An artificial intelligence-based construction engineering safety monitoring method and system

Abstract

The application relates to the field of building engineering safety monitoring data processing and discloses a building engineering safety monitoring method and system based on artificial intelligence, which comprises the following steps: receiving an original video code stream and writing the original video code stream into a ring buffer; establishing a bypass listening flow process to analyze data packet header information, extract motion vector data, and calculate direction angle data; calculating a vector field direction discrete value; performing logical gate control decision; when a modulus sum exceeds a threshold value and the direction discrete value represents consistent distribution, shielding triggering is performed; only when discrete distribution occurs, an awakening interrupt is generated and a time stamp is recorded; according to the time stamp backtracking, an original video code stream segment is locked and delivered to a deep neural network inference engine; and the application realizes negative logic filtering of high-energy-consumption computing tasks by distinguishing wind-induced global disturbance from independent target motion from a mathematical statistical level through a vector field space consistency verification mechanism.

Description

Technical Field

[0001] This invention relates to a method and system for monitoring the safety of building engineering based on artificial intelligence, belonging to the field of data processing technology for building engineering safety monitoring. Background Technology

[0002] Currently, in the field of digital supervision of construction projects, the deployment of high-definition video acquisition equipment and high-frequency physical sensors to acquire multi-source heterogeneous data streams has become a standard technical path for realizing the monitoring of tower crane group collision prevention and illegal operation. The industry generally adopts a full-time, full-volume data processing method, with edge computing nodes configured to continuously and completely decode the input high-bandwidth video stream and sensor data stream and perform deep neural network inference. Although this logic ensures the continuity of monitoring, it encounters the bottleneck of rigidly limited edge computing resources in actual engineering applications. Construction sites are in a low-information-entropy silent or safe state for most of the time windows. The full-volume processing mode causes the processor to maintain high load operation for a long time to process a large amount of invalid data, resulting in a surge in equipment power consumption and a decrease in thermal stability.

[0003] Conventional algorithm optimization struggles to overcome energy efficiency bottlenecks. For example, Chinese invention patent CN118069469B discloses an AI-based method and system for monitoring building safety. While this solution utilizes gamma correction and the YOLOv5 model attention mechanism to improve the accuracy of moving target detection, the technical path still follows a linear serial processing mode of acquisition followed by inference. Regardless of whether the site is in a low-value, silent state, the system continuously calls high-energy-consuming deep neural networks to perform full calculations on each frame of image. Facing the need for all-weather monitoring, this mechanism exacerbates the ineffective load on edge computing nodes, making it impossible to achieve a balance between energy efficiency and response speed in embedded environments with limited computing power. The core contradiction lies in the fact that sudden safety accidents are often accompanied by instantaneous concurrent changes in data from multiple sensors. When the processor's computing resources are filled with conventional invalid data streams, the system struggles to allocate sufficient instantaneous computing power to respond to interruption requests from sudden events, leading to congestion in the data processing queue and uncontrollable decision-making delays. If a threshold-based sleep-wake mechanism is adopted, the physical delay between hardware startup and context switching causes the loss of critical data frames in the early stages of the accident, making it impossible to form a complete chain of evidence.

[0004] Therefore, the technical problem to be solved by this invention is how to resolve the mutually exclusive technical contradiction between the high computing power consumption of real-time analysis of full high-dimensional data and the zero-latency requirement for capturing sudden transient events by reconstructing the data flow processing logic under the premise of limited edge computing resources. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A method for monitoring the safety of building engineering based on artificial intelligence, comprising the following steps:

[0006] It receives the raw video stream input from the construction site acquisition equipment and writes the raw video stream into the circular buffer in memory in real time. The circular buffer temporarily stores historical data for a preset duration according to the first-in-first-out rule and automatically overwrites the old data when there is no external instruction.

[0007] In the process of writing the raw video bitstream into the circular buffer, a bypass listening process independent of the full decoding path is established.

[0008] In the bypass monitoring process, the data packet header information of the original video bitstream is directly parsed to extract the motion vector data of all macroblocks in the current frame, and the corresponding orientation angle data is calculated based on the motion vector data;

[0009] The system performs a logic-gated decision based on the consistency of the vector field space, and calculates the directional dispersion statistics of all non-zero motion vectors in the current frame. If the sum of the magnitudes of the motion vector data exceeds the preset motion amplitude threshold, but the directional dispersion statistics indicate that the directional distribution of the motion vector data has a consistent characteristic, then the current state is determined to be a global background disturbance and the wake-up interrupt signal is blocked. When the sum of the magnitudes of the motion vector data exceeds the preset motion amplitude threshold, and the directional dispersion statistics indicate that the directional distribution of the motion vector data has a discrete characteristic, a wake-up interrupt signal is generated and the trigger timestamp is recorded.

[0010] In response to the wake-up interrupt signal, the original video stream segment corresponding to the time window is backtracked from the circular buffer and locked according to the trigger timestamp;

[0011] The locked original video stream segment is sent to the deep neural network inference engine to perform full decoding and target recognition calculations, and outputs security monitoring results.

[0012] Preferably, the method further includes a step of establishing a load feedback loop to monitor the resource load status of the processor used to run the deep neural network inference engine in real time; dynamically adjusting the preset motion amplitude threshold in the logic gating decision according to the resource load status; and increasing the preset motion amplitude threshold according to the preset mapping relationship when the resource load status indicates an increase in processor occupancy, so as to filter low-amplitude motion vector data fluctuations and prioritize the processing resources for high-amplitude mutation data.

[0013] Preferably, the method further includes a heterogeneous signal timing interlocking step: synchronously receiving a mechanical vibration signal stream heterogeneous with the original video bitstream, and calculating the short-time energy value of the mechanical vibration signal stream within a sliding window; when a logic gating decision generates a wake-up interrupt signal, and the rate of change of the short-time energy value simultaneously exceeds a preset transient response threshold, performing an operation to backtrack from the circular buffer and lock the original video bitstream segment corresponding to the time window; wherein the heterogeneous signal timing interlocking step verifies the time correlation based on the following inequality relationship: ,in, To determine the time point at which the wake-up interrupt signal is generated. This refers to the time point at which the rate of change of the short-term energy value exceeds a preset transient response threshold. This is a preset physical propagation delay interlock time window.

[0014] Preferably, the calculation of the directional dispersion statistics of all non-zero motion vectors in the current frame includes: collecting the directional angle data of all non-zero motion vectors in the current frame and constructing a directional angle histogram; calculating the variance of the directional angle histogram as the directional dispersion statistics; determining that the current state is a global background disturbance, specifically including: comparing the variance with a preset consistency threshold, and when the variance is lower than the consistency threshold, identifying the presence of a single large-scale motion source in the image and performing the operation of masking the wake-up interrupt signal.

[0015] Preferably, parsing the data packet header information of the original video stream includes: identifying the compression encoding format in the original video stream; directly extracting the motion vector difference between P-frames and B-frames and the reference frame index from the slice header information or macroblock layer information to reconstruct the motion vector data; for I-frames, setting the associated motion vector data to zero or reusing the data from the previous P-frame to maintain the temporal continuity of the bypass listening process.

[0016] Preferably, the method further includes an adaptive baseline suppression step: performing a long-period moving average operation on the sum of the magnitudes of the motion vector data to update the dynamic base value representing the level of ambient background noise in real time; calculating the difference between the current instantaneous value of the sum of magnitudes and the dynamic base value; setting a preset motion amplitude threshold as the sum of the dynamic base value and the fixed sensitivity increment; and using logic gating to determine whether the kinetic energy triggering condition is met based on whether the difference exceeds the fixed sensitivity increment.

[0017] Preferably, the system also includes a fallback inspection step, which sets a timer independent of the logic gating decision. When the timer reaches the preset inspection cycle, a wake-up interrupt signal is forcibly generated to trigger the deep neural network inference engine to analyze the original video stream segment at the current moment. After outputting the security monitoring result, a state reset step is also included: after the deep neural network inference engine completes a calculation task, the memory occupied by the original video stream segment is immediately released; it is detected whether the logic gating decision still generates a wake-up interrupt signal at the current moment; if not, the system's main frequency is reduced and it returns to a low-power standby state that only executes the bypass listening process.

[0018] Preferably, the duration of the original video stream segment corresponding to the time window is determined by the following rules: taking the trigger timestamp as the center, a first preset duration is extracted forward and a second preset duration is extracted backward; the first preset duration is used to cover the cause process of the event, and the second preset duration is used to cover the subsequent evolution process of the event; the sum of the first preset duration and the second preset duration is less than the total capacity of the circular buffer.

[0019] Preferably, the mechanical vibration signal stream originates from an acceleration sensor installed at a key node of the tower crane structure; the short-time energy value is calculated using a time-domain sum-of-squares integration algorithm, which only calculates the amplitude of the sampling point in the time domain and does not perform frequency domain transformation.

[0020] An artificial intelligence-based building safety monitoring system includes:

[0021] The data acquisition and buffering module is used to receive the raw video stream input from the on-site acquisition equipment of the construction project and write the raw video stream into the circular buffer in memory in real time. The circular buffer temporarily stores historical data for a preset duration according to the first-in-first-out rule and automatically overwrites the old data when there is no external instruction.

[0022] The bypass listening and feature extraction module is used to establish a bypass listening process independent of the full decoding path during the process of writing the original video bitstream into the circular buffer. The bypass listening process directly parses the data packet header information of the original video bitstream, extracts the motion vector data of all macroblocks in the current frame, and calculates the corresponding orientation angle data based on the motion vector data.

[0023] The logic gating decision module is used to perform logic gating decisions based on the consistency of the vector field space: calculate the directional dispersion statistics of all non-zero motion vectors in the current frame; if the sum of the magnitudes of the motion vector data exceeds the preset motion amplitude threshold, but the directional dispersion statistics indicate that the directional distribution of the motion vector data has a consistent characteristic, then the current state is determined to be a global background disturbance and the wake-up interrupt signal is blocked; when the sum of the magnitudes of the motion vector data exceeds the preset motion amplitude threshold, and the directional dispersion statistics indicate that the directional distribution of the motion vector data has a discrete characteristic, a wake-up interrupt signal is generated and the trigger timestamp is recorded.

[0024] The asynchronous backtracking control module is used to respond to the wake-up interrupt signal, backtrack from the circular buffer according to the trigger timestamp, and lock the original video stream segment of the corresponding time window;

[0025] The deep neural network inference module is used to receive the locked original video stream segment, perform full decoding and target recognition calculations, and output security monitoring results.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] 1. This invention establishes a circular buffer at the data receiving end to synchronously run a low-dimensional proxy data analysis bypass, and constructs a storage, listening, backtracking, and asynchronous processing architecture. It decouples the high-computing-power deep neural network inference task from the real-time data acquisition task time axis. When the entropy of the proxy data representation information exceeds a preset threshold, the main processor is awakened by an interrupt signal to extract the corresponding time window historical data fragments from the circular buffer for fine processing with zero copy. This transforms the original continuous linear full-volume decoding and inference load into a discrete event-driven pulse-type computing load. Without reducing the monitoring time resolution, it significantly reduces the average processor utilization and thermal power consumption of edge computing nodes, and solves the resource contention bottleneck when deploying multiple high-frequency monitoring tasks on industrial-grade microcontrollers or embedded devices.

[0028] 2. This invention utilizes macroblock motion vector data from video coding standards as proxy features to directly parse data packet header information to construct a vector field reflecting dynamic changes in the scene. It calculates the statistical value of the sum of the temporal magnitudes of the vector field and the directional consistency of the spatial dimension. The logic decision-maker distinguishes between global background uniform displacement caused by strong winds or base resonance and discrete local displacement caused by independent target movement from a mathematical statistical perspective. The compression domain directly executes feature filtering logic to avoid wasting computational power on pixel-level decoding and image reconstruction of invalid background images. This ensures that the system maintains a stable and silent listening state under severe weather conditions and accurately allocates limited computing resources to sudden events with real physical significance.

[0029] 3. This invention utilizes the accompanying characteristics of the visual dimension and mechanical vibration dimension of physical events to construct a time-window-based heterogeneous signal logic interlocking mechanism. It requires that the generation of the trigger signal must simultaneously satisfy the sudden change in the form of video proxy data and the transient response of vibration signal energy, limiting the time difference between the two to within a preset physical propagation delay window. Multi-source heterogeneous data and gate logic operations filter out single-dimensional false signal interference, including pure visual disturbances caused by birds flying past the lens or electromagnetic interference that cause sensor value jumps. This improves the confidence of safety monitoring results in complex electromagnetic and physical environments and reduces the overhead of invalid data back-transmission and storage due to false alarms. Attached Figure Description

[0030] Figure 1 This is a flowchart of the asynchronous monitoring logic based on bypass listening and heterogeneous interlocking in this invention.

[0031] Figure 2 This is a comparison diagram of the motion vector magnitude and direction dispersion response under typical working conditions of the present invention;

[0032] Figure 3 This is a diagram of the edge computing node architecture that integrates bypass listening and backtracking control in this invention. Detailed Implementation

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] This invention discloses an artificial intelligence-based method and system for monitoring safety in construction projects, including data acquisition and buffering, bypass feature extraction, logic gating decision, heterogeneous signal interlocking, load adaptive adjustment, and asynchronous backtracking inference. In the data acquisition and buffering stage, the data acquisition module receives the raw video stream input from the construction site monitoring equipment. This raw video stream typically uses H.264 or H.265 compression encoding formats. The processor uses direct memory access technology to write the received raw video stream into a pre-allocated circular buffer in physical memory in real time. This circular buffer follows a first-in, first-out (FIFO) rule, and its storage space is divided into contiguous memory blocks for temporarily storing historical stream data of a preset duration. When the buffer write pointer catches up with the read pointer, the system automatically overwrites the oldest written data block. The system maintains a dynamically updated time sliding window. During the bypass feature extraction stage, while writing the original video bitstream into the circular buffer, a parallel bypass listening path is established. In this path, the processor directly parses the network abstraction layer units of the original video bitstream, extracts the data packet header information and macroblock layer data, and does not perform full decoding operations such as inverse quantization, inverse transform, and pixel reconstruction. For P-frames and B-frames, the system directly extracts the motion vector data and reference frame index of each macroblock from the compressed bitstream. For I-frames, the associated motion vector data is set to zero or the data of the previous P-frame is reused to maintain the continuity of the time series. Based on the extracted motion vector data, the processor calculates the sum of the magnitudes of all non-zero motion vectors in the current frame and the corresponding orientation angle data to generate a low-dimensional proxy data stream that characterizes the scene's kinetic energy level and motion trend.

[0035] During the logic gating decision phase, the system executes verification logic based on the spatial consistency of the vector field to distinguish between valid target motion and environmental interference. The processor statistically analyzes the direction angle data of all non-zero motion vectors in the current frame, constructs a direction angle histogram, and calculates the statistical variance of the direction angle histogram to characterize the spatial dispersion of the vector field. The logic decision unit executes the following deterministic decision procedure: If the sum of the magnitudes of the motion vector data exceeds a preset motion amplitude threshold, and the calculated direction dispersion statistical value is lower than a preset consistency threshold, the current state is determined to be wind-induced global disturbance or base resonance, and the wake-up interrupt signal is blocked; when the sum of the magnitudes of the motion vector data exceeds a preset motion amplitude threshold, and the direction dispersion statistical value is higher than a preset consistency threshold, it is determined that there is discrete independent target motion in the picture, a preliminary wake-up signal is generated, and the video trigger timestamp is recorded. During the heterogeneous signal interlocking phase, the system synchronously receives mechanical vibration signal streams from acceleration sensors at key nodes of the tower crane structure. The processor performs time-domain sum-of-squares integration on these mechanical vibration signal streams to calculate the short-time energy value within a sliding window. No frequency-domain transformation is performed. The logic decision unit continuously monitors the rate of change of the short-time energy value. When this rate of change exceeds a preset transient response threshold, the vibration trigger time point is marked. The system verifies the time correlation between video triggering and vibration triggering, and determines whether the time difference between the two satisfies the inequality. ,in, The system uses a preset physical propagation delay interlock time window. When this inequality holds, the system confirms that a real physical collision or structural anomaly has occurred and generates a valid wake-up interrupt signal.

[0036] During the load adaptive adjustment phase, the system establishes a negative feedback control loop based on the computing resource load status. The processor monitors the length of the task queue or the remaining space of the circular buffer read / write pointer in real time. When the processor utilization rate increases or the buffer space falls below a preset warning threshold, the logic decision unit increases the preset motion amplitude threshold according to a preset inverse mapping relationship, filtering out low-amplitude minor motion vector fluctuations and reserving computing resources to prioritize processing high-amplitude abrupt data. In addition, the system performs adaptive baseline suppression, performs a long-period moving average calculation on the sum of the magnitudes of the motion vector data, and updates the dynamic base value representing the environmental background noise level in real time. The logic decision unit calculates the current instantaneous value of the sum of modulo values. With dynamic basis value The difference between them is used to set the kinetic energy triggering condition to be greater than a fixed sensitivity increment; during the asynchronous backtracking inference phase, the asynchronous backtracking control module responds to the wake-up interrupt signal and, based on the recorded trigger timestamp... The system backtracks from the circular buffer and locks the original video stream segment corresponding to the time window. The length of the time window is determined by the first preset duration of the segment before the trigger timestamp and the second preset duration of the segment after the trigger timestamp. The locked original video stream segment is sent to the deep neural network inference engine to perform full decoding and target recognition calculation, and outputs the security monitoring results. After completing one inference task, if no new interrupt signal is generated, the system releases memory resources and controls the processor to enter a low-power standby state. The system is equipped with an independent timer, which forcibly triggers a full analysis when the preset inspection cycle is reached.

[0037] Example 1: In a high-altitude tower crane group monitoring scenario under severe convective weather conditions, the video acquisition equipment experiences continuous periodic oscillations due to gust wind loads, while the tower structure experiences high-frequency mechanical vibrations. Under these conditions, the background of the video image undergoes global displacement, resulting in a large number of high-modulus non-zero motion vectors in the original video stream. This high-energy environmental noise masks the characteristics of discrete safety events such as broken hook wire ropes or illegal swaying at the data level. The system writes the received original video streams into a circular buffer in real time, maintaining a historical data window of a preset duration using a first-in-first-out rule. Simultaneously, the bypass listening process directly parses the data packet header of the stream, extracts the motion vector data of the current frame, and calculates all non-zero motion vectors in the current frame based on the global motion characteristics caused by wind-induced jitter. When a gust of wind causes the overall image to shift, although the sum of the magnitudes of the motion vector data exceeds the preset motion amplitude threshold, the calculated direction dispersion statistics are lower than the preset consistency threshold because all vector directions converge. The current high-energy signal is determined to be a global background disturbance, and the wake-up interrupt signal is kept shielded, allowing the deep neural network inference engine to remain in low-power silence. When a steel cable breaks, causing the suspended object to fall abnormally, a discrete vector field appears in the image that is inconsistent with the background motion direction, causing the direction dispersion statistics to suddenly change and exceed the consistency threshold. At this time, the system combines the synchronously received mechanical vibration signal stream to perform heterogeneous signal timing interlock verification. The processor calculates the short-time energy change rate of the vibration signal. When this change rate exceeds the preset transient response threshold, and the vibration trigger time point... With video feature mutation timestamp Satisfying inequalities At that time, the logic gating decision module generates a valid wake-up interrupt signal. In response to this signal, the asynchronous backtracking control module determines the wake-up interrupt based on the timestamp. The original video stream segment containing the cause and evolution of the accident is locked and extracted from the circular buffer, and then sent to the inference engine for full decoding and target recognition.

[0038] Example 2: To verify the practical effectiveness, anti-interference capability, and rationality of the core parameter settings of this invention in complex engineering environments, a hardware verification platform including edge computing nodes, high-definition surveillance cameras, and high-frequency vibration accelerometers was built. The test data sources were divided into two parts: video data used the construction site high-altitude operation dataset (CS-HAO-2024), which was injected with sinusoidal global motion noise with a frequency of 0.5Hz to 2.5Hz and a pixel displacement amplitude of 10 pixels to 80 pixels through an affine transformation algorithm to simulate the periodic swaying and irregular shaking of a tower crane under strong wind load; vibration data was acquired by installing a triaxial accelerometer on a programmable excitation table. The excitation table output the background vibration waveform of the tower structure measured on site, superimposed with Gaussian white noise with a signal-to-noise ratio of 15dB and 50Hz power frequency interference to construct a high-fidelity heterogeneous signal test environment; regarding the physical propagation delay interlocking time window in the heterogeneous signal timing interlocking step. The setting is determined based on the latency measurement results of the system link. The main factors affecting the value of this parameter include the video encoding and decoding time. And the time consumed by vibration signal transmission and interruption response The technical trade-off lies in: if Setting the threshold too narrow will cause the system to fail to match real concurrent events caused by jitter, resulting in missed detections; setting it too wide will introduce non-causal random overlap interference, increasing the false alarm rate. Through 1000 concurrent trigger tests on the experimental platform, the average time difference between the two signals arriving at the logic decision unit was measured to be 125.4 ms, with a standard deviation of 18.2 ms. Based on 3... Statistical principles, to cover 99.7% of truly relevant events, will... The value is set to 180ms. This value is not an empirical estimate, but a statistical limit based on the system's time delay distribution characteristics.

[0039] Three comparative groups were designed to verify the synergistic effect mechanism and performance boundary of the technical solution. Control group A adopted a single threshold triggering logic based on the sum of motion vector magnitudes to simulate traditional motion detection technology. Control group B introduced the vector field direction dispersion statistics for spatial consistency verification, but did not enable heterogeneous signal interlocking. The sample group of this invention enabled a complete logic gating decision and heterogeneous signal timing interlocking mechanism to simulate two typical scenarios: strong wind causing large background swaying (pure noise condition) and strong wind causing the suspended object to fall due to wire rope breakage (signal plus noise condition). The processor parsed the motion vector of the P frame of the H.264 bitstream in real time and calculated the relevant statistical features. Table 1 shows the comparison of key intermediate feature data and final triggering judgment results of each group under different conditions.

[0040] Table 1: Comparison of Key Feature Values ​​and Triggering Results for Different Processing Logics under Strong Wind Interference

[0041]

[0042] Referring to Table 1, the data shows that under pure strong wind disturbance conditions, the image exhibits drastic uniform displacement, resulting in a total MV magnitude of 0.88, exceeding the preset motion amplitude threshold (set to 0.5), causing a false alarm in control group A. However, the directional dispersion statistics at this time are only 0.12, far below the preset uniformity threshold (set to 0.4), indicating a high degree of convergence of motion vectors. Control group B and the present invention sample group successfully shielded this interference accordingly. It is worth noting that under extreme gust conditions, sudden changes in wind load cause the vibration energy change rate to reach 2.15 J / s, approaching the transient response threshold, and the image magnitude is extremely large. At this time, single-dimensional verification faces the risk of failure, but the present invention sample group, through logic AND gate operation... The calculation requires that the rate of change of vibration energy must simultaneously exhibit a sudden change of orders of magnitude, such as 18.42 J / s during a real fall, in order to maintain the system's silent state. Further analysis revealed that the technical effect exhibits obvious nonlinear inflection point characteristics at the parameter boundaries. When the simulated wind speed exceeds the threshold of 30 m / s, the camera bracket undergoes non-rigid deformation, causing disordered random disturbances in the motion vector field of the video image. The directional dispersion statistical value abnormally increases to over 0.55, causing the suppression capability of the control group B to fail, and the false alarm rate surges from 1.2% to 28.5%. However, the sample group of this invention, thanks to the physical interlocking limitation of the mechanical vibration signal, can still maintain the false alarm rate at a low level of 3.4% under this extreme condition.

[0043] Example 3: This example combines Figures 1 to 3 This document describes a method and system for monitoring the safety of construction projects based on artificial intelligence. Figure 1 As shown, the bitstream contains H.264 or H.265 compressed data packets. On one hand, the bitstream is written to a circular buffer, which follows the rules of first-in-first-out and automatic overwriting of old data to extract historical data. On the other hand, it enters the bypass listening process, parsing the packet header and extracting motion vectors in a decoding-free state. The generated motion vector data is input to the logic gating decision module to perform vector field space consistency verification to generate an initial wake-up signal. At the same time, the mechanical vibration signal stream from the accelerometer is used as heterogeneous data input to the heterogeneous signal timing interlock module to verify the time correlation between video and vibration. When the logic verification passes, a wake-up interrupt and timestamp are generated. The asynchronous backtracking control module responds to the timestamp locking and extracts the video segment, which is sent to the deep neural network inference module for full decoding, target recognition and high-energy-consumption calculation. Finally, the safety monitoring result is output. The system also includes a load feedback loop to dynamically adjust the motion amplitude threshold according to the resource load status.

[0044] like Figure 2As shown, this figure depicts the characteristic changes under five typical conditions: a silent baseline, pure strong wind disturbance, extreme gusts, actual fall, and strong wind plus fall. The solid line represents the sum of normalized MV magnitudes, and the dashed line represents the directional dispersion calculated as variance. Under the pure strong wind disturbance condition, although the sum of MV magnitudes is high, the directional dispersion is extremely low, indicating that the vector field has spatial consistency. Under the actual fall condition, both values ​​increase, indicating that the vector field exhibits a discrete distribution characteristic. Figure 3 As shown, this system includes a tower crane physical carrier at the front end, equipped with a high-definition monitoring camera as the video acquisition source and a three-axis accelerometer as the vibration acquisition source. The output video data stream and acceleration signal converge to an edge computing terminal node, which is an industrial-grade embedded host. The operating environment includes an embedded operating system and a memory management unit. The host integrates six core components, including a ring buffer system for temporary storage of memory data, a bypass listening component for feature extraction and vector analysis, a logic decision component for performing consistency verification and interlocking, a backtracking control component for locking and reading historical data, an artificial intelligence inference engine for performing deep neural network target recognition, and a load adaptive module for resource monitoring and threshold adjustment. The system output is connected to a monitoring display terminal to present real-time alarm pop-ups and to a safety management server to store the data in an accident evidence database.

[0045] Example 4: This example constructs an adaptive baseline suppression and floating threshold update mechanism based on dynamic statistical characteristics. The core is to transform static absolute threshold determination into dynamic incremental determination relative to the real-time environmental baseline, thereby achieving automatic decoupling from the time-varying characteristics of the environment. In a construction site scenario with continuous 24 / 7 monitoring, the background motion noise level captured by the video acquisition equipment exhibits slow but drifting characteristics with the alternation of day and night and weather changes. If a fixed threshold is used, the system will face a dilemma: too high a threshold will lead to missed detections of minor incidents, while too low a threshold will cause continuous false alarms in severe weather. Therefore, the system embeds a background noise statistics thread in the logic gating decision module, with a maintenance length of [missing information]. A long-period sliding window is used to sample the sum of motion vector magnitudes output by the bypass feature extraction module in real time.

[0046] For each frame of input data, the processor does not directly perform threshold comparisons. Instead, it performs background basis update operations. The system uses a quantile-based statistical filtering algorithm to calculate the th quantile of the sum of motion vector magnitudes within the current sliding window. Percentile value, defined as the current dynamic base value This statistic can robustly characterize the peak level of background noise in the current environment, while filtering out occasional transient interference. The system calculates the sum of the magnitudes of the current instantaneous motion vector. With dynamic basis value The difference between The calculation formula is: The trigger condition for the logic decision unit is refactored to: when the difference... Exceeding the preset relative sensitivity increment Only when this time is it determined to be a valid mutation, among which, This is a fixed constant characterizing the system's sensitivity to abnormal events. Its value depends on the minimum kinetic energy characteristic of the target to be detected and is independent of the environmental background. Through this differential operation, the system is actually performing a constant false alarm rate (CFAR) detection. Regardless of how the background noise floor slowly rises with increasing rainfall or wind, the trigger threshold always remains above a fixed increment of the background floor, ensuring the consistency of the system's detection capability under different signal-to-noise ratio environments. Furthermore, to prevent the dynamic floor value from being affected by sudden high-energy events... The erroneous rise masked the subsequent real incident. To address this, the system introduces a base update freeze logic. When the logic decision maker determines the current state is triggered, updates to the data within the sliding window are paused, maintaining... The base value is locked at the level prior to the event until the trigger state is released, ensuring that the base value always purely reflects the environmental background and avoiding the risk of desensitization caused by adaptive drift.

[0047] Example 5: To ensure key parameters in the heterogeneous signal timing interlocking mechanism The setup is based on rigorous physical principles and is reproducible. This embodiment constructs a standardized offline calibration and time delay distribution measurement procedure. Before the system is officially deployed, engineers build a closed-loop test environment in a controlled laboratory, including a video encoder, network transmission link, and vibration sensor acquisition module. When the test starts, a light pulse signal of a preset frequency is simultaneously projected into the field of view of the camera lens through a high-precision signal generator, and a mechanical excitation signal of the same frequency is applied to the accelerometer to simulate a concurrent physical event with zero time difference. The system backend records the timestamp of the light pulse signal parsed from the video bitstream. Timestamp of the mechanical excitation signal trigger threshold This test process is repeated at least 1000 times to obtain a statistically significant sample set, by calculating all time differences in the sample set. Based on the distribution characteristics, the system fits the Gaussian distribution curve of the time delay difference, and according to... Statistical principles will be used to interlock the physical propagation delay time window. Determined as the mean Add three standard deviations This summation avoids parameter setting uncertainties caused by hardware differences.

[0048] To address the complex and variable vibration environment at construction sites and resolve false alarms caused by differences in the inherent frequencies of different tower crane structures, this embodiment establishes a pre-deployment calibration procedure. After the monitoring system is installed, a 24-hour silent background data acquisition period is required. During this period, the system does not output alarms but focuses on collecting the mechanical vibration signal stream of the tower crane under normal load conditions. The processor performs time-domain statistical analysis on the acquired long-series signals, extracting feature vectors including short-time energy mean, peak factor, and kurtosis to construct a benchmark vibration fingerprint model for the specific tower crane. The logic decision unit automatically adjusts the preset transient response threshold, setting it to 1.5 to 2.0 times the peak value of the background noise energy. This pre-calibration process ensures that the monitoring system can adapt to the structural characteristics of different tower cranes, avoiding false alarms caused by differences in base stiffness or inherent resonance.

[0049] Example 6: In the actual engineering scenario of deploying this invention on a newly installed tower crane, a standardized on-site pre-calibration and threshold optimization procedure was established. This procedure was executed after the system was initially powered on and before it was officially put into monitoring. The aim was to eliminate parameter uncertainties introduced by differences in camera focal length, installation height, and lens distortion. The calibration process used the natural swing of the tower crane hook as the standard excitation source. In a controlled environment with no wind or light wind, the operator controlled the hook to perform a single-degree-of-freedom reciprocating motion. During this period, the system collected video streams of a preset duration as a calibration sample set. The processor extracted motion vectors for each frame in the sample set and calculated the statistical distribution of the magnitude values ​​of all non-zero vectors. For different combinations of focal length and resolution, the system statistically determined the magnitude quantiles that could effectively distinguish between background noise and target motion, and defined them as the reference motion amplitude threshold for this specific deployment environment. To further determine the physical propagation delay interlocking time window in the heterogeneous signal timing interlocking logic, To determine the optimal operating point, the procedure executes a closed-loop time delay characteristic measurement. The system generates a step-like mechanical impact by controlling the tower crane's hoisting mechanism. This impact simultaneously triggers a sudden change in the visual image and a response from the accelerometer. The system backend records the timestamps of the motion vector magnitude jumps in the video stream. The timestamp of the vibration signal energy crossing the base noise The impact test is repeated at least 20 times to obtain a statistical sample, and the system calculates the time difference sample for each test. And calculate the mean. with standard deviation The calculation formula is: Based on statistical principles, the system interlocks time windows. Automatically set to This allows for the maximum compression of the admission window for non-causal interference while covering 99.7% of the system's inherent jitter range.

[0050] Example 7: Preset Motion Amplitude Threshold On-site Initialization Calibration Before formally connecting to the alarm circuit, environmental base noise quantization is performed. The camera is controlled to point at the non-operational, stationary building site, and a 300-second video stream is continuously acquired. The processor extracts the bypass motion vector data of each frame and calculates the sum of magnitudes. Histogram statistics are performed on the sequence of sums of magnitudes, and the value corresponding to a cumulative probability distribution reaching 95% is defined as the on-site background noise base. The preset motion amplitude threshold is set to the background noise floor. 1.2 to 1.5 times; physical propagation delay interlocking time window in heterogeneous signal timing interlocking steps. Based on the results of the on-site closed-loop impact response test, a single instantaneous mechanical impact was applied to a key node of the tower crane structure at the center of the camera's field of view, simultaneously triggering a high-brightness flash signal. The backend recorded the time stamps of the step values ​​of the motion vector magnitude in the video stream. The accelerometer output signal exceeds the trigger level time stamp The process is repeated 20 times to build a time difference sample set, and the processor calculates the time difference in the sample set. mean with standard deviation The physical propagation delay interlock time window Determined as The resource load status dynamically adjusts the preset motion amplitude threshold according to a piecewise linear mapping strategy. The processor reads the operating system task queue length register or idle cycle counter to obtain the real-time load rate. When the value is below 80%, the preset motion amplitude threshold is maintained at the initial calibration value. constant; When the linear adjustment range is between 80% and 95%, dynamic compensation calculation is performed, and the real-time threshold is set. ; When the threshold exceeds 95%, the preset motion amplitude threshold will be forcibly locked. 1.5 times.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for monitoring the safety of construction projects based on artificial intelligence, characterized in that, Includes the following steps: It receives the raw video stream input from the construction site acquisition equipment and writes the raw video stream into the circular buffer in memory in real time. The circular buffer temporarily stores historical data for a preset duration according to the first-in-first-out rule and automatically overwrites the old data when there is no external instruction. In the process of writing the raw video bitstream into the circular buffer, a bypass listening process independent of the full decoding path is established. In the bypass monitoring process, the data packet header information of the original video stream is directly parsed. Parsing the data packet header information of the original video stream includes: identifying the compression encoding format in the original video stream; directly extracting the motion vector difference between P-frames and B-frames and the reference frame index from the slice header information or macroblock layer information to reconstruct the motion vector data; for I-frames, the associated motion vector data is set to zero or the data of the previous P-frame is reused to maintain the temporal continuity of the bypass monitoring process; the motion vector data of all macroblocks in the current frame is extracted, and the corresponding orientation angle data is calculated based on the motion vector data. The system executes a logic-gated decision based on the consistency of the vector field space, calculating the directional dispersion statistics of all non-zero motion vectors in the current frame. This includes: collecting the directional angle data of all non-zero motion vectors in the current frame and constructing a directional angle histogram; calculating the variance of the directional angle histogram as the directional dispersion statistics; if the sum of the magnitudes of the motion vector data exceeds a preset motion amplitude threshold, but the directional dispersion statistics indicate that the directional distribution of the motion vector data exhibits a consistent characteristic, then the current state is determined to be a global background disturbance, and the wake-up interruption signal is blocked. Specifically, this includes: comparing the variance with a preset consistency threshold; when the variance is lower than the consistency threshold, it is determined that there is a single large-scale motion source in the image, and the wake-up interruption signal is blocked; when the sum of the magnitudes of the motion vector data exceeds a preset motion amplitude threshold, and the directional dispersion statistics indicate that the directional distribution of the motion vector data exhibits a discrete characteristic, a wake-up interruption signal is generated and the trigger timestamp is recorded. This includes a heterogeneous signal timing interlocking step: synchronously receiving a mechanical vibration signal stream heterogeneous with the original video stream, and calculating the short-time energy value of the mechanical vibration signal stream within a sliding window; when a logic gate decision generates a wake-up interrupt signal, and the rate of change of the short-time energy value simultaneously exceeds a preset transient response threshold, performing an operation to backtrack from the circular buffer and lock the original video stream segment corresponding to the time window; the heterogeneous signal timing interlocking step verifies time correlation based on the following inequality relationship: ,in, To determine the time point at which the wake-up interrupt signal is generated. This refers to the time point at which the rate of change of the short-term energy value exceeds a preset transient response threshold. The preset physical propagation delay interlock time window; In response to the wake-up interrupt signal, the original video stream segment corresponding to the time window is backtracked from the circular buffer and locked according to the trigger timestamp; The locked original video stream segment is sent to the deep neural network inference engine to perform full decoding and target recognition calculations, and outputs security monitoring results.

2. The method for monitoring the safety of building construction based on artificial intelligence according to claim 1, characterized in that, It also includes the step of establishing a load feedback loop to monitor the resource load status of the processor used to run the deep neural network inference engine in real time; dynamically adjust the preset motion amplitude threshold in the logic gating decision according to the resource load status; when the resource load status indicates that the processor utilization rate is rising, increase the preset motion amplitude threshold according to the preset mapping relationship to filter low-amplitude motion vector data fluctuations.

3. The method for monitoring building engineering safety based on artificial intelligence according to claim 1, characterized in that, It also includes an adaptive baseline suppression step: performing a long-period moving average operation on the sum of the magnitudes of the motion vector data to update the dynamic baseline value characterizing the level of ambient background noise in real time; Calculate the difference between the current instantaneous value of the sum of moduli and the dynamic base value; The preset motion amplitude threshold is set as the sum of the dynamic base value and the fixed sensitivity increment; the logic gating decision is based on whether the difference exceeds the fixed sensitivity increment to determine whether the kinetic energy triggering condition is met.

4. The method for monitoring the safety of building construction projects based on artificial intelligence according to claim 1, characterized in that, It also includes a fallback inspection step, setting a timer independent of the logic gating decision; when the timer reaches the preset inspection cycle, a wake-up interrupt signal is forcibly generated to trigger the deep neural network inference engine to analyze the original video stream segment at the current moment and output the security monitoring results. After that, a state reset step is also included: after the deep neural network inference engine completes a calculation task, the memory occupied by the original video stream segment is immediately released; it checks whether the logic gating decision still generates a wake-up interrupt signal at the current moment; if not, it controls the system main frequency to decrease and returns to a low-power standby state that only executes the bypass listening process.

5. The method for monitoring the safety of building construction projects based on artificial intelligence according to claim 1, characterized in that, The duration of the original video stream segment corresponding to the time window is determined by the following rules: taking the trigger timestamp as the center, a first preset duration is extracted forward and a second preset duration is extracted backward; the first preset duration is used to cover the cause process of the event, and the second preset duration is used to cover the subsequent evolution process of the event; the sum of the first preset duration and the second preset duration is less than the total capacity of the circular buffer.

6. The method for monitoring the safety of building construction based on artificial intelligence according to claim 1, characterized in that, The mechanical vibration signal stream originates from the acceleration sensor installed at the key node of the tower crane structure; the short-time energy value is calculated using the time-domain sum-of-squares integration algorithm, which only calculates the amplitude of the sampling point in the time domain and does not perform frequency domain transformation.

7. An artificial intelligence-based building engineering safety monitoring system, used to implement the method of claim 1, characterized in that, include: The data acquisition and buffering module is used to receive the raw video stream input from the on-site acquisition equipment of the construction project and write the raw video stream into the circular buffer in memory in real time. The circular buffer temporarily stores historical data for a preset duration according to the first-in-first-out rule and automatically overwrites the old data when there is no external instruction. The bypass listening and feature extraction module is used to establish a bypass listening process independent of the full decoding path during the process of writing the original video bitstream into the circular buffer. The bypass monitoring process directly parses the data packet header information of the original video stream. This parsing includes: identifying the compression encoding format in the original video stream; directly extracting the motion vector difference between P-frames and B-frames and the reference frame index from the slice header information or macroblock layer information to reconstruct the motion vector data; for I-frames, setting the associated motion vector data to zero or reusing the data from the previous P-frame to maintain the temporal continuity of the bypass monitoring process; extracting the motion vector data of all macroblocks in the current frame; and calculating the corresponding orientation angle data based on the motion vector data. The logic gating decision module is used to perform logic gating decisions based on the consistency of the vector field space: calculating the directional dispersion statistics of all non-zero motion vectors in the current frame, including: collecting the direction angle data of all non-zero motion vectors in the current frame and constructing a direction angle histogram; calculating the variance of the direction angle histogram as the directional dispersion statistics; if the sum of the magnitudes of the motion vector data exceeds a preset motion amplitude threshold, but the directional dispersion statistics indicate that the directional distribution of the motion vector data has a consistent characteristic, then the current state is determined to be a global background disturbance and the wake-up interruption signal is blocked; specifically, it includes: comparing the variance value with a preset consistency threshold, and when the variance value is lower than the consistency threshold, it is determined that there is a single large-scale motion source in the picture, and the operation of blocking the wake-up interruption signal is performed; when the sum of the magnitudes of the motion vector data exceeds a preset motion amplitude threshold, and the directional dispersion statistics indicate that the directional distribution of the motion vector data has a discrete characteristic, a wake-up interruption signal is generated and the trigger timestamp is recorded; The heterogeneous signal timing interlock module is used to synchronously receive a mechanical vibration signal stream that is heterogeneous with the original video bitstream, and calculate the short-time energy value of the mechanical vibration signal stream within a sliding window. When the logic gate decision generates a wake-up interrupt signal, and the rate of change of the short-time energy value simultaneously exceeds a preset transient response threshold, the module performs an operation to backtrack from the circular buffer and lock the original video bitstream segment corresponding to the time window. The heterogeneous signal timing interlock step verifies the time correlation based on the following inequality: ,in, To determine the time point at which the wake-up interrupt signal is generated. This refers to the time point at which the rate of change of the short-term energy value exceeds a preset transient response threshold. The preset physical propagation delay interlock time window; The asynchronous backtracking control module is used to respond to the wake-up interrupt signal, backtrack from the circular buffer according to the trigger timestamp, and lock the original video stream segment of the corresponding time window; The deep neural network inference module is used to receive the locked original video stream segment, perform full decoding and target recognition calculations, and output security monitoring results.

Images