Thermal imaging-based analysis method and system for uniformity of recycled asphalt mortar mixing

Abstract

The application discloses a recycled asphalt mortar mixing uniformity analysis method and system based on thermal imaging, relates to the technical field of computer vision and industrial automation, and introduces multi-scale temperature feature analysis and time sequence dynamic modeling, so that the detection capability for microscopic unevenness is remarkably improved; non-contact thermal imaging collection is adopted, image processing and machine learning algorithms are combined, precise identification and quantitative evaluation of temperature field abnormalities in the mixing process are realized; the application effectively enhances the response sensitivity to the edges of small RAP agglomerates by extracting multi-scale temperature gradient features; the correlation model of pixel-level temperature response and material thermal conductivity characteristics is established by introducing time sequence temperature residual features, which can distinguish the abnormality caused by real agglomeration from ordinary thermal noise, and the reliability of the detection result is greatly improved; and through unsupervised clustering and adaptive threshold segmentation technology, the change of different mixing conditions and material proportions can be adapted, and the misjudgment risk of the fixed threshold is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and industrial automation, in particular to a method and system for analyzing the uniformity of recycled asphalt mortar based on thermal imaging. BACKGROUND

[0002] In the recycling of asphalt pavement, the large-scale application of recycled asphalt pavement material (RAP) has become an industry trend; however, during the milling, crushing and storage of RAP, it is easy to form agglomerated particles that wrap the internal aged asphalt, and these agglomerates are difficult to be completely dispersed and infiltrated by new asphalt during the mixing stage, resulting in material inhomogeneity at the microscale in the recycled mortar.

[0003] Currently, the main method for analyzing mixing uniformity is based on thermal imaging technology, which measures the temperature field distribution of the material flow during mixing by non-contact means, and calculates the regional temperature statistics using image processing algorithms to evaluate uniformity; some systems further introduce machine learning algorithms to identify temperature field patterns to monitor the mixing state.

[0004] However, the existing technology has limited ability to detect micro-inhomogeneity caused by RAP agglomeration; thermal imaging systems tend to mask the temperature signals of small agglomerates in the overall temperature background of the surrounding material, making it difficult to effectively identify such local defects relying solely on macro temperature statistical characteristics; in addition, the differences in thermal conductivity between RAP agglomerates and surrounding new materials are not fully modeled, making it challenging to accurately infer material uniformity based on temperature uniformity. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides a method and system for analyzing the uniformity of recycled asphalt mortar based on thermal imaging, which solves the problem that existing thermal imaging methods cannot identify micro-inhomogeneity caused by RAP agglomeration, as local temperature signals are easily masked by macro averages, and there is a lack of in-depth modeling of thermal conductivity characteristics.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] In a first aspect, the present application provides a method for analyzing the uniformity of recycled asphalt mortar based on thermal imaging, which includes:

[0009] Step S1: non-contact thermal imaging acquisition of the material field of view during mixing to obtain a sequence of thermal images arranged in time order; wherein the acquisition includes emissivity setting and temperature calibration;

[0010] Step S2, bad point correction, denoising and radiation temperature correction are performed on the thermal image sequence, and disparity / occlusion correction is performed based on the set blender geometry and camera pose, to obtain a corrected thermal image sequence;

[0011] Step S3, temperature features including at least one of the following are extracted from the corrected thermal image sequence:

[0012] Multi-scale temperature gradient features, local temperature variation coefficient and time series temperature residual features; wherein the time series temperature residual features are constructed based on the temperature response difference of the same pixel point over time;

[0013] Step S4, threshold segmentation and morphological processing are performed on the feature map to obtain a set of abnormal candidate regions;

[0014] Step S5, unsupervised clustering is performed on the abnormal candidate regions, and pseudo-abnormal filtering is performed based on the temperature features and morphological features of the regions, to obtain a set of abnormal regions of suspected RAP clusters;

[0015] Step S6, the number, area proportion, distribution density and intensity weighted index of the abnormal regions are counted, the blender uniformity index is calculated, and the evaluation result is output.

[0016] As a preferred scheme of the regenerative asphalt mortar blending uniformity analysis method based on thermal imaging according to the application, wherein: the multi-scale temperature gradient features are obtained by calculating the edge response in different window scales, and the edge operator includes at least one of Sobel, Laplace-Gaussian or Canny.

[0017] As a preferred scheme of the regenerative asphalt mortar blending uniformity analysis method based on thermal imaging according to the application, wherein: the time series temperature residual features are obtained by time registration of the thermal image sequence, fitting of the pixel temperature sequence using a background heat conduction response model, and taking the statistical quantity of the fitting residual as an abnormal score;

[0018] The fitting and residual construction steps of the background heat conduction response model include:

[0019] A) After completing the time registration, a first-order background model driven by global thermal load is established at the pixel layer, and a prediction residual is formed based on this, which is used as an abnormality measure:

[0020] ,

[0021] ,

[0022] Wherein, represents the background prediction temperature of pixel at discrete time , represents the background prediction temperature of pixel a baseline constant term, denotes a pixel an inertia coefficient for the prediction of the temperature at the previous time instant, denotes a pixel a gain coefficient for the global drive, denotes the time instant a global temperature drive sequence, denotes a pixel a prediction residual at the time instant denotes a pixel a measured temperature at the time instant denotes a pixel index, denotes a time frame index;

[0023] B) the ratio of the residual peak to a relatively robust scale is taken as the pixel anomaly score, and the scale is estimated in the form of a Gaussian equivalent of the median absolute deviation:

[0024] ,

[0025] ,

[0026] wherein, denotes the anomaly score of a pixel denotes the operation that takes the maximum over the time index denotes the absolute value operation, denotes a pixel a robust scale estimator of the residual sequence, denotes the median operator, denotes the median of the residual sequence over all time frames for a pixel a constant for converting the MAD to a Gaussian equivalent standard deviation;

[0027] C) the pixel anomaly score set is statistically evaluated within a low gradient background mask, and a threshold is formed as a quantile:

[0028] ,

[0029] wherein, denotes the pixel-level anomaly score threshold, denotes the quantile operator that returns the quantile of the input set at a probability level denotes the target false alarm rate, denotes the background pixel set consisting of low gradient and stable regions;

[0030] ​​​​​D) Huber-weighted recursive least squares is adopted for parameter estimation of step A), initialized by smoothing from median of time instants Take the initial steady-state frame of mixing, limited to , Take the non-negative, get pixels, and then perform connected component aggregation.

[0031] As a preferred scheme of the hot imaging-based recycled asphalt mortar mixing uniformity analysis method of the present application, wherein: the threshold segmentation adopts an adaptive threshold based on noise level estimation, and the noise level is estimated from a uniform black body calibration frame or an initial steady-state frame of mixing.

[0032] As a preferred scheme of the hot imaging-based recycled asphalt mortar mixing uniformity analysis method of the present application, wherein: the unsupervised clustering adopts density clustering, and the density distance threshold and the minimum sample number are adaptively set according to the scale distribution of abnormal candidates and the spatial resolution of the field of view.

[0033] As a preferred scheme of the hot imaging-based recycled asphalt mortar mixing uniformity analysis method of the present application, wherein: the pseudo-abnormal filtering at least includes: screening out according to the area lower limit, the boundary contrast lower limit and the circularity upper limit, and re-screening with a trajectory continuity index between adjacent frames.

[0034] As a preferred scheme of the hot imaging-based recycled asphalt mortar mixing uniformity analysis method of the present application, wherein: the mixing uniformity index is obtained by weighted combination of the abnormal area proportion, the distribution density and the abnormal intensity weighted item, and is corresponding to a process determination threshold, for generating mixing process adjustment suggestions;

[0035] The weight setting and normalization method of the weighted combination are as follows:

[0036] E) Quantize the abnormal region set obtained in step S5, first construct area proportion, distribution density and intensity three indicators, and normalize to dimensionless score with saturation mapping:

[0037] ,

[0038] ,

[0039] wherein, represents the normalized score of the index , represents the original index, represents the bending coefficient of the corresponding index, represents the total area proportion of abnormal regions, represents the abnormal region distribution density, anomaly intensity weighted term, a set of indices representing anomaly regions, a pixel area or corresponding physical area of the i-th anomaly region, a field of view area, a number of anomaly regions, an area weighted mean of pixel anomaly scores within the i-th anomaly region, a set of symbols to refer to the three specific indicators themselves;

[0040] F) Reliability is set according to the local sensitivity of each indicator to the pixel-level threshold and normalized into weights:

[0041] ,

[0042] ,

[0043] ,

[0044] wherein, represents the adaptive weight of the indicator , represents the reliability of the indicator , represents the finite difference sensitivity of the indicator at the threshold , represents a small positive number, represents the amplitude of symmetric perturbation to the threshold , represents the normalized result obtained in step E) at a given threshold,

[0045] The three scores are aggregated in additive form, and the uniformity level is represented by an inverse score:

[0046] ,

[0047] wherein, represents the blending uniformity indicator, respectively represent the adaptive weights of area proportion, distribution density and intensity score, respectively represent the corresponding normalized scores, and the constant 1 represents the upper bound, so that the indicator range converges to .

[0048] ​As a preferred scheme of the hot imaging-based analysis method for uniformity of recycled asphalt mortar mixing, wherein: the mixing process adjustment suggestion comprises adjusting at least one of mixing temperature, mixing time, stirring speed and RAP preheating strategy, and the adjustment is generated according to the deviation degree of the uniformity index relative threshold value classification.

[0049] In a second aspect, the present application provides a hot imaging-based analysis system for uniformity of recycled asphalt mortar mixing, comprising:

[0050] A thermal imaging acquisition module is configured to acquire a thermal image sequence of the mixing process and perform emissivity setting and temperature calibration.

[0051] A preprocessing and correction module is configured to perform denoising, radiation temperature correction and time / space registration on the thermal image sequence.

[0052] A feature construction module is configured to generate at least one of a multi-scale temperature gradient feature, a local temperature variation coefficient and a time-series temperature residual feature.

[0053] An analysis and identification module is configured to perform threshold segmentation, abnormal candidate extraction, unsupervised clustering and pseudo-abnormal filtering, and output an abnormal region suspected of RAP clumping.

[0054] An evaluation output module is configured to calculate a mixing uniformity index and generate a mixing process adjustment suggestion.

[0055] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement any step of the hot imaging-based analysis method for uniformity of recycled asphalt mortar mixing according to the first aspect of the present application.

[0056] The present application has the following advantages: the present application introduces multi-scale temperature feature analysis and time-series dynamic modeling, which significantly improves the detection ability of micro-uniformity phenomena. Non-contact thermal imaging acquisition is combined with image processing and machine learning algorithms to realize accurate identification and quantitative evaluation of temperature field anomalies in the mixing process. By extracting multi-scale temperature gradient features, the present application effectively enhances the response sensitivity to small RAP clumping edges, overcoming the missed detection problem caused by signal dilution in traditional methods. By introducing time-series temperature residual features, a correlation model between pixel-level temperature response and material thermal conductivity characteristics is established, which can distinguish between real clumping-induced anomalies and ordinary thermal noise, significantly improving the reliability of the detection results.

[0057] The application can adapt to changes of different mixing conditions and material proportions by using unsupervised clustering and adaptive threshold segmentation technology, and avoids the misjudgment risk caused by fixed threshold. The pseudo-anomaly filtering mechanism comprehensively uses morphological and motion trajectory characteristics, effectively eliminates the interference caused by temporary thermal disturbance, and ensures the accuracy of anomaly area identification.

[0058] The mixing uniformity index provided by the application comprehensively considers multi-dimensional information such as the number, distribution density and intensity of the abnormal area, can comprehensively reflect the mixing quality of the mortar, and provides accurate guidance for process adjustment. Full automation processing from data acquisition to result output is realized, the manual intervention intensity is significantly reduced, and reliable technical support is provided for high-quality production of recycled asphalt mortar. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope of the present application.

[0060] Figure 1 The flowchart of the recycled asphalt mortar mixing uniformity analysis method in the embodiment.

[0061] Figure 2 The framework diagram of the recycled asphalt mortar mixing uniformity analysis system in the embodiment. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0063] All terms used in the present application (including technical and scientific terms) have meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted to have meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0064] For example, the terms "first", "second", etc. used in the present application are only used to distinguish similar objects, to distinguish the first object from another object, and are not used to describe a specific order or sequence, nor can be understood as indicating or implying relative importance.

[0065] The present application proposes a recycled asphalt mortar mixing uniformity analysis method based on thermal imaging, which combines Figure 1 As shown in the figure, the method comprises:

[0066] Step S1, non-contact thermal imaging acquisition is performed on the field of view area of the material during mixing to obtain a thermal image sequence arranged in time sequence; wherein, the acquisition includes emissivity setting and temperature calibration;

[0067] In this embodiment, the field of view area is the infrared visible area of the material in the mixing container, the ROI is determined by the camera pose and locked at the beginning of the batch; the emissivity refers to the equivalent specific radiation rate of the material surface to infrared radiation, which is obtained by black body attachment or reference target comparison. The emissivity is 0.95 by default, which can be adjusted within 0.90-0.98, and is set by expert rules according to the oil and stone ratio of the mixing material and the apparent roughness; the acquisition frame rate is 30 Hz by default, which can be matched according to the stirring line speed within 10-60 Hz; the temperature calibration adopts double-point calibration, which takes one steady point in the mixing temperature interval by default, and is updated once at the beginning of each batch. Alternatively, the ROI can be drawn manually once and used in subsequent batches. If the black body reference frame is missing, the calibration parameters of the last batch are temporarily used and marked as incomplete calibration in the uniformity evaluation.

[0068] Step S2, bad pixel correction, denoising and radiation temperature correction are performed on the thermal image sequence, and parallax / occlusion correction is performed based on the set mixer geometry and camera pose to obtain a corrected thermal image sequence; specifically, the bad pixels are continuously saturated or zero response pixels, which are marked by time consistency detection; the denoising adopts space-time joint filtering to suppress isolated high-frequency points caused by splashing while maintaining the edges. The bad pixel criterion is default to be continuous ≥5 frames above or below the threshold, and the threshold is set to be 3-5 times the noise equivalent temperature difference NETD of the sensor; the time window of the space-time filtering is 5 frames by default, which can be adjusted within 3-9 frames; the parallax / occlusion correction is updated every 20 frames to match the camera micro-vibration. Optionally, when the stirring frequency is stable, it can be degenerated to only time domain filtering. If the registration residual exceeds the upper limit of 0.6px pixel level, it will be rolled back to the last effective registration matrix and the time window will be reduced to 3 frames.

[0069] Step S3, at least one of the following temperature features is extracted from the corrected thermal image sequence:

[0070] Multi-scale temperature gradient features, local temperature coefficient of variation, and time-series temperature residual features; wherein the time-series temperature residual features are constructed based on the temperature response differences of the same pixel points over time; for example, time registration employs sparse feature tracking based on brightness conservation and is optimized with global affine constraints to ensure that the same pixel trajectory corresponds to the same physical small domain. The upper limit of registration optimization iterations is 30 by default, and the convergence threshold is sub-pixel 0.2px; the residual statistical update period is consistent with the sampling period. The estimation of pixel inertia and global driving gain is obtained online by a recursive algorithm, and the initial value comes from the proportional relationship between the temperature variance of the initial mixing steady-state frame and the global median change rate, which is typically in the range of (0, 1) and a small non-negative amount. If the registration failure rate exceeds 5%, the window median is used to replace the residual construction in this section and the subsequent threshold sensitivity is lowered.

[0071] Step S4, threshold segmentation and morphological processing are performed on the feature map to obtain a set of abnormal candidate regions; further, threshold segmentation is performed on the feature map, and morphological processing is used to connect adjacent fragments and remove isolated noise points. Adaptive thresholding is used for binaryzation, with a default local block window size of 31x31 pixels and a step size of 8 pixels; open-close operation default structure element radius is 2-3 pixels, which can be scaled proportionally with the spatial resolution of the field of view. Optionally, when large-scale bias is caused by material surface reflection, global background fitting can be performed before segmentation to stabilize the candidate size. If the candidate area proportion instantaneously exceeds 20% of the field of view, it is judged as a global thermal disturbance, and the candidate output of this frame is skipped to avoid false positives.

[0072] Step S5, unsupervised clustering is performed on the abnormal candidate regions, and pseudo-anomaly filtering is performed based on the temperature features and morphological features of the regions to obtain an abnormal region set of suspected RAP clusters; similarly, unsupervised clustering is performed in the centroid coordinate and intensity feature space of the candidate regions to merge adjacent and similar candidates. The density distance threshold is set to 1.5-2.5 pixels equivalent physical size corresponding to the field of view resolution by default; the minimum sample size is 5-12 by default and is adaptive to the candidate density. Optionally, when the number of candidates is small, clustering can be degenerated into a connected domain merging rule to maintain the continuity of the output. If the clustering returns an empty set but the total area proportion of the candidates is steadily increasing, a one-time parameter relaxation (distance threshold increased by 10%, minimum sample size decreased by 2) is triggered and only takes effect for the current batch.

[0073] Step S6, count the number, area proportion, distribution density and intensity weighted index of the abnormal area, calculate the mixing uniformity index and output the evaluation result; in the embodiment, the intensity weighted index is expressed by the area weighted mean of the pixel abnormal score in the region to emphasize the defects with significant temperature difference but small scale. The uniformity index output period defaults to rolling once every 1s, and the window length is 5s; the evaluation result includes the index value and the grade label, and the label threshold is set according to the process history data quantile and automatically updated after each batch ends. If the valid frames in the output window are less than 50%, the last valid index is frozen and a data insufficient identifier is attached.

[0074] In one embodiment, the multi-scale temperature gradient feature is obtained by calculating the edge response at different window scales, and the edge operator includes at least one of Sobel, Laplacian-Gaussian or Canny; specifically, the multi-scale implementation is to calculate the edge response at 1-3 scales and fuse the maximum response to consider the edge saliency of small particles and larger agglomerates. The scale factor defaults to 1:√2:2; the high-to-low threshold ratio of Canny defaults to 2:1, and the threshold is automatically converted based on the noise level. Optionally, when noise dominates, Laplacian-Gaussian is preferred; when the edge contrast is low, Canny is used to enhance stability.

[0075] In one embodiment, the time-series temperature residual feature is obtained by time registration of the thermal map sequence, fitting the pixel temperature sequence using a background heat conduction response model, and taking the statistical quantity of the fitting residual as the abnormal score; optionally, the residual statistical quantity is robustly scaled to reduce the influence of temperature drift, and a pixel-level threshold is adaptively set within the background mask to meet the target false alarm rate. The background mask defaults to the intersection of the 10% quantile of the gradient amplitude and the 30% quantile of the time-series variance; the false alarm rate defaults to 0.01 and can be adjusted within 0.005-0.02 according to the good batch distribution. The background-driven sequence is taken from the time series of the median or low gradient mean in the field of view, and the smoothing is given by a fixed-length moving average, typically with a window length of several frames.

[0076] The fitting and residual construction steps of the background heat conduction response model include:

[0077] A) After completing the time registration, a first-order background model driven by global thermal load is established at the pixel level, and a predicted residual is formed therefrom for subsequent abnormality measurement:

[0078] ,

[0079] ,

[0080] wherein, represents the pixel at discrete time background prediction temperature, representing a pixel baseline constant term, representing a pixel inertial coefficient on the previous time step prediction temperature, representing a pixel gain coefficient on the global drive, representing the time global temperature drive sequence, composed of the field median or low gradient region mean, representing a pixel prediction residual at time , representing a pixel measured temperature at time , representing a pixel index, representing a time frame index;

[0081] B) adopting the ratio of the residual peak to the relatively robust scale as the pixel anomaly score, and estimating the scale in the form of the Gaussian equivalent of the median absolute deviation:

[0082] ,

[0083] ,

[0084] where, representing the pixel anomaly score, representing the operation of taking the maximum over the time index , representing the absolute value operation, representing the pixel robust scale estimator of the residual sequence, representing the median operator, representing the pixel residual sequence median over all time frames, constant for converting the MAD to the Gaussian equivalent standard deviation;

[0085] C) statistically aggregating the pixel anomaly scores within the low gradient background mask and forming a threshold in the form of a quantile, enabling the adaptation to batch-to-batch noise and load differences:

[0086] ,

[0087] where, representing the pixel-level anomaly score threshold, representing the quantile operator returning the quantile of the input set at the probability level , representing the target false alarm rate, represents a background pixel set consisting of low gradient and stable regions;

[0088] D) Parameter estimation of step A) can employ recursive least squares with Huber weighting to reduce outlier influence, Smoothed from time median to suppress splash disturbance, initialize Take initial steady frame, Restrict in To meet the first order stable inertia characteristics, Take non-negative to match the heating-cooling physical response, get The pixels are connected domain aggregation and subsequent link;

[0089] Further, the parameter estimation employs a recursive method with robust weighting to achieve online updating, ensuring that the sensitivity to anomalies is maintained under slow drift of mixing thermal load. The recursive update step is 0.05 by default, which can be adjusted within 0.01-0.1 according to noise intensity; the robust weighting threshold interval is automatically determined according to residual scale. Optionally, when the computing resources are limited, the recursive least squares can be degraded to single-step update with exponential weighting, but the same input and output are maintained.

[0090] Specifically, this step link takes the pixel layer time series as the object, first establishes a first-order background model with thermal inertia and global load driving, describes the ubiquitous heating-cooling dynamic with a small number of parameters, while keeping the key physical properties under controlled computational load; On this basis, the ratio of residual peak to robust scale is used as the anomaly score, so that sudden warming or delayed cooling forms a significant contrast in the score, and is insensitive to occasional noise and slow drift; Subsequently, the quantile threshold based on the background set is introduced, and the explicit false alarm rate parameter is used to adapt to the statistical distribution difference of different batches and different conditions, avoiding the cross-scene deviation caused by fixed threshold; Further, the probability level of the self-adaptive threshold is set according to the process target false alarm rate, and the robust statistics of the background set keeps the consistent false alarm control under different batches and different load conditions. The target false alarm rate is 0.01 by default, which can be fine-tuned within 0.005-0.02 according to the balance of qualified batch false alarm cost and false alarm cost; the background set is re-estimated every 100 frames to track statistical drift.

[0091] In one embodiment, the threshold segmentation employs an adaptive threshold based on noise level estimation, which is estimated from a uniform blackbody calibration frame or a mixing initial steady frame; in this embodiment, the noise level is characterized by the short window variance of temperature readings, and is modified by the NETD provided by the device. The short window length is set to 21 frames by default, and can be set in the range of 11-31 frames; the amplification coefficient of the adaptive threshold is set to 3.5 by default, and can be selected in the range of 3-5 according to the on-site reflection environment. If the steady frame judgment fails, the median of the variance in the adjacent time period is used to replace the estimation to maintain the stability of the threshold.

[0092] In one embodiment, the unsupervised clustering employs density clustering, and the density distance threshold and the minimum sample number are adaptively set according to the scale distribution of the abnormal candidate and the spatial resolution of the field of view; specifically, the adaptivity of the density distance threshold and the minimum sample number is based on the nearest neighbor distance statistics and the sample density distribution of the candidate, to ensure that the candidate is not excessively split when it is sparse, and not excessively merged when it is dense. The nearest neighbor distance quantile is set to 25% quantile as the reference amount of the distance threshold by default, and fine tuning is allowed within 20%-35% quantile; the minimum sample number is rounded down according to the logarithmic function of the total number of candidates, and is not less than 5. If the adaptive calculation does not converge, the fixed parameters of the previous batch are used as a fallback, and a stabilization prompt is given.

[0093] In one embodiment, the pseudo abnormality filtering at least includes: screening out according to the area lower limit of the region, the boundary contrast lower limit and the circularity upper limit, and reselecting according to the trajectory continuity index between adjacent frames;

[0094] For example, the trajectory continuity is jointly measured by the displacement of the region centroid in adjacent frames and the area change rate, which is used to eliminate transient false points caused by splashes or thermal reflections. The continuity criterion requires that the candidate exists at least in 3 frames, the adjacent displacement is not more than 5mm equivalent physical, and the area change rate is not more than 50% / frame, and the three conditions must be met at the same time. Optionally, when the stirring intensity is high, the displacement upper limit can be proportionally adjusted according to the stirrer linear speed.

[0095] In one embodiment, the mixing uniformity index is obtained by weighted combination of the abnormal region area proportion, the distribution density and the abnormal intensity weighted item, and is corresponding to the process determination threshold, which is used to generate mixing process adjustment suggestions;

[0096] The weight setting and normalization method of the weighted combination are as follows:

[0097] E) Quantifying the abnormal region set obtained in step S5, first constructing area proportion, distribution density and intensity three indicators, and normalizing to dimensionless score by saturation mapping:

[0098] ,

[0099] ,

[0100] wherein, denotes the normalized score of the index , denotes the original index, denotes the bending coefficient of the corresponding index, denotes the total area proportion of abnormal regions, denotes the distribution density of abnormal regions, denotes the abnormal intensity weighted term, denotes the index set of abnormal regions, denotes the pixel area or the corresponding physical area of the th abnormal region, denotes the area of the field of view region, denotes the number of abnormal regions, denotes the area weighted mean of pixel abnormality scores in the th abnormal region, the symbol set is used to refer to the three specific indexes themselves;

[0101] F) To reduce the instability caused by threshold fluctuation, the reliability is set according to the local sensitivity of each index to the pixel-level threshold and is normalized to a weight:

[0102] ,

[0103] ,

[0104] ,

[0105] wherein, denotes the adaptive weight of the index , denotes the reliability of the index , denotes the finite difference sensitivity of the index at the threshold , denotes a small positive number to avoid zero denominator, denotes the amplitude of symmetric perturbation to the threshold , denotes the normalized result obtained in step E) under the given threshold,

[0106] The three scores are aggregated in additive form, and the uniformity level is represented by the inverse score:

[0107] ,

[0108] wherein, denotes the blending uniformity index, respectively represent the adaptive weights of area ratio, distribution density and intensity fraction, respectively represent the corresponding normalized fractions, and the constant 1 represents the upper bound, so that the index range converges to ; Similarly, the index output is accompanied by a quality identifier in addition to the divisor value, which is used to prompt the uncertainty brought by threshold sensitivity and guide process decision-making. The adaptive disturbance amplitude of the weight is defaulted to be 1% of the threshold, and the two-way disturbance step length of the sensitivity calculation is the same. When the variance of any one of the three fractions is significantly higher than its historical P75 within the window, the upper limit of the weight of this fraction is limited to 0.4. If the three fractions are missing at the same time within the window, the index inherits the value of the last window and is marked as missing.

[0109] When there is no need for adaptation, the can be empirically set to highlight the impact of area ratio, The fraction on the process qualified batch, is of the order of , The is taken to stabilize the value;

[0110] Optionally, when sufficient historical data cannot be provided on site, the weight can be temporarily fixed according to the empirical proportion and updated to a new empirical value after each batch ends in a leave-one-out cross-validation manner; when the threshold disturbance step length setting is insufficient to cause weight jitter, a first-order smoothing with an exponential smoothing coefficient of 0.2-0.4 is performed on the weight time series to improve stability.

[0111] Specifically, the three types of regional level observables are used as the basis signal, and a saturation type normalization mapping is used to construct dimensionless fractions, avoiding the domination effect of extreme large values on the weighted sum; Then a weight adaptive strategy based on threshold local sensitivity is introduced, by slightly disturbing the threshold in both directions, the change rate of each fraction is calculated and the weight is inversely proportional to it, the item with high sensitivity gets smaller weight, the item with low and stable sensitivity gets larger weight, and the robustness of the overall weighted threshold setting is improved; Finally, the uniformity index is given in the form of reverse score, so that the three fractions form a stronger punishment on the final index when they are near the upper bound, thereby widening the distinction between different mixing states;

[0112] ​In one embodiment, the mixing process adjustment suggestion includes adjusting at least one of the mixing temperature, the mixing time, the stirring speed, and the RAP preheating strategy, and the adjustment is generated according to the degree of deviation of the uniformity index from the relative threshold value; in this embodiment, the classification of the process adjustment is determined by the deviation of the uniformity index from the threshold value and the duration thereof, so as to avoid overreaction to short-term fluctuations. The minimum duration is 10 s to trigger a first-level adjustment, and the deviation exceeds the second threshold value and lasts for ≥30 s to trigger a second-level adjustment; when the index returns to within the threshold value and lasts for ≥20 s, the suggestion is automatically removed. If the evaluation output is determined to be insufficient data, only an observation-level prompt is generated without triggering a substantive adjustment suggestion.

[0113] The application also proposes a recycled asphalt mortar mixing uniformity analysis system, which combines Figure 2 as shown, comprising:

[0114] a thermal imaging acquisition module for acquiring a thermal image sequence of the mixing process and performing emissivity setting and temperature calibration;

[0115] a preprocessing and correction module for denoising, radiation temperature correction, and time / space registration of the thermal image sequence;

[0116] a feature construction module for generating at least one of a multi-scale temperature gradient feature, a local temperature variation coefficient, and a time-series temperature residual feature;

[0117] an analysis and recognition module for threshold segmentation, abnormal candidate extraction, unsupervised clustering, and pseudo-abnormal filtering, and outputting an abnormal region suspected of RAP clumping;

[0118] an evaluation output module for calculating a mixing uniformity index and generating a mixing process adjustment suggestion.

[0119] The embodiment also provides a storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method for analyzing uniformity of recycled asphalt mortar mixing based on thermal imaging as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage devices or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0120] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the above embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

[0121] In addition, those skilled in the art can understand that although some embodiments herein include certain features included in other embodiments rather than other features, the combination of features of different embodiments means to be within the scope of the present application and form different embodiments. For example, all the above embodiments can be used in any combination. The information disclosed in the background section is only intended to deepen the understanding of the overall background of the present application, and should not be regarded as acknowledging or implying in any form that the information constitutes prior art known to those skilled in the art.

Claims

1. A method for analyzing the uniformity of a hot-mix asphalt mixture based on thermography, characterized in that, Comprising: Step S1, non-contact thermal imaging acquisition of the material field of view area during mixing, obtaining a thermal image sequence arranged in time sequence; wherein, the acquisition includes emissivity setting and temperature calibration; Step S2, bad point correction, denoising and radiation temperature correction are performed on the thermal image sequence, and disparity / occlusion correction is performed based on the set mixer geometry and camera pose, to obtain a corrected thermal image sequence; Step S3, at least one of the following temperature features is extracted from the corrected thermal image sequence: Multi-scale temperature gradient feature, local temperature variation coefficient and time series temperature residual feature; wherein, the time series temperature residual feature is constructed based on the temperature response difference of the same pixel point over time; Step S4, threshold segmentation and morphological processing are performed on the feature map to obtain a set of abnormal candidate regions; Step S5, unsupervised clustering is performed on the abnormal candidate regions, and pseudo-anomaly filtering is performed based on the temperature features and morphological features of the regions, to obtain a set of abnormal regions suspected of RAP clusters; Step S6, the number, area ratio, distribution density and intensity weighted index of the abnormal regions are counted, the mixing uniformity index is calculated, and the evaluation result is output; The time series temperature residual feature is fitted by using a background heat conduction response model after time registration of the thermal image sequence, and the statistical quantity of the fitting residual is taken as the abnormal score; The fitting and residual construction steps of the background heat conduction response model include: A) After time registration is completed, a first-order background model driven by global heat load is established at the pixel layer, and a predicted residual is formed based on the model, which is used as an abnormality measure: ， ， wherein, represents a pixel at a discrete time instant a background prediction temperature, represents a pixel a baseline constant term, represents a pixel an inertia coefficient for a prediction temperature of a previous time instant, represents a pixel a gain coefficient for a global drive, represents a global temperature drive sequence at a time instant represents a prediction residual of a pixel at a time instant represents a measured temperature of a pixel at a time instant represents a pixel index, represents a time frame index; B) The ratio of the relative robust scale of the residual peak value is taken as the pixel abnormal score, and the median absolute deviation is estimated in the form of Gaussian equivalent to estimate the scale: ， ， wherein, represents the pixel anomaly score, represents the operation taking the maximum over the time index represents the absolute value operation, represents the pixel robust scale estimator of the residual sequence, represents the median operator, represents the pixel median of the residual sequence over all time frames, constant used to convert the MAD to a Gaussian equivalent standard deviation;​ C) The pixel abnormal score set is counted within the low gradient background mask, and a threshold is formed in the form of quantile: ， wherein, represents a pixel-level anomaly score threshold, represents a quantile operator that returns the quantile of the input set at a probability level , represents a target false alarm rate, represents a set of background pixels consisting of low gradients and stable regions; D) Huber-weighted recursive least squares is applied to the parameter estimates of step A), initialized by smoothing the median of the time instants Take the initial mixing steady-state frame, limited to , Take the non-negative, get pixels after the connected domain aggregation and.

2. The thermography-based analysis method of the uniformity of the regenerative asphalt mortar mixing according to claim 1, characterized in that, The multi-scale temperature gradient feature is obtained by calculating the edge response at different window scales, and the edge operator includes at least one of Sobel, Laplace-Gaussian or Canny.

3. The thermography-based analysis method of the uniformity of the regenerative asphalt mortar mixing according to claim 1, characterized in that, Threshold segmentation uses an adaptive threshold based on noise level estimation, which is estimated from a uniform blackbody calibration frame or a stable initial mixing frame.

4. The thermography-based analysis method of the uniformity of the regenerative asphalt mortar mixing according to claim 1, characterized in that, Unsupervised clustering uses density clustering, and the density distance threshold and the minimum sample number are adaptively set according to the scale distribution of the abnormal candidate and the spatial resolution of the field of view.

5. The thermography-based analysis method of the uniformity of the regenerative asphalt mortar mixing as claimed in claim 1, characterized in that, Pseudo-anomaly filtering at least includes: screening out according to the area lower limit, boundary contrast lower limit and circularity upper limit, and reselecting according to the trajectory continuity index between adjacent frames.

6. The thermography-based analysis method of the uniformity of the regenerative asphalt mortar mixing as claimed in claim 1, characterized in that, The mixing uniformity index is obtained by weighted combination of the area ratio, distribution density and abnormal intensity weighted items of the abnormal regions, and is corresponding to the process determination threshold, which is used to generate mixing process adjustment suggestions; The weight setting and normalization method of the weighted combination are as follows: E) Quantify the abnormal region set obtained in step S5, first construct the area ratio, distribution density and intensity, and normalize the three scores to dimensionless scores in the form of saturation mapping: ， ， wherein, denotes the normalized score of the index , denotes the original index, denotes the bending coefficient of the corresponding index, denotes the total area proportion of the abnormal area, denotes the abnormal area distribution density, denotes the abnormal intensity weighted term, denotes the index set of the abnormal area, denotes the pixel area or the corresponding physical area of the th abnormal area, denotes the field of view area, denotes the number of abnormal areas, denotes the area weighted mean of the pixel abnormal score in the th abnormal area, and the symbol set is used to refer to the three specific indexes themselves; F) Setting the reliability of the local sensitivity of each index to the pixel-level threshold and normalizing to a weight: ， ， ， wherein represents an index of the adaptive weight, represents an index of the reliability, represents an index of the finite difference sensitivity at a threshold value, represents a small positive number, represents the amplitude of the symmetric perturbation of the threshold value, represents the normalized result obtained at step E) at a given threshold value, Aggregate the three scores in additive form, and use the reverse score to represent the uniformity level: ， wherein, denotes the mixing uniformity index, denotes the adaptive weights of area ratio, distribution density and intensity fraction, respectively, denotes the corresponding normalized fraction, and the constant 1 denotes the upper bound, making the index range converge to .

7. The thermography-based analysis method of the uniformity of the regenerative asphalt mortar mixing as claimed in claim 1, characterized in that, The blending process adjustment suggestion includes adjusting at least one of the blending temperature, the blending time, the stirring speed, and a RAP preheating strategy, the adjustment being generated according to a grading of a deviation degree of the uniformity index relative threshold value.

8. A thermal imaging based analysis system for the homogeneity of the mixture of recycled asphalt mortar, based on the thermal imaging based analysis method for the homogeneity of the mixture of recycled asphalt mortar according to any one of claims 1 to 7, characterized in that, The method comprises the following steps: a thermal imaging acquisition module for acquiring a thermal image sequence of the blending process and performing emissivity setting and temperature calibration; a preprocessing and correction module for denoising, radiometric temperature correction, and time / space registration of the thermal image sequence; a feature construction module for generating at least one of a multi-scale temperature gradient feature, a local temperature variation coefficient, and a time-series temperature residual feature; an analysis and identification module for threshold segmentation, abnormal candidate extraction, unsupervised clustering, and pseudo-abnormal filtering, and outputting an abnormal region of suspected RAP clumps; an evaluation output module for calculating a blending uniformity index and generating a blending process adjustment suggestion.

9. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program, when executed by a processor, implements the steps of the thermal imaging-based reclaimed asphalt mortar blending uniformity analysis method of any one of claims 1-7.

Images