A coating quality online evaluation method and system based on cooling temperature curve

Abstract

The application discloses a coating quality online evaluation method and system based on a cooling temperature curve, wherein the method comprises collecting temperature data with geometric vertex and geometric valley as double characteristic points and constructing time-temperature difference series and temperature difference curves of the two, extracting a multi-dimensional feature vector containing intensity, timing, shape and spectral characteristics based on the temperature difference curve, and synchronously and real-timely obtaining coating thickness and bonding force grade prediction values through a physical guidance-data driven hybrid inversion model. The application mines the geometric characteristics of a periodic structure workpiece and its heat dissipation difference characteristics, constructs a temperature difference dynamics curve of the geometric vertex and the geometric valley during spraying and extracts multi-dimensional correlation characteristics, converts into an effective information source for evaluating coating quality and performs synchronous hybrid inversion, realizes synchronous, nondestructive and second-level online prediction of coating thickness and bonding force double indexes, and has the advantages of stability and reliability, precision and rapidness, and is especially suitable for high-sensitivity evaluation of coating quality of complex structure workpieces.

Description

Technical Field

[0001] This invention belongs to the field of material coating quality assessment technology, specifically relating to an online coating quality assessment method and system based on cooling temperature curves, which is particularly suitable for the synchronous, online, non-destructive assessment of coating thickness and adhesion of workpieces with periodic undulation structures, such as threaded steel bars. Background Technology

[0002] Metal components such as reinforcing bars are among the core materials for ensuring the strength of structures, and all require surface protection treatments such as spraying before being put into use. In the process of metal surface treatment, especially the thermal spraying of anti-corrosion coatings (such as aluminum coatings) for threaded reinforcing bars, coating thickness and coating-substrate interface adhesion are the two most critical indicators for measuring coating quality and determining its protective life.

[0003] The existing methods for evaluating coating quality based on these two indicators in industrial production, and their problems, are as follows:

[0004] (1) Destructive and offline sampling inspection evaluation methods: The quantitative evaluation of coating adhesion relies heavily on destructive testing, such as hydraulic adhesion testers (e.g., PosiTest), which require the coating to be pulled off the substrate. This evaluation method can only rely on offline sampling inspection, which cannot achieve full inspection, has the risk of missing inspection, and causes waste of materials and time. In particular, the surface texture of threaded steel bars will cause slight differences in coating thickness and adhesion, making it difficult to cover all feature areas, and the obtained quality evaluation results have a large deviation.

[0005] (2) Single-parameter and offline detection and evaluation methods: This type of method relies on offline, single-parameter damage detection technology to achieve quality evaluation. Among them, the evaluation method based on coating thickness is usually carried out after the coating has cooled to room temperature. It cannot reflect the instantaneous quality fluctuations during the coating process and cannot evaluate the bonding force. The indirect evaluation method based on bonding force is easily affected by the geometry of the workpiece (such as threads), has poor applicability to complex structural parts such as threaded steel bars, and has insufficient accuracy and reliability. The quality evaluation results are not ideal.

[0006] In addition, there are methods that use active thermal excitation to heat the coating after spraying (such as flash lamp thermal excitation) and then analyze its thermal response to assess thickness or defects. This not only increases the number of inspection steps and energy consumption, but also cannot be synchronized with the spraying process, resulting in low efficiency. Furthermore, it is difficult to avoid damage to the existing coating when the coating is subjected to secondary thermal excitation.

[0007] Therefore, there is an urgent need to find an online detection method that can simultaneously, non-destructively, and quantitatively assess the two key quality indicators of coating thickness and interfacial adhesion during the spraying production process. Summary of the Invention

[0008] This invention proposes an online coating quality evaluation method and system based on cooling temperature curves, aiming to solve the technical problem that existing technologies cannot simultaneously, online, and non-destructively quantitatively evaluate coating thickness and interfacial adhesion during thermal spraying.

[0009] This invention is achieved through the following technical solution:

[0010] This invention proposes an online coating quality evaluation method based on cooling temperature curves, comprising the following steps:

[0011] S1, Dual Feature Point Temperature Data Acquisition: Taking a workpiece with a periodic surface structure as the spraying target, within one feature cycle of the spraying target, accurately locate two feature points with geometric and heat dissipation differences; taking the spray gun spraying to the feature cycle as the zero point of time, synchronously acquire infrared temperature-time series data of the two feature points for no less than 30s during the cooling process.

[0012] S2, Temperature difference data series and temperature difference curve construction: Based on infrared temperature-time series data, construct the original time-temperature difference sequence ΔT between two feature points. raw (t), and then filtered to construct the temperature difference curve ΔT(t);

[0013] S3, High-dimensional physical feature extraction: from the original time-temperature difference sequence ΔT raw From the temperature difference curve ΔT(t), extract the multidimensional feature vector F, which includes intensity features, temporal features, morphological features and spectral features;

[0014] S4, Hybrid Model Synchronous Inversion: The multidimensional feature vector F is input into a pre-trained hybrid inversion model; the hybrid inversion model synchronously outputs the predicted coating thickness T for the current monitoring location. final and the probability distribution of coating adhesion level [P] high , P mid , P low This serves as the result of coating quality assessment.

[0015] Based on the above methods, by utilizing the periodic surface structure and inherent heat dissipation differences of the workpiece surface, a temperature difference curve feature is constructed with spraying as the starting point. Corresponding features with coating thickness and adhesion are selectively extracted from multiple dimensions such as time domain, frequency domain, and curve morphology. The temperature difference dynamics process is comprehensively characterized and synchronous hybrid inversion is performed. This enables the synchronous, online, non-destructive, and rapid acquisition of the two key quality indicators of coating thickness and adhesion during the spraying process, overcoming the limitations of traditional methods such as destructive, offline, and single-parameter detection.

[0016] Further, in step S1, the method for locating the feature points is as follows: at least two feature periods are covered within the field of view of the infrared thermal imager, and the geometric vertices and valleys within the same feature period are automatically located as feature points using an edge recognition algorithm within the field of view; the temperature value collected for each feature point is the average temperature of the small pixel area centered on it, so as to creatively utilize the heat dissipation differences generated by the workpiece's own geometric structure, and transform the interference of the geometric structure into an effective information source for coating quality assessment, which reduces the amount of data processing and can better cover the temperature changes within the feature period, thereby achieving specific and highly sensitive detection of the coating quality of complex structural parts, and is particularly suitable for coating quality detection of complex structural parts such as threaded steel bars.

[0017] Furthermore, in step S3, the extracted multidimensional feature vector F includes multiple features from the following:

[0018] Intensity characteristics: peak temperature difference F1, steady-state temperature difference F2, where the steady-state temperature difference F2 is the mean temperature difference of the curve fitting over t∈[25,30]s;

[0019] Timing characteristics: peak time F3, full width at half maximum (F4), decay time constant F5;

[0020] Morphological characteristics: initial upward slope F6, integral area of ​​t∈[0,5]s - F7, integral area of ​​t∈[5,15]s - F8, morphological skewness F9.

[0021] Spectral characteristics: dominant frequency F10, spectral entropy F11, where the temperature difference curve is converted into a power spectrum by fast Fourier transform, the dominant frequency F10 is the frequency corresponding to the maximum peak value of the power spectrum, and the spectral entropy F11 is the Shannon entropy of the power spectrum.

[0022] Auxiliary features: maximum temperature at geometric vertices F12, maximum temperature at geometric valleys F13, cooling rate at geometric vertices F14, ambient temperature F15.

[0023] Based on the above scheme, specific and computable feature sets are provided from different dimensions. By utilizing the sensitivity of different features to thickness and bonding strength, the two quality indicators can be separated and accurately evaluated.

[0024] Preferably, in step S4, the hybrid inversion model is a two-stage model: the first stage is a physical-guided model, used to quickly calculate the rough estimate T of the coating thickness from the feature vector F based on the physical principles of thermal conduction and the initial rule base for judging adhesion. phy And a preliminary classification of the bonding force state is performed. phy The second stage is a data-driven model based on machine learning, using the feature vector F and the output [T] of the first stage. phy , S phyAs an enhancement feature, it is input into a machine learning model to perform precise regression of the thickness value and probabilistic classification of the binding strength level, respectively, to obtain the final prediction result T. final and [P] high , P mid , P low This model utilizes a physics-guided model to ensure the interpretability of the method and allows for preliminary judgment with a small number of samples. It then uses a data-driven model to achieve high-precision fitting of complex nonlinear relationships, combining robustness and accuracy.

[0025] Furthermore, in the initial physical guidance assessment of the first stage, the coarse thickness estimate T phy Through formula T phy = a(F8 / F12) + bF5 + c is calculated, where F8 is the integral area of ​​t∈[5,15]s, F12 is the highest temperature at the geometric vertex, F5 is the decay time constant, and a, b, and c are calibration coefficients.

[0026] Furthermore, in the initial physical guidance judgment of the first stage, S is obtained. phy The initial rule base for determining the binding force includes at least the following two rules:

[0027] Rule 1: If the peak temperature difference F1 exceeds the upper limit of the threshold and the peak time F3 exceeds the lower limit of the threshold, the coating is judged to be too thin or poorly bonded, and the bonding strength is marked as "low".

[0028] Rule 2: If the geometric vertex cooling rate F14 is significantly greater than the historical benchmark, the substrate is considered to have good thermal conductivity, and the adhesion is marked as "high," or the coating is considered extremely thin, and the adhesion T is also considered high. phy Determine the label level.

[0029] Furthermore, the second stage, the training steps for the data-driven model based on machine learning, include:

[0030] S4201, Take N sets of workpieces and perform spraying operations with preset parameters, and collect temperature data and obtain multi-dimensional feature vector F synchronously according to steps S1-S3. i And obtain the rough thickness estimate T in the first stage of step S4. phy and bonding strength rating S phy ;

[0031] S4202, After spraying, the actual coating thickness T is obtained by offline testing of the sprayed sample. true i and bonding force value C true i And define the range of bonding force thresholds;

[0032] S4203, with multidimensional feature vector F iAnd the T obtained in the first phase phy As input, with T true i As a label, the thickness is used to accurately estimate the value of T. final To obtain the output, a machine learning model is trained to obtain a thickness regression model;

[0033] S4204, with multidimensional feature vector F i And the S obtained in the first phase phy As input, in C true i As the label, with the binding probability [P] high , P mid , P low The output is used to train a machine learning model, resulting in a binding force classification model.

[0034] Preferably, the aforementioned method further includes an online calibration step: periodically selecting evaluation points, and after completing the online evaluation, performing actual measurement verification using an offline measurement device; based on the deviation between the online predicted value and the offline measured value, when the thickness deviation exceeds a preset threshold or the bonding strength evaluation crosses levels, initiating the deviation correction of the hybrid inversion model; this step establishes a closed-loop verification mechanism between the online evaluation system and the offline standard, ensuring the stability and reliability of the evaluation results during long-term operation and preventing model drift.

[0035] Preferably, the aforementioned method further includes a model adaptive update step: using the offline verified test results and the corresponding multidimensional feature vector F as new samples, and when the new sample library reaches a preset number, the incremental training update of the hybrid inversion model is automatically started, so that the system has continuous learning capabilities, can accumulate production data and adaptively optimize the model, adapt to changes in process parameters, raw material batches, etc., and continuously improve the evaluation accuracy and generalization ability.

[0036] This invention also provides an online coating quality evaluation system based on a cooling temperature curve, comprising an infrared thermal imager, a processing unit, and an early warning unit; the infrared thermal imager is used to acquire temperature data at dual feature points in real time, the processing unit is used to execute any of the methods described above; the early warning unit is used to detect when the predicted thickness value T... final Exceeding the preset target thickness by ±20μm, or P low An alarm is triggered when the value exceeds 0.7, enabling real-time, automatic alerts for serious quality defects, facilitating timely intervention and preventing batch defects.

[0037] Compared with the prior art, at least one technical solution of the present invention has the following significant technical effects:

[0038] (1) Simultaneous online evaluation of dual indicators for complex geometric structures: The heat dissipation difference brought about by the geometric features (rib tops and valley bottoms) of workpieces with periodic undulations such as threaded steel bars is creatively explored, and this "interference factor" is transformed into an effective information source for evaluating coating quality. Furthermore, the infrared temperature field information during the natural cooling process after spraying is creatively utilized. By analyzing the physical correlation between the characteristic quantities of the temperature difference dynamic curves of the dual feature points of the rib top (geometric apex) and the valley bottom (geometric valley point) and the coating thickness and bonding force, the inversion method is used during the spraying process to achieve synchronous, non-destructive, and second-level online prediction of the two key quality indicators of coating thickness and interface bonding force. This solves the limitations of traditional methods that are destructive, offline, and single-parameter. Moreover, it has strong generalization ability and is especially suitable for the specific and highly sensitive detection of coating quality of complex structural parts.

[0039] (2) Deep integration of physical mechanism and data-driven approach: A hybrid inversion model architecture of "physical guidance-data-driven" is proposed. The first-stage physical model provides interpretable preliminary judgment results and feature enhancement based on the principle of heat conduction. The second-stage machine learning model is responsible for fitting complex nonlinear relationships to achieve accurate prediction. This architecture has both physical interpretability and high-precision generalization ability, which is more reliable than the pure black box model and has higher evaluation accuracy and stability.

[0040] (3) Closed-loop quality control and adaptive capability: The online calibration mechanism ensures long-term detection stability, and the model adaptive update enables continuous learning, so that the system can adapt to changes in raw materials, environment, etc., forming an intelligent closed-loop quality control of perception-evaluation-feedback-optimization, which significantly improves the consistency and reliability of production quality. Attached Figure Description

[0041] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0042] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0043] To facilitate a better understanding of this solution, the following detailed description of the invention is based on an example of thermal spraying aluminum coating on threaded steel bars, but the scope of protection of the invention is not limited thereto.

[0044] Example 1

[0045] This embodiment provides an online coating quality evaluation method based on a cooling temperature profile, including the following steps:

[0046] S1, Temperature data acquisition at dual feature points:

[0047] A medium-wave or long-wave high-speed infrared thermal imager is aimed at the sprayed area of ​​the threaded steel bar, and the infrared thermal imager's field of view covers at least two complete thread cycles; within the field of view, a general edge recognition algorithm is used to automatically locate the geometric vertices (rib tops) and geometric valleys (valley bottoms) within one thread cycle as feature points.

[0048] Using the moment when the spray gun reaches the selected thread cycle as the zero point, the temperature data T of the rib top and valley areas are simultaneously recorded at a frequency of 50Hz for the following 30 seconds. top (t) and T valley (t), and simultaneously record the ambient temperature T. env Calculate the infrared temperature-time series data ΔT relative to the ambient temperature rise. top (t), ΔT valley (t), where T top (t) and T valley (t) represents the average temperature of a 3×3 pixel region centered on the rib top or valley bottom.

[0049] S2, Temperature Difference Data Series and Temperature Difference Curve Construction:

[0050] Based on the infrared temperature-time series data ΔT obtained in step S1 top (t), ΔT valley (t), generating the original rib-valley temperature difference sequence ΔT raw (t) = ΔT top (t) - ΔT valley (t);

[0051] The original series is subjected to Savitzky-Golay filtering with a window length of 11 and a third-order polynomial to obtain a smooth ΔT(t) curve, which is the temperature difference curve.

[0052] S3, High-dimensional physical feature extraction: from the original time-temperature difference sequence ΔT raw In the temperature difference curve ΔT(t), the multidimensional feature vector F containing intensity features, temporal features, morphological features and spectral features is accurately calculated and extracted;

[0053] Furthermore, the multidimensional feature vector F = [F1, F2, ..., F15] is used, and each sub-feature is specifically selected based on its physical correlation with coating thickness and adhesion. The specific selection criteria are shown in Tables 1 to 3.

[0054] Table 1: Calculation methods and physical significance of extracted feature quantities

[0055]

[0056] Table 2: Comparison of Physical Correlation Between Extracted Feature Quantities and Coating Thickness and Adhesion

[0057]

[0058] Table 3: Differences in the response of extracted feature quantities to coating thickness and adhesion.

[0059]

[0060] S4, Hybrid Model Synchronous Inversion: The multidimensional feature vector F extracted in step S3 is input into the pre-trained hybrid inversion model; the hybrid inversion model synchronously outputs the predicted coating thickness T for the current monitoring location. final And the probability distribution of coating adhesion level [P] high , P mid , P low The coating quality assessment result is used as the result of the hybrid inversion model. Preferably, the hybrid inversion model adopts a two-stage model: the first stage is a physical-guided model, which is used to quickly calculate the rough estimate T of the coating thickness from the feature vector F based on the physical principles of thermal conduction and the initial rule base for judging adhesion. phy And a preliminary classification of the bonding force state is performed. phy The second stage is a data-driven model based on machine learning, using the feature vector F and the output [T] of the first stage. phy , S phy As an enhancement feature, it is input into a machine learning model to perform precise regression of the thickness value and probabilistic classification of the binding strength level, respectively, to obtain the final prediction result T. final and [P] high , P mid , P low ].

[0061] Specifically, the process of constructing the hybrid inversion model includes:

[0062] First, a batch of steel bars with known coating thickness (measured offline using an eddy current thickness gauge) and bonding strength (destructive testing offline using a hydraulic adhesion tester) were collected. The corresponding feature vector F was then extracted according to steps S1-S3. i And obtain the measured coating thickness T true i and bonding force value C true i At the same time, the bonding strength level threshold is divided according to the measured value (e.g., "high" level > 40MPa, "medium" level is set to 30-40MPa, "low" level < 30MPa); (corresponding to steps S4201 and S4202)

[0063] Secondly, in the initial physical guidance assessment in the first stage, a small number of samples are used to establish an inversion physical guidance model:

[0064] (1) Thickness inversion model: Based on the analysis of Tables 2 and 3, the response of the cooling temperature curve to the coating thickness mainly includes features related to heat capacity, such as decay time constant F5, early and middle integrated area F7 and F8, and features related to heat conduction path, such as peak time F3 and half peak width F4.

[0065] Based on this, with T phy =a(F8 / F12)+b*F5+c constructs a rough estimate of thickness T phy The physical inversion model is used, where F8 / F12 characterizes the medium-term thermal storage capacity per unit fin top temperature, and F5 directly reflects the heat capacity. The combination of the two can obtain a high-precision thickness estimate. a, b, and c are obtained by linear fitting of a small number of samples.

[0066] (2) Preliminary model for bonding force: Based on the analysis in Tables 2 and 3, the characteristic quantities that are sensitive to the bonding force response of the cooling temperature curve are selected, and the level is initially judged by combining the preliminary judgment rule base, and S is output. phy The initial classification identifier for ∈{high, medium, low} should include at least the following two rules in the initial judgment rule base:

[0067] Rule 1: If the peak temperature difference F1 exceeds the upper limit of the threshold and the peak time F3 exceeds the lower limit of the threshold, the coating is judged to be too thin or poorly bonded, and the bonding strength is marked as "low".

[0068] Rule 2: If the geometric vertex cooling rate F14 is significantly greater than the historical benchmark, the substrate is considered to have good thermal conductivity, and the adhesion is marked as "high," or the coating is considered extremely thin, and the adhesion T is also considered high. phy Determine the label level.

[0069] Finally, in the second stage of data-driven model refinement based on machine learning, all collected samples are used as training samples to construct and optimize machine learning models and ensemble models for thickness and binding force classification:

[0070] (1) Thickness regression model (S4203): The corresponding multidimensional feature vector F is obtained. i And the T obtained in the first phase phy As input, with T true i Using the Huber loss as the objective function and the labels as the target, a GBDT thickness regression model is trained to obtain the thickness prediction T. gbdt and with T final = 0.9T gbdt + 0.1T phy Output the final predicted coating thickness T final ;

[0071] (2) Associative force classification model (S4204): based on multidimensional feature vector F i And the S obtained in the first phasephy As input, in C true i For each label, a GBDT multi-classifier model is trained, and combined with the associative strength level threshold, the probabilities of the three levels are output [P]. high ,P mid , P low ] serves as an indicator of the bonding strength level.

[0072] This embodiment is based on the above method and has the following advantages:

[0073] Achieving simultaneous online evaluation of dual indicators for complex geometric structures: This method creatively exploits the inherent heat dissipation differences in the geometric features (rib tops, valley bottoms) of periodically undulating structural workpieces such as threaded steel bars, transforming this "interference factor" into an effective information source for evaluating coating quality. A temperature difference curve is constructed starting from real-time spraying. By analyzing the physical correlation between the temperature difference dynamics curves of the two feature points (rib tops, geometric apexes, and valley bottoms) and coating thickness and adhesion, key feature quantities are selectively extracted from multiple dimensions, including time domain, frequency domain, and curve morphology. This comprehensively characterizes the temperature difference dynamics process and performs synchronous hybrid inversion, enabling simultaneous, second-level online, and non-destructive acquisition of the two key quality indicators—coating thickness and adhesion—during the spraying process. This achieves specific and highly sensitive detection of coating quality for complex structural parts, effectively overcoming the limitations of traditional destructive, offline, and single-parameter detection methods.

[0074] Example 2

[0075] This embodiment provides an online coating quality evaluation method based on a cooling temperature curve, which differs from Embodiment 1 in that:

[0076] Based on Example 1, an online calibration step is also included:

[0077] The online calibration process is initiated at specific intervals, such as every 50 steel bars produced or every batch of materials changed.

[0078] Take any one threaded steel bar in the current cycle and randomly select any three threaded cycles with dual feature points as monitoring points.

[0079] After completing the online evaluation of Example 1 and obtaining the predicted coating thickness and adhesion level probability values, an offline measurement device was used for actual measurement verification, and the deviations ΔT and ΔC between the online predicted values ​​and the offline measured values ​​were calculated.

[0080] When |ΔT|>15μm or ΔC crosses levels, the offline detection value and the corresponding multidimensional feature vector F are used as feature samples. The hybrid inversion model is retrained according to step S4 of Example 1 to achieve bias correction.

[0081] The advantage of this embodiment is that it establishes a closed-loop verification mechanism between the online evaluation system and the offline standard, ensuring the stability and reliability of the evaluation results during long-term operation and preventing model drift.

[0082] Example 3

[0083] This embodiment provides an online coating quality evaluation method based on a cooling temperature curve, which differs from Embodiments 1 and 2 in that:

[0084] Based on Example 2, a model adaptive update step is also included:

[0085] All offline verification test results and corresponding multidimensional feature vectors F collected in Example 2 are stored as new samples in the verification sample library. When the verification sample library reaches a preset number (e.g., 100 groups), the incremental training and update of the hybrid inversion model is automatically started during idle periods. The incremental training process is the same as step S4 in Example 1.

[0086] The advantage of this embodiment is that it enables the assessment and prevention system to have continuous learning capabilities, accumulate production data and adaptively optimize the model to adapt to changes in spraying process parameters, raw material batches, etc., thereby improving the assessment accuracy and generalization ability.

[0087] Example 4

[0088] This embodiment provides an online coating quality evaluation system based on a cooling temperature curve, including an infrared thermal imager, a processing unit, and an early warning unit;

[0089] Among them, the infrared thermal imager is used to collect temperature data of dual feature points in real time. Preferably, the infrared thermal imager is a medium-wave (3-5μm) or long-wave (8-14μm) high-speed thermal imager with a spatial resolution of not less than 640×512 pixels, thermal sensitivity (NETD) <50mK, and full frame rate of not less than 100Hz; and is equipped with a macro lens with a field of view sufficient to cover at least two spiral cycles.

[0090] The processing unit is used to execute the methods of Embodiments 1 to 3;

[0091] The early warning unit is used when the predicted thickness value T final Exceeding the preset target thickness by ±20μm, or P low An alarm is triggered when the value exceeds 0.7, enabling real-time, automatic alerts for serious quality defects, facilitating timely intervention and preventing batch defects.

[0092] Based on the online evaluation method and system described in Examples 1 to 4, HRB400 steel bars with a diameter of 20mm were sprayed and the coating quality was evaluated. Thirty groups of samples with measured coating thickness of 150-350μm and bonding strength of 25-50MPa were prepared.

[0093] The accuracy of the evaluation method of this invention is as follows:

[0094] (1) Coating thickness prediction: The mean absolute error (MAE) is 11.5 μm and the coefficient of determination R²>0.94, achieving high accuracy and high reliability in online thickness prediction;

[0095] (2) Binding strength classification: In the three-level classification task, the macro average score reached 0.89; among them, the recall rate of the "low" binding strength category was as high as 0.95, which effectively ensured that serious defects were not missed.

[0096] (3) Online application accuracy: The coating quality of 200 steel bars was continuously evaluated online on the production line. The system issued an early warning for 5 defect areas with suspected "low" bonding strength. After offline destructive verification, the bonding strength of 4 of them was indeed lower than 30MPa. The early warning accuracy rate reached 80%, which is significantly better than the traditional sampling inspection method.

[0097] In summary, the online evaluation method and system of this invention have high stability and high accuracy, providing a feasible solution for the simultaneous, online, and non-destructive testing of two key quality evaluation indicators, coating thickness and coating adhesion, during the spraying production process. It achieves second-level online evaluation of coating quality, effectively solving the problems of missed detection and incomplete coverage that exist with reliance on destructive sampling inspection, as well as the defects that restrict production efficiency and cannot achieve completely non-destructive testing when performing single-parameter offline testing or secondary thermal excitation testing after spraying is completed. It has advantages such as strong adaptability to complex workpieces, fast, accurate, convenient, and non-destructive testing.

[0098] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for online evaluation of coating quality based on cooling temperature profiles, characterized in that, Includes the following steps: S1, Dual Feature Point Temperature Data Acquisition: Taking a workpiece with a periodic surface structure as the spraying target, within one feature cycle of the spraying target, accurately locate two feature points with geometric and heat dissipation differences; taking the spray gun spraying to the feature cycle as the zero point of time, synchronously acquire infrared temperature-time series data of the two feature points for no less than 30s during the cooling process. S2, Temperature difference data series and temperature difference curve construction: Based on infrared temperature-time series data, construct the original time-temperature difference sequence ΔT between two feature points. raw (t), and then filtered to construct the temperature difference curve ΔT(t); S3, High-dimensional physical feature extraction: from the original time-temperature difference sequence ΔT raw From the temperature difference curve ΔT(t), extract the multidimensional feature vector F, which includes intensity features, temporal features, morphological features and spectral features; S4, Hybrid Model Synchronous Inversion: The multidimensional feature vector F is input into a pre-trained hybrid inversion model; the hybrid inversion model synchronously outputs the predicted coating thickness T for the current monitoring location. final And the probability distribution of coating adhesion level [P] high , P mid , P low This serves as the result of the coating quality assessment. The hybrid inversion model is a two-stage model: the first stage is a physical-guided model, which is used to quickly calculate a rough estimate of the coating thickness T based on the physical principles of thermal conduction and the initial rule base for judging adhesion, using a multi-dimensional feature vector F. phy The bonding force states were initially classified, and the bonding force level identifier S was obtained. phy ; The second stage is a data-driven model based on machine learning, using a multidimensional feature vector F and the output [T] from the first stage. phy , S phy As an enhancement feature, it is input into a machine learning model to perform precise regression of the thickness value and probabilistic classification of the binding strength level, respectively, to obtain the final prediction result T. final and [P] high , P mid ,P low ].

2. The method according to claim 1, characterized in that, In step S1, the method for locating the feature points is as follows: at least two feature periods are covered within the field of view of the infrared thermal imager, and the geometric vertices and geometric valleys within the same feature period are automatically located as feature points using an edge recognition algorithm within the field of view; the acquisition temperature value of each feature point is the average temperature of the small pixel area centered on it.

3. The method according to claim 2, characterized in that, In step S3, the extracted multidimensional feature vector F includes multiple features from the following: Intensity characteristics: peak temperature difference F1, steady-state temperature difference F2, where the steady-state temperature difference F2 is the mean temperature difference of the curve fitting over t∈[25,30]s; Timing characteristics: peak time F3, half-peak width F4, decay time constant F5; Morphological characteristics: initial upward slope F6, integral area of ​​t∈[0,5]s - F7, integral area of ​​t∈[5,15]s - F8, morphological skewness F9. Spectral characteristics: dominant frequency F10, spectral entropy F11, where the temperature difference curve is converted into a power spectrum by fast Fourier transform, the dominant frequency F10 is the frequency corresponding to the maximum peak value of the power spectrum, and the spectral entropy F11 is the Shannon entropy of the power spectrum. Auxiliary features: maximum temperature at geometric vertices F12, maximum temperature at geometric valleys F13, cooling rate at geometric vertices F14, ambient temperature F15.

4. The method according to claim 3, characterized in that, In the initial physical guidance assessment of the first stage, the rough estimate of thickness T is... phy Through formula T phy = a(F8 / F12) + bF5 + c is calculated, where F8 is the integral area of ​​t∈[5,15]s, F12 is the highest temperature at the geometric vertex, F5 is the decay time constant, and a, b, and c are calibration coefficients.

5. The method according to claim 3, characterized in that, In the initial physical guidance and judgment of the first stage, S is obtained. phy The initial rule base for determining the binding force includes the following two rules: Rule 1: If the peak temperature difference F1 exceeds the upper limit of the threshold and the peak time F3 exceeds the lower limit of the threshold, the coating is judged to be thin or poorly bonded, and the bonding strength is marked as "low". Rule 2: If the geometric vertex cooling rate F14 is greater than the historical benchmark, the substrate is judged to have good thermal conductivity, and the adhesion is marked as "high" level; or the coating is judged to be thin, and the adhesion T is marked as high. phy Determine the label level.

6. The method according to claim 1, characterized in that, The second stage involves training a data-driven model based on machine learning, including the following steps: S4201, Take N groups of workpieces and perform spraying operations with preset parameters. Collect temperature data and obtain multi-dimensional feature vector F simultaneously according to steps S1-S3, and obtain the rough thickness estimate T in the first stage of step S4. phy and bonding strength rating S phy ; S4202, After spraying, the actual coating thickness T is obtained by offline testing of the sprayed sample. true i and bonding force value C true i And define the range of bonding force thresholds; S4203, using the multidimensional feature vector F and T obtained in the first stage phy As input, with T true i For the label, use the predicted thickness value T final To obtain the output, a machine learning model is trained to obtain a thickness regression model; S4204, using the multidimensional feature vector F and S obtained in the first stage phy As input, in C true i As the label, with the binding probability [P] high , P mid , P low The output is used to train a machine learning model, resulting in a binding force classification model.

7. The method according to claim 1, characterized in that, It also includes an online calibration step: periodically selecting evaluation points, and after completing the online evaluation, using an offline measurement device for actual measurement verification; based on the deviation between the online predicted value and the offline measured value, when the thickness deviation exceeds the preset threshold or the bonding strength evaluation crosses the level, the deviation correction of the hybrid inversion model is initiated.

8. The method according to claim 7, characterized in that, It also includes a model adaptive update step: the offline verified test results and the corresponding multidimensional feature vector F are used as new samples. When the new sample library reaches a preset number, the incremental training and update of the hybrid inversion model is automatically started.

9. An online coating quality evaluation system based on cooling temperature profiles, characterized in that, It includes an infrared thermal imager, a processing unit, and an early warning unit; the infrared thermal imager is used to collect temperature data of dual feature points in real time, and the processing unit is used to execute the method as described in any one of claims 1-8; The early warning unit is used when the thickness prediction value T final Exceeding the preset target thickness by ±20μm, or P low An alarm will be issued when the value is greater than 0.7.

Images