A method and system for underwater blasting construction based on real-time video image processing

Abstract

This invention provides an underwater blasting construction method and system based on real-time video image processing. The method includes deploying multiple underwater cameras and a near-infrared supplementary lighting array to acquire synchronous video. After improved dark channel-guided filtering for defogging, cracks, bubble plumes, and turbid areas are segmented at the pixel level. A risk index I is calculated based on crack length, plume density, and turbidity area. This risk index is then fed into a second convolutional neural network model, which outputs the explosive charge and detonation delay for each blast hole. When I changes beyond a threshold or reaches a preset period, an electric detonator network is triggered to execute the blast. The system consists of an underwater imaging subsystem, an underwater control box, a shore-based GPU edge server, and a blasting PLC, achieving integrated environmental perception, intelligent decision-making, and closed-loop detonation. This invention improves segmentation accuracy and real-time decision-making, reduces explosive consumption and turbidity plume, and enhances construction safety and environmental friendliness.

Description

Technical Field

[0001] This invention relates to the field of blasting construction technology, specifically to an underwater blasting construction method and system based on real-time video image processing. Background Technology

[0002] Underwater blasting is a common method for rock breaking and obstacle clearing, widely used in engineering scenarios such as channel dredging, submarine pipeline laying, reservoir hazard mitigation, and near-shore port expansion. Traditional construction processes typically rely on blasting designers calculating charge quantity, borehole spacing, and detonation timing offline based on preliminary geological surveys, diver visual inspections, and limited sonar or pressure sensor data. During construction, divers or remotely operated vehicles then verify the borehole depths on-site and complete the detonation connection. This approach suffers from the following technical drawbacks: Delayed information acquisition: Insufficient underwater lighting and turbid water conditions mean sonar can only roughly reflect rock surface undulations, and divers' visibility 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. Lack of adaptability in charge quantity and timing: Blasting parameters are set once on land and cannot be adjusted in real time according to the rapidly changing rock integrity or water flow velocity at the construction site, often resulting in wasted explosives or suboptimal blasting effects. High safety and environmental risks: Divers need to dive multiple times for retesting, and are exposed to high pressure, low temperature and sediment-laden water for extended periods; non-real-time control also makes it difficult to suppress the instantaneous turbid plume and underwater shock wave from the blast, which can easily damage aquatic ecosystems or nearby structures. Difficult data traceability: Traditional methods lack continuous video or high-resolution sensor records, making it difficult to trace responsibility based on fragmented data should quality disputes or environmental complaints arise later.

[0003] In recent years, rapid progress has been made in waterproof high-frame-rate cameras, laser-assisted lighting, and deep learning vision algorithms, providing new perception methods for underwater operations. Existing research has utilized underwater cameras to perform 3D reconstruction of rock walls before blasting, or used convolutional networks to segment the particle size distribution of rock fragments after blasting to assess the effect; some teams have also attempted to use real-time image analysis to monitor bubble plumes and estimate the blast center location. However, most of these studies are limited to offline evaluation: images are processed in batches by shore-based workstations after the operation is completed, failing to provide feedback during blasting and unable to form a closed loop with the blasting control system. Furthermore, current published literature typically focuses only 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 mostly use PLCs to trigger delayed detonators at fixed times, lacking interfaces for linkage with visual recognition results. In summary, a holistic solution of "real-time multi-modal visual perception + risk quantification + adaptive detonation control" has not yet been formed, hindering the upgrading of underwater blasting towards digitalization, precision, and safety.

[0004] To overcome the aforementioned 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. This system should be able to acquire clear images in highly turbid and low-light environments, and perform pixel-level identification of cracks, plumes, and turbid areas; assess blasting environmental risks in real time through continuous frame time-series information and sliding window statistics; automatically adjust explosive charge and detonation delay when risk indicators change significantly or the decision cycle is reached, achieving closed-loop control; and effectively improve construction accuracy by incorporating the risk index into the calculation of explosive charge and blasting timing. Simultaneously, the entire process of video and control commands should be documented to enhance safety traceability. This invention is based on these needs and proposes a four-level system of "video perception—risk assessment—deep reinforcement learning—detonation execution" to achieve intelligent, adaptive, and traceable 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 problems in existing technologies, this invention provides an underwater blasting construction method and system based on real-time video image processing. The method deploys multiple underwater cameras and a near-infrared supplementary lighting array to acquire synchronous video. After defogging using an improved dark channel-guided filtering method, a first convolutional neural network incorporating a temporal attention mechanism is used to perform pixel-level segmentation of cracks, bubble plumes, and turbidity areas. A risk index I is calculated based on crack length, plume density, and turbidity area. The defogging RGB vector, labeled grayscale vector, and I are concatenated into a joint vector and input into a second convolutional neural network model containing a risk modulation activation function, outputting the explosive charge and detonation delay for each blast hole. When the change in I exceeds a threshold or reaches a preset period, the electric detonator network is triggered to execute the blast. The system consists of an underwater imaging subsystem, an underwater control box, a shore-based GPU edge server, and a blasting PLC, achieving integrated environmental perception, intelligent decision-making, and closed-loop detonation. This invention improves segmentation accuracy and real-time decision-making, reduces explosive consumption and turbidity plume, and enhances construction safety and environmental friendliness.

[0006] This application provides an underwater blasting construction method based on real-time video image processing, including the following steps:

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

[0008] S2: Perform dark channel prior transmittance estimation on each frame of the synchronous video stream, and refine the transmittance map using guided filtering in a 5×5 window to reconstruct the dehazed frame sequence. Expand each dehazed frame into an RGB three-channel vector by row to obtain the first video stream frame vector.

[0009] S3: The video frames after dehazing are processed by the first convolutional neural network containing the temporal attention mechanism to obtain labeled video frame images of rock cracks, bubble plumes and water turbidity areas. The pixel gray values ​​of the labeled video frame images are expanded into single-channel vectors by row 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, obtain the total crack length, average plume density, and total turbidity area of ​​the video stream, and calculate the risk index I.

[0011] S5: The first video stream frame vector, the second video stream frame vector and the risk index I are fused to generate a joint vector. The joint vector is then input into the second convolutional neural network model to generate the explosive charge and detonation sequence.

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

[0013] Preferably, step S2 further includes: for each frame of the synchronized video stream, for each pixel in the frame with different color channel brightness in a preset local neighborhood, selecting pixels with brightness values ​​less than a set threshold 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 ratio coefficient; using the original frame image as a guide image, performing guided filtering and thinning 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 a set lower limit threshold of transmittance to generate a dehazed frame image; and organizing all dehazed frame images into a dehazed frame sequence and outputting it in chronological order.

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

[0015] Preferably, the step 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; and finally appending the risk index I, which contains only a single scalar element; thereby obtaining a continuous one-dimensional data sequence with the first segment being dehazed frame data, the second segment being labeled frame data, and the last segment being 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; summing the normalized crack length with a weight of 40%, the normalized plume density with a weight of 35%, and the normalized turbidity zone area with a weight of 25%, and summing the three weighted results to obtain the risk index I.

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

[0018] The synchronous video stream acquisition module deploys N underwater cameras and corresponding near-infrared LED supplementary light arrays in the blasting area to acquire synchronous video streams.

[0019] The dehazing and first video stream frame vector generation module performs dark channel prior transmittance estimation on each frame of the synchronous video stream, and uses guided filtering to refine the transmittance map in a 5×5 window to reconstruct the dehazing frame sequence. Each dehazing frame is expanded into an RGB three-channel vector by row to obtain the first video stream frame vector.

[0020] The second video stream frame vector acquisition module processes the dehazed video frames through a first convolutional neural network containing a temporal attention mechanism to obtain labeled video frame images of rock cracks, bubble plumes, and turbid water areas. The pixel gray values ​​of the labeled video frame images are expanded into single-channel vectors by rows 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 area of ​​the turbidity zone of the video stream based on the crack length, plume density, and turbidity zone area of ​​each frame of the video stream, and calculates the Risk Index I.

[0022] The explosive charge and detonation timing generation module 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 the second convolutional neural network model to generate the explosive charge and detonation timing.

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

[0024] Preferably, the dehazing and first video stream frame vector generation module further includes: for each frame image in the synchronous video stream, for the different color channel brightness of each pixel in the frame within a preset local neighborhood, selecting pixels with brightness values ​​less than a set threshold 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 ratio coefficient; using the original frame image as a guide image, performing guided filtering and thinning 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 a set lower limit threshold of transmittance to generate a dehazed frame image; and organizing all dehazed frame images into a dehazed frame sequence in chronological order and outputting it.

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

[0026] Preferably, the step 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; and finally appending the risk index I, which contains only a single scalar element; thereby obtaining a continuous one-dimensional data sequence with the first segment being dehazed frame data, the second segment being labeled frame data, and the last segment being 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; summing the normalized crack length with a weight of 40%, the normalized plume density with a weight of 35%, and the normalized turbidity zone area with a weight of 25%, and summing the three weighted results to obtain the risk index I.

[0028] This 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. This invention organizes the process into a sequential sequence of "dehazing-semantic segmentation-risk quantification-intelligent decision-making," which logically decouples the first convolutional neural network (responsible for pixel-level segmentation) and the second convolutional neural network (responsible for parameter decision-making) while maintaining data coherence. The first network only receives the original dehazed image and outputs four types of labeled images. The system concatenates the dehazed RGB one-dimensional vector, the labeled grayscale vector, and the real-time risk index into a fixed-length joint vector as the input to the second network. This ensures the integrity of the state description and significantly compresses the feature dimension (from 18.6MB to 5MB), reducing inference latency by approximately 32%. The newly introduced risk index I integrates crack length, plume density, and turbidity area with weights of 40% / 35% / 25%. The quantification result is continuously differentiable from 0 to 1, providing a unified and safe metric 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 cycle, avoiding high-frequency jitter of "one decision per frame"; The PLC first performs a safety link self-check and a 5-second audible and visual countdown, and then triggers the electric detonator according to the delay sequence, achieving millisecond-level closed-loop and minute-level safety confirmation dual-layer protection; Compared with the fixed charge scheme, the average blasting unit consumption is reduced by 12.8%, the peak radius of the turbid plume is reduced by 26%, and the construction cycle is shortened by 20%.

[0031] 3. The present invention improves the scene 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, which improves the mIoU of crack and plume detection by 1.4pp and reduces the false detection rate by 14%. The second convolutional neural network model uses R-Sigmoid in the last layer of the Actor, which directly writes the risk index into the activation slope. The output saturates faster in high-risk scenarios, and the explosive charge automatically converges to the safe lower limit. Neither activation function adds any weights, with only one multiplication overhead, keeping the end-to-end latency unchanged. At the same time, learnable parameters τ and r are provided, enabling the network to automatically calibrate under different sea conditions. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

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

[0034] Figure 2This is a schematic diagram of an underwater blasting construction system based on real-time video image processing according to the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Example 1:

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

[0038] S1: Deploy N underwater cameras and corresponding near-infrared LED supplementary lighting arrays in the blasting area to obtain synchronized video streams. In shallow-water port dredging scenarios, four 4K / 60fps waterproof cameras (approximately 120° field of view, 20m pressure resistance) and corresponding 850nm near-infrared LED supplementary lighting arrays can be selected. During deployment, tripods are placed along the four cardinal directions (north, south, east, and west) on a horizontal plane with a radius of 6m, with the top of the tripods approximately 0.5m above the seabed. Each camera is tilted downwards at 15° to ensure a field of view overlap of no less than 70%. Four sets of LED light strips are installed circumferentially around the camera housing, with each strip having a 120° opening angle, providing an illumination of no less than 40lx at 3m. The cameras and light array share a single DC24V / Gigabit Ethernet composite cable connected to a shore-based PoE switch, with the length of a single cable controlled within 80m. Synchronization uses IEEE-1588PTP, with the shore server as the master clock and the cameras as slave clocks, maintaining a frame timestamp error within 3ms. The entire deployment process was completed by two technicians working with the ROV, taking approximately 45 minutes. During the operation, the cable tension was remotely checked every 2 hours. If a sudden peak in turbidity occurred, the LED power could be temporarily increased by 20% to maintain image quality.

[0039] In a deep-sea subsea pipeline trench blasting scenario, six 1080p / 50fps low-light cameras (90° field of view, 50m pressure resistance) and a 940nm focused LED array can be used. The cameras are suspended by steel cables on equally angled ring supports with a ring radius of 10m and the bottom of the supports 0.8m above the seabed. The camera depression angle is 20°, the field of view overlap is maintained at over 60%, and three cameras are weighted 3kg on top to reduce sway. LED supplementary lights are installed on two 30° outward-polarizing lamp arms on the left and right sides to reduce forward scattering, providing an illumination of no less than 25lx at 5m. Power and data are transmitted to the underwater control box via AC48V / single-mode fiber optic composite cable, then routed back to shore via gigabit Ethernet through the buoy; the PTP master clock is deployed on the buoy, with a frame synchronization error of no more than 5ms. Deployment is completed by three technicians and one ROV operator, taking approximately 90 minutes. The camera housing temperature is monitored every hour during operation; the steel cable is automatically tightened to stabilize the viewing angle when the ocean current speed exceeds 0.7 m·s⁻¹. Both implementations use IP68 quick-connect plugs for easy and rapid equipment replacement, and 2A fuses are configured on the power supply side to ensure system redundancy and safety.

[0040] S2: Perform dark channel prior transmittance estimation frame by frame on the synchronized video stream, and refine the transmittance map using guided filtering in a 5×5 window to reconstruct the dehazed frame sequence. Expand each dehazed frame row by row into an RGB three-channel vector to obtain the first video stream frame vector. In this embodiment, the acquired 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, and the core steps are as follows:

[0041] 1. Frame reading and three-channel splitting: The OpenCV video decoding interface cv2.VideoCapture is used for frame-by-frame reading. Each frame read is immediately converted to RGB color space and cached in a NumPy array img_rgb. The resolution remains 1920×1080, and the data type is uint8.

[0042] 2. Dark Channel Image Generation: To enhance processing speed, only the local minimum luminance value of the RGB three channels within a 15×15 pixel neighborhood is calculated, and a single-channel dark channel image (dark_channel) is output. In this embodiment, cv2.erode is used in conjunction with a custom minimum pooling kernel to complete the calculation, which takes approximately 8ms per frame.

[0043] 3. Initial Transmittance Estimation and Atmospheric Light Sampling: In the dark channel image, the original color pixels corresponding to the top 0.1% of the pixels in brightness are selected, and their average RGB value is calculated as the atmospheric light A_rgb. Subsequently, the dark channel image is reverse-mapped using a fixed coefficient of 0.95 to obtain a coarse transmittance image t_coarse (still a uint8 grayscale image). This step takes 4ms / frame.

[0044] 4. For 5×5 window guided filtering refinement, the `guided_filter` function from the `Guided-Filter-Py` library is called. The original RGB image `img_rgb` is used as the guide image, and `t_coarse` is used 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, a smooth transmittance image `t_refined` is obtained. The computation time is 11ms / frame, and the GPU usage is approximately 15%.

[0045] 5. Dehazed frame reconstruction: A lower limit of 0.1 was set for the excessively small transmittance value in t_refined to avoid forcibly amplifying noise; then, the haze-free frame img_clear was reconstructed at the pixel level according to the dark channel prior recovery model. Reconstruction time was 6ms / frame. Actual comparison showed that the average brightness of the dehazed frame increased by 28%, and the color cast was significantly reduced.

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

[0047] 7. Real-time performance and caching strategy: Jetson's end-to-end processing time is approximately 29–30ms / frame, stably supporting 30fps real-time dehazing. To ensure smoothness, the system employs dual caching: the main thread continuously writes the next frame, while worker threads use CUDA streams to compute the current frame in parallel, ensuring peak video memory usage is <1GB.

[0048] 8. Abnormalities and fault tolerance: If a frame fails to be read or zero transmittance is detected during defogging, the system will automatically skip the frame and record a timestamp in the log; if five consecutive frames are skipped, an alarm will be triggered, the LED supplementary light power will be increased by 20%, and the camera will be reset.

[0049] S3: The dehazing video frames are processed using a first convolutional neural network incorporating a temporal attention mechanism to obtain labeled video frame images of rock cracks, bubble plumes, and turbid water areas. The pixel grayscale values ​​of the labeled video frame images are expanded row by row into single-channel vectors to obtain the second video stream frame vector. The first convolutional neural network for the temporal attention mechanism adopts a two-stage structure of "Deform-Conv+TC-Former". The first stage is a 4-layer deformable convolutional encoder that extracts multi-scale features of 64×128×128, 128×64×64, 256×32×32, and 512×16×16 frame by frame. The second stage uses a 5-frame sliding window to feed the features of 5 consecutive frames of the same scale into a temporal cross-frame Transformer (TC-Former), and uses multi-head attention to calculate key-value weights in the time dimension to capture the dynamic correlation between crack propagation and plume diffusion. The network outputs four types of pixel probability maps, which are then processed pixel-by-pixel argmax to obtain a single-channel labeled image of 1920×1080 resolution. Grayscale encoding: 0 represents the 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 composite images, all manually labeled with the four mask types. The AdamW optimizer is used with an initial learning rate of 3×10⁻⁶. -4 With a batch size of 8 (5 frames / sample), training on a 2×NVIDIA A40 for 60 epochs achieved a mIoU of 92.3%. During deployment, 30% pruning was performed, and quantization was implemented to INT8. Single-frame inference time was 18ms, and with a 5-frame sliding window, the overall latency was approximately 90ms, sufficient for a 10Hz decision frequency. The inference-generated annotation map was expanded in row-scanning order: first, the first row of 1920 pixels of grayscale values ​​was read, then rows 2 through 1080 were read sequentially, and 2,073,600 grayscale bytes were written sequentially into a one-dimensional array, forming the second video stream frame vector. This vector, concatenated with the RGB vector of the same frame and the real-time risk index, was then input into the second convolutional neural network model to dynamically adjust the detonation 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 obtained, and the risk index I is calculated. The 4K camera used at the blasting site was calibrated with an underwater laser ruler, and the pixel-to-meter conversion factor p2m≈0.002m / px was obtained. In the single-channel labeled grayscale image output by the network inference, grayscale value 85 represents crack pixels, 170 represents bubble plume pixels, and 255 represents turbidity area pixels. The system processes each frame continuously at 30fps and generates the risk index I in real time according to the following steps: Calculate the crack length frame by frame, perform thinning on the pixel set with grayscale value of 85, and then extract the skeleton using the 8-connected tracking algorithm; multiply the number of skeleton pixels by p2m to obtain the crack length L_f of that frame. Calculate the plume density frame by frame, count the number of pixels with grayscale value of 170 and divide by the total number of pixels in the whole frame to obtain the plume density D_f (value 0–1). The area of ​​the hazy region is calculated frame by frame, and the number of grayscale values ​​of 255 pixels, A_px, is counted and multiplied by p2m. 2 Convert to turbidity area A_f (unit: m) 2 The data is summarized using a sliding window method, summing or averaging over 30 frames (1 second sliding window) to obtain the real-time total crack length L_sum, average plume density D_avg, and total area of ​​the turbid zone A_sum, and updated every second. Normalization is then performed by dividing L_sum, D_avg, and A_sum by preset maximum allowable values: L_max = 6m, D_max = 0.25, A_max = 1.5m. 2 This yields three 0–1 normalized indices: L, D, and A. These are then weighted and summed, with L, D, and A added together 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, I = 0.40 × 0.53 + 0.35 × 0.72 + 0.25 × 0.60 ≈ 0.61. For risk classification and triggering, the system limits I to the range of 0–1: I < 0.30 is considered low risk; 0.30–0.60 is considered medium risk; and I ≥ 0.60 is considered 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 classification 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 the risk indicators at the millisecond level under 30fps conditions, while the decision-making level only outputs a new blasting plan when there is a substantial change in the environment or when 1 second has elapsed. This avoids the unreasonable high-frequency control of "one decision per frame" and ensures that the blasting parameters remain sensitive to environmental fluctuations.

[0051] S5: The first video stream frame vector, the second video stream frame vector, and the risk index I are fused to generate a joint vector. The joint vector is then input into the second convolutional neural network model to generate the explosive charge and detonation timing. The acquisition end triggers a decision every 0.2 seconds (i.e., 6 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 from these 6 frames as follows: Vector compression: The RGB vector is fed into the encoding end 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 encoding end of another 1×1 convolutional autoencoder and compressed into a 256-dimensional vector v2; both encoders are fine-tuned with 50,000 dehazed / annotated images after ImageNet transfer learning, with a PSNR loss of <0.2dB. For temporal aggregation, element-wise averaging is performed on v1 and v2 obtained from 6 frames to obtain frame window average vectors V1 and V2 (dimensions remain 1024 and 256). This preserves the scene trend within 1 second while significantly reducing noise. Joint vector concatenation places V1 in the first segment and V2 in the second segment, finally appending a single scalar risk index I (double-precision floating-point and normalized to 0–1) to obtain a joint vector of length 1281: state = [V1||V2||I]. The joint vector is written into a pre-allocated gRPC message and sent to the DRL inference endpoint. The second convolutional neural network structure is a one-dimensional convolutional-attention hybrid policy network. The first two layers use 1DConv (128 channels, 256 channels, kernel width 5, stride 1) to extract local correlation features. The third layer uses multi-head self-attention (4 heads, hidden dimension 512) to capture cross-dimensional correlations, then flattens and splits into two fully connected branches:

[0052] Actor branch: FC(512→256→80), the first 40 nodes are mapped to 0–0.5kg via Sigmoid as the explosive charge m for the 40 boreholes. i The last 40 nodes are mapped to adjacent detonation delay Δt via Sigmoid × 100ms. i Critic branch: FC (512→256→1) outputs state value V(s). Offline training and online inference: In the offline phase, PPO (clip=0.2) is used to train 3×10 6 The reward for each step is a weighted average of four factors: rock particle size target, explosive consumption, peak turbidity, and vibration acceleration. After convergence, the Actor branch inference latency on NVIDIA Orin is approximately 4ms. Example of parameter output: When the risk index I = 0.48 in the joint vector state, and the scene is classified as medium risk, the Actor outputs:

[0053] • Charge quantity m = [0.28, 0.25, 0.29, ..., 0.22] kg (40 value);

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

[0055] The system writes the detonation table to the PLC in real time. If the difference between the next decision cycle I and the current I is ≤0.05, the table is used and not reissued. Through this embodiment, the joint vector length is fixed, and a small edge GPU can output a complete set of explosive charge and delay sequences for each blast hole in real time at a decision rate of 10Hz. This avoids the redundancy of "one decision per frame" and ensures that a new adaptive blasting scheme can be generated within ≤0.3 seconds when environmental risks change abruptly.

[0056] 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 the blasting according to the explosive charge and detonation sequence. The blasting control system operates on a 1-second decision cycle. Before the operation, the engineer sets the risk difference threshold ΔI_thr to 0.05 and writes a three-state logic control program into the PLC: monitoring—ready—detonation. The following uses the measured logs of four consecutive cycles as an example to illustrate the triggering and execution process.

[0057] In cycle 0 (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 Simultaneously, the baseline risk index I_prev = 0.42 is recorded. The PLC caches the parameters in the "download register," but the system is in a safety check phase and does not immediately detonate.

[0058] In the first cycle (T1 = +1s), the new risk index I_now = 0.45, |I_now – I_prev| = 0.03 < 0.05. The PLC maintains the "ready" state: the charge quantity and delay table remain the same as the previous cycle; the LED warning light is constantly yellow, indicating that the blasting unit is on standby; the audible and visual alarms and the water-based 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: Parameters are reissued: The DRL module generates a new version of m′1…m′ within 0.25s. 40 ,Δt′1…Δt′ 40The PLC overwrites the register and locks the write port; Safety confirmation: The system polls the "diver away from the area" sensor, the resistance of the fiber optic pyrotechnic circuit, and the UPS voltage. After all return a green signal, the status changes to "detonation preparation"; 5-second countdown: During the countdown, the float's audible and visual alarm flashes intensely to warn surrounding vessels. T2+5s (=+7s) Detonation execution: The PLC triggers the main relay, sequentially injecting 3A, 24V current into 40 electric detonators; within the first 40ms, Δt′1…Δt′ is executed. 40 All explosions were completed at intervals. The high voltage was then immediately cut off, and the data marker "Blast#17" was activated.

[0060] In the third cycle (T3 = +8s), the sensor detected debris dust and pressure fluctuations, and I_now temporarily jumped to 0.65. Since the blasting was complete, this cycle transitioned to "safe cooling" mode, recording only data and temporarily not generating a new charge table. After 3 minutes, the system automatically downgraded to "hole reset" state. This example illustrates the practical role of the risk index differential threshold in triggering detonation: only when the environmental risk changes significantly relative to the previous decision cycle (|I_now – I_prev| > 0.05) will the system recalculate and immediately execute a new charge and delay plan; otherwise, the established parameters are maintained to avoid excessively frequent detonation commands, while ensuring that safety checks and audible / visual warnings are complete and traceable.

[0061] In some embodiments, step S2 further includes: for each frame of the synchronous video stream, for each pixel in the frame with different color channel brightness in a preset local neighborhood, selecting pixels with brightness values ​​less than a set threshold 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 ratio coefficient; 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 a set lower limit threshold of transmittance to generate a dehazed frame image; organizing all dehazed frame images into a dehazed frame sequence in chronological order and outputting it. In some embodiments, in a video stream processing of 1920×1080 pixels at 30fps, the system performs dehazing enhancement on each frame according to the following fixed parameters. First, a 15×15 local neighborhood is taken centered on each pixel, and the brightness of the R, G, and B channels of the neighborhood is compared; if the brightness is less than 30 (0–255 scale), the pixel is written into the dark channel image. After all dark channel pixels within a frame are extracted, the system selects the top 0.1% of pixels in brightness from the dark channel image, reads the RGB values ​​of the corresponding positions in the original frame, and averages them to obtain the global atmospheric brightness A≈[218,221,226]. Then, according to a preset scaling factor of 0.95, the entire dark channel image is reverse-mapped to generate an initial transmittance map; typical pixel transmittance falls between 0.2 and 0.9. Using the original frame image as a guide image, a guided filter is applied to the initial transmittance map within a 5×5 window to refine it, resulting in a smooth transmittance map; the regularization coefficient ε is fixed at 1×10⁻³ during the filtering process. To avoid over-brightening of excessively dark areas, the system sets a lower limit threshold of 0.1 for transmittance per pixel, and then fuses it pixel-by-pixel with the original frame according to the classic dark channel recovery formula to generate a dehazed frame. The entire processing takes approximately 29ms / frame on JetsonOrin and can support 30fps in real-time. The processed dehazed frames are written to a circular buffer in chronological order, aggregated into a dehazed frame sequence, and output to the subsequent semantic segmentation network in real-time.

[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 dehazed frame are read sequentially from left to right in a row-scanning order, and then continuously appended in the R, G, B order. The length of the resulting sequence is three times the number of pixels in the original frame. The second video stream frame vector is also a one-dimensional data sequence. The single-channel grayscale values ​​of the corresponding pixels in the labeled frame are read sequentially 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 sequence is a delay sequence between adjacent blast holes, with the master control trigger time recorded as 0. The dehazed frame output by the camera has a resolution of 1920×1080 pixels. The system processes this frame in a row-scanning order, generating two vectors and providing a demonstration of the detonation delay.

[0063] The first video stream frame vector (RGB vector) reads the first 1920 pixels of the first row: first, take the red channel R... 1-1 …R 1-1920 Write sequentially to the buffer; then read the second line of red channel and append it as well; continue until the 1080th line of red channel is written. At this point, the buffer length is 1920 × 1080 = 2073600 bytes; return to the first line and write the green channel G sequentially; then complete the writing of the blue channel B for all lines. The final one-dimensional vector length = 2073600 × 3 = 6220800 bytes (or floating-point value), and the order is always "all R → all G → all B".

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

[0065] Example of detonation sequence; this batch has 40 boreholes. The main control trigger time is denoted as t0 = 0ms. Delay sequence Δt. i (i = 1…40) are output by the reinforcement learning network, for example: 0ms, 12ms, 23ms, 35ms, 49ms, 65ms, 82ms, …, 421ms. The system generates absolute detonation times online by accumulating the values: t1 = 0ms, t2 = 12ms, t3 = 35ms, t4 = 70ms…t 40 =421ms. The PLC loads these times into the detonation control table and triggers them sequentially according to the delay detonator circuit order; if the risk index changes by less than 0.05 in the next decision cycle, the same delay table is used instead of being reissued. This ensures frame-level data accuracy while avoiding 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 as the first segment of the joint vector in a predetermined order; then sequentially appending the second video stream frame vector; and finally appending the risk index I, which contains only a single scalar element. This results in a continuous one-dimensional data sequence where the first segment is dehazed frame data, the second segment is labeled frame data, and the last segment is the risk index, which is the joint vector. In this embodiment, the dehazed frame resolution is 1920×1080, stored using the uint8 type; the risk index I is represented by float32. The system performs the following operations sequentially in GPU memory: allocating a contiguous buffer, first requesting a float32 buffer state_buf (6220800+2073600+1=8294401) with a length of 8294401 elements in the video memory. Based on 4 bytes per second, the total usage is approximately 31.6MB, which can be reused in one go. Copy the first video stream frame vector. The dehazed frame RGB vector was generated and normalized to 0-1 in the previous step, with a length of 6220800. Call CUDAcudaMemcpyAsync to write it to state_buf(0:6220800) in its original order. Append the second video stream frame vector. The grayscale value vector of the labeled frame is converted to float32 after mapping (0→0, 85→1, 170→2, 255→3), with a length of 2073600. Write it to state_buf(6220800:8294400). Append the risk index I. The latest risk index I (e.g., 0.48) is directly written to the last slot state_buf[8294400]. Thus, 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 the single scalar I. The one-dimensional vector is formed, and `state_buf` is now the joint vector. Since the three data segments are arranged consecutively in physical address, the subsequent network can directly use `float32[8294401]` as a single tensor input without additional copying or rearranging. The inference call is made through TensorRT's asynchronous inference interface, using the binding name "state" to pass in the joint vector; the network input layer dimension is fixed at (1, 8294401). The inference latency is approximately 4ms, and the output is 40 charge quantities and 40 detonation delays.

[0067] In some embodiments, calculating the risk index I includes: comparing the total crack length, average plume density, and total area of ​​the turbidity zone in the video stream 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%, the normalized plume density with a weight of 35%, and 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 1-second decision cycle. The system first accumulates three field indicators according to the aforementioned method: total crack length L_sum = 3.8m; average plume density D_avg = 0.17 (i.e., plume pixels account for 17% of the total frame pixels); total turbidity zone area A_sum = 0.92m. 2 Before the operation commenced, the supervision unit set the three maximum allowable values ​​as L_max = 6m, D_max = 0.25m, and A_max = 1.5m, based on environmental protection red lines and safety requirements. 2 The system then performs normalization and weighting based on this:

[0068] The normalized crack value Ln = L_sum ÷ L_max = 3.8 ÷ 6 ≈ 0.63

[0069] The normalized value of the plume is Dn = D_avg ÷ D_max = 0.17 ÷ 0.25 = 0.68

[0070] The normalized value of turbidity, An, is calculated as follows: An = A_sum ÷ A_max = 0.92 ÷ 1.5 ≈ 0.61

[0071] The risk index I is 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 limits the result to the range of 0–1 and compares it with the classification 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 exceeds the preset threshold of 0.05. The control logic immediately triggers the deep reinforcement learning module to recalculate the explosive charge and detonation delay for the new round, and executes 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 composition and connection method include: an underwater imaging and attitude acquisition subsystem; four underwater cameras: 4K / 60fps, 120° field of view, 30m withstand voltage, with RJ-45 gigabit Ethernet ports and DC24V power supply ports sharing a composite trunk line; four near-infrared LED fill light arrays: 850nm / 15000lm, with four LED strips evenly distributed around the circumference of each camera's casing, the LED strip power supply sharing the DC core of the composite trunk line with the cameras; four inertial measurement units (IMUs): three-axis gyroscopes + three-axis accelerometers, connected to the underwater sub-control box via a CAN-FD bus; two underwater acoustic positioning transponders: mounted on diagonal camera mounts, used for 3D positioning by measuring distance with the float base station; and an underwater sub-control box (IP68 aluminum-titanium alloy casing), internally integrating an 8-port gigabit PoE switch (4 x cameras + 1 x uplink), a CAN-FD hub, a PTP slave clock module, and a 48V→24V DC-DC power supply. The switch's uplink port is connected to the pontoon communication cabin via a composite optical-electric cable (single-mode fiber + AC48V). The pontoon communication cabin houses a PTP master clock (GPS time synchronization, nanosecond-level stability) and a gigabit fiber optic router. The fiber optic cable from the underwater sub-control box is converted to RJ-45 and transmitted to the shore-based central control room via a 200m roaming Ethernet cable; AC48V is transmitted down from the shore-based UPS via the same composite cable. The shore-based edge computing and blasting control area uses a GPU edge server: NVIDIA Jetson AGX Orin (32GB GPU), receiving 4 RTSP streams via a gigabit switch. The blasting PLC: a Siemens S7-1500F with high-speed I / O, connected to the same Profinet switch; it transmits and receives charge quantity and detonation delay in real time with the Jetson via gRPC. The electric detonator triggering module: 40 independent optocoupler isolated outputs, driven by the PLC's 24VDC pulse; each output is connected in series with a 3A fast fuse. UPS: 6kVA online type, supplying power to PLC, Jetson, switch, and buoy, and monitoring remaining power. Auxiliary sensing and safety interface: Hydrophone array ×4: 15m around the blast zone, RS-485 bus connected to PLC for monitoring shock wave. Diver gate sensor: Fiber Bragg grating switch, detecting the evacuation status of construction personnel; signal access to PLC safety chain. Connection topology and signal path: Camera → (composite cable) → underwater sub-control box PoE switch. Underwater sub-control box fiber optic uplink → buoy communication cabin → gigabit cable → shore-based switch. Jetson receives current from the switch, runs defogging, segmentation, risk assessment, and DRL model; results are sent to PLC via gRPC. After satisfying the logic "divers evacuated + electric detonator circuit self-test OK + |I_now–I_prev|>0.05 or reaches 1s cycle", PLC injects pulses into 40 electric detonators through trigger module according to DRL output table. After the blast was completed, the hydrophone reported the peak value of the shock wave to the PLC; Jetson wrote the event tag and uploaded the entire video and parameters to the NAS for blockchain storage.This hardware architecture achieves high bandwidth, low latency, and strong synchronization for multi-machine video backhaul and closed-loop blasting control through "optical-electric composite cable + PTP synchronization + Ethernet gRPC". At the same time, it digests PoE power supply and CAN / IMU data at the underwater end, and centrally completes AI inference and safety interlocking at the shore-based end, meeting the real-time and reliability requirements of blasting operations in various scenarios such as port areas and deep water.

[0073] The synchronous video stream acquisition module deploys N underwater cameras and corresponding near-infrared LED supplementary light arrays in the blasting area to acquire synchronous video streams.

[0074] The dehazing and first video stream frame vector generation module performs dark channel prior transmittance estimation on each frame of the synchronous video stream, and uses guided filtering to refine the transmittance map in a 5×5 window to reconstruct the dehazing frame sequence. Each dehazing frame is expanded into an RGB three-channel vector by row to obtain the first video stream frame vector.

[0075] The second video stream frame vector acquisition module processes the dehazed video frames through a first convolutional neural network containing a temporal attention mechanism to obtain labeled video frame images of rock cracks, bubble plumes, and turbid water areas. The pixel gray values ​​of the labeled video frame images are expanded into single-channel vectors by rows 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 area of ​​the turbidity zone of the video stream based on the crack length, plume density, and turbidity zone area of ​​each frame of the video stream, and calculates the Risk Index I.

[0077] The explosive charge and detonation timing generation module 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 the second convolutional neural network model to generate the explosive charge and detonation timing.

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

[0079] The first convolutional neural network (FrameSemanticSegmentationCNN) takes a single frame of dehazed RGB image as input, with dimensions of 1920×1080×3. The output is a single-channel labeled image of the same size (0 = background, 1 = crack, 2 = bubble plume, 3 = hazy area).

[0080] The core structure consists of an encoder: a 4-layer DeformableConv (3×3) + Batch-Norm + ReLU, with channel counts of 64 / 128 / 256 / 512 respectively; each layer is followed by 2×2 downsampling. The deformable convolution focuses on the elongated shape of cracks or plumes at the learned offset positions. The temporal attention module (TC-Former) stacks features of the same scale using a 5-frame sliding window, and calculates weights in the temporal dimension through 4 multi-head self-attention blocks (key / value dimension 512, head count = 4), capturing the dynamic correlation between crack propagation and plume diffusion. The decoder uses a top-down feature pyramid (FPN); each level is upsampled and skipped, and finally a 1×1 convolution is mapped to 4 class channels and softmaxed. First, deformable convolution is used to extract multi-scale shape features from a single frame. Then, TC-Former focuses attention along the time dimension on "newly appearing or expanding regions from the previous frame to the current frame," significantly improving the detection rate of crack filaments and sparse plumes. Finally, FPN is used to restore the resolution step by step to obtain a pixel-level segmentation map. The network is trained under supervision on manually labeled data using cross-entropy + boundary IoU loss.

[0081] The second convolutional neural network (State-to-Action 1DCNN-Agent) takes the following input: a one-dimensional joint vector of length 8,294,401 (1024-dimensional dehazed RGB encoding + 256-dimensional labeled grayscale encoding + 1-dimensional risk index). The output is an action vector m. i (Explosive charge for 40 boreholes, 0–0.5 kg); Delay vector Δt i(40 adjacent detonation delays, 0–100ms); Value V(s) (evaluates the current state value for PPO algorithm updates). Core structure: 1DConv feature extraction: Conv1d128×k5→ReLU→Conv1d256×k5→ReLU, step size 1, used to capture local coupling relationships between different dimensions (e.g., the correspondence between "new crack area" and "increased risk index"). Temporal-dimensional self-attention: Multi-head self-attention (512 hidden dimensions, 4 heads), globally models the entire vector, automatically focusing on the dimension combination most relevant to the charge quantity. Dual-branch output: Actor (FC512→256→80) is mapped to charge quantity and delay respectively via sigmoid; Critic (FC512→256→1) outputs the state value for PPO to calculate the advantage function. This network is essentially a lightweight policy / value joint network (Actor-Critic). It treats the environmental vectors from the visual encoder and the real-time risk index as "states," first using convolution to find local patterns (such as "crack area 0.6 + risk 0.7"), then using self-attention to integrate global features, and finally outputting the charge amount and detonation delay that satisfy the reward function of "maximum efficiency and minimum risk." During the training phase, PPO is used to interact with the environment in a high-fidelity simulator for learning; during the deployment phase, only forward inference is performed, with 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] Where, α t This is the average global attention value output by the temporal attention module for this frame, ranging from (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, and τ is a learnable scalar with an initial value of 0.5 and a value range of [0,2]. Adding α... t Subsequently, the same convolutional kernel has a larger effective slope in "keyframes," amplifying the gradient and making it easier to preserve subtle crack or sparse plume features; while in "non-keyframes," the slope shrinks, actively suppressing redundant noise. Experiments show that mIoU improved from 92.3% to 93.7%, and the false positive rate decreased by 14%.

[0085] The activation function f2(x) used in 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 injects the risk index into the activation slope, directly modulating the action distribution using the environmental safety level. The risk index I is directly written into the sigmoid slope, maintaining a wide threshold in low-risk (I≈0) scenarios; the curve becomes steeper in high-risk (I→1) scenarios, and the output is more prone to saturation: the explosive charge automatically approaches the lower limit, and the delay interval is increased. Actual measurements show that under the same blast zone, the explosive consumption per unit is reduced by 12.8%, and the peak radius of the turbid plume is reduced by 26%.

[0087] In some embodiments, the dehazing and first video stream frame vector generation module further includes: for each frame image in the synchronous video stream, for each pixel in the frame with different color channel brightness in a preset local neighborhood, selecting pixels with brightness values ​​less than a set threshold to generate a dark channel image; generating an initial transmittance image based on the dark channel image and the scene's global atmospheric brightness according to a preset ratio coefficient; using the original frame image as a guide image, performing guided filtering and thinning 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 a set lower limit threshold of transmittance to generate a dehazed frame image; and organizing all dehazed frame images into a dehazed frame sequence in chronological order and outputting it.

[0088] In some embodiments, the first video stream frame vector is a one-dimensional data sequence, in which the red, green, and blue channel values ​​of each pixel in the dehazed frame are read sequentially from left to right according to the row scanning order, and continuously appended in the order of R, G, B, and 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, in which the single-channel grayscale values ​​of the corresponding pixels in the labeled frame are read sequentially from left to right and continuously appended, and 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.

[0089] 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 as the first segment of the joint vector in a predetermined order; then sequentially appending the second video stream frame vector; and finally appending the risk index I, which contains only a single scalar element; thereby obtaining a continuous one-dimensional data sequence with the first segment being dehazed frame data, the second segment being labeled frame data, and the last segment being 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 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; summing the normalized crack length with a weight of 40%, the normalized plume density with a weight of 35%, and the normalized turbidity zone area with a weight of 25%, and summing the three weighted results to obtain the risk index I.

[0091] This 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. This invention significantly improves the visualization accuracy of blasting operations. Employing an improved dark channel-guided filtering defogging algorithm, the average peak signal-to-noise ratio (PSNR) is increased by 9 dB in a sea trial environment with a line-of-sight distance of 1.5 m and a suspended particle concentration >50 mg L⁻¹. The Transformer-CNN network achieves a pixel-level segmentation mIoU of 92.8% for cracks, bubble plumes, and turbid areas, approximately 11 percentage points higher than the traditional U-Net. This invention introduces a risk index I and two convolutional neural network models for processing. Under the same blasting zone, while maintaining the target particle size requirement, the average explosive consumption is reduced by 13.4%, and the detonation delay error converges from ±5 ms to ±1 ms. When a sudden increase in crack propagation speed (>20% / s) is detected online, the system can complete parameter recalculation and issue a new detonation sequence within 0.3 s, achieving millisecond-level closed-loop adjustment. With improved safety and environmental performance, the maximum turbidity plume diffusion radius was reduced by 27% and the peak turbidity of the water was reduced by 31% after adjusting the amount of explosive based on real-time plume density feedback, effectively reducing the impact on adjacent coral reefs and aquaculture areas.

[0093] 2. The number of dives per blasting zone is reduced from an average of 4 to 1, used only for initial hole layout confirmation. The distance between personnel and the blast site is consistently greater than 300m throughout the operation, significantly reducing personnel safety risks. Regarding construction efficiency and cost advantages, multi-scenario comparative tests show that this invention shortens the average construction cycle of a single blasting zone by 18%–22%, and saves over 15% in overall costs (including explosives, labor, and downtime). The video stream and control commands are stored in a continuous chain, providing reliable evidence for quality acceptance, insurance claims, and environmental assessments, saving on subsequent auditing and resumption costs.

[0094] The above provides a detailed description of an underwater blasting construction method and system based on real-time video image processing. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas and methods of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for underwater blasting construction based on real-time video image processing, characterized in that, The method comprises the steps of: S1: arranging N underwater cameras and corresponding near-infrared LED light arrays in the blasting area to obtain a synchronous video stream; S2: performing dark channel prior transmission rate estimation on the synchronous video stream frame by frame, and refining the transmission map under a 5*5 window using guided filtering to obtain a sequence of de-fogging frames, and expanding each de-fogging frame into an RGB three-channel vector to obtain a first video stream frame vector; S3: processing the video frames after de-fogging processing through a first convolutional neural network containing a timing attention mechanism to obtain labeled video frame images marking rock mass cracks, bubble plumes and water turbidity areas, and expanding the pixel gray values of the labeled video frame images into a single-channel vector to obtain a second video stream frame vector; S4: obtaining the total length of cracks, the average density of plumes and the total area of turbidity areas in the video stream according to the crack length, the plume density and the turbidity area of each frame of the video stream, and calculating a risk index I; 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 timing; S6: when the difference between the latest risk index I and the risk index of the last decision cycle exceeds a preset threshold, controlling an electric detonator network to perform blasting according to the explosive charge and the detonation timing.

2. A method of underwater blasting construction based on real-time video image processing according to claim 1, characterized in that, Step S2 further comprises: for each frame of the synchronous video stream, selecting pixels with luminance values less than a set threshold for different color channel luminances of the pixels in a preset local neighborhood to generate a dark channel image; generating an initial transmission rate image according to the dark channel image and the global atmospheric light luminance of the scene by a preset proportion coefficient; performing guided filtering refinement processing on the initial transmission rate image in a 5*5 pixel window with the original frame image as the guide image to obtain a smooth transmission rate image; performing pixel-by-pixel fusion on the smooth transmission rate image and the original frame image according to a set transmission rate lower threshold to generate a de-fogging frame image; and organizing all de-fogging frame images into a de-fogging frame sequence in chronological order and outputting the de-fogging frame sequence.

3. A method of underwater blasting construction 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, and the red, green and blue channel values of each pixel of the de-fogging frame are read from left to right in row scanning order and sequentially appended in the order of R, G and B, and the length of the obtained sequence is three times the number of original frame pixels; the second video stream frame vector is a one-dimensional data sequence, and the single-channel gray values of the corresponding pixels of the labeled frame are sequentially read from left to right and appended, and the length of the obtained sequence is equal to the number of original frame pixels; and the detonation timing is a delay sequence between adjacent blast holes, and the main control trigger time is recorded as 0.

4. A method of underwater blasting based on real-time video image processing according to claim 1, characterized in that, The fusion of the first video stream frame vector, the second video stream frame vector and the risk index I to generate a joint vector comprises: 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; and finally appending the risk index I which only contains a single scalar element; thereby obtaining a continuous one-dimensional data sequence with the first segment being de-fogging frame data, the second segment being labeled frame data and the end being a risk index, which is the joint vector.

5. A method of underwater blasting based on real-time video image processing as claimed in claim 1, wherein, The calculation risk index I includes: the total length of the video stream crack, the average density of the plume and the total area of the turbidity area are compared with the respective preset maximum allowable value to obtain three normalized evaluation values between zero and one; the normalized crack length is accumulated with a weight of 40%, the normalized plume density is accumulated with a weight of 35%, and the normalized turbidity area is accumulated with a weight of 25%, and the three weighted results are summed to obtain the risk index I.

6. A real-time video image processing based underwater blasting construction system, characterized in that, It comprises: A synchronous video stream acquisition module, N underwater cameras and corresponding near-infrared LED light supplement arrays are arranged in the blasting area to obtain synchronous video streams; A defogging and first video stream frame vector generation module, which performs dark channel prior transmittance estimation on the synchronous video stream frame by frame, and refines the transmittance map under a 5*5 window using guided filtering to reconstruct a defogging frame sequence, and expands each defogging frame into an RGB three-channel vector to obtain a first video stream frame vector; A second video stream frame vector acquisition module, which processes the video frame after defogging by a first convolutional neural network containing a timing attention mechanism to obtain a labeled video frame image of rock mass cracks, bubble plumes and water turbidity areas, and expands the pixel gray value of the labeled video frame image into a single-channel vector to obtain a second video stream frame vector; A risk index I calculation module, which obtains the total length of the video stream crack, the average density of the plume and the total area of the turbidity area according to the crack length, the plume density and the turbidity area of each frame of the video stream image, and calculates the risk index I; An explosive charge and initiation timing 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 initiation timing; An execution blasting module, which controls the electric detonator network to execute blasting according to the explosive charge and initiation timing when the difference between the latest risk index I and the risk index of the last decision cycle exceeds a preset threshold.

7. A real-time video image processing based underwater blasting system as claimed in claim 6, wherein, The defogging and first video stream frame vector generation module further comprises: for each frame image in the synchronous video stream, selecting pixels with luminance values less than a set threshold for different color channel luminances of the pixels in a preset local neighborhood to generate a dark channel image; generating an initial transmittance image according to the dark channel image and the scene global atmospheric light luminance with a preset proportion coefficient; performing guided filtering refinement processing on the initial transmittance image in a 5*5 pixel window to obtain a smooth transmittance image, taking the original frame image as a guide image; performing pixel-by-pixel fusion on the smooth transmittance image and the original frame image according to a set transmittance lower threshold to generate a defogging frame image; organizing all defogging frame images into a defogging frame sequence in chronological order and outputting.

8. A real-time video image processing based underwater blasting system as claimed in claim 6, wherein, 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 a left-to-right scanning order and sequentially appended in the order of R, G and B, and the length of the obtained sequence is equal to three times the number of pixels of the original frame; the second video stream frame vector is a one-dimensional data sequence, and the single-channel gray values of the corresponding pixels of the labeled frame are sequentially read from left to right and sequentially appended, and the length of the obtained sequence is equal to the number of pixels of the original frame; and the detonation timing sequence is a delay sequence between adjacent blast holes, and the main control trigger time is recorded as 0.

9. A real-time video image processing based underwater blasting system as claimed in claim 6, wherein, The fusion of the first video stream frame vector, the second video stream frame vector and the risk index I to generate a joint vector comprises the following steps: directly taking the first video stream frame vector as the first section of the joint vector in a predetermined order; subsequently sequentially appending the second video stream frame vector; and finally appending the risk index I containing only a single scalar element; thereby obtaining a continuous one-dimensional data sequence with the first section being defogging frame data, the second section being labeled frame data and the end being a risk index, which is the joint vector.

10. A real-time video image processing based underwater blasting system as claimed in claim 6, wherein, The calculation of the risk index I comprises the following steps: comparing the total length of the video stream crack, the average density of the plume and the total area of the turbid region with the respective preset maximum allowable values to obtain three normalized evaluation values between zero and one; adding the normalized crack length with a weight of 40%, adding the normalized plume density with a weight of 35%, adding the normalized turbid region area with a weight of 25%, and summing the three weighted results to obtain the risk index I.

Images