations or even structural failure in severe cases.

[0003] Current track inspection methods mainly rely on periodic manual inspections and contact sensor arrays, which have the following prominent problems: First, manual inspections are greatly affected by subjective factors and have long inspection cycles, making it impossible to capture transient thermal deformation events during operational breaks; Second, contact strain gauges require a large number of sensors to be pre-embedded on the track, resulting in high installation costs, high maintenance difficulties, and insufficient coverage of the entire track section due to sparse measurement points; Third, existing methods generally lack quantitative modeling of the fatigue weakening effect of track materials, making it difficult to provide reliable early warnings in the early stages of deformation anomalies, and problems are often only discovered after macroscopic damage occurs, missing the best time for intervention.

[0004] Infrared thermal imaging technology, with its advantages of non-contact and full-field measurement, has been widely used in fields such as aviation structural health monitoring and electrical equipment fault diagnosis. However, there is still a lack of methods to apply it to the full-process thermal deformation monitoring of roller coaster tracks in high-dynamic scenarios and to combine it with deep learning to achieve adaptive hierarchical early warning. Summary of the Invention

[0005] The purpose of this invention is to address the problems of existing roller coaster track thermal deformation monitoring methods, such as narrow coverage, poor real-time performance, lack of fatigue weakening modeling and graded early warning capabilities. It provides a roller coaster track thermal deformation monitoring method based on infrared feature extraction, aiming to achieve non-contact, real-time, and high-precision thermal deformation perception of the entire track. Through multi-feature fusion and adaptive anomaly detection, it identifies potential structural risks in their nascent stage, thereby providing reliable technical support for ensuring the safe operation of amusement facilities.

[0006] The specific technical solution is as follows: A method for monitoring thermal deformation of roller coaster tracks based on infrared feature extraction, the method comprising: Under conditions of no track load and stable ambient temperature, a full-track infrared thermal image sequence was collected. After radiometric calibration correction, a reference temperature field model for each track segment was constructed as a reference baseline for subsequent thermal deformation comparison.

[0007] Infrared images of the track surface are acquired in real time during train operation. Non-uniformity correction, atmospheric transmittance compensation, and motion blur deconvolution processing are performed on the image sequence to obtain a thermal radiation distribution map.

[0008] Macroscopic temperature gradient features, local hot spot topological features, and thermal pulse response features are extracted from the thermal radiation distribution map. The extracted thermal features are input into a pre-trained temperature strain conversion model, which integrates the thermal expansion coefficient of the orbital material, geometric constraints, and historical load spectrum to infer the real-time thermal deformation and direction vector of each node.

[0009] An anomaly detector combining the isolated forest algorithm and long short-term memory network is used to identify thermal deformation anomalies in the time series of thermal deformation of each track section. A comprehensive risk index is calculated by combining the amount of thermal deformation, deformation rate, historical defect records of the track section, and current operating status. Yellow, orange, and red three-level early warning measures are triggered according to preset adaptive thresholds.

[0010] Furthermore, the radiometric calibration correction method specifically involves radiometric calibration of the original grayscale image acquired by the infrared thermal imager, converting the grayscale values ​​into radiance values ​​and equivalent target temperatures.

[0011] The radiation calibration formula is: In the formula, The original grayscale values ​​were acquired by the infrared thermal imager. The value is the radiance, expressed in W / (m²·sr·μm). This is the gain coefficient; This is the bias coefficient; and All were determined by standard blackbody calibration experiments.

[0012] Target temperature The following can be obtained by inversion from Planck's radiation law: In the formula, It is the first radiation constant; It is the second radiation constant; Equivalent wavelength, unit: meters; The emissivity of the orbital material ranges from 0.8 to 0.95.

[0013] Furthermore, the method for constructing the reference temperature field model is as follows: A sequence of M consecutive infrared thermal images collected when the orbit is unloaded and the ambient temperature is stable is selected, and the average temperature value at the same spatial location is taken to construct the reference temperature field model. In the formula, The total number of steady-state frames participating in the mean calculation; For the first The temperature value at spatial coordinates (x, y, z) of a frame of infrared thermal image, in Kelvin; the variance of the mean temperature of M consecutive frames. When the value is ≤0.1K, the reference field is considered to have reached a steady state; if the variance exceeds this threshold, we continue to wait until the steady-state condition is met before constructing the reference temperature field model.

[0014] Furthermore, the method for extracting the macroscopic temperature gradient features is as follows: [The text abruptly shifts to a different topic] ... thermal radiation distribution map... implement Two-dimensional discrete wavelet decomposition, in the first layer layer( ) Obtain detail values ​​in the horizontal, vertical, and diagonal directions respectively. , , ; The amplitude of the temperature gradient field is synthesized from multi-scale detail coefficients: In the formula, For wavelet decomposition layer index; This represents the total number of decomposition levels. Let the scale weights be those of the j-th layer. The local heat flux density distribution diagram is approximately obtained from the heat conduction equation: In the formula, The thermal conductivity of the rail steel is given, with a value ranging from 50. 60 W / (m·K); Local heat flux density: Exceeding the hot spot detection threshold At that time, the area was marked as a hot spot and entered the local hot spot topology feature analysis process; It is determined by the thermal fatigue limit of the track material and the safety factor.

[0015] Furthermore, the extraction of the thermal pulse response characteristics employs phase-locked thermal imaging technology, using the periodic mechanical vibrations generated by the train's passage and the braking heat release as natural periodic excitation sources, with an excitation frequency of... Determined by the frequency of train passage; for infrared image sequences containing P frames. Perform phase-locked integration to extract in-phase components. and orthogonal components : ; ; In the formula, The total number of image frames involved in the integration calculation; The time interval between adjacent frames, in seconds; For frame number, Thermal response amplitude diagram Phase diagram They are respectively: ; Phase diagram It can effectively reveal the 5 subsurface layers below the orbital surface. Internal structural defects within a depth range of 20mm; areas of phase anomaly are identified as suspected cracks or delamination defects, triggering a close-range re-inspection process.

[0016] Furthermore, the temperature-strain conversion model maps thermal characteristics to thermal strain, and the mapping relationship is as follows: In the formula, These are the position coordinates on the track; For position exist Thermal strain at any moment; The effective thermal expansion coefficient, which is related to temperature, is calculated using piecewise linear interpolation: In the formula, Reference temperature The reference thermal expansion coefficient at that location, in Kelvin. - ¹; Temperature gradient of thermal expansion coefficient, unit K. - ²; and These represent the lower and upper limits of the model's effective temperature range. As a nodal constraint correction term, a local stiffness reduction factor is introduced at bolted connections, weld zones, and support frame locations. Correcting the redistribution effect of thermal strain at the constraint boundary; For the train dynamic load correction item, the dynamic additional strain is calculated based on the train's real-time speed, axle load and position, and the thermal strain component is separated from the total strain. This is a historical fatigue correction term.

[0017] Furthermore, the calculation formula for the historical fatigue correction term is as follows: In the formula, The fatigue weakening coefficient reflects the effect of microscopic damage on the coefficient of thermal expansion, and is expressed in Kelvin. - ¹; For position exist The cumulative number of load cycles at any given time; The fatigue index, with a value range of 0 < ≤1, determined by fitting fatigue test data of track materials; when When the preset fatigue warning threshold is exceeded, the track section is marked as a priority maintenance node.

[0018] Furthermore, the combined anomaly detector of the isolated forest algorithm and LSTM network operates in a parallel voting manner; the isolated forest algorithm evaluates the thermal deformation feature vector at the k-th time step. Calculate abnormal scores for: In the formula, thermal deformation feature vector The expected average path length in a set of isolated trees; Here is a normalization factor for the average path length of an isolated tree with n samples: In the formula, Number of training samples for the isolated tree; For harmonic numbers, ; An LSTM network performs a single-step prediction of the thermal deformation time series of the k-th track segment, and the anomaly score is calculated. Defined as the standardized value of the prediction residual: In the formula, This represents the single-step prediction of thermal strain at time t by the LSTM network. is the standard deviation of the historical prediction residuals for the k-th track segment, reflecting the normal fluctuation baseline of that track segment.

[0019] The fusion anomaly score is: In the formula, The weighting coefficients for fusing the anomaly scores of the Isolation Forest algorithm and the LSTM network; when Output high-confidence warnings in real time; when Output low-confidence suspected alerts and continuously track them; among which and These are the threshold values ​​for high and low confidence warnings, respectively; the LSTM network updates its parameters online daily to adapt to data distribution drift caused by equipment aging and environmental changes.

[0020] Furthermore, the calculation method for the comprehensive risk index is as follows: for the first... For each track segment, the comprehensive risk index is calculated using the following formula. : In the formula, This is the maximum allowable thermal deformation. The maximum permissible thermal deformation rate; This represents the cumulative damage value of historical defects in the k-th track segment; The reference damage value is determined by material fracture mechanics analysis; The operational status coefficient reflects the impact of current train speed and load on risk; the weighting coefficient satisfies... It is determined from historical operating data using the Analytic Hierarchy Process (AHP).

[0021] Risk level classification criteria: <0.4 indicates a green and safe state; 0.4≤ <0.6 triggers a yellow alert; 0.6≤ <0.8 triggers an orange alert; A value ≥0.8 triggers a red alert and initiates closed-loop control; the time from alert triggering to train response is no more than 3 seconds.

[0022] Furthermore, the adaptive threshold is calculated as follows: warning threshold Based on the basic threshold, dynamic corrections are made by comprehensively considering seasonal temperature, train load, and material fatigue state. In the formula, For the first The basic threshold of the track section is determined by material mechanics calculations and a safety factor; The seasonal temperature correction factor is calculated based on the deviation between the current ambient temperature and the annual average temperature baseline, reflecting the impact of diurnal and seasonal temperature fluctuations on the normal thermal expansion range. This is the load correction factor, calculated based on the ratio of the current real-time train speed to the axle load relative to the rated value, with a value range of [0,1].

[0023] The material fatigue correction factor is calculated using the following formula: In the formula, The fatigue sensitivity coefficient; The design life cycle number is set for the k-th track segment; the weighting coefficients satisfy... The threshold correction range is guaranteed to be no more than 30% of the base threshold; the threshold update cycle is no more than 1 hour; and the system automatically switches to conservative threshold mode under extreme high temperature weather or abnormal load conditions.

[0024] Compared with the prior art, the beneficial effects of this invention are: This invention achieves automatic compensation for environmental fluctuations through radiation calibration and reference temperature field modeling, eliminating the systematic errors introduced by environmental temperature differences in traditional methods. The fusion of multi-scale wavelet decomposition and phase-locked thermal imaging technology can simultaneously perceive macroscopic temperature gradient distribution, millimeter-level hot spot topology, and internal defects with a depth of 5-20mm, with detection sensitivity significantly superior to single-sensor modes. The temperature strain conversion model introduces three types of compensation terms: node constraint correction, train dynamic load correction, and historical fatigue weakening correction, ensuring that the accuracy of thermal deformation inference remains stable throughout the entire life cycle. The parallel voting anomaly detection mechanism of isolated forest and LSTM network has the characteristics of strong noise resistance and adaptive online updating. The comprehensive risk index combined with AHP weights can achieve three-level accurate early warning, with the train response time not exceeding 3 seconds from early warning triggering, effectively avoiding the lag of manual inspection and significantly improving the operational safety level of amusement facilities. Attached Figure Description

[0025] Figure 1 This is a flowchart of a roller coaster track thermal deformation monitoring method based on infrared feature extraction according to the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0027] like Figure 1 The diagram shown is a flowchart of a roller coaster track thermal deformation monitoring method based on infrared feature extraction according to the present invention, which includes the following steps: Step S1: Construction of reference thermal map; Under the condition of no orbit and stable ambient temperature, a full orbit infrared thermal map sequence is collected, and after radiometric calibration correction, a reference temperature field model for each orbit segment is constructed as a reference baseline for subsequent thermal deformation comparison.

[0028] In practical implementation, taking a roller coaster currently in operation at a theme park as an example, the facility's track is approximately 450 m long, divided into 32 sections, with a minimum curve radius of 8 m and a maximum operating speed of 120 km / h. Step S1 is completed within 60 minutes before the park opens each operating day, at which time the track is unloaded, and the ambient temperature fluctuation is confirmed by meteorological sensors to be within ±0.5℃. Eight uncooled focal plane infrared thermal imagers are evenly spaced along the track. The detectors have a resolution of 640×512 pixels, a spectral response range of 8–14 μm, a thermal sensitivity (NETD) better than 0.05 K, and a frame rate set to 25 Hz. The fields of view of adjacent thermal imagers overlap by 10% to ensure seamless stitching across the entire section. Each thermal imager is aligned with the BeiDou timing signal through a time synchronization module, with an inter-frame timing error not exceeding 1 ms, ensuring that the multi-channel thermal images are strictly aligned in the time dimension before stitching.

[0029] The radiometric calibration correction method specifically involves radiometrically calibrating the original grayscale image acquired by the infrared thermal imager, converting the grayscale values ​​into radiance values ​​and equivalent target temperatures. The radiation calibration formula is: In the formula, The original grayscale values ​​were acquired by the infrared thermal imager. The value is the radiance, expressed in W / (m²·sr·μm). This is the gain coefficient; This is the bias coefficient; and All were determined by standard blackbody calibration experiments; Target temperature The following can be obtained by inversion from Planck's radiation law: In the formula, It is the first radiation constant; It is the second radiation constant; Equivalent wavelength, unit: meters; The emissivity of the orbital material ranges from 0.8 to 0.95.

[0030] The method for constructing the reference temperature field model is as follows: A sequence of M consecutive infrared thermal images collected when the orbit is unloaded and the ambient temperature is stable is selected; the average temperature value at the same spatial location is taken to construct the reference temperature field model. In the formula, The total number of steady-state frames participating in the mean calculation; For the first The temperature value at spatial coordinates (x, y, z) of a frame of infrared thermal image, in Kelvin; the variance of the mean temperature of M consecutive frames. When the value is ≤0.1K, the reference field is considered to have reached a steady state; if the variance exceeds this threshold, we continue to wait until the steady-state condition is met before constructing the reference temperature field model.

[0031] The process of establishing the reference temperature field is illustrated using the 12th track segment (curved area, radius 10 m) as an example. Radiation calibration was performed at two reference temperature points, 333 K and 373 K, on ​​a standard blackbody, and the gain coefficient was obtained through fitting. Bias coefficient .collection A frame steady-state thermal image (sampled at 25 Hz for 12 s) was used to construct a reference temperature field by averaging the temperature values ​​at the same spatial coordinates. The steady-state test results show variance. ,satisfy The criteria for judgment are that the reference temperature field is effective. Under an ambient temperature of 22℃, the average surface temperature of the track section is 24.3℃, the temperature in the weld area is about 0.4℃ higher, and the temperature at the bolted connection joint is about 0.2℃ higher due to changes in the local heat conduction path. These patterns are consistent with expectations and can be used as a reliable reference baseline.

[0032] Step S2, Dynamic thermal image acquisition and preprocessing: Infrared images of the track surface are acquired in real time during train operation, and non-uniformity correction, atmospheric transmittance compensation and motion blur deconvolution processing are performed on the image sequence to obtain a thermal radiation distribution map.

[0033] Non-uniformity correction (NUC) employs a two-point calibration method based on high temperature. and low temperature The gain matrix is ​​calculated pixel-by-pixel based on the response values ​​of the two reference black bodies. and bias matrix The corrected numerical value is: Taking a certain measured frame as an example, there is about 3% response non-uniformity in the center and edge regions of the original image. After two-point NUC processing, it is reduced to less than 0.3%, which meets the accuracy requirements for subsequent temperature estimation.

[0034] Atmospheric transmittance compensation is based on the Lowtran-7 atmospheric radiative transfer model. Inputs include the current relative humidity (e.g., 65%RH), air temperature (e.g., 25℃), and the path length (0.5–3 m) for each pixel. Atmospheric transmittance is calculated pixel-by-pixel. Under typical operating conditions The corrected radiance is: Taking the distance from the track to the farthest point of the thermal imager as an example, the radiation brightness error before compensation was about 0.8%, and the equivalent temperature error after compensation was reduced from about 0.3 K to less than 0.05 K, which significantly improved the temperature measurement accuracy of the far track section.

[0035] Motion blur deconvolution was performed using the Wiener filtering method. The parameters were set at a train speed of 120 km / h and a frame exposure time of... For example, the blurred displacement of the train in the track direction within a single frame is approximately 33 mm (about 5 pixels). A linear motion point spread function (PSF) is constructed using this displacement and direction. With a noise power ratio (NSR) of 0.02, Wiener inverse filtering is performed. After processing, the spatial resolution is restored from approximately 50 mm before blurring to better than 10 mm. The thermal boundaries of track joints and bolt locations are clearly restored, laying the foundation for accurate extraction of subsequent local hotspot topological features. For deceleration and braking phases where train speeds are below 30 km / h, the blurred displacement is less than 1 pixel, and the convolution step can be skipped directly to save computational resources.

[0036] Step S3, thermal feature extraction and thermal deformation inference: Extract macroscopic temperature gradient features, local hot spot topological features and thermal pulse response features from the thermal radiation distribution map, and input the extracted thermal features into a pre-trained temperature-strain conversion model. The temperature-strain conversion model integrates the thermal expansion coefficient of the orbital material, geometric constraints and historical load spectrum to infer the real-time thermal deformation and direction vector of each node.

[0037] In the specific implementation of macroscopic temperature gradient feature extraction, the following methods are adopted: The Daubechies-4 (db4) wavelet performs two-dimensional discrete wavelet decomposition, with scale weights for each layer. The coefficients are selected sequentially as 1, 2, 4, and 8. Taking a thermal map of a track curve under high-temperature conditions (ambient temperature 38℃) in summer as an example, the first layer of detail coefficients mainly reflects the high-frequency temperature difference (amplitude approximately 0.1℃) of the micro-processed texture of the track surface. The third and fourth layers can effectively capture the macroscopic thermal expansion gradient (amplitude approximately 1.2℃ / m) spanning multiple sleepers. The amplitude of the temperature gradient field synthesized by combining the coefficients of the four layers is... The maximum value in this curve region reaches 3.6℃ / m. Taking the thermal conductivity... Corresponding local heat flux density Exceeding the hot spot detection threshold The area was then subjected to local hot spot topology analysis, and was ultimately identified as a heat accumulation zone formed by the continuous frictional heat of the train on the inner side of the curve rail.

[0038] In the specific implementation of thermal pulse response feature extraction, the train passing frequency is used as an example. (The train's running cycle is approximately 12 seconds) serves as the excitation source, for items containing... Frame (frame interval) Phase-locked integration is performed on the infrared sequence to obtain the in-phase component. and orthogonal components Taking track segment 18 (a straight section, adjacent to the high-speed braking zone) as an example, the phase diagram... An area of ​​phase anomaly was detected at a depth of approximately 12 mm from the track surface, with a phase lag of about 25°, exceeding the tolerance band of ±5° for the normal reference phase. This area was identified as a suspected internal delamination defect, triggering a close-range re-inspection process. Subsequent ultrasonic phased array testing confirmed the presence of an early fatigue crack approximately 18 mm long at this location, validating the effective detection capability of phase-locked thermal imaging technology for internal defects.

[0039] In the specific implementation of the temperature strain conversion model, taking Q345B rail steel as an example, the reference temperature is... Lower reference thermal expansion coefficient Temperature gradient coefficient The effective temperature range of the model is 253 K to 373 K. The local stiffness reduction factor at the bolted connection joint is taken as... Take samples from the weld area When the current temperature at a node in the 7th track segment is 316 K and the reference temperature is 295 K, the effective thermal expansion coefficient is... Basic thermal strain After superimposing the constraint correction term and the dynamic load correction term, it is finally inferred that the thermal deformation of the node is about 0.24 mm, and the direction vector points to the track axis (the direction of thermal expansion).

[0040] The method for extracting the macroscopic temperature gradient features is as follows: [The method involves analyzing the thermal radiation distribution map.] implement Two-dimensional discrete wavelet decomposition, in the first layer layer( ) Obtain detail values ​​in the horizontal, vertical, and diagonal directions respectively. , , ; The amplitude of the temperature gradient field is synthesized from multi-scale detail coefficients: In the formula, For wavelet decomposition layer index; This represents the total number of decomposition levels. Let the scale weights be those of the j-th layer. The local heat flux density distribution diagram is approximately obtained from the heat conduction equation: In the formula, The thermal conductivity of the rail steel is given, with a value ranging from 50. 60 W / (m·K); Local heat flux density: Exceeding the hot spot detection threshold At that time, the area was marked as a hot spot and entered the local hot spot topology feature analysis process; It is determined by the thermal fatigue limit of the track material and the safety factor.

[0041] The extraction of the thermal pulse response characteristics employs phase-locked-array thermal imaging technology, using the periodic mechanical vibrations generated by the train's passage and the release of braking heat as natural periodic excitation sources, with an excitation frequency of... Determined by the frequency of train passage; for infrared image sequences containing P frames. Perform phase-locked integration to extract in-phase components. and orthogonal components : ; ; In the formula, The total number of image frames involved in the integration calculation; The time interval between adjacent frames, in seconds; For frame number, Thermal response amplitude diagram Phase diagram They are respectively: ; Phase diagram It can effectively reveal the 5 subsurface layers below the orbital surface. Internal structural defects within a depth range of 20mm; areas of phase anomaly are identified as suspected cracks or delamination defects, triggering a close-range re-inspection process.

[0042] The temperature-strain conversion model maps thermal characteristics to thermal strain, and the mapping relationship is as follows: In the formula, These are the position coordinates on the track; For position exist Thermal strain at any moment; The effective thermal expansion coefficient, which is related to temperature, is calculated using piecewise linear interpolation: In the formula, Reference temperature The reference thermal expansion coefficient at that location, in Kelvin. - ¹; Temperature gradient of thermal expansion coefficient, unit K. - ²; and These represent the lower and upper limits of the model's effective temperature range. As a nodal constraint correction term, a local stiffness reduction factor is introduced at bolted connections, weld zones, and support frame locations. Correcting the redistribution effect of thermal strain at the constraint boundary; For the train dynamic load correction item, the dynamic additional strain is calculated based on the train's real-time speed, axle load and position, and the thermal strain component is separated from the total strain. This is a historical fatigue correction term.

[0043] The formula for calculating the historical fatigue correction term is as follows: In the formula, The fatigue weakening coefficient reflects the effect of microscopic damage on the coefficient of thermal expansion, and is expressed in Kelvin. - ¹; For position exist The cumulative number of load cycles at any given time; The fatigue index, with a value range of 0 < ≤1, determined by fitting fatigue test data of track materials; when When the preset fatigue warning threshold is exceeded, the track section is marked as a priority maintenance node.

[0044] Step S4, Anomaly Detection and Graded Early Warning: An anomaly detector combining the isolated forest algorithm and the long short-term memory network (LSTM) is used to identify thermal deformation anomalies in the thermal deformation time series of each track section. The comprehensive risk index is calculated by combining the thermal deformation amount, deformation rate, historical defect records of the track section and the current operating status. Yellow, orange and red three-level early warning measures are triggered according to the preset adaptive threshold.

[0045] The combined anomaly detector of the isolated forest algorithm and LSTM network operates in a parallel voting manner; the isolated forest algorithm evaluates the thermal deformation feature vector at the k-th time step. Calculate abnormal scores for: In the formula, thermal deformation feature vector The expected average path length in a set of isolated trees; Here is a normalization factor for the average path length of an isolated tree with n samples: In the formula, Number of training samples for the isolated tree; For harmonic numbers, ; An LSTM network performs a single-step prediction of the thermal deformation time series of the k-th track segment, and the anomaly score is calculated. Defined as the standardized value of the prediction residual: In the formula, This represents the single-step prediction of thermal strain at time t by the LSTM network. is the standard deviation of the historical prediction residuals for the k-th track segment, reflecting the baseline of normal fluctuations for that track segment; The fusion anomaly score is: In the formula, The weighting coefficients for fusing the anomaly scores of the Isolation Forest algorithm and the LSTM network; when Output high-confidence warnings in real time; when Output low-confidence suspected alerts and continuously track them; among which and These are the threshold values ​​for high and low confidence warnings, respectively; the LSTM network updates its parameters online daily to adapt to data distribution drift caused by equipment aging and environmental changes.

[0046] The comprehensive risk index is calculated as follows: for the first... For each track segment, the comprehensive risk index is calculated using the following formula. : In the formula, This is the maximum allowable thermal deformation. The maximum permissible thermal deformation rate; This represents the cumulative damage value of historical defects in the k-th track segment; The reference damage value is determined by material fracture mechanics analysis; The operational status coefficient reflects the impact of current train speed and load on risk; the weighting coefficient satisfies... It is determined from historical operating data using the Analytic Hierarchy Process (AHP); Risk level classification criteria: <0.4 indicates a green and safe state; 0.4≤ <0.6 triggers a yellow alert; 0.6≤ <0.8 triggers an orange alert; A value ≥0.8 triggers a red alert and initiates closed-loop control; the time from alert triggering to train response is no more than 3 seconds.

[0047] The adaptive threshold is calculated as follows: warning threshold Based on the basic threshold, dynamic corrections are made by comprehensively considering seasonal temperature, train load, and material fatigue state. In the formula, For the first The basic threshold of the track section is determined by material mechanics calculations and a safety factor; The seasonal temperature correction factor is calculated based on the deviation between the current ambient temperature and the annual average temperature baseline, reflecting the impact of diurnal and seasonal temperature fluctuations on the normal thermal expansion range. This is the load correction factor, calculated based on the ratio of the current real-time train speed to the axle load relative to the rated value, with a value range of [0, 1]. The material fatigue correction factor is calculated using the following formula: In the formula, The fatigue sensitivity coefficient; The design life cycle number is set for the k-th track segment; the weighting coefficients satisfy... The threshold correction range is guaranteed to be no more than 30% of the base threshold; the threshold update cycle is no more than 1 hour; and the system automatically switches to conservative threshold mode under extreme high temperature weather or abnormal load conditions.

[0048] The following is a complete early warning response case illustrating the collaborative working process of each step in this invention. During the afternoon peak operating hours of a theme park in summer, the comprehensive risk index of track segment 7 (straight section, immediately upstream of the braking zone) continuously increased over three consecutive train passing cycles. Isolated forest anomaly score. LSTM prediction of standardized outlier scores in residuals Take the fusion weight , fusion of abnormal scores Exceeding the high confidence warning threshold Output a high-confidence warning. Further calculate the comprehensive risk index: current thermal deformation. thermal deformation rate Historical defects and cumulative damage Operational status coefficient (Full load, high speed), take the weight determined by AHP. Comprehensive risk index If the current ambient temperature falls within the orange alert range (0.6 ≤ R < 0.8), the system automatically sends an orange alert to the control room and initiates encrypted monitoring (every 30 seconds). Simultaneously, adaptive threshold correction is performed: the current ambient temperature is 12℃ higher than the annual average baseline. Full load operation corresponds Fatigue correction factor Corrected threshold The correction range was 11.8%, within 30%, thus meeting the constraints. After the day's operations concluded, maintenance personnel conducted a focused inspection of the 7th rail section and discovered a minor thermal fatigue crack (approximately 14 mm long) at the rail joint. This was promptly repaired by grinding, confirming the effectiveness and accuracy of the graded early warning mechanism of this invention.

[0049] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring the thermal deformation of roller coaster tracks based on infrared feature extraction, characterized in that, The method includes: Under conditions of no track load and stable ambient temperature, a full-track infrared thermal image sequence was collected. After radiometric calibration correction, a reference temperature field model for each track segment was constructed as a reference baseline for subsequent thermal deformation comparison. Infrared images of the track surface are collected in real time during train operation. Non-uniformity correction, atmospheric transmittance compensation, and motion blur deconvolution processing are performed on the image sequence to obtain a thermal radiation distribution map. Macroscopic temperature gradient features, local hot spot topological features, and thermal pulse response features are extracted from the thermal radiation distribution map. The extracted thermal features are input into a pre-trained temperature strain conversion model. The temperature strain conversion model integrates the thermal expansion coefficient of the orbital material, geometric constraints, and historical load spectrum to infer the real-time thermal deformation and direction vector of each node. An anomaly detector combining the isolated forest algorithm and long short-term memory network is used to identify thermal deformation anomalies in the time series of thermal deformation of each track section. A comprehensive risk index is calculated by combining the amount of thermal deformation, deformation rate, historical defect records of the track section, and current operating status. Yellow, orange, and red three-level early warning measures are triggered according to preset adaptive thresholds.

2. The method according to claim 1, characterized in that, The radiometric calibration correction method specifically involves radiometrically calibrating the original grayscale image acquired by the infrared thermal imager, converting the grayscale values ​​into radiance values ​​and equivalent target temperatures. The radiation calibration formula is: In the formula, The original grayscale values ​​were acquired by the infrared thermal imager. This refers to the radiance value. This is the gain coefficient; This is the bias coefficient; Target temperature The following can be obtained by inversion from Planck's radiation law: In the formula, It is the first radiation constant; It is the second radiation constant; Equivalent wavelength, unit: meters; For orbital material emissivity.

3. The method according to claim 2, characterized in that, A sequence of M consecutive infrared thermal images was collected when the orbit was unloaded and the ambient temperature was stable. The average temperature values ​​at the same spatial location were then used to construct a reference temperature field model. In the formula, The total number of steady-state frames participating in the mean calculation; For the first The temperature value of a frame of infrared thermal image at spatial coordinates (x, y, z), in K; When the variance of the mean temperature of M consecutive frames When the value is ≤0.1K, the reference field is considered to have reached a steady state; if the variance exceeds this threshold, we continue to wait until the steady-state condition is met before constructing the reference temperature field model.

4. The method according to claim 3, characterized in that, The method for extracting the macroscopic temperature gradient features is as follows: [The method involves analyzing the thermal radiation distribution map.] Perform two-dimensional discrete wavelet decomposition to obtain detail coefficients in the horizontal, vertical, and diagonal directions, and synthesize the temperature gradient field amplitude from the multi-scale detail coefficients. The local heat flux density distribution diagram is approximately obtained from the heat conduction equation: In the formula, The value is the thermal conductivity of the rail steel.

5. The method according to claim 4, characterized in that, The thermal pulse response characteristics use the periodic mechanical vibrations generated by the train passing and the release of braking heat as natural periodic excitation sources; for infrared image sequences Perform phase-locked integration to extract in-phase components. and orthogonal components : ； ； In the formula, The total number of image frames involved in the integration calculation; The time interval between adjacent frames, in seconds; For frame number, ; Thermal response amplitude diagram Phase diagram They are respectively: ; .

6. The method according to claim 5, characterized in that, The temperature-strain conversion model maps thermal characteristics to thermal strain, and the mapping relationship is as follows: In the formula, These are the position coordinates on the track; For position exist Thermal strain at any moment; The effective thermal expansion coefficient, which is related to temperature, is calculated using piecewise linear interpolation: In the formula, Reference temperature The reference thermal expansion coefficient at that location, in Kelvin. - ¹; The temperature gradient represents the coefficient of thermal expansion. and These represent the lower and upper limits of the model's effective temperature range; For node constraint correction terms; For train dynamic load correction items; This is a historical fatigue correction term.

7. The method according to claim 6, characterized in that, The formula for calculating the historical fatigue correction term is as follows: In the formula, This is the material fatigue weakening coefficient, reflecting the influence of microscopic damage on the coefficient of thermal expansion. For position exist The cumulative number of load cycles at any given time; The fatigue index, with a value range of 0 < ≤1, determined by fitting fatigue test data of track materials.

8. The method according to claim 7, characterized in that, The combined anomaly detector of the isolated forest algorithm and LSTM network operates in a parallel voting manner. Isolation Forest Algorithm for the first Thermal deformation feature vector at each time step Calculate abnormal scores ;LSTM network for the first Single-step prediction of thermal deformation time series of each track segment, anomaly score Defined as the standardized value of the prediction residual; The fusion anomaly score is: In the formula, The weighting coefficients for fusing the anomaly scores of the Isolation Forest algorithm and the LSTM network; when Output high-confidence warnings in real time; when It outputs low-confidence suspected alerts and continuously tracks them; in and These are the threshold values ​​for determining high and low confidence level warnings, respectively.

9. The method according to claim 8, characterized in that, The comprehensive risk index is calculated as follows: for the first... For each track segment, the comprehensive risk index is calculated using the following formula. : In the formula, This is the maximum allowable thermal deformation. The maximum permissible thermal deformation rate; For the first Cumulative damage value of historical defects in the track section; For reference damage values; The operational status coefficient is affected; the weighting coefficient satisfies... ; <0.4 indicates a green and safe state; 0.4≤ <0.6 triggers a yellow alert; 0.6≤ <0.8 triggers an orange alert; A value ≥0.8 triggers a red alert and initiates closed-loop control; the time from alert triggering to train response is no more than 3 seconds.

10. The method according to claim 9, characterized in that, The adaptive threshold is dynamically adjusted based on the base threshold, taking into account seasonal temperature, train load, and material fatigue state. In the formula, For the first The basic threshold for the track segment; This is a seasonal temperature correction factor. This is the load correction factor; This is a material fatigue correction factor; the weighting coefficients satisfy... The threshold update cycle does not exceed 1 hour.

Images