Underwater blasting construction method and system based on real-time video image processing

Abstract

The invention provides an underwater blasting construction method and system based on real-time video image processing, and the method comprises the steps: arranging a plurality of underwater cameras and a near-infrared light supplement array to collect a synchronous video, and carrying out the pixel-level segmentation of cracks, bubble plumes and turbid regions after improved dark channel-guided filtering defogging; calculating a risk index I based on the crack length, the plume density and the turbid area, inputting the joint vector into a second convolutional neural network model, and outputting the explosive charge and the detonation delay of each blast hole; and when the I change exceeds a threshold value or reaches a preset period, the electric detonator network is triggered to perform blasting. The system is composed of an underwater imaging subsystem, an underwater sub-control box, a shore-based GPU edge server and a blasting PLC, and integration of environment perception, intelligent decision making and closed-loop blasting is achieved. According to the method, the segmentation precision and decision real-time performance are improved, the explosive unit consumption and turbid plume are reduced, and the construction safety and environmental protection performance are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of blasting construction, and in particular to an underwater blasting construction method and system based on real-time video image processing. Background Art

[0002] Underwater blasting is a common method for rock demolition and obstacle removal, widely used in projects such as channel dredging, submarine pipeline laying, reservoir hazard removal, and nearshore port expansion. Traditionally, blasting designers rely on offline calculations of charge weights, blasthole spacing, and detonation sequences based on preliminary geological surveys, visual inspections by divers, and limited sonar or pressure sensor data. During construction, divers or remotely operated vehicles then verify blasthole depths and complete detonation connections on-site. This model suffers from the following technical pain points: Information acquisition lags: Underwater lighting is insufficient and the water is turbid, sonar can only roughly capture rock surface fluctuations. Divers' line of sight is often less than 1 meter, making it impossible to monitor crack propagation, bubble plumes, and turbidity changes in the blasting area in real time. Charge and sequence lack adaptability: Blasting parameters are set once onshore and cannot be adjusted in real time to the rapidly changing rock integrity or water velocity at the construction site, often resulting in wasted explosives and suboptimal blasting results. High safety and environmental risks: Divers must undergo multiple underwater retests, exposing themselves to high pressure, low temperatures, and sandy water for extended periods. Without real-time control, it's difficult to suppress the turbidity plume and underwater shock waves generated by the blast, potentially damaging aquatic ecosystems or nearby structures. Data traceability is difficult: Traditional methods lack continuous video or high-resolution sensor recordings, making it difficult to trace responsibility based on fragmented data should quality disputes or environmental complaints arise.

[0003] In recent years, rapid progress has been made in waterproof high-frame-rate cameras, laser-assisted illumination, and deep learning vision algorithms, providing new perception tools for underwater operations. Some studies have used underwater cameras to perform three-dimensional reconstruction of rock walls before blasting, or segmented the particle size distribution of rock blocks after blasting using convolutional networks to assess the effect. Other teams have attempted to use real-time image analysis to monitor bubble plumes and estimate the location of the blast center. However, most of these studies are limited to offline assessments: images are batch-processed by onshore workstations after the operation is completed, unable to provide feedback during blasting and forming a closed-loop with the blasting control system. Furthermore, currently available literature often focuses on single indicators (such as particle size, plume, or turbidity), lacking quantification and decision-making logic for the comprehensive risk of multiple indicators. Current industrial control systems also often use PLCs to trigger delayed detonators with fixed timing, lacking interfaces for linking with visual recognition results. Overall, a comprehensive solution combining "real-time multimodal visual perception + risk quantification + adaptive detonation control" has yet to be established, hindering the advancement of underwater blasting towards digitalization, refinement, and safety.

[0004] In order to overcome the above shortcomings, there is an urgent need for an underwater blasting construction method and system that combines high-frame-rate video acquisition, image enhancement and dehazing, semantic segmentation, risk index calculation, and deep reinforcement learning decision-making: it can obtain clear images in high turbidity and low-light environments, and perform pixel-level recognition of cracks, plumes, and turbid areas; through continuous frame timing information and sliding window statistics, it can assess the blasting environment risk in real time; when the risk index changes significantly or reaches the decision cycle, it can automatically adjust the explosive charge and detonation delay to achieve closed-loop control; and by incorporating the risk index into the calculation process of blasting explosive quantity and blasting timing selection to effectively improve the construction accuracy, it is urgent to solve the problem. At the same time, the whole process video and control instructions are recorded to improve the safety traceability capability. The present invention is proposed based on the above needs. Through the four-level system of "video perception-risk assessment-deep reinforcement learning-detonation execution", it realizes the intelligence, adaptability and traceability of underwater blasting construction, providing a new technical approach for the safe, efficient and green development of marine engineering. Summary of the Invention

[0005] To address the aforementioned issues in the prior art, the present invention provides a method and system for underwater blasting construction based on real-time video image processing. This method employs multiple underwater cameras and a near-infrared fill-light array to capture synchronized video. After dehazing using an improved dark channel-guided filtering algorithm, a first convolutional neural network (CNN) with a temporal attention mechanism performs pixel-level segmentation of cracks, bubble plumes, and turbid areas. A risk index (I) is calculated based on crack length, plume density, and turbid area. The dehazed RGB vector, the labeled grayscale vector, and I are concatenated into a joint vector and input into a second convolutional neural network model with a risk-modulated activation function. This outputs the explosive charge and detonation delay for each blasthole. When I changes beyond a threshold or reaches a preset period, the electric detonator network is triggered to initiate blasting. The system, comprised of an underwater imaging subsystem, an underwater subcontrol box, a shore-based GPU edge server, and a blasting PLC, integrates environmental perception, intelligent decision-making, and closed-loop detonation. This invention improves segmentation accuracy and real-time decision-making, reduces explosive consumption and turbid plume generation, and enhances construction safety and environmental performance.

[0006] The present application provides an underwater blasting construction method based on real-time video image processing, comprising the steps of:

[0007] S1: Deploy N underwater cameras and corresponding near-infrared LED fill light arrays in the blasting area to obtain synchronized video streams;

[0008] S2: performing dark channel prior transmittance estimation on each frame of the synchronized video stream, and refining the transmission map using guided filtering under a 5×5 window to reconstruct a defogging frame sequence. Each defogging frame is expanded row by row into an RGB three-channel vector to obtain a first video stream frame vector;

[0009] S3: The dehazed video frames are processed by the first convolutional neural network with a temporal attention mechanism to obtain annotated video frame images with rock cracks, bubble plumes, and turbid water areas marked. The pixel grayscale values ​​of the annotated video frame images are expanded row by row into a single-channel vector to obtain the second video stream frame vector.

[0010] S4: Based on the crack length, plume density and turbidity area of ​​each frame of the video stream, the total crack length, average plume density and total turbidity area of ​​the video stream are obtained, and the risk index I is calculated;

[0011] S5: fusing the first video stream frame vector, the second video stream frame vector, and the risk index I to generate a joint vector, and inputting the joint vector into a second convolutional neural network model to generate an explosive charge and a detonation sequence;

[0012] S6: When the difference between the latest risk index I and the risk index of the previous decision cycle exceeds a preset threshold, the electric detonator network is controlled to execute blasting according to the explosive charge and detonation sequence.

[0013] Preferably, step S2 also includes: for each frame image in the synchronous video stream, for the brightness of different color channels of each pixel in the frame within a preset local neighborhood, selecting pixels whose brightness values ​​are less than a set threshold to generate a dark channel image; generating an initial transmittance image according to a preset proportional coefficient based on the dark channel image and the global atmospheric light brightness of the scene; using the original frame image as a guide image, performing guided filtering refinement processing on the initial transmittance image within a 5×5 pixel window to obtain a smooth transmittance image; fusing the smooth transmittance image and the original frame image pixel by pixel according to the set transmittance lower limit threshold to generate a defogging frame image; organizing all defogging frame images into a defogging frame sequence in chronological order and outputting them.

[0014] Preferably, the first video stream frame vector is a one-dimensional data sequence, and the red, green, and blue channel values ​​of each pixel of the defogging frame are read in sequence from left to right in row scanning order, and are continuously appended in the order of R, G, and B. The length of the resulting sequence is equal to three times the number of pixels in the original frame; the second video stream frame vector is a one-dimensional data sequence, and the single-channel grayscale values ​​of the corresponding pixels of the marked frame are read in sequence from left to right and are continuously appended. The length of the resulting sequence is equal to the number of pixels in the original frame; the detonation timing is a delay sequence between adjacent blast holes, and the master control triggering moment is recorded as 0.

[0015] Preferably, the fusion of the first video frame vector, the second video frame vector and the risk index I to generate a joint vector includes: directly taking the first video stream frame vector as the first segment of the joint vector in a predetermined order; then sequentially appending the second video stream frame vector thereafter; and finally appending the risk index I containing only a single scalar element; thereby obtaining a continuous one-dimensional data sequence whose first segment is defogging frame data, the second segment is labeled frame data, and the end is the risk index, which is the joint vector.

[0016] Preferably, the calculation of the risk index I includes: comparing the total length of the video stream crack, the average density of the plume and the total area of ​​the turbidity zone with their respective preset maximum allowable values ​​to obtain three normalized evaluation values ​​between zero and one; accumulating the normalized crack length with a weight of 40%, accumulating the normalized plume density with a weight of 35%, and accumulating the normalized turbidity zone area with a weight of 25%, and summing the three weighted results to obtain the risk index I.

[0017] The present application also provides an underwater blasting construction system based on real-time video image processing, comprising:

[0018] Synchronous video stream acquisition module, deploying N underwater cameras and corresponding near-infrared LED fill light arrays in the blasting area to acquire synchronous video streams;

[0019] A defogging and first video stream frame vector generation module performs dark channel prior transmittance estimation on each frame of the synchronized video stream, refines the transmission map using guided filtering within a 5×5 window, reconstructs a defogging frame sequence, and expands each defogging frame row-wise into an RGB three-channel vector to obtain the first video stream frame vector.

[0020] The second video stream frame vector acquisition module processes the dehazed video frames using a first convolutional neural network with a temporal attention mechanism to obtain annotated video frame images with rock cracks, bubble plumes, and turbid water areas marked. The pixel grayscale values ​​of the annotated video frame images are expanded row by row into a single-channel vector to obtain the second video stream frame vector.

[0021] The risk index I calculation module obtains the total crack length, average plume density, and total turbidity area of ​​the video stream based on the crack length, plume density, and turbidity area of ​​each frame of the video stream, and calculates the risk index I;

[0022] An explosive charge and detonation sequence generation module, which fuses the first video stream frame vector, the second video stream frame vector, and the risk index I to generate a joint vector, and inputs the joint vector into a second convolutional neural network model to generate an explosive charge and a detonation sequence;

[0023] The blasting execution module controls the electric detonator network to execute blasting according to the explosive charge and detonation sequence when the difference between the latest risk index I and the risk index of the previous decision cycle exceeds a preset threshold.

[0024] Preferably, the defogging and first video stream frame vector generation module also includes: for each frame image in the synchronized video stream, for the brightness of different color channels of each pixel in the frame within a preset local neighborhood, pixels with brightness values ​​less than a set threshold are selected to generate a dark channel image; an initial transmittance image is generated according to a preset proportional coefficient based on the dark channel image and the global atmospheric light brightness of the scene; using the original frame image as a guide image, guided filtering refinement processing is performed on the initial transmittance image within a 5×5 pixel window to obtain a smooth transmittance image; the smooth transmittance image and the original frame image are fused pixel by pixel according to the set transmittance lower limit threshold to generate a defogging frame image; all defogging frame images are organized into a defogging frame sequence in chronological order and output.

[0025] Preferably, the first video stream frame vector is a one-dimensional data sequence, and the red, green, and blue channel values ​​of each pixel of the defogging frame are read in sequence from left to right in row scanning order, and are continuously appended in the order of R, G, and B. The length of the resulting sequence is equal to three times the number of pixels in the original frame; the second video stream frame vector is a one-dimensional data sequence, and the single-channel grayscale values ​​of the corresponding pixels of the marked frame are read in sequence from left to right and are continuously appended. The length of the resulting sequence is equal to the number of pixels in the original frame; the detonation timing is a delay sequence between adjacent blast holes, and the master control triggering moment is recorded as 0.

[0026] Preferably, the fusion of the first video frame vector, the second video frame vector and the risk index I to generate a joint vector includes: directly taking the first video stream frame vector as the first segment of the joint vector in a predetermined order; then sequentially appending the second video stream frame vector thereafter; and finally appending the risk index I containing only a single scalar element; thereby obtaining a continuous one-dimensional data sequence whose first segment is defogging frame data, the second segment is labeled frame data, and the end is the risk index, which is the joint vector.

[0027] Preferably, the calculation of the risk index I includes: comparing the total length of the video stream crack, the average density of the plume and the total area of ​​the turbidity zone with their respective preset maximum allowable values ​​to obtain three normalized evaluation values ​​between zero and one; accumulating the normalized crack length with a weight of 40%, accumulating the normalized plume density with a weight of 35%, and accumulating the normalized turbidity zone area with a weight of 25%, and summing the three weighted results to obtain the risk index I.

[0028] The present invention provides an underwater blasting construction method and system based on real-time video image processing, which can achieve the following beneficial technical effects:

[0029] 1. The present invention organizes the sequential steps of "defogging-semantic segmentation-risk quantification-intelligent decision-making" to logically decouple the first convolutional neural network (responsible for pixel-level segmentation) and the second convolutional neural network (responsible for parameter decision-making) from each other and connect them in terms of data. The first network only receives the original defogging image and outputs four types of labeled images. The system concatenates the defogging RGB one-dimensional vector + the labeled grayscale vector + the real-time risk index into a fixed-length joint vector as the input of the second network, which not only ensures the integrity of the state description, but also significantly compresses the feature dimension (from 18.6MB to 5MB), reducing the inference delay by about 32%. The newly introduced risk index I fuses the crack length, plume density and turbidity area with weights of 40% / 35% / 25%. The quantization result is continuously differentiable from 0 to 1, providing a unified and secure measurement for the subsequent decision-making network.

[0030] 2. Adaptive scheduling of blasting parameters and execution timing: The second network re-outputs 40 sets of explosive charges and adjacent detonation delays every 1 second only when |I_now–I_prev|>0.05 or reaches a fixed period, avoiding the high-frequency jitter of "one decision per frame". The PLC first performs a safety link self-check and a 5-second sound and light countdown, and then triggers the electric detonator according to the delay sequence, achieving a dual-layer protection of millisecond-level closed-loop and minute-level safety confirmation. Compared with the fixed charge scheme, the average blasting unit consumption is reduced by 12.8%, the turbidity plume peak radius is reduced by 26%, and the construction period is shortened by 20%.

[0031] 3. The present invention improves the scenario coupling of the activation function. The first convolutional neural network model uses the T-Swish activation function: the slope is dynamically adjusted by the temporal attention score α_t, the mIoU of crack and plume detection is improved by 1.4pp, and the false detection rate is reduced by 14%; the second convolutional neural network model uses R-Sigmoid at the last layer of the Actor, and the risk index is directly written into the activation slope. In high-risk scenarios, the output saturates faster, and the explosive charge automatically converges to the safe lower limit. Both activation methods do not add weights and only have a single multiplication overhead, keeping the end-to-end delay unchanged. At the same time, learnable parameters τ and r are provided, enabling the network to automatically calibrate under different sea conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 This is a flowchart of the steps of an underwater blasting construction method based on real-time video image processing of the present invention;

[0034] Figure 2It is a schematic diagram of an underwater blasting construction system based on real-time video image processing of the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0036] Example 1:

[0037] In order to solve the above-mentioned technical problems mentioned in the prior art, the following Figure 1 As shown: This application provides an underwater blasting construction method based on real-time video image processing, comprising the steps of:

[0038] S1: N underwater cameras and corresponding near-infrared LED fill-light arrays are deployed in the blasting area to capture synchronized video streams. For shallow harbor dredging scenarios, four 4K / 60fps waterproof cameras (approximately 120° field of view, pressure-resistant up to 20m) and corresponding 850nm near-infrared LED fill-light arrays can be used. The deployment involves placing tripods in four directions, east, west, south, and north, on a horizontal plane with a radius of 6m, centered around the explosion center. The tripod tops are approximately 0.5m above the seabed. Each camera is tilted downward 15° to ensure at least 70% overlap of the field of view. Four LED light strips are installed around the camera housing, each with a 120° angle, providing at least 40lx of illumination at a distance of 3m. The cameras and light arrays share a single DC24V / Gigabit Ethernet composite cable connected to a shore-based PoE switch. Each cable length is limited to 80m. IEEE-1588PTP is used for synchronization, with the shore server serving as the master clock and the cameras as slaves. Frame timestamp error is maintained within 3ms. The entire deployment process was completed by two technicians working with an ROV and took approximately 45 minutes. During the operation, the cable tension was remotely checked every 2 hours. If a sudden turbidity peak occurred, the LED power could be temporarily increased by 20% to maintain image quality.

[0039] In deepwater submarine pipeline trench blasting scenarios, six 1080p / 50fps low-light cameras (90° field of view, 50m pressure resistance) and a 940nm focused LED array are used. The cameras are suspended by steel cables from equally spaced circular supports with a 10m radius and a base height of 0.8m from the seabed. The cameras have a 20° pitch angle, maintaining field of view overlap of at least 60%, and a 3kg counterweight atop each camera to minimize sway. LED fill lights are mounted on two 30° outward-biased arms to reduce forward scatter, providing a minimum illumination of 25lx at 5m. Power and data are transmitted via an AC48V / single-mode fiber composite cable to an underwater control box, where they are converted to Gigabit Ethernet and routed back to shore via a buoy. A PTP master clock is deployed on the buoy, with frame synchronization error within 5ms. The deployment was completed by three technicians and one ROV operator in approximately 90 minutes. The camera housing temperature is monitored every hour during operation. When current speed exceeds 0.7 m / s⁻¹, the cable is automatically tightened to stabilize the viewing angle. Both implementations utilize IP68 quick-connect plugs for quick device replacement, and a 2A fuse is installed on the power supply side to ensure system redundancy and safety.

[0040] S2: Perform a priori dark channel transmittance estimation on each frame of the synchronized video stream, and refine the transmission map using guided filtering within a 5×5 window to reconstruct a dehazed frame sequence. Each dehazed frame is expanded row-wise into a RGB three-channel vector to obtain the first video stream frame vector. In this embodiment, the collected synchronized video stream is a 1920×1080@30fps MP4 file located at / data / burst_001.mp4. Processing is performed on a small edge server equipped with an NVIDIA Jetson AGX Orin (GPU 32GB). The entire process is executed sequentially for each frame. The core steps are as follows:

[0041] 1. Frame reading and three-channel splitting. Use OpenCV's video decoding interface, cv2.VideoCapture, to read frames by frame. Each frame is converted to RGB color space and cached into the NumPy array img_rgb. The resolution remains at 1920×1080, and the data type is uint8.

[0042] 2. Dark channel image generation: To increase processing speed, we only calculate the local minimum brightness value for the RGB channels within a 15×15 pixel neighborhood, outputting a single-channel dark channel image, dark_channel. This example uses cv2.erode with a custom minimum pooling kernel, and the calculation takes approximately 8ms per frame.

[0043] 3. Initial transmittance estimation and atmospheric light sampling: In the dark channel image, take the original color pixels corresponding to the top 0.1% of pixels in brightness and calculate their average RGB value as the atmospheric light A_rgb. Subsequently, the dark channel image is reverse-mapped by a fixed coefficient of 0.95 to obtain a rough transmittance image t_coarse (still a uint8 grayscale image). This step takes 4ms per frame.

[0044] 4. For 5×5 window guided filtering refinement, call the guided_filter function from the Guided-Filter-Py library, using the original RGB image img_rgb as the guide image and t_coarse as the image to be refined. The window radius is set to 2 (i.e., 5×5), and the regularization coefficient ε is set to 1e-3. After refinement, the smoothed transmittance image t_refined is obtained. The computation takes 11ms per frame, and the GPU utilization is approximately 15%.

[0045] 5. Dehazed frame reconstruction: A lower threshold of 0.1 is set for excessively small transmittance values ​​in t_refined to avoid forced noise amplification. Then, using the dark channel prior recovery model, the haze-free frame img_clear is reconstructed at the pixel level. Reconstruction takes 6ms per frame. Actual comparisons show that the average brightness of the dehazed frame is increased by 28%, and color cast is significantly reduced.

[0046] 6. Frame-level vectorization is performed. Img_clear is expanded in a top-down, left-to-right row scan order: first, the R channel values ​​of all pixels are written sequentially; then, the G channel values ​​of all pixels are written sequentially; and finally, the B channel values ​​are written. The resulting one-dimensional array vec_rgb is fixed at 1920 × 1080 × 3 = 6220800 elements, of data type uint8. This vector is the "first video stream frame vector" and is subsequently concatenated with the second frame vector at the same time step and the risk index in subsequent steps.

[0047] 7. Real-time Performance and Caching Strategy: End-to-end processing on JETSON takes approximately 29–30ms per frame, enabling stable real-time dehazing at 30fps. To ensure smoothness, the system employs dual caching: the main thread continuously writes the next frame, while worker threads use CUDA streams to parallelize the current frame, ensuring peak video memory usage is less than 1GB.

[0048] 8. Abnormalities and fault tolerance: If a frame fails to be read or the transmittance is detected to be zero during the defogging process, the system automatically skips the frame and records the timestamp in the log. If the frame is skipped five times in a row, an alarm is triggered, the LED fill light power is increased by 20%, and the camera is reset.

[0049] S3: The dehazed video frames are processed by a first convolutional neural network (CNN) with a temporal attention mechanism to produce annotated video frames with rock cracks, bubble plumes, and turbid water areas. The pixel grayscale values ​​of the annotated video frames are expanded row-wise into single-channel vectors to generate the second video stream frame vector. The first convolutional neural network with the temporal attention mechanism uses a two-stage "Deform-Conv + TC-Former" architecture. The first stage consists of a four-layer deformable convolutional encoder, extracting multi-scale features of 64×128×128, 128×64×64, 256×32×32, and 512×16×16, frame by frame. The second stage feeds the same-scale features from five consecutive frames into a temporal cross-frame transformer (TC-Former) using a five-frame sliding window. Multi-head attention is used to calculate key-value weights in the temporal dimension to capture the dynamic relationship between crack propagation and plume diffusion. The network outputs a four-category pixel probability map, which is then subjected to pixel-by-pixel argmax to produce a single-channel labeled image of 1920×1080 pixels. Grayscale values ​​are encoded as follows: 0 represents background, 85 represents cracks, 170 represents bubble plumes, and 255 represents turbid water areas. The training data consists of 12,000 sea trial images and 6,000 simulated synthetic images, all manually annotated with four-category masks. The AdamW optimizer is used, with an initial learning rate of 3×10 -4 , batch size = 8 (5 frames / sample), trained on 2x NVIDIA A40s for 60 epochs, achieving a mIoU of 92.3%. During deployment, 30% pruning and quantization to INT8 were applied. Single-frame inference took 18ms, and with a 5-frame sliding window, the overall latency was approximately 90ms, meeting a 10Hz decision frequency. The resulting annotation graph is expanded in row-by-row order: the grayscale values ​​of the first row of 1920 pixels are read first, followed by rows 2 through 1080, and the 2,073,600 grayscale bytes are sequentially written into a one-dimensional array, forming the second video stream frame vector. This vector is concatenated with the RGB vector of the same frame and the real-time risk index, and then fed into the second convolutional neural network model to dynamically adjust the blasting parameters.

[0050] S4: Based on the crack length, plume density, and turbidity area of ​​each frame in the video stream, the total crack length, average plume density, and total turbidity area of ​​the video stream are calculated, and the risk index I is calculated. The 4K camera used at the blasting site was calibrated with an underwater laser ruler, resulting in a pixel-to-meter conversion factor p2m ≈ 0.002 m / px. In the single-channel annotated grayscale image output by network inference, a grayscale value of 85 represents crack pixels, 170 represents bubble plume pixels, and 255 represents turbidity pixels. The system continuously processes each frame at 30 fps and generates the risk index I in real time according to the following steps: Frame-by-frame crack length calculation: Thinning is performed on the set of pixels with a grayscale of 85, and then a skeleton is extracted using the 8-connected tracking algorithm. The number of skeleton pixels is multiplied by p2m to obtain the crack length L_f for that frame. Frame-by-frame plume density calculation: Counting the number of pixels with a grayscale of 170 and dividing it by the total number of pixels in the entire frame yields the plume density D_f (ranging from 0 to 1). Calculate the turbidity area frame by frame, count the grayscale as 255 pixels A_px, and multiply it by p2m 2 Converted to turbidity area A_f (unit m 2 ). Sliding window summary, sum or average with a 1s sliding window (30 frames) to obtain the real-time total crack length L_sum, plume average density D_avg, and turbidity zone total area A_sum, and update once per second. Normalization processing, divide L_sum, D_avg, and A_sum by the preset maximum allowable values: L_max = 6m, D_max = 0.25, A_max = 1.5m 2 , we get three 0–1 normalized indicators L, D, and A. We perform weighted summation, adding L, D, and A with weights of 0.40, 0.35, and 0.25, respectively, to obtain the risk index I. For example: if L_sum = 3.2m, D_avg = 0.18, and A_sum = 0.9m 2 , then L = 0.53, D = 0.72, A = 0.60, and I = 0.40 × 0.53 + 0.35 × 0.72 + 0.25 × 0.60 ≈ 0.61. For risk grading and triggering, the system limits I to the range of 0–1: I < 0.30 indicates low risk; 0.30–0.60 indicates medium risk; and I ≥ 0.60 indicates high risk. If the difference between the current I and the previous decision cycle I_prev |I–I_prev| > 0.05, or if I crosses the grading threshold, the deep reinforcement learning agent is triggered to recalculate the explosive charge and detonation delay; otherwise, the previous set of blasting parameters is used. Through the above process, the system updates risk indicators at millisecond speeds at 30fps. The decision-making stage only outputs a new blasting plan when the environment changes substantially or after a full second. This avoids the unreasonable high-frequency control of "one decision per frame" while ensuring that blasting parameters remain sensitive to environmental fluctuations.

[0051] S5: The first and second video stream frame vectors are fused together to generate a joint vector. This joint vector is then fed into a second convolutional neural network model to generate the explosive charge and detonation sequence. The acquisition end triggers a decision every 0.2 seconds (i.e., six frames). The system first processes the first video stream frame vector (RGB vector, 6,220,800 bytes), the second video stream frame vector (single-channel grayscale vector, 2,073,600 bytes), and the real-time risk index I obtained for each of these six frames as follows: The RGB vector is fed into the encoder of an offline-trained 1×1 convolutional autoencoder and compressed into a 1024-dimensional floating-point vector v1. The grayscale vector is fed into the encoder of another 1×1 convolutional autoencoder and compressed into a 256-dimensional vector v2. Both encoders are fine-tuned using 50,000 dehazed / annotated images after ImageNet transfer learning, achieving a PSNR loss of <0.2dB. Aggregate the time dimension and perform element-wise averaging on v1 and v2 obtained from 6 frames to obtain the frame window average vectors V1 and V2 (the dimensions are still 1024 and 256). This not only preserves the scene trend within 1 second, but also significantly reduces noise. Joint vector concatenation, place V1 in the first segment and V2 in the second segment, and finally append a single scalar risk index I (double-precision floating point and normalized to 0–1) to obtain a joint vector state = [V1||V2||I] of length 1281. The joint vector is written into the pre-allocated gRPC message and sent to the DRL inference end. The second convolutional neural network structure is a one-dimensional convolution-attention hybrid strategy network. The first two layers use 1DConv (128 channels, 256 channels, kernel width 5, stride 1) to extract local correlation features. The third layer is multi-head self-attention (4 heads, hidden dimension 512) to capture cross-dimensional correlations, which are then flattened and divided into two fully connected branches:

[0052] Actor branch: FC (512→256→80), the first 40 nodes are mapped to 0–0.5 kg as the explosive charge m of 40 blastholes through Sigmoid i , the last 40 nodes are mapped to adjacent detonation delays Δt by Sigmoid×100ms i ; Critic branch: FC (512→256→1) outputs state value V(s). Offline training and online inference, PPO (clip=0.2) is used in the offline stage to train 3×10 6 The reward is a weighted combination of four factors: rock particle size target, explosive consumption per unit, turbidity peak, and vibration acceleration. After convergence, the Actor branch has an inference latency of approximately 4ms on NVIDIA Orin. Here's an example of parameter output: when the joint vector state has a medium risk index of I = 0.48 and the scene is judged to be medium risk, the Actor output is:

[0053] Charge mass m = [0.28, 0.25, 0.29, …, 0.22] kg (40 values);

[0054] Detonation delay Δt = [0, 12, 23, 34, ..., 421] ms (40 values);

[0055] The system instantly writes the PLC detonation table; if the next decision cycle I differs from the current I by ≤0.05, the table is used without re-issuing it. With this implementation, the length of the joint vector is fixed, and a small edge GPU can output the explosive charge and delay sequence for an entire set of blastholes in real time at a 10Hz decision rate. This avoids the redundancy of "one decision per frame" and ensures that a new adaptive blasting plan can be generated within ≤0.3 seconds when environmental risks change suddenly.

[0056] S6: When the difference between the latest risk index I and the risk index from the previous decision cycle exceeds a preset threshold, the electric detonator network is controlled to execute the blast based on the explosive charge and detonation sequence. The blasting control system uses a one-second decision cycle. Before the operation, engineers set the risk differential threshold ΔI_thr to 0.05 and programmed a three-state logic control program into the PLC: monitor—ready—detonate. The following example illustrates the triggering and execution process using a four-cycle measured log.

[0057] In the 0th cycle (T0 = 00:00:00), the deep reinforcement learning module outputs the explosive charge table m1…m 40 With the detonation delay sequence Δt1…Δt 40 The baseline risk index, I_prev, is also recorded as 0.42. The PLC caches the parameters in the "distribution register," but the system is in the safety check phase and does not immediately initiate the detonation.

[0058] During the first cycle (T1 = +1s), the new risk index, I_now, calculated in real time, is 0.45, with |I_now – I_prev| = 0.03 < 0.05. The PLC remains in the "Ready" state: the charge and delay tables retain the results from the previous cycle; the LED warning light remains solid yellow, indicating that the blasting unit is on standby; the audible and visual alarms and the water buzzer remain silent.

[0059] In the second cycle (T2 = +2s), I_now rises to 0.49, and the difference |0.49–0.42| = 0.07 > ΔI_thr. The logic triggers the following process: Re-issuance of parameters: The DRL module generates a new version of m′1…m′ within 0.25s. 40 , Δt′1…Δt′ 40, the PLC overwrites the register and locks the write port; Safety confirmation: The system polls the "diver out of zone" sensor, the fiber optic pyrotechnic circuit resistance, and the UPS voltage. When all return green signals, the system changes the state to "detonation ready"; 5-second countdown: During the countdown, the buoy's sound and light alarm flashes at high intensity to warn surrounding ships. T2+5s (=+7s) detonation execution PLC triggers the main relay, sequentially injecting 3A, 24V current into 40 electric detonators; within the first 40ms, press Δt'1...Δt' 40 All blasts were completed at intervals. The high voltage was then immediately cut off and the data marker "Blast #17" was activated.

[0060] During the third cycle (T3 = +8s), sensors detected dust and pressure fluctuations, causing I_now to temporarily jump to 0.65. Since the blasting was complete, this cycle entered "safety cooling" mode, recording only data and temporarily suspending the generation of a new charge table. After three minutes, the system automatically downgraded to "hole layout reset" mode. This example illustrates the practical role of the risk index differential threshold in triggering detonation execution: only when the environmental risk has significantly changed compared to the previous decision cycle (|I_now – I_prev| > 0.05) does the system recalculate and immediately execute the new charge and delay plan; otherwise, it maintains the established parameters, avoiding excessively frequent detonation commands while ensuring the integrity and traceability of safety checks and audible and visual warnings.

[0061] In some embodiments, step S2 further includes: for each frame in the synchronized video stream, selecting pixels with brightness values ​​less than a set threshold for the brightness of different color channels within a preset local neighborhood of each pixel in the frame to generate a dark channel image; generating an initial transmittance image based on the dark channel image and the global atmospheric brightness of the scene according to a preset proportional coefficient; performing guided filtering and refinement on the initial transmittance image within a 5×5 pixel window using the original frame image as a guide image to obtain a smooth transmittance image; pixel-by-pixel fusion of the smooth transmittance image with the original frame image according to a set transmittance lower threshold to generate a dehazed frame image; and organizing all dehazed frames into a dehazed frame sequence in chronological order and outputting the result. In some embodiments, when processing a video stream captured at 1920×1080 pixels and 30 fps, the system performs dehaz enhancement on each frame using the following fixed parameters. First, a 15×15 local neighborhood is taken at each pixel and the brightness of the R, G, and B channels of the neighborhood is compared. If the brightness is less than 30 (on a 0–255 scale), the pixel is written to the dark channel image. After extracting all dark channel pixels within a frame, the system selects the top 0.1% of pixel locations in the dark channel image, reads the RGB values ​​from the corresponding locations in the original frame, and averages them to obtain the global atmospheric brightness A≈[218,221,226]. The dark channel image is then reverse-mapped to generate an initial transmittance map using a preset scaling factor of 0.95; typical pixel transmittances range from 0.2 to 0.9. Using the original frame image as a guide map, the initial transmittance map is refined using a guided filter within a 5×5 window to produce a smoothed transmittance map. The regularization coefficient ε is fixed at 1×10⁻³ during the filtering process. To prevent over-brightening of dark areas, the system sets a pixel-by-pixel lower threshold of 0.1 for transmittance. The system then fuses the transmittance pixel-by-pixel with the original frame using the classic dark channel restoration formula to generate a dehazed frame. The entire process takes approximately 29ms per frame on Jetson Orin, capable of supporting 30fps in real time. The processed dehazed frames are written to a chronological ring buffer, aggregated into a dehazed frame sequence, and output in real time to the subsequent semantic segmentation network.

[0062] In some embodiments, the first video stream frame vector is a one-dimensional data sequence. The red, green, and blue channel values ​​of each pixel in the defogging frame are read sequentially from left to right in row scan order and appended consecutively in the order of R, G, and B. The resulting sequence length is equal to three times the number of pixels in the original frame. The second video stream frame vector is a one-dimensional data sequence. The single-channel grayscale values ​​of the corresponding pixels in the annotated frame are read sequentially from left to right and appended consecutively. The resulting sequence length is equal to the number of pixels in the original frame. The detonation timing sequence is a delay sequence between adjacent blastholes, with the master control trigger time being recorded as 0. The defogging frame output by the camera has a resolution of 1920×1080 pixels. The system processes this frame in row scan order, generating two vectors and providing a demonstration of the detonation delay.

[0063] The first video stream frame vector (RGB vector) reads 1920 pixels of the first row: first take the red channel R 1-1 …R 1-1920 Write the buffer sequentially. Next, read the red channel from the second row and append it. This continues until the 1080th row has been written. The buffer is now 1920 × 1080 = 2073600 bytes long. Return to the first row and write the green channel G sequentially. Then, complete the blue channel B for all rows. The final one-dimensional vector length = 2073600 × 3 = 6220800 bytes (or floating-point values). The order is always "all R → all G → all B".

[0064] The second video stream frame vector (single-channel grayscale / label vector) uses a single-channel grayscale encoding: 0 = background, 85 = crack, 170 = plume, 255 = turbidity. The system scans the entire frame of pixels again from top to bottom and left to right, writing each of the 2,073,600 grayscale values ​​into another one-dimensional buffer, the length of which is the number of pixels in the original frame.

[0065] Example of detonation sequence. The number of blastholes in this batch is 40. The master control triggering time is recorded as t0 = 0ms. Delay sequence Δt i (i=1…40) is output by the reinforcement learning network, for example: 0ms, 12ms, 23ms, 35ms, 49ms, 65ms, 82ms, …, 421ms. The system accumulates the absolute detonation time online: t1=0ms, t2=12ms, t3=35ms, t4=70ms…t 40 = 421ms. The PLC loads these times into the detonation control table, triggering them sequentially according to the delayed detonator circuit sequence. If the risk index changes by less than 0.05 in the next decision cycle, the same delay table is used without reissuing it. This ensures frame-level data accuracy while avoiding the unreasonable high-frequency control of "one detonation per frame."

[0066] In some embodiments, fusing the first video frame vector, the second video frame vector, and the risk index I to generate a joint vector includes: directly taking the first video stream frame vector in a predetermined order as the first segment of the joint vector; subsequently appending the second video stream frame vector sequentially thereafter; and finally appending the risk index I, which contains only a single scalar element. This results in a continuous one-dimensional data sequence, with the first segment consisting of defogging frame data, the second segment consisting of annotated frame data, and the final segment consisting of the risk index, which is the joint vector. In this embodiment, the defogging frame has a resolution of 1920×1080 and is stored using the uint8 type; the risk index I is represented as a float32. The system performs the following operations in GPU memory: allocating a continuous buffer, first requesting a float32 buffer state_buf (6220800+2073600+1=8294401) with a length of 8294401 elements in video memory. Based on 4 bytes, this total occupies approximately 31.6MB and can be reused at one time. Copy the first video stream frame vector. The dehazed frame RGB vector was generated in the previous step and normalized to 0-1, with a length of 6220800. Call CUDAcudaMemcpyAsync to write it in original order to state_buf(0:6220800). Append the second video stream frame vector. The labeled frame grayscale value vector is mapped (0→0, 85→1, 170→2, 255→3) and converted to float32, with a length of 2073600. Write it to state_buf(6220800:8294400). Append the risk index I. The latest risk index I (for example, 0.48) is written directly to the last slot state_buf[8294400]. At this point, the first segment 0–6220799 is the dehazed frame RGB data; the second segment 6220800–8294399 is the labeled frame category data; and the last segment 8294400 is a single scalar I. The one-dimensional vector is formed, and state_buf is now the joint vector. Since the three segments of data are physically contiguous, the subsequent network can directly input the float32[8294401] tensor as a single tensor without additional copying or reordering. Inference calls are made through TensorRT's asynchronous inference interface, passing the joint vector using the binding name "state". The network input layer dimension is fixed to (1,8294401). Inference latency is approximately 4ms, outputting 40 charge sizes and 40 detonation delays.

[0067] In some embodiments, the calculation of the risk index I includes: comparing the total length of cracks in the video stream, the average density of plumes, and the total area of ​​turbidity zones with their respective preset maximum allowable values, to obtain three normalized evaluation values ​​between zero and one; accumulating the normalized crack length with a weight of 40%, accumulating the normalized plume density with a weight of 35%, and accumulating the normalized turbidity zone area with a weight of 25%, and summing the three weighted results to obtain the risk index I. A total of 30 frames of video stream are processed within a 1s decision cycle. The system first accumulates three on-site indicators according to the above method: total crack length L_sum = 3.8m; average plume density D_avg = 0.17 (i.e., plume pixels account for 17% of the pixels in the entire frame); total area of ​​turbidity zone A_sum = 0.92m 2 Before the operation starts, the supervision unit sets the three maximum allowable values ​​as L_max=6m, D_max=0.25, and A_max=1.5m according to the environmental protection red line and safety requirements. 2 The system then normalizes and weights the values ​​accordingly:

[0068] Crack normalization value Ln=L_sum÷L_max=3.8÷6≈0.63

[0069] Plume normalization value Dn = D_avg ÷ D_max = 0.17 ÷ 0.25 = 0.68

[0070] Turbidity normalization value An=A_sum÷A_max=0.92÷1.5≈0.61

[0071] The risk index, I, was then calculated using a fixed weighting of 40%:35%:25%: I = 0.40 × 0.63 + 0.35 × 0.68 + 0.25 × 0.61 = 0.252 + 0.238 + 0.153 = 0.643. The system then bounded the result to a range of 0–1 and compared it to a grading threshold: I ≥ 0.60 indicates high risk. Because the risk index recorded in the previous decision cycle was 0.57, the difference |0.643–0.57| = 0.073 exceeded the preset threshold of 0.05. The control logic immediately triggered the deep reinforcement learning module to recalculate the new explosive charge and detonation delay, and then executed the blast after a 5-second audible and visual warning countdown.

[0072] This application also provides an underwater blasting construction system based on real-time video image processing, such as Figure 2As shown, the system hardware components and connections include: an underwater imaging and attitude acquisition subsystem, four underwater cameras (4K / 60fps, 120° field of view, and 30m pressure resistance), an RJ-45 Gigabit Ethernet port and a shared composite cable for the 24V DC power supply. Four near-infrared LED fill light arrays (850nm / 15,000lm), four light strips evenly distributed around the circumference of each camera housing, the light strip power supply and the camera share the DC core of the composite cable. Four inertial measurement units (IMUs): a triaxial gyroscope and a triaxial accelerometer, connected to the underwater control box via the CAN-FD bus. Two hydroacoustic positioning transponders: mounted on diagonal camera mounts, used for distance measurement with the buoy base station for 3D positioning. The underwater control box (IP68 aluminum-titanium alloy housing) integrates an 8-port Gigabit PoE switch (4 cameras + 1 uplink), a CAN-FD hub, a PTP slave clock module, and a 48V to 24V DC-DC power supply. The switch's uplink port is connected to the buoy's communications cabin via an optical fiber composite cable (single-mode fiber + AC48V). The buoy's communications cabin houses a PTP master clock (GPS timing, nanosecond-level stability) and a gigabit fiber router. The fiber optic cable uploaded from the underwater sub-control box is converted to RJ-45 and sent to the shore-based centralized control room via a 200m roaming Ethernet cable. The AC48V is transmitted downstream from the shore-based UPS via the same composite cable. In the shore-based edge computing and blasting control area, the GPU edge server: NVIDIA Jetson AGX Orin (32GB GPU) receives four RTSP streams via a gigabit switch. The blasting PLC: Siemens S7-1500F with high-speed I / O, connected to the same switch via Profinet; real-time communication between the Jetson and the charge quantity and detonation delay is carried out via gRPC. The electric detonator trigger module: 40 independent optocoupler-isolated outputs, driven by the PLC's 24VDC pulse; each output is connected in series with a 3A fast fuse. A 6kVA online UPS supplies power to the PLC, Jetson, switches, and buoys, and monitors remaining battery power. Auxiliary sensing and safety interfaces include a hydrophone array of four: 15 meters around the explosion zone, connected to the PLC via an RS-485 bus for shock wave monitoring. Diver gate sensors: Fiber Bragg grating switches detect the evacuation of construction personnel; signals are fed into the PLC safety chain. Connection topology and signal path: camera → (composite cable) → underwater sub-control box PoE switch. Fiber uplink from the underwater sub-control box → buoy communication cabin → Gigabit cable → shore-based switch. Jetson receives traffic from the switch, performs defogging, segmentation, risk assessment, and DRL models; the results are transmitted to the PLC via gRPC. When the PLC meets the following logic: "diver evacuated + detonator circuit self-test OK + |I_now – I_prev| > 0.05 or reaches a 1s period," the trigger module injects pulses into the 40 detonators according to the DRL output table. After the blast is completed, the hydrophone feeds back the shock wave peak value to the PLC; Jetson writes the event tag and uploads the entire video and parameters to the NAS for blockchain storage.This hardware architecture uses "optical composite cable + PTP synchronization + Ethernet gRPC" to achieve high-bandwidth, low-latency, and strongly synchronized multi-machine video backhaul and closed-loop blasting control. At the same time, it digests PoE power supply and CAN / IMU data at the underwater end, and centrally completes AI reasoning and safety interlocking at the shore-based end, meeting the real-time and reliability requirements of blasting operations in multiple scenarios such as port areas and deep waters.

[0073] Synchronous video stream acquisition module, deploying N underwater cameras and corresponding near-infrared LED fill light arrays in the blasting area to acquire synchronous video streams;

[0074] A defogging and first video stream frame vector generation module performs dark channel prior transmittance estimation on each frame of the synchronized video stream, refines the transmission map using guided filtering within a 5×5 window, reconstructs a defogging frame sequence, and expands each defogging frame row-wise into an RGB three-channel vector to obtain the first video stream frame vector.

[0075] The second video stream frame vector acquisition module processes the dehazed video frames using a first convolutional neural network with a temporal attention mechanism to obtain annotated video frame images with rock cracks, bubble plumes, and turbid water areas marked. The pixel grayscale values ​​of the annotated video frame images are expanded row by row into a single-channel vector to obtain the second video stream frame vector.

[0076] The risk index I calculation module obtains the total crack length, average plume density, and total turbidity area of ​​the video stream based on the crack length, plume density, and turbidity area of ​​each frame of the video stream, and calculates the risk index I;

[0077] An explosive charge and detonation sequence generation module, which fuses the first video stream frame vector, the second video stream frame vector, and the risk index I to generate a joint vector, and inputs the joint vector into a second convolutional neural network model to generate an explosive charge and a detonation sequence;

[0078] The blasting execution module controls the electric detonator network to execute blasting according to the explosive charge and detonation sequence when the difference between the latest risk index I and the risk index of the previous decision cycle exceeds a preset threshold.

[0079] The first convolutional neural network (FrameSemanticSegmentationCNN) takes as input a single-frame dehazed RGB image of size 1920×1080×3. It outputs a single-channel annotation map of the same size (0 = background, 1 = crack, 2 = bubble plume, 3 = turbid area).

[0080] Core architecture: Encoder: 4 layers of DeformableConv (3×3) + Batch-Norm + ReLU, with 64 / 128 / 256 / 512 channels, respectively; each layer is followed by 2×2 downsampling. Deformable convolution focuses on the elongated morphology of cracks or plumes at learned offsets. Temporal Attention Module (TC-Former): Stacks features of the same scale using a 5-frame sliding window. Four multi-head self-attention blocks (key / value dimension 512, number of heads = 4) calculate weights in the temporal dimension to capture the dynamic correlation between crack propagation and plume diffusion. Decoder: Top-down Feature Pyramid (FPN); upsampling and skipping at each level, culminating in a 1×1 convolution mapping to 4 class channels and softmax. Deformable convolution is used to extract multi-scale shape features from a single frame. TC-Former then focuses on regions that have newly appeared or expanded from the previous frame to the current one, significantly improving the detection rate of crack filaments and sparse plumes. Finally, FPN is used to gradually restore the resolution to produce a pixel-level segmentation map. The network is trained using a cross-entropy and boundary intersection over union loss on manually annotated data.

[0081] The second convolutional neural network (State-to-Action1DCNN-Agent) has an input of 8294401-dimensional joint vector (1024-dimensional dehazing RGB encoding + 256-dimensional labeled grayscale encoding + 1-dimensional risk index). Output: action vector m i (40 blastholes with explosive charge, 0–0.5 kg); delay vector Δt i(40 adjacent detonation delays, 0–100ms); value V(s) (evaluates the current state value for updating the PPO algorithm). Core structure: 1D Conv feature extraction: Conv1d128×k5 → ReLU → Conv1d256×k5 → ReLU, stride 1, used to capture local coupling relationships between different dimensions (for example, the correspondence between "new crack region" and "increased risk index"). Temporal-dimensional self-attention: Multi-head self-attention (hidden dimension 512, head number 4) globally models the entire vector, automatically focusing on the dimension combination most relevant to the charge. Dual-branch output: Actor (FC512→256→80) is mapped to charge and delay via sigmoids; Critic (FC512→256→1) outputs the state value for PPO to calculate the advantage function. This network is essentially a lightweight joint policy / value network (Actor-Critic). It considers the environment vector from the visual encoder and the real-time risk index as "states." It first uses convolution to identify local patterns (e.g., "crack area 0.6 + risk 0.7"), then uses self-attention to integrate global features. Finally, it outputs the charge and detonation delay that satisfy the "maximum efficiency, minimum risk" reward function. During the training phase, PPO is used to interactively learn with the environment in a high-fidelity simulator. During deployment, only forward inference is performed, resulting in a latency of approximately 4ms.

[0082] The activation function f1(x) used in the first convolutional neural network model is expressed as follows:

[0083] f1(x)=Tanh((1+τ*α t )x);

[0084] Among them, α t is the average global attention value output by the temporal attention module for this frame, with a value range of (0,1). It reflects the "temporal importance" of the current frame to the 5-frame sliding window. x represents the input value of the current neuron in the first convolutional neural network, τ is a learnable scalar with an initial value of 0.5 and a value range of [0,2]. t Afterwards, the same convolution kernel has a larger effective slope in keyframes, amplifying the gradient to easily preserve subtle cracks or sparse plume features; while the slope shrinks in non-keyframes to actively suppress redundant noise. Experiments show that the mean Intersection Over Union (MIoU) increases from 92.3% to 93.7%, and the false positive rate decreases by 14%.

[0085] The activation function f2(x) used by the second convolutional neural network is expressed as follows:

[0086] Where z represents the input value of the current neuron in the second convolutional neural network, e is the base of the natural logarithm, r is a learnable positive number (in some embodiments, r = 0.6, r∈[0.3, 1]), and I is the risk index. The second convolutional neural network model incorporates the risk index into the activation slope, directly modulating the action distribution with the environmental safety level. By directly incorporating the risk index I into the Sigmoid slope, a wide threshold is maintained in low-risk scenarios (I≈0). At high risk (I→1), the curve steepens, and the output is more likely to saturate: the explosive charge automatically approaches the lower limit, and the delay interval is increased. Measured results show a 12.8% reduction in explosives consumption per unit and a 26% reduction in the peak radius of the turbidity plume for the same blast area.

[0087] In some embodiments, the defogging and first video stream frame vector generation module also includes: for each frame image in the synchronized video stream, for the brightness of different color channels of each pixel in the frame within a preset local neighborhood, pixels with brightness values ​​less than a set threshold are selected to generate a dark channel image; an initial transmittance image is generated according to a preset proportional coefficient based on the dark channel image and the global atmospheric light brightness of the scene; using the original frame image as a guide image, guided filtering refinement processing is performed on the initial transmittance image within a 5×5 pixel window to obtain a smooth transmittance image; the smooth transmittance image and the original frame image are fused pixel by pixel according to the set transmittance lower limit threshold to generate a defogging frame image; all defogging frame images are organized into a defogging frame sequence in chronological order and output.

[0088] In some embodiments, the first video stream frame vector is a one-dimensional data sequence, and the red, green, and blue channel values ​​of each pixel of the defogging frame are read in sequence from left to right in a row scanning order, and are continuously appended in the order of R, G, and B. The length of the resulting sequence is equal to three times the number of pixels in the original frame; the second video stream frame vector is a one-dimensional data sequence, and the single-channel grayscale values ​​of the corresponding pixels of the marked frame are read in sequence from left to right and are continuously appended. The length of the resulting sequence is equal to the number of pixels in the original frame; the detonation timing is a delay sequence between adjacent blast holes, and the master control triggering moment is recorded as 0.

[0089] In some embodiments, the fusion of the first video frame vector, the second video frame vector and the risk index I to generate a joint vector includes: directly taking the first video stream frame vector as the first segment of the joint vector in a predetermined order; then sequentially appending the second video stream frame vector thereafter; and finally appending the risk index I containing only a single scalar element; thereby obtaining a continuous one-dimensional data sequence whose first segment is defogging frame data, the second segment is labeled frame data, and the end is the risk index, which is the joint vector.

[0090] In some embodiments, the calculation of the risk index I includes: comparing the total length of the cracks in the video stream, the average density of the plume, and the total area of ​​the turbidity zone with their respective preset maximum allowable values ​​to obtain three normalized evaluation values ​​between zero and one; accumulating the normalized crack length with a weight of 40%, accumulating the normalized plume density with a weight of 35%, and accumulating the normalized turbidity zone area with a weight of 25%, and summing the three weighted results to obtain the risk index I.

[0091] The present invention provides an underwater blasting construction method and system based on real-time video image processing, which can achieve the following beneficial technical effects:

[0092] 1. The present invention significantly improves blasting visualization accuracy. Using an improved dark channel-guided filtering dehazing algorithm, the average peak signal-to-noise ratio (PSNR) increased by 9dB in a sea trial environment with a viewing distance of 1.5m and a suspended particle concentration >50mgL⁻¹. The Transformer-CNN network achieved a pixel-level segmentation mIoU of 92.8% for cracks, bubble plumes, and turbidity, an improvement of approximately 11 percentage points over the traditional U-Net. By introducing a risk index (I) and two convolutional neural network models, the present invention reduced explosive consumption per unit by an average of 13.4% while maintaining the target particle size in the same blasting area, and the detonation delay error converged from ±5ms to ±1ms. When a sudden increase in crack propagation speed (>20% / s) is detected online, the system can recalculate parameters and issue a new detonation sequence within 0.3s, achieving millisecond-level closed-loop regulation. Safety and environmental performance have been improved. After adjusting the charge based on real-time plume density feedback, the maximum turbidity plume diffusion radius has been reduced by 27%, and the peak water turbidity has been reduced by 31%, effectively reducing the impact on nearby coral reefs and aquaculture areas.

[0093] 2. The number of diver visits per blasting area has been reduced from an average of four to just one, used only for initial hole placement confirmation. During operations, the distance between operator and blaster remains >300m, significantly reducing personnel safety risks. Construction efficiency and cost advantages: Comparative tests in multiple scenarios have shown that this solution shortens the average construction period for a single blasting area by 18%–22%, saving over 15% in overall costs (including explosives, labor, and downtime). The video stream and control command chain are fully documented, providing a reliable basis for quality inspection, insurance claims, and environmental assessments, saving on post-audit and work resumption costs.

[0094] The above describes in detail an underwater blasting construction method and system based on real-time video image processing. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only intended to help understand the core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas and methods of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. An underwater blasting construction method based on real-time video image processing, characterized in that: Including steps: S1: Deploy N underwater cameras and corresponding near-infrared LED fill light arrays in the blasting area to obtain synchronized video streams; S2: performing dark channel prior transmittance estimation on each frame of the synchronized video stream, and refining the transmission map using guided filtering under a 5×5 window to reconstruct a defogging frame sequence. Each defogging frame is expanded row by row into an RGB three-channel vector to obtain a first video stream frame vector; S3: The dehazed video frames are processed by the first convolutional neural network with a temporal attention mechanism to obtain annotated video frame images with rock cracks, bubble plumes, and turbid water areas marked. The pixel grayscale values ​​of the annotated video frame images are expanded row by row into a single-channel vector to obtain the second video stream frame vector. S4: Based on the crack length, plume density and turbidity area of ​​each frame of the video stream, the total crack length, average plume density and total turbidity area of ​​the video stream are obtained, and the risk index I is calculated; S5: fusing the first video stream frame vector, the second video stream frame vector, and the risk index I to generate a joint vector, and inputting the joint vector into a second convolutional neural network model to generate an explosive charge and a detonation sequence; S6: When the difference between the latest risk index I and the risk index of the previous decision cycle exceeds a preset threshold, the electric detonator network is controlled to execute blasting according to the explosive charge and detonation sequence.

2. The underwater blasting construction method based on real-time video image processing according to claim 1, characterized in that: Step S2 also includes: for each frame image in the synchronized video stream, for the brightness of different color channels of each pixel in the frame within a preset local neighborhood, selecting pixels whose brightness values ​​are less than a set threshold to generate a dark channel image; generating an initial transmittance image according to a preset proportional coefficient based on the dark channel image and the global atmospheric brightness of the scene; using the original frame image as a guide image, performing guided filtering and refinement processing on the initial transmittance image within a 5×5 pixel window to obtain a smooth transmittance image; fusing the smooth transmittance image with the original frame image pixel by pixel according to the set transmittance lower limit threshold to generate a defogging frame image; and organizing all the defogging frame images into a defogging frame sequence in chronological order and outputting them.

3. The underwater blasting construction method based on real-time video image processing according to claim 1, characterized in that: The first video stream frame vector is a one-dimensional data sequence. The red, green, and blue channel values ​​of each pixel in the defogging frame are read from left to right in row scanning order, and are continuously appended in the order of R, G, and B. The length of the resulting sequence is equal to three times the number of pixels in the original frame. The second video stream frame vector is a one-dimensional data sequence. The single-channel grayscale values ​​of the corresponding pixels in the labeled frame are read from left to right and continuously appended. The length of the resulting sequence is equal to the number of pixels in the original frame. The detonation timing is a delay sequence between adjacent blast holes, and the master control trigger time is recorded as 0.

4. The underwater blasting construction method based on real-time video image processing according to claim 1, characterized in that: The method of fusing the first video frame vector, the second video frame vector and the risk index I to generate a joint vector includes: directly taking the first video stream frame vector as the first segment of the joint vector in a predetermined order; then sequentially appending the second video stream frame vector thereafter; and finally appending the risk index I containing only a single scalar element; thereby obtaining a continuous one-dimensional data sequence whose first segment is defogging frame data, the second segment is labeled frame data, and the end is the risk index, which is the joint vector.

5. The underwater blasting construction method based on real-time video image processing according to claim 1, characterized in that: The calculation of the risk index I includes: comparing the total length of the video stream crack, the average density of the plume, and the total area of ​​the turbidity zone with their respective preset maximum allowable values ​​to obtain three normalized evaluation values ​​between zero and one; accumulating the normalized crack length with a weight of 40%, accumulating the normalized plume density with a weight of 35%, and accumulating the normalized turbidity zone area with a weight of 25%, and summing the three weighted results to obtain the risk index I.

6. An underwater blasting construction system based on real-time video image processing, characterized in that: include: Synchronous video stream acquisition module, deploying N underwater cameras and corresponding near-infrared LED fill light arrays in the blasting area to acquire synchronous video streams; A defogging and first video stream frame vector generation module performs dark channel prior transmittance estimation on each frame of the synchronized video stream, refines the transmission map using guided filtering within a 5×5 window, reconstructs a defogging frame sequence, and expands each defogging frame row-wise into an RGB three-channel vector to obtain the first video stream frame vector. The second video stream frame vector acquisition module processes the dehazed video frames using a first convolutional neural network with a temporal attention mechanism to obtain annotated video frame images with rock cracks, bubble plumes, and turbid water areas marked. The pixel grayscale values ​​of the annotated video frame images are expanded row by row into a single-channel vector to obtain the second video stream frame vector. The risk index I calculation module obtains the total crack length, average plume density, and total turbidity area of ​​the video stream based on the crack length, plume density, and turbidity area of ​​each frame of the video stream, and calculates the risk index I; An explosive charge and detonation sequence generation module, which fuses the first video stream frame vector, the second video stream frame vector, and the risk index I to generate a joint vector, and inputs the joint vector into a second convolutional neural network model to generate an explosive charge and a detonation sequence; The blasting execution module controls the electric detonator network to execute blasting according to the explosive charge and detonation sequence when the difference between the latest risk index I and the risk index of the previous decision cycle exceeds a preset threshold.

7. The underwater blasting construction system based on real-time video image processing according to claim 6, characterized in that: The defogging and first video stream frame vector generation module also includes: for each frame image in the synchronized video stream, for the brightness of different color channels of each pixel in the frame within a preset local neighborhood, pixels with brightness values ​​less than a set threshold are selected to generate a dark channel image; an initial transmittance image is generated according to a preset proportional coefficient based on the dark channel image and the global atmospheric brightness of the scene; using the original frame image as a guide image, the initial transmittance image is subjected to guided filtering refinement processing within a 5×5 pixel window to obtain a smooth transmittance image; the smooth transmittance image and the original frame image are fused pixel by pixel according to the set transmittance lower limit threshold to generate a defogging frame image; all defogging frame images are organized into a defogging frame sequence in chronological order and output.

8. The underwater blasting construction system based on real-time video image processing according to claim 6, characterized in that: The first video stream frame vector is a one-dimensional data sequence. The red, green, and blue channel values ​​of each pixel in the defogging frame are read from left to right in row scanning order, and are continuously appended in the order of R, G, and B. The length of the resulting sequence is equal to three times the number of pixels in the original frame. The second video stream frame vector is a one-dimensional data sequence. The single-channel grayscale values ​​of the corresponding pixels in the labeled frame are read from left to right and continuously appended. The length of the resulting sequence is equal to the number of pixels in the original frame. The detonation timing is a delay sequence between adjacent blast holes, and the master control trigger time is recorded as 0.

9. The underwater blasting construction system based on real-time video image processing according to claim 6, characterized in that: The method of fusing the first video frame vector, the second video frame vector and the risk index I to generate a joint vector includes: directly taking the first video stream frame vector as the first segment of the joint vector in a predetermined order; then sequentially appending the second video stream frame vector thereafter; and finally appending the risk index I containing only a single scalar element; thereby obtaining a continuous one-dimensional data sequence whose first segment is defogging frame data, the second segment is labeled frame data, and the end is the risk index, which is the joint vector.

10. The underwater blasting construction system based on real-time video image processing according to claim 6, characterized in that: The calculation of the risk index I includes: comparing the total length of the video stream crack, the average density of the plume, and the total area of ​​the turbidity zone with their respective preset maximum allowable values ​​to obtain three normalized evaluation values ​​between zero and one; accumulating the normalized crack length with a weight of 40%, accumulating the normalized plume density with a weight of 35%, and accumulating the normalized turbidity zone area with a weight of 25%, and summing the three weighted results to obtain the risk index I.

Images