A Deep Learning-Based Intelligent Concrete Slump Detection Method

Abstract

This invention belongs to the interdisciplinary field of artificial intelligence and visual inspection, specifically disclosing a deep learning-based intelligent method for detecting concrete slump. The method includes: acquiring a thermal infrared video sequence of the concrete slump process; extracting three-dimensional spatiotemporal temperature field data; generating a multi-channel dynamic feature map containing temperature change rate, gradient magnitude, and entropy value; inputting the map into a convolutional-temporal hybrid neural network model for slump prediction; outputting the results and evaluating the confidence level, triggering retesting if necessary. By integrating thermodynamic evolution features with deep learning modeling, this invention can predict slump with high accuracy and automation, improving detection robustness and engineering applicability, and meeting the real-time quality control needs of construction sites.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and visual inspection technology, specifically relating to a deep learning-based intelligent detection method for concrete slump. Background Technology

[0002] With the continuous advancement of intelligent construction and infrastructure monitoring technologies, real-time assessment of concrete workability has become a crucial link in ensuring project quality and construction efficiency. Slump, as a core indicator for measuring the fluidity of fresh concrete, is traditionally tested primarily through manual visual inspection or mechanical slump cone tests. These methods are highly subjective, inefficient, and only provide static, discrete measurement results, failing to reflect the dynamic rheological characteristics of concrete in real-world environments. Especially under complex working conditions, the flow behavior of concrete is influenced by multiple factors, including the release of heat of hydration, fluctuations in ambient temperature, and interactions among material components, placing higher demands on the timeliness, continuity, and physical perception dimensions of testing methods.

[0003] Current slump testing technologies generally neglect the dynamic regulation of concrete flowability by the temperature field, failing to capture the correlation mechanism between heat transfer and paste structure evolution during hydration. Furthermore, conventional visual or mechanical sensing methods focus only on geometric deformation or surface texture, lacking non-contact perception of the internal thermodynamic state of concrete, making it difficult to establish a mapping relationship between flowability and the spatiotemporal distribution of the temperature field. In addition, existing analytical models are mostly based on steady-state assumptions, unable to handle the nonlinear and non-uniform temperature-rheological coupling evolution characteristics during slump, leading to significant deviations between evaluation results and actual performance.

[0004] Therefore, there is an urgent need for an intelligent method for detecting concrete slump that integrates thermal infrared sensing and deep learning, so as to achieve high-precision and dynamic joint analysis of the temperature field evolution and flow characteristics throughout the slump process. Summary of the Invention

[0005] To achieve the above objectives, this invention provides a deep learning-based intelligent method for detecting concrete slump, comprising the following steps:

[0006] To obtain thermal infrared video sequences of the concrete slump process, an infrared thermal imager was used to continuously capture the entire process of fresh concrete from being lifted from the slump cone to the point of flow stabilization during the standard slump test, thus obtaining thermal infrared video sequences containing information on time dimension and spatial temperature distribution.

[0007] Extract time-series temperature field data, perform inter-frame alignment and radiometric correction on the thermal infrared video sequence, extract the temperature value of each pixel on the concrete surface frame by frame, and construct a three-dimensional spatiotemporal temperature field dataset, with two spatial dimensions corresponding to image coordinates and the third dimension being time.

[0008] A dynamic feature map of the temperature field is generated. Based on the three-dimensional spatiotemporal temperature field dataset, the temperature change rate, temperature gradient magnitude, and temperature field entropy value at each spatial location are calculated and fused to form a multi-channel dynamic feature map.

[0009] The slump prediction is performed by inputting the deep neural network model. The multi-channel dynamic feature map is used as input and fed into a pre-trained convolutional-temporal hybrid neural network model. The convolutional-temporal hybrid neural network model extracts local thermodynamic features through spatial convolutional layers and models the time evolution of the temperature field through gated cyclic unit layers, and finally outputs the slump value of concrete.

[0010] Output the detection results and perform confidence assessment. Generate the collapse prediction value based on the output of the convolutional-temporal hybrid neural network model, and simultaneously calculate the confidence interval and uncertainty index of the prediction results. If the uncertainty is greater than the preset threshold, trigger the retest mechanism or manual review process.

[0011] Preferably, the inter-frame alignment adopts a phase-correlation-based sub-pixel level image registration algorithm, and the radiation correction is based on Planck's blackbody radiation law combined with on-site measured environmental parameters to compensate for non-uniformity, so as to control temperature measurement error.

[0012] Preferably, the temperature change rate is obtained by smoothing the time series using a high-order Savitzky-Golay filter and then taking its derivative; the temperature gradient magnitude is calculated in the spatial domain using the Sobel operator; and the temperature field entropy is defined as the Shannon entropy of the probability density function of the temperature distribution in the current frame.

[0013] Preferably, the convolutional-temporal hybrid neural network model includes multiple spatial convolutional blocks, each convolutional block consisting of a convolutional layer, a batch normalization layer, and a ReLU activation function, followed by multiple layers of bidirectional gated recurrent units. The output layer uses a fully connected layer with a linear activation function, and the model training uses a weighted combination of mean squared error loss function and Huber loss.

[0014] Preferably, the confidence assessment is based on the Monte Carlo Dropout method, which performs multiple random sampling predictions on the same input during the inference phase, calculates the standard deviation of the predicted values ​​as an uncertainty index, and determines the result as low confidence when the standard deviation is greater than a preset distance threshold.

[0015] Preferably, the method further includes an environmental interference suppression step: before acquiring the thermal infrared video sequence, ambient light and background scene images are simultaneously acquired using a visible light camera, and a semantic segmentation network is used to identify and eliminate thermal signals from non-concrete areas to prevent interference from false heat sources caused by solar radiation, equipment heating, or personnel movement.

[0016] Preferably, the semantic segmentation network is a DeepLabv3+ architecture, the backbone network uses ResNet-50, and the output categories include concrete, ground, equipment, people and sky. Pixels that are not in the concrete category are marked as invalid in the infrared image.

[0017] Preferably, the method further includes a multimodal data fusion step: pixel-level alignment of the thermal infrared video sequence with the synchronously acquired visible light video, extraction of geometric deformation features in the visible light image, channel splicing with the temperature field dynamic feature map, and inputting them together into a two-stream neural network architecture.

[0018] Preferably, the geometric deformation features are calculated using the Lucas-Kanade optical flow method to calculate the displacement field between adjacent frames, and the maximum flow radius and collapse height attenuation rate are extracted and encoded into a single-channel spectrum to participate in feature fusion.

[0019] Preferably, the method is deployed on an edge computing device, with an infrared thermal imager and a visible light camera integrated on the slump test support. The entire process of data acquisition, processing, and prediction is completed within a predetermined time period after a single test, meeting the real-time detection needs of the construction site.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. Breaking through the limitations of traditional detection methods that rely solely on geometric deformation, this method is the first to introduce thermal infrared imaging technology into the field of slump testing. By capturing the dynamic evolution of the surface temperature field caused by the release of hydration heat and environmental heat exchange, a direct mapping relationship between flow properties and thermodynamic state is established. This method can reflect implicit information such as the internal hydration process of concrete, the uniformity of aggregate distribution, and the effect of admixtures, significantly improving the consistency between test results and actual working performance.

[0022] 2. The constructed convolutional-temporal hybrid neural network model deeply integrates spatial thermal characteristics and temporal evolution patterns, demonstrating excellent generalization ability for concrete under different environmental temperatures, mix proportions, and admixture systems. On the measured dataset, the mean absolute error of slump prediction is controlled within the preset error range, significantly lower than that of traditional visual methods, and maintains stable output even under complex conditions such as strong sunlight, rain, and fog.

[0023] 3. By integrating infrared and visible light dual-modal sensing, edge computing, and automatic confidence assessment mechanisms, end-to-end automation from data acquisition to result output is achieved, eliminating the need for manual intervention or subjective judgment. The time required for a single test does not exceed a predetermined time threshold, allowing for seamless integration into ready-mixed concrete factory inspection or on-site quality control processes. This significantly improves testing efficiency and standardization, providing reliable data support for intelligent construction. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0025] Figure 2 This is a schematic diagram of the core principle framework of the convolutional-temporal hybrid neural network model in this invention;

[0026] Figure 3 This is a flowchart illustrating the logical flow of thermal infrared video sequence processing and three-dimensional spatiotemporal temperature field construction in this invention.

[0027] Figure 4 This is a logical flowchart of the temperature field dynamic feature map generation and multimodal data fusion mechanism in this invention;

[0028] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the infrared and visible light dual-mode sensing system in this invention. Detailed Implementation

[0029] Example 1: Reference Figures 1 to 5 In the intelligent detection method for concrete slump based on deep learning, step 1, which acquires the thermal infrared video sequence of the concrete slump process, specifically includes the following operation process: a dedicated support system is deployed at the standard slump test site. This dedicated support system integrates an infrared thermal imager and a visible light camera. The two share the same optical axis and are fixed by rigid connectors to ensure that the field of view overlap error is less than 0.5 pixels.

[0030] The infrared thermal imager uses an uncooled microbolometer (VOx) focal plane array sensor with a thermal sensitivity (NETD) of no more than 0.03℃, a spatial resolution of 640×512 pixels, a frame rate of 30Hz, and a temperature measurement range covering -20℃ to +150℃, in order to fully capture the dynamic changes in surface temperature of fresh concrete from initial filling to flow stabilization.

[0031] Before the formal data collection, an on-site environmental parameter calibration procedure was performed: First, a high-precision blackbody radiation source was used to perform two-point temperature calibration on the infrared thermal imager; then, the ambient temperature, relative humidity, atmospheric pressure, and wind speed were simultaneously measured using a handheld weather station and input into the radiation transfer model built into the infrared thermal imager; at the same time, the ambient reflection temperature was measured using a diffuse reflection aluminum foil plate, and the emissivity of the concrete surface was determined according to the ASTM E1933 standard, with a typical value of 0.92±0.03.

[0032] After calibration, the slump test was initiated: freshly mixed concrete was poured into the slump cone in three layers according to GB / T50080 standard, with each layer tamped 25 times. After leveling the top surface, the slump cone was lifted vertically at a uniform speed. Simultaneously, an infrared thermal imager continuously acquired a sequence of no less than 120 frames of thermal infrared images at a frame rate of 30Hz, forming a sequence that includes the time dimension. with spatial coordinates Complete thermal infrared video sequence , , Simultaneously, the visible light camera captures visible light video sequences at the same frame rate. It is used for subsequent environmental interference suppression and multimodal fusion. It represents the set of real numbers.

[0033] In the above method, step 2, extracting time-series temperature field data, is specifically implemented as follows: First, for... Perform inter-frame alignment processing. Because concrete undergoes overall displacement and local deformation during collapse, directly extracting temperature values ​​frame by frame would lead to spatiotemporal misalignment.

[0034] Therefore, a sub-pixel level image registration algorithm based on phase correlation is adopted: for each frame of image With reference frame (Usually, the 10th frame is selected, at which point the concrete has completely detached from the slump cone but has not yet spread significantly.) Perform a two-dimensional Fourier transform, calculate the normalized cross-power spectrum, obtain the peak position through the inverse Fourier transform, and then solve for the translation vector. The accuracy can reach 0.1 pixels. Subsequently, bicubic interpolation is used... Perform subpixel translation correction to generate an aligned sequence. Next, radiation correction is performed: according to Planck's blackbody radiation law, the infrared detector output signal... With target radiance The relationship between them is , and These are the gain and bias parameters. Combined with the ambient reflection temperature measured on-site. Atmospheric transmittance (Calculated from humidity, distance, and CO2 concentration) and concrete emissivity The non-uniformity compensation for each pixel is performed using the radiative transfer equation, and its core formula is:

[0035] ;

[0036] This is the corrected actual surface temperature of the concrete. This is the equivalent atmospheric temperature. After this correction, the temperature value error of each pixel is controlled within ±0.2℃. Finally, the corrected value... Construct a three-dimensional spatiotemporal temperature field dataset by stacking data in chronological order. , , , Each element Indicates at time Spatial location The concrete surface temperature value at that location.

[0037] In the above method, step 3, generating a dynamic feature map of the temperature field, specifically includes the following sub-operations:

[0038] Calculation of temperature change rate map: for three-dimensional temperature field along For each spatial location Its temperature time series is processed independently. To suppress high-frequency noise and preserve the true thermal evolution trend, a 5th-order Savitzky-Golay filter was used for smoothing, with a window length of 15 frames (corresponding to 0.5 seconds). The smoothed sequence was then numerically differentiated to obtain the temperature change rate.

[0039] Calculate the temperature gradient magnitude map: For each time frame, plot the two-dimensional temperature field Treating it as a scalar field, the partial derivatives in the x and y directions are calculated using the Sobel operator. and The convolution kernels are [-1, 0, 1] and their transposes. After normalization, the gradient magnitude is... Defined as:

[0040] ;

[0041] Calculate the temperature field entropy map: For each time frame, statistically analyze the temperature distribution histogram of all valid pixels in the current frame (i.e., pixels identified as concrete areas after semantic segmentation), divide the frame into 64 temperature intervals, and calculate the probability of each interval. Shannon entropy Defined as:

[0042] ;

[0043] Entropy reflects the spatial uniformity of the temperature field: low entropy indicates that the temperature distribution is concentrated and highly uniform, while high entropy indicates the presence of significant hot spots or cold zones, which may originate from aggregate segregation or uneven hydration.

[0044] Feature fusion: The three feature maps mentioned above are concatenated along the channel dimension to form a multi-channel dynamic feature map. Channel 1 is Channel 2 is The third channel is a frame-by-frame copy. (That is, the entire frame shares the same entropy value).

[0045] Furthermore, if a multimodal data fusion mechanism is enabled, geometric deformation feature extraction is performed simultaneously on the visible light video sequence: first, the concrete region is identified using the U-Net semantic segmentation network, and the background is removed; then, the Lucas-Kanade optical flow method is used to calculate the displacement field between adjacent frames, extracting geometric features such as the maximum flow radius and slump height attenuation rate, and encoding them into a single-channel map, which is then compared with... The features are stitched together along the channel dimension to form an enhanced four-channel feature map. .

[0046] In the above method, step 4, inputting a deep neural network model for collapse prediction, is implemented as follows: The convolutional-temporal hybrid neural network model used consists of a front-end spatial encoder and a back-end temporal modeler cascaded together. The spatial encoder contains four spatial convolutional blocks, each consisting of a 3×3 convolutional layer (output channels of 32, 64, 128, and 256 respectively), a batch normalization layer (BatchNorm2d), and a ReLU activation function, with a stride of 2, achieving spatial downsampling to 1 / 16 of the original size (i.e., 32×40).

[0047] for First, it unfolds along the time dimension as Each of the 10 independent samples is processed by a spatial encoder to obtain a feature vector. Then, global average pooling is used to compress the vectors into spatially independent feature vectors. .

[0048] Then, the sequence The input is fed into a back-end time series modeler, which consists of two layers of bidirectional gated recurrent units (Bi-GRU), with 128 hidden units in each layer, capable of simultaneously capturing the forward and backward dependencies of the temperature field evolution. The final hidden state output by the Bi-GRU is shown below. The predicted collapse value is obtained by mapping through a fully connected layer (output dimension 1) and a linear activation function. .

[0049] The model training phase uses a weighted combination loss function. , , Mean square error, The Huber loss function is used to balance large error penalty with small error sensitivity. The training dataset contains 12,000 labeled samples. Each sample covers thermal infrared video sequences and their corresponding measured slump values ​​(measured with a standard ruler, accuracy ±2mm) under different cement types (P·O42.5, P·II52.5), water-cement ratios (0.35–0.55), mineral admixtures (fly ash, mineral powder, dosage 0–30%), ambient temperatures (5–40℃), and hydration ages (0–120 minutes). Training uses 5-fold cross-validation, and the prediction determination coefficient on the validation set is... The average absolute error is ≤8mm.

[0050] In the above method, step 5 outputs the detection results and performs confidence evaluation, which is specifically implemented as follows: During the inference phase, Monte Carlo Dropout sampling is performed on the same input feature map: the forward propagation is repeated while all Dropout layers in the network (with a retention rate of 0.5) remain active. Each time, a slump prediction value is obtained. .

[0051] Calculate the mean of the prediction set As the final slump output value, the standard deviation As an indicator of uncertainty. Preset uncertainty threshold. ,like If the result is too high, it is considered a low-confidence result, and the system automatically triggers a retest mechanism: a complete slump test is performed again and new data is collected; if the threshold is exceeded after two consecutive retests, a manual review task is pushed to the mobile quality inspection terminal, along with the original thermal infrared video, feature spectrum, and uncertainty analysis report. In addition, the system simultaneously outputs a 95% confidence interval. This is provided for reference in engineering decision-making.

[0052] To further enhance robustness, this embodiment integrates an environmental interference suppression module: Before acquiring thermal infrared video in step 1, a visible light camera is simultaneously activated to obtain ambient lighting and background scene images. A pre-trained DeepLabv3+ semantic segmentation network (with a ResNet-50 backbone, outputting 19 Cityscapes labels) is used to perform pixel-level classification of the visible light images, identifying categories such as "concrete," "ground," "equipment," "personnel," and "sky." Masks for non-"concrete" categories are mapped onto the infrared image, forcibly setting the temperature values ​​of the corresponding pixels to invalid (NaN) labels, thereby eliminating false high temperatures caused by direct sunlight, heat dissipation from mechanical equipment, or instantaneous thermal interference caused by personnel movement. The masks are effective throughout the inter-frame alignment in step 2 and the feature calculation in step 3, ensuring that only valid regions participate in modeling.

[0053] This embodiment is also deployed on an edge computing device: employing the NVIDIA Jetson AGX Orin embedded platform, integrating an 8-core ARM CPU and a 2048-core GPU, with 32GB of memory. An infrared thermal imager and a visible light camera are connected via a GigEVision interface, and data acquisition is scheduled in real-time by a customized Linux driver. The entire processing pipeline—from video acquisition, frame alignment, radiometric correction, feature generation to neural network inference—is completed within 8 seconds after a single test, meeting the real-time requirements of "instant results" at construction sites. System power consumption is controlled below 45W, supports IP65 protection, and can operate stably in environments ranging from -10℃ to +50℃.

[0054] To verify the actual effectiveness of this embodiment, the following specific application example was constructed: In the factory inspection area of ​​a ready-mixed concrete batching plant, C30 pumped concrete (mixture ratio: cement 320 kg / m³, fly ash 80 kg / m³, sand ratio 42%, water-reducing agent dosage 1.2%) was tested. The ambient temperature was 28℃ and the relative humidity was 65%. A standard slump test was performed, and an infrared thermal imager recorded 120 frames (4 seconds) of thermal infrared video.

[0055] After inter-frame alignment, it was found that the temperature in the central region of the concrete rose from 26.5℃ to 28.1℃, while the edge region remained at 25.8℃ due to evaporative cooling, forming a significant radial temperature gradient. The temperature change rate map showed that the heating rate reached 0.8℃ / s in the first 2 seconds, then slowed down. The temperature field entropy value decreased from an initial 1.8 bits to a stable 1.2 bits, indicating that the temperature tended to be uniform during the flow process. The multi-channel feature map was input into the trained model, and after 50 Monte Carlo samplings, the predicted slump was 182mm with a standard deviation of 6.3mm (<10mm), indicating a high confidence result. The measured slump was 185mm, with an error of only 3mm. The system completed the entire process within 7.2 seconds, and the results were automatically uploaded to the quality management system.

[0056] Another example involves high-flowability self-compacting concrete (C50, slump spread 650mm): ambient temperature 12℃, humidity 80%. Due to the low temperature delaying hydration, the temperature field changes gradually, with a maximum temperature rise of only 0.9℃. However, the geometric deformation is drastic, and visible light images show rapid paving. The multimodal fusion mechanism effectively combines weak thermal signals with significant geometric features, and the model outputs a slump of 228mm (measured 230mm), with a standard deviation of 7.8mm, still within the high confidence interval. If only the thermal infrared single-mode is used, the predicted value deviation reaches 15mm, demonstrating the necessity of multimodal fusion for extreme working conditions.

[0057] Example 2: In another embodiment, the temporal modeler of the convolutional-temporal hybrid neural network model uses a Transformer encoder instead of a Bi-GRU. Specifically, the feature sequence output by the spatial encoder is... After linear projection, given query Q, key K, and value V, calculate the self-attention weights: A stacked 6-layer encoder, with each layer containing multi-head attention (8 heads) and a feedforward network, is better at capturing long-range temporal dependencies and is suitable for ultra-long-term observations (such as...). (Frame) Scene. A learning rate warmup strategy is used during training, with an initial learning rate of 10. -4 The batch size was 16. On concrete samples containing retarders, the MAE of the convolutional-temporal hybrid neural network model was reduced to 6.5 mm, which was better than the 8.2 mm of the Bi-GRU.

[0058] Example 3: In another embodiment, the environmental interference suppression module employs a joint strategy of physical modeling and deep learning: in addition to semantic segmentation, a heat conduction inversion model is introduced. Based on the temperature distribution of non-concrete areas in the infrared image, the steady-state heat conduction equation is solved. The method estimates background heat flux and subtracts interference terms from the boundary conditions of the concrete region. Under strong sunlight (solar radiation intensity >800W / m²), the temperature measurement error is reduced from ±1.5℃ to ±0.4℃, improving the quality of the characteristic spectrum.

[0059] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent method for detecting concrete slump, characterized in that, Includes the following steps: To obtain thermal infrared video sequences of the concrete slump process, an infrared thermal imager was used to continuously capture the entire process of fresh concrete from being lifted from the slump cone to the point of flow stabilization during the standard slump test, thus obtaining thermal infrared video sequences containing information on time dimension and spatial temperature distribution. Extract time-series temperature field data, perform inter-frame alignment and radiometric correction on the thermal infrared video sequence, extract the temperature value of each pixel on the concrete surface frame by frame, and construct a three-dimensional spatiotemporal temperature field dataset, with two spatial dimensions corresponding to image coordinates and the third dimension being time. A dynamic feature map of the temperature field is generated. Based on the three-dimensional spatiotemporal temperature field dataset, the temperature change rate, temperature gradient magnitude, and temperature field entropy value at each spatial location are calculated and fused to form a multi-channel dynamic feature map. The slump prediction is performed by inputting the deep neural network model. The multi-channel dynamic feature map is used as input and fed into a pre-trained convolutional-temporal hybrid neural network model. The convolutional-temporal hybrid neural network model extracts local thermodynamic features through spatial convolutional layers and models the time evolution of the temperature field through gated cyclic unit layers, and finally outputs the slump value of concrete. Output the detection results and perform confidence assessment. Generate the collapse prediction value based on the output of the convolutional-temporal hybrid neural network model, and simultaneously calculate the confidence interval and uncertainty index of the prediction results. If the uncertainty is greater than the preset threshold, trigger the retest mechanism or manual review process.

2. The intelligent concrete slump detection method based on deep learning according to claim 1, characterized in that, The inter-frame alignment employs a phase-correlation-based sub-pixel level image registration algorithm, while the radiation correction uses Planck's blackbody radiation law combined with on-site measured environmental parameters to compensate for non-uniformity and control temperature measurement errors.

3. The intelligent concrete slump detection method based on deep learning according to claim 2, characterized in that, The temperature change rate is obtained by smoothing the time series using a high-order Savitzky-Golay filter and then taking its derivative. The temperature gradient magnitude is calculated in the spatial domain using the Sobel operator. The temperature field entropy is defined as the Shannon entropy of the probability density function of the temperature distribution in the current frame.

4. The intelligent concrete slump detection method based on deep learning according to claim 3, characterized in that, The convolutional-temporal hybrid neural network model contains multiple spatial convolutional blocks. Each convolutional block consists of a convolutional layer, a batch normalization layer, and a ReLU activation function, followed by multiple bidirectional gated recurrent units. The output layer uses a fully connected layer with a linear activation function. The model training uses a weighted combination of mean squared error loss and Huber loss.

5. The intelligent concrete slump detection method based on deep learning according to claim 4, characterized in that, The confidence assessment is based on the Monte Carlo Dropout method, which performs multiple random sampling predictions on the same input during the inference phase and calculates the standard deviation of the predicted values ​​as an uncertainty index. When the standard deviation is greater than a preset distance threshold, it is judged as a low confidence result.

6. The intelligent concrete slump detection method based on deep learning according to claim 5, characterized in that, The method also includes an environmental interference suppression step: before acquiring the thermal infrared video sequence, ambient light and background scene images are simultaneously acquired using a visible light camera, and a semantic segmentation network is used to identify and eliminate thermal signals in non-concrete areas to prevent interference from false heat sources caused by solar radiation, equipment heating, or personnel movement.

7. The intelligent concrete slump detection method based on deep learning according to claim 6, characterized in that, The semantic segmentation network uses the DeepLabv3+ architecture, with ResNet-50 as the backbone network. The output categories include concrete, ground, equipment, people, and sky. Pixels that are not in the concrete category are marked as invalid in the infrared image.

8. The intelligent concrete slump detection method based on deep learning according to claim 7, characterized in that, The method also includes a multimodal data fusion step: the thermal infrared video sequence is pixel-level aligned with the synchronously acquired visible light video, the geometric deformation features in the visible light image are extracted, and the channel is stitched together with the temperature field dynamic feature map, and then input into a dual-stream neural network architecture.

9. The intelligent concrete slump detection method based on deep learning according to claim 8, characterized in that, The geometric deformation features are calculated using the Lucas-Kanade optical flow method to determine the displacement field between adjacent frames, and the maximum flow radius and collapse height attenuation rate are extracted and encoded into a single-channel spectrum for feature fusion.

10. The intelligent concrete slump detection method based on deep learning according to claim 9, characterized in that, The method is deployed on edge computing devices, with infrared thermal imagers and visible light cameras integrated on the slump test support. The entire process of data acquisition, processing, and prediction is completed within a predetermined time period after a single test, meeting the real-time detection needs of construction sites.

Images