Construction monitoring method and system based on artificial intelligence

Abstract

The invention relates to the technical field of data processing, and discloses a construction monitoring method and system based on artificial intelligence. The method comprises the following steps: collecting construction operation behaviors through a high-speed camera, and performing frequency division filtering on data according to a vibration spectrum to obtain standardized data; inputting the data into a construction behavior feature genetic algorithm, and optimizing a recognition network through behavior feature coding to obtain a lightweight model; carrying out probability calculation on the abnormal behavior according to the model; the probability value is transmitted to edge equipment, and model parameters are compressed through weight pruning to obtain a reasoning result; and early warning signals are obtained through three-level early warning threshold grading judgment. The technical problems of unstable data quality, lack of construction scene adaptability of an AI algorithm, delay of cloud processing and insufficient early warning decision reliability of an existing construction monitoring technology in a complex environment are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a construction monitoring method and system based on artificial intelligence. BACKGROUND

[0002] The existing construction monitoring method mainly relies on manual inspection and fixed sensor monitoring. Manual inspection is carried out by safety management personnel who regularly patrol the construction site to observe the work behavior and safety equipment wearing of construction personnel, and correct and record the site after finding the violation. Fixed sensor monitoring uses cameras, vibration sensors and other devices to continuously monitor the construction site, and the sensor data is transmitted to the monitoring center through wired or wireless network, and the professional personnel analyzes the monitoring data and judges whether there is a safety hazard. Some advanced construction sites also use basic image recognition technology to identify whether the construction personnel wear safety helmets through simple pattern matching algorithm, or to judge whether the personnel enter the dangerous area through regional intrusion detection algorithm.

[0003] However, the existing technology has significant deficiencies: manual inspection is inefficient and has monitoring blind spots, cannot achieve 24-hour uninterrupted monitoring, and manual judgment is easily affected by subjective factors, with the risk of missed detection and misjudgment. Although fixed sensor monitoring solves the problem of continuous monitoring, the traditional image recognition algorithm has low accuracy in complex construction environment, especially in the case of light changes, shielding interference and multiple target coexistence, the recognition effect is greatly reduced. The existing basic AI technology lacks optimization design for construction scenes, and the general algorithm is difficult to adapt to the complex and variable characteristics of the construction environment, and most systems rely on cloud processing, which has network delay problem and cannot meet the real-time early warning demand of construction safety accidents.

[0004] Based on the above analysis, the progressive technical problems to be solved in the prior art include: first, how to obtain high-quality multi-modal monitoring data in a complex and variable construction environment and perform effective preprocessing to eliminate the influence of environmental noise and interference on data quality; second, how to design an intelligent recognition algorithm specifically for construction behavior characteristics to improve the recognition accuracy and generalization ability of construction abnormal behavior through optimization of network structure; third, how to deploy a lightweight model on an edge computing device to realize millisecond-level real-time inference through model compression and optimization technology to solve the delay problem of cloud processing; and finally, how to establish a reliable multi-level early warning decision mechanism to ensure the accuracy and reliability of the early warning results through probability theory and uncertainty analysis to avoid the negative impact of false positives and false negatives on construction safety. SUMMARY

[0005] The application provides a construction monitoring method and system based on artificial intelligence, which is used to solve the technical problems of unstable data quality of existing construction monitoring technology in complex environment, lack of construction scene adaptability of AI algorithm, delay of cloud processing and insufficient reliability of early warning decision.

[0006] In a first aspect, the application provides a construction monitoring method based on artificial intelligence, which comprises: collecting the operation behavior of construction personnel in real time through a high-speed camera arranged on a construction operation surface, performing frequency filtering processing on original construction site data according to the vibration frequency spectrum of construction machinery to obtain standardized construction behavior data; inputting the standardized construction behavior data into a construction behavior feature genetic algorithm, optimizing the identification network through behavior feature coding of aerial operation and hoisting operation to obtain a lightweight construction abnormality identification model; performing probability calculation processing on the construction abnormal behavior of not wearing a safety helmet and illegally entering a dangerous area according to the lightweight construction abnormality identification model to obtain a construction abnormality probability value; transmitting the construction abnormality probability value to a construction site edge computing device, compressing the model parameters through weight pruning to obtain an inference result; and performing hierarchical judgment processing on the construction abnormality probability through a three-level early warning threshold according to the inference result to obtain a construction safety early warning signal.

[0007] In a second aspect, the application provides a construction monitoring system based on artificial intelligence, which comprises:

[0008] The acquisition module is configured to collect the operation behavior of construction personnel in real time through a high-speed camera arranged on a construction operation surface, perform frequency filtering processing on original construction site data according to the vibration frequency spectrum of construction machinery, and obtain standardized construction behavior data.

[0009] The optimization module is configured to input the standardized construction behavior data into a construction behavior feature genetic algorithm, optimize the identification network through behavior feature coding of aerial operation and hoisting operation, and obtain a lightweight construction abnormality identification model.

[0010] The calculation module is configured to perform probability calculation processing on the construction abnormal behavior of not wearing a safety helmet and illegally entering a dangerous area according to the lightweight construction abnormality identification model, and obtain a construction abnormality probability value.

[0011] The compression module is configured to transmit the construction abnormality probability value to a construction site edge computing device, compress the model parameters through weight pruning, and obtain an inference result.

[0012] The hierarchical module is configured to perform hierarchical judgment processing on the construction abnormality probability through a three-level early warning threshold according to the inference result, and obtain a construction safety early warning signal.

[0013] In a third aspect, an artificial intelligence-based construction monitoring device is provided, comprising a memory and at least one processor, the memory having instructions stored therein; the at least one processor invoking the instructions in the memory to cause the artificial intelligence-based construction monitoring device to perform the artificial intelligence-based construction monitoring method described above.

[0014] In a fourth aspect, a computer-readable storage medium is provided, having instructions stored therein, which, when executed on a computer, cause the computer to perform the artificial intelligence-based construction monitoring method described above.

[0015] In the technical solution provided in the present application, the high-speed camera arranged on the construction operation surface is used to collect the construction personnel's operation behavior in real time, and the original construction site data is processed by frequency filtering in combination with the construction machinery vibration frequency spectrum, effectively solving the problem of unstable data quality in the traditional monitoring method. The standardized construction behavior data obtained significantly improves the recognition accuracy of the subsequent AI algorithm. The standardized construction behavior data is input into the construction behavior feature genetic algorithm, and the recognition network is optimized by behavior feature coding of high-altitude operation and hoisting operation, and a lightweight construction anomaly recognition model is obtained. Compared with the general AI model, the lightweight construction anomaly recognition model has stronger construction scene adaptability, greatly improved recognition accuracy and significantly reduced model complexity. According to the lightweight construction anomaly recognition model, the construction anomaly probability of not wearing a safety helmet and illegally entering a dangerous area is calculated and processed, and the construction anomaly probability value accurately reflects the actual risk level of the construction site, providing a reliable data basis for subsequent early warning decision-making. The construction anomaly probability value is transmitted to the construction site edge computing device, the model parameters are compressed by weight pruning, and the inference result avoids the network delay of cloud processing, realizes millisecond-level real-time response, and meets the timeliness requirement of construction safety monitoring. According to the inference result, the construction anomaly probability is processed by three-level early warning threshold grading determination, and the construction safety early warning signal obtained has hierarchy and pertinence, which not only avoids frequent false reports to interfere with normal construction, but also ensures the timely discovery and processing of key safety risks.

[0016] In the field of construction monitoring, the construction behavior feature genetic algorithm of the present application is specially designed for the particularity of high-altitude operation and hoisting operation, the fitness function of the algorithm comprehensively considers three key indicators of identification accuracy, reasoning speed and model size, so that the optimized network structure not only ensures high-precision identification of construction abnormal behavior, but also meets the resource limitation requirements of edge computing devices. The grading determination mechanism of the three-level early warning threshold is designed according to the construction safety specification and risk level, different early warning levels correspond to different response measures, from gentle prompt to forced shutdown, ensuring the rationality and effectiveness of the early warning response. The weight pruning algorithm removes the network connections with low contribution to the construction monitoring task on the premise of retaining the key network path of construction abnormality identification, and this targeted model compression strategy better balances the relationship between model performance and computing efficiency compared with general compression methods, providing technical support for real-time construction monitoring in the edge computing environment. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings.

[0018] Figure 1 An embodiment of the construction monitoring method based on artificial intelligence in the present application is shown in the figure.

[0019] Figure 2 An embodiment of the construction monitoring system based on artificial intelligence in the present application is shown in the figure.

[0020] Figure 3 The structure schematic diagram of the construction monitoring device based on artificial intelligence in the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0021] The embodiment of the present application provides a construction monitoring method and system based on artificial intelligence. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0022] For ease of understanding, the specific process of the embodiment of the present application is described below. Please refer to Figure 1 One embodiment of the construction monitoring method based on artificial intelligence in the embodiment of the present application comprises:

[0023] Step S101, the working behavior of the construction personnel is collected in real time by a high-speed camera arranged on the construction operation surface, the original construction site data is processed by frequency filtering according to the vibration frequency spectrum of the construction machinery, and standardized construction behavior data is obtained;

[0024] Step S102, the standardized construction behavior data is input into a construction behavior feature genetic algorithm, and the identification network is optimized by behavior feature coding of high-altitude operation and hoisting operation, to obtain a lightweight construction abnormality identification model;

[0025] Step S103, the construction abnormal behavior of not wearing a safety helmet and illegally entering a dangerous area is processed by probability calculation according to the lightweight construction abnormality identification model, to obtain a construction abnormality probability value;

[0026] Step S104, the construction abnormality probability value is transmitted to a construction site edge computing device, the model parameters are compressed by weight pruning, and an inference result is obtained;

[0027] Step S105, the construction abnormality probability is processed by three-level early warning threshold according to the inference result, to obtain a construction safety early warning signal.

[0028] It can be understood that the execution subject of the present application can be a construction monitoring system based on artificial intelligence, and can also be a terminal or a server, which is not limited here. The embodiment of the present application takes the server as the execution subject for example.

[0029] Specifically, the image sequence of the construction personnel's working behavior is collected by deploying a high-speed camera on the construction operation surface, and the vibration signal of the construction site is obtained by configuring an acceleration sensor, forming multi-modal original construction site data. The high-speed camera is set to a high-frequency sampling mode, continuously capturing image frames of the construction operation surface, and the acceleration sensor synchronously collects the vibration waveform signal of the construction site. In the data preprocessing stage, the collected data is divided into mechanical vibration frequency band and personnel activity frequency band according to the vibration spectrum characteristics of the construction machinery. The mechanical vibration frequency band mainly contains the vibration signals generated by the operation of heavy equipment such as tower cranes and excavators, and the personnel activity frequency band contains the low-frequency vibration generated by the walking and operating tools of the construction personnel. The adaptive band-pass filter separates and processes the signals according to the characteristic parameters of different frequency bands. The passband range of the filter is dynamically adjusted according to the actual equipment type of the construction site, effectively eliminating environmental noise and mutual interference between equipment. The image data is denoised to eliminate noise points caused by changes in light and dust, and the light compensation processing adjusts the brightness and contrast of the image according to the natural light conditions at different times. The vibration data is converted into amplitude-frequency feature vectors through fast Fourier transform. The amplitude reflects the vibration intensity, and the frequency reflects the vibration type. The preprocessed construction data set is classified according to the construction operation type labels such as high-altitude operation, lifting operation and ground operation. The data range is standardized to the standard interval through the Z-score standardization method, ensuring that different types of data have the same numerical scale.

[0030] The standardized construction behavior data is input into the construction behavior feature genetic algorithm for network optimization. The construction behavior feature genetic algorithm is specially designed for the construction scene, and its fitness function comprehensively considers the three key indicators of recognition accuracy, reasoning speed and model size. First, the action sequence of high-altitude workers and the operation trajectory of hoisting equipment are extracted from the standardized construction behavior data as feature samples. The action sequence of high-altitude workers includes continuous actions such as climbing, moving and operating, and the operation trajectory of hoisting equipment includes parameters such as the rotation angle of the boom and the lifting height of the hook. The construction behavior semantic analysis labels the feature samples according to the construction safety specifications, clearly distinguishing between normal operation behavior and abnormal violation behavior. The construction behavior feature genetic algorithm encodes the number of layers, the number of neurons in each layer, and the type of activation function into chromosomes, and the length of the chromosome is dynamically adjusted according to the complexity of the network. Each individual in the initial population represents a network structure configuration, and the population size is set to a moderate number to balance search efficiency and diversity. The semantic-guided crossover operation ensures that the network structure after crossover can still effectively extract construction behavior features, and the crossover point selection considers the functional relevance between network layers. The mutation operation based on construction safety constraints limits the range of network structure changes, preventing the generation of network configurations that do not meet the construction monitoring requirements. The optimized network structure is converted into specific neural network parameter configurations, including the initialization method of weight matrices and bias vectors. The construction abnormal behavior fitness function evaluates the overall performance of the network in the construction anomaly identification task, and the optimal network structure is gradually selected through multiple generations of evolutionary selection. The lightweight construction anomaly identification model is used to calculate the probability of construction abnormal behavior. The lightweight construction anomaly identification model uses a convolutional neural network architecture that is specifically designed for optimizing image features in the construction scene. The model input layer receives standardized construction behavior data, and the data format includes image pixel matrices and vibration feature vectors. The convolutional layer uses multiple convolution kernels of different sizes to extract features from the input image. Small-size convolution kernels capture detailed features such as safety helmet edge contours, and large-size convolution kernels capture overall features such as body posture. For the head region of construction personnel, the convolutional layer identifies the contour features of the safety helmet through edge detection operators and identifies the standard color of the safety helmet through color feature extractors. For body posture features, the convolutional layer locates the joint positions of the human body through key point detection algorithms and analyzes the limb angles and motion trajectories. The feature map generated by convolutional operation retains spatial position information and feature intensity information, forming a construction personnel behavior feature map. The pooling layer reduces the dimension of the feature map, with max pooling retaining the most significant feature response and average pooling calculating the average feature intensity in the region. The safety helmet wearing state feature is compressed into a fixed-length vector through the pooling layer, and the personnel position coordinate feature retains accurate spatial positioning information. The spatial relationship between the dangerous area boundary coordinates and the personnel position coordinates is calculated using the geometric distance algorithm, which calculates the shortest distance from the personnel to the dangerous area boundary and determines whether there is a region crossing behavior.The spatial relationship calculation result is fused with other features to form a construction abnormal behavior feature vector. A fully connected layer performs nonlinear transformation on the feature vector, and a multilayer perceptron structure maps the high-dimensional features to an abnormal probability space. The ReLU function is used to process the hidden layer output, and the Sigmoid function maps the output to the probability interval. The probability of not wearing a safety helmet and the probability of entering a dangerous area in violation of regulations correspond to different output neurons, and the Softmax function normalizes the probabilities of multiple abnormal behaviors to ensure that the sum of all probability values is 1, and outputs a comprehensive construction abnormality probability value.

[0031] The construction abnormality probability value is transmitted to the edge computing device through the construction site wireless communication network. The construction site wireless communication network uses 5G or WiFi6 protocol to ensure the real-time and reliability of data transmission. The edge computing device is equipped with a dedicated AI processing chip and sufficient memory space, and the network parameters of the lightweight construction abnormality recognition model are loaded into the device memory in binary format. The network parameters include convolution kernel weights, fully connected layer weights, bias parameters, etc., and the parameter loading process uses memory mapping technology to ensure loading efficiency. Based on the weight distribution analysis of the importance of construction behavior, the network connections with low contribution to construction abnormality recognition are identified, and the weight pruning algorithm judges the importance of the connection according to the size of the weight value and the gradient information. Low-contribution network connections are marked as redundant connections and removed from the network, and the remaining network paths ensure that the model can still accurately identify key construction abnormal behaviors. The compressed model parameter set significantly reduces the storage space requirement and computational complexity. The floating-point parameters are converted to 8-bit or 16-bit integer parameters through quantization algorithms, and the quantization process uses linear quantization or nonlinear quantization strategies, and selects the appropriate quantization method according to the parameter distribution characteristics. The parallel reconstruction of the model structure changes the original serial calculation process to a parallel computing mode, making full use of the multi-core processing capability of the edge device. The edge-optimized model performs local inference calculation on the construction abnormality probability values received in real time, and the inference process includes forward propagation, activation function calculation, output layer processing, etc. The abnormality determination result is output through the forward propagation of the model.

[0032] Based on the inference results, a three-level warning decision is made. The inference results include the specific probability and confidence level of the construction anomaly. The three-level warning thresholds are determined based on construction safety standards and historical accident data. Level 1 warning corresponds to low-risk anomalies, level 2 warning corresponds to medium-risk anomalies, and level 3 warning corresponds to high-risk anomalies. The construction anomaly probability is compared with the preset threshold using an interval determination method, and the initial warning level is determined based on the threshold interval within which the anomaly probability falls. Bayesian decision theory calculates the posterior probability of each warning level using prior probabilities and likelihood probabilities. The prior probability is derived from a statistical analysis of historical construction anomaly events, while the likelihood probability reflects the reliability of the current observation data. The probability confidence calculation for level 1, level 2, and level 3 warnings takes into account the uncertainty of the model prediction and environmental interference factors. The uncertainty factor is quantified using information entropy theory to reflect the reliability of the warning decision. Alarm time series correlation analysis establishes a trend model for the anomaly probability within a time window. Sliding window statistics and trend fitting algorithms are used to identify upward or downward trends in the anomaly probability. When a sustained upward trend in the anomaly probability is detected, the warning system automatically lowers the warning threshold to enhance the sensitivity of the warning. Dynamic early warning determinations comprehensively consider current abnormality probabilities, historical trends, and environmental factors to inform early warning decisions. Early warning signal format conversion converts digitized early warning results into corresponding control commands. The construction safety communication protocol defines corresponding response measures for different warning levels. Level 1 warnings use a gentle tone to remind personnel to be vigilant. Level 2 warnings use an audible and visual alarm, combining a warning light and audible alarm to draw attention. Level 3 warnings use a forced shutdown command to directly cut off power to equipment or halt hazardous operations to ensure personnel safety.

[0033] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0034] The high-speed camera is set to high-frequency sampling mode to continuously capture images of the construction work surface. At the same time, the acceleration sensor is used to obtain the vibration signal of the construction site to obtain multi-modal raw construction site data.

[0035] Based on the spectral characteristics of construction machinery vibration, the multimodal raw construction site data is divided into mechanical vibration frequency band and human activity frequency band. The signals of different frequency bands are separated and processed through adaptive bandpass filters to obtain the construction signal combination after frequency division.

[0036] The image data in the frequency-divided construction signal combination is subjected to denoising and illumination compensation processing, and the vibration data is converted into amplitude-frequency feature vectors to obtain the preprocessed construction data set.

[0037] The pre-processed construction dataset is normalized and standardized according to the construction operation type label, and the data range is unified to a standard interval through a standardization method to obtain standardized construction behavior data.

[0038] Specifically, when the high-speed camera in the high-frequency sampling mode continuously collects images of the construction site, the image sensor of the camera captures pixel data of the construction site at fixed time intervals. Each frame of image contains a pixel matrix of RGB three color channels, and the pixel value range is usually between 0 and 255. The acceleration sensor detects the mechanical vibration of the construction site through the built-in piezoelectric ceramic element. The analog voltage signal output by the sensor is converted into a digital signal by an analog-to-digital converter, forming a time series of acceleration values. The multi-modal raw construction site data includes the pixel matrix of the image data and the time series of the vibration data. The two data types are collected synchronously on the time axis, and the correspondence between the data is established through the time stamp.

[0039] The construction machinery vibration spectrum feature analysis converts the time-domain vibration signal into a frequency-domain signal through fast Fourier transform, and the spectrum analysis result shows the characteristic frequency distribution of different construction equipment. The vibration frequency generated by the operation of the tower crane is concentrated in a specific frequency band, the vibration frequency generated by the operation of the excavator is distributed in another frequency band, and the vibration frequency generated by the walking and operating tools of the construction personnel is relatively low. The adaptive band-pass filter divides the multi-modal data into a mechanical vibration frequency band and a personnel activity frequency band according to the preset frequency threshold. The passband parameters of the filter are dynamically adjusted according to the type of equipment in the current construction site. The mechanical vibration frequency band retains the high-frequency vibration signal generated by the operation of heavy equipment, and the personnel activity frequency band retains the low-frequency human activity signal. The signal combination after filtering processing includes separated image data stream and corresponding vibration data stream.

[0040] The image denoising processing adopts a median filter algorithm to eliminate random noise. The algorithm replaces the original pixel value by calculating the median value in the pixel neighborhood, effectively removing salt and pepper noise and impulse interference. The illumination compensation processing adjusts the brightness distribution of the image through histogram equalization. The algorithm calculates the cumulative distribution function of the image and remaps the pixel value to a uniformly distributed brightness range. The vibration data is converted into a frequency-domain representation through fast Fourier transform. The amplitude component reflects the intensity of the vibration, and the frequency component reflects the periodic characteristics of the vibration. The amplitude-frequency feature vector converts the time-domain vibration signal into a frequency-domain feature representation. The pre-processed construction dataset includes the denoised and illumination compensated image matrix and the amplitude-frequency feature vector. The data format is unified as a numerical matrix to facilitate subsequent algorithm processing.

[0041] The construction operation type label is defined according to construction safety specifications and operation process, including labels of categories such as high-altitude operation, hoisting operation, ground operation, and welding operation. Normalization converts data of different dimensions into dimensionless data, and the image pixel value is divided by the maximum pixel value to obtain a normalized value between 0 and 1, and the vibration amplitude is divided by the maximum vibration amplitude to obtain a corresponding normalized value. The standardization processing adopts the Z-score method to calculate the mean and standard deviation of the data, and each data point is subtracted from the mean and divided by the standard deviation to obtain a standardized value. The standardized data have a mean of 0 and a standard deviation of 1. Standard interval unification maps all feature data into the same numerical range, eliminating the dimensional differences and numerical scale differences between different features, and the standardized construction behavior data form a data matrix in a unified format.

[0042] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0043] The action sequence of the high-altitude operation personnel and the operation trajectory of the hoisting equipment in the standardized construction behavior data are taken as feature samples, and the feature samples are labeled by behavior type through construction behavior semantic analysis to obtain a labeled construction behavior feature sample set;

[0044] Based on the safety helmet wearing state feature and the dangerous area position feature in the construction behavior feature sample set, the number of layers and the number of neurons of the neural network are encoded through the construction behavior feature genetic algorithm to obtain a construction anomaly recognition network chromosome structure;

[0045] According to the construction anomaly recognition network chromosome structure, the chromosome gene fragments are recombined through construction behavior semantic guided crossover operation, and the chromosome is mutated based on the construction safety constraint condition to obtain an optimized network structure code;

[0046] The optimized network structure code is converted into specific neural network parameter configuration, the network performance is evaluated through the construction abnormal behavior fitness function to obtain a construction anomaly recognition performance score;

[0047] Based on the construction anomaly recognition performance score, multiple generations of evolution selection are performed, and the network structure with high fitness is retained and the low-performance structure is eliminated to perform iterative optimization processing, and a lightweight construction anomaly recognition model is obtained.

[0048] Specifically, the high-altitude worker action sequence extraction in standardized construction behavior data identifies the continuous actions of construction workers during high-altitude work such as scaffolding and tower cranes through image sequence analysis. The action sequence includes time-dependent behavior patterns such as climbing, moving, and operating tools. The hoisting equipment operation trajectory extraction tracks the movement path of the tower crane jib and hook, records the position changes and motion parameters of the equipment in three-dimensional space. The construction behavior semantic analysis semantically labels the extracted feature samples according to construction safety regulations and work standards, classifies the action sequence and operation trajectory into normal work behavior and potential abnormal behavior, and the behavior type labeling includes safety regulation work, minor violation, and serious violation category labels. The labeled construction behavior feature sample set forms the training data for supervised learning. The hard hat wearing state feature extracts the color, shape, and texture features of the construction worker's head region from the image, and quantifies the existence and standardization of the hard hat through feature descriptors. The dangerous area position feature defines the boundary of the dangerous area in the construction site through a spatial coordinate system, and calculates the relative distance and spatial relationship between the personnel position and the dangerous area. The construction behavior feature genetic algorithm encodes the neural network structure parameters as chromosomes, with each gene site corresponding to a network structure parameter. The network layer is encoded as an integer sequence, the number of neurons is encoded as an integer value corresponding to each layer, and the activation function type is encoded as an enumeration value. The construction anomaly recognition network chromosome structure contains complete network architecture information, and the chromosome length is determined according to the network complexity. The gene encoding uses binary or real number encoding to represent network parameters.

[0049] The construction behavior semantic-oriented crossover operation ensures that the network structure after crossover still has construction anomaly recognition capability. The crossover point selection considers the functional dependence relationship between network layers to avoid producing invalid network configurations. The chromosome gene segment recombination is performed through single-point crossover, multi-point crossover, or uniform crossover, and the crossover probability is dynamically adjusted according to the fitness of the parent chromosomes. The construction safety constraint condition limits the variation range of the network structure to ensure that the mutated network still meets the actual needs of construction monitoring. The mutation operation includes random change of gene sites and local search, and the mutation probability is adaptively adjusted with the evolution generation. The optimized network structure code is checked and repaired to ensure that the corresponding network structure is computationally feasible and effective in construction applications.

[0050] The network structure encoding conversion converts the abstract chromosome representation into a specific neural network parameter configuration, including the dimension of the weight matrix, the length of the bias vector, the inter-layer connection mode, and other detailed specifications. The construction abnormal behavior fitness function comprehensively evaluates the accuracy, recall rate, reasoning speed, and model complexity of the network in the construction anomaly identification task. The fitness calculation considers the generalization ability and robustness of the network in different construction scenarios. The performance evaluation tests the network performance on multiple construction datasets through cross-validation methods. The construction anomaly identification performance score reflects the comprehensive ability index of the network. The multi-generation evolution selection adopts the elite reservation strategy and roulette wheel selection method. The network structure with high fitness is directly reserved to the next generation, and the remaining individuals are randomly selected according to the fitness ratio. The elimination mechanism removes the network structures with low fitness to release computing resources for exploring better network configurations. Iterative optimization gradually improves the network structure by repeatedly performing selection, crossover, and mutation operations. The evolution process monitors the convergence of the fitness and the population diversity. When the fitness no longer significantly improves, the evolution process is terminated. The lightweight construction anomaly identification model is selected from the final evolved network structure. The model structure minimizes the number of parameters and computational complexity while ensuring identification accuracy.

[0051] In a specific embodiment, the process of performing step S103 can specifically include the following steps:

[0052] The input layer of the lightweight construction anomaly identification model receives standardized construction behavior data, and the convolutional layer performs feature extraction processing on the head region and body posture features of the construction personnel to obtain a construction personnel behavior feature spectrum.

[0053] Based on the construction personnel behavior feature spectrum, the pooling layer performs dimensionality reduction processing on the safety helmet wearing state feature and the personnel position coordinate feature, and simultaneously performs spatial relationship calculation processing on the dangerous area boundary coordinates and the personnel position coordinates to obtain a construction abnormal behavior feature vector.

[0054] The construction abnormal behavior feature vector is input into the fully connected layer for non-linear transformation processing, and the activation function is used to map the probability of not wearing a safety helmet and the probability of entering a dangerous area illegally to obtain a construction abnormal behavior probability distribution.

[0055] The construction abnormal behavior probability distribution is normalized by the Softmax function to convert the multi-class abnormal behavior probability into a single construction abnormal comprehensive probability value, and a construction abnormal probability value is obtained.

[0056] Specifically, when the input layer of the lightweight construction anomaly identification model receives standardized construction behavior data, the data format includes an image pixel matrix and a corresponding vibration feature vector. The input layer rearranges these multi-dimensional data into a tensor format required by the neural network. The convolutional layer uses multiple convolution kernels of different sizes to filter the input image. Each convolution kernel is used to detect specific visual features. For head region feature extraction, the convolutional layer uses small-size convolution kernels to detect the edge contour and color features of the safety helmet. The edge detection operator is used to identify the boundary lines between the head and the safety helmet. For body posture feature extraction, the convolutional layer uses large-size convolution kernels to capture the overall contour and joint position of the human body. The feature response map is used to locate key body parts such as shoulders, arms, and waists. The output of the convolution operation forms a multi-layer feature map, each layer of which retains visual feature information at different levels of abstraction. Low-level feature maps contain basic features such as edges and textures, while high-level feature maps contain complex features such as object shapes and semantic information. These feature maps combine to form a construction worker behavior feature map.

[0057] The pooling layer reduces the dimensionality of the construction worker behavior feature map by using max pooling or average pooling algorithms to compress the spatial size of the feature map. Max pooling selects the maximum feature value within the pooling window, while average pooling calculates the arithmetic mean of the feature values within the pooling window. The safety helmet wearing state feature is compressed into a fixed-length feature vector by the pooling layer. Each element of the feature vector corresponds to a different attribute feature of the safety helmet, including color intensity, shape matching degree, position accuracy, and other quantitative indicators. The personnel position coordinate feature retains spatial positioning information through the pooling layer. The coordinate data represents the specific position of the personnel in the image in the form of pixel position. The spatial relationship between the dangerous area boundary coordinates and the personnel position coordinates is calculated using the Euclidean distance formula, which calculates the shortest straight-line distance from the personnel's current position to the dangerous area boundary and determines whether the personnel have crossed the predefined safety boundary line. The spatial relationship calculation result is concatenated with the pooled feature vector to form a comprehensive feature representation that includes visual features and spatial relationships. This comprehensive feature representation is the construction anomaly behavior feature vector.

[0058] After the fully connected layer receives the construction abnormal behavior feature vector, it performs linear transformation on the input features through the weight matrix and bias vector. The linear transformation maps the high-dimensional feature space to the decision space, and each neuron corresponds to a specific abnormal behavior pattern. The non-linear transformation introduces a non-linear factor through the activation function, which maps the output of the linear transformation to a specified numerical range. Common activation functions include ReLU function, Sigmoid function, Tanh function, etc. The calculation of the probability of not wearing a safety helmet is completed by a special output neuron, whose input weight has learned the importance of safety helmet-related features through training. The neuron output value reflects the possibility of not wearing a safety helmet in the current input data. The calculation of the probability of violating the safety regulations in the dangerous area is completed by another output neuron, which is specifically designed to process spatial position-related features. The output value reflects the possibility of violating safety regulations in the dangerous area. The activation function maps the original output value of the neuron to a probability range between zero and one, forming a construction abnormal behavior probability distribution.

[0059] When the Softmax function normalizes the construction abnormal behavior probability distribution, it first calculates the exponential values of all class probabilities, and then divides the exponential value of each class by the sum of all class exponential values to ensure that the normalized probability distribution sums to one. The multi-class abnormal behavior probability includes the probability values of not wearing a safety helmet, violating safety regulations in a dangerous area, and improper use of equipment. Softmax normalization ensures that these probability values are within the range of zero to one and the sum is one. The construction abnormal comprehensive probability value is calculated by weighted summation. Different abnormal types are assigned different weight coefficients according to their severity. Serious safety violations are assigned higher weights, and minor violations are assigned lower weights. The final output construction abnormal probability value represents the comprehensive possibility of any type of abnormal behavior occurring in the current construction scene.

[0060] In a specific embodiment, the process of performing step S104 can specifically include the following steps:

[0061] The construction abnormal probability value is transmitted to the edge computing device through the construction site wireless communication network, and the network parameters of the lightweight construction abnormality recognition model are loaded into the edge device memory to obtain an edge device model configuration;

[0062] Based on the construction behavior importance weight distribution in the edge device model configuration, the low-contribution network connections are removed through the weight pruning algorithm, while the key network paths for construction abnormality recognition are retained, to obtain a compressed model parameter set;

[0063] The floating-point parameters in the compressed model parameter set are converted to integer parameters through the quantization algorithm, and the model structure is reconstructed in parallel to obtain an edge-optimized model;

[0064] The edge optimization model is used to locally infer and calculate the construction abnormality probability value received in real time, and the abnormality determination result is output through model forward propagation to obtain the inference result.

[0065] Specifically, when the construction abnormality probability value is transmitted to the edge computing device through the construction site wireless communication network, the data is transmitted in the form of a digital signal through the WiFi6 or 5G network protocol, and the transmission process includes data encapsulation, routing selection, transmission confirmation and other communication steps. The construction site wireless communication network uses the TCP / IP protocol stack to ensure the reliability of data transmission, and the data packet carries a timestamp and a sequence number during transmission. The edge computing device performs integrity checking and sorting processing after receiving the data packet. The network parameters of the lightweight construction abnormality identification model include weight matrix, bias vector, activation function parameters and other model configuration information, which are stored in the memory of the edge device in the form of a binary file. The network parameter loading process uses memory mapping technology to directly map the parameter file to the device memory address space, avoiding repeated data copy operations. The edge device model configuration includes complete network structure information and trained parameter values, which prepares for subsequent model optimization and inference calculation. The construction behavior importance weight distribution is obtained by analyzing the contribution of each connection in the network to the construction abnormality identification task. The weight distribution reflects the importance difference of different network connections in identifying abnormal behaviors such as safety helmet wearing and dangerous area intrusion. The weight pruning algorithm judges the importance of network connections according to the absolute value of the weight value and the gradient information. The connections with smaller weight absolute values have less impact on the model output and are marked as low-contribution connections. The pruning operation removes low-contribution network connections from the model, which is specifically manifested as setting the corresponding weight parameters to zero and updating the connection matrix structure of the network. The key network paths of construction abnormality identification include convolution kernel connections sensitive to safety helmet features and fully connected layer connections sensitive to spatial positions, which are retained during the pruning process. The compressed model parameter set significantly reduces the number of original model parameters, reducing the storage space requirement and computational complexity of the model, while maintaining the ability to identify construction abnormal behaviors.

[0066] The quantization algorithm converts the thirty-two-bit floating-point parameters in the compressed model parameter set into eight-bit or sixteen-bit integer parameters. The quantization process uses a linear quantization method to map floating-point numbers to integer space by determining the quantization range and quantization step. The calculation process of linear quantization includes determining the maximum and minimum values of the floating-point numbers, calculating the quantization range, and then dividing each floating-point number by the quantization step after subtracting the minimum value to obtain the corresponding integer value. Integer parameters occupy less storage space than floating-point parameters, and integer operations on edge devices are faster than floating-point operations. The parallel reconstruction of the model structure reorganizes the originally serially executed network layers into a parallel computing mode. Parallel computing is used between multiple convolution kernels of the convolution layer, and the matrix operations of the fully connected layer are decomposed into parallel operations of multiple sub-matrices. Parallel reconstruction fully utilizes the parallel computing capabilities of multi-core processors and dedicated AI acceleration chips on edge devices. The edge-optimized model significantly improves the inference speed while maintaining recognition accuracy.

[0067] When the edge-optimized model performs local inference calculation on the real-time received construction anomaly probability values, the input data is first loaded into the input layer of the model, and then forward propagation calculation is performed in sequence according to the network structure. Local inference calculation avoids communication delay with the cloud server, and all calculation processes are completed inside the edge device, with inference delay controlled within milliseconds. Model forward propagation includes convolution operations, activation function calculations, pooling operations, and fully connected layer matrix operations. Each step is executed in parallel on the processor of the edge device. The abnormality determination result is obtained through the model output layer. The values of the output layer are determined by threshold judgment to determine the abnormality category and risk level. The inference result includes abnormal type identification, confidence score, risk level evaluation, and other information.

[0068] For example, the hoist basket operation area of the outer facade construction site of a high-rise building, the construction anomaly probability value includes the probability value of workers not wearing safety ropes and operating the hoist basket controller in violation of regulations. The data is transmitted to the edge computing device installed on the construction elevator through the 5G base station deployed on the site. The edge computing device adopts an eight-core processor with ARM architecture and a dedicated neural network acceleration chip. The device memory capacity is eight GB, and the storage capacity is one hundred and twenty-eight GB. The network parameter file size of the lightweight construction anomaly identification model is fifty MB, including convolution layer weight matrix, fully connected layer weight matrix, bias vector and other parameter data. The parameter loading process uses direct memory access technology to map the parameter file to a specific address region in the device memory. The construction behavior importance weight distribution analysis found that the convolution kernel weight absolute value responsible for detecting safety rope features is larger, indicating that these connections have a higher contribution to the safety rope identification task, while the convolution kernel weight absolute value responsible for background texture identification is smaller, marked as low contribution connections. The weight pruning algorithm removes network connections with weight absolute values less than a certain threshold. The number of model parameters after pruning is reduced from two million to one hundred and twenty thousand, and the sparsity of network connections is significantly increased. The quantization algorithm converts the remaining floating-point weight parameters to sixteen-bit integer parameters. The quantization process first analyzes the numerical value distribution range of each layer of parameters, then determines the appropriate quantization step, and finally maps each floating-point weight to the corresponding integer value. Parallel reconstruction decomposes the originally sequentially executed convolution operation into multiple parallel sub-convolution operations, and divides the large matrix operation of the fully connected layer into multiple parallel operations of small matrices, fully utilizing the eight CPU cores and the parallel computing units of the neural network acceleration chip. After receiving the real-time data containing the current work state of the worker, the edge optimization model performs forward propagation calculation locally. The convolution layer extracts the worker's body posture and safety equipment wearing features, the pooling layer compresses the feature dimension, and the fully connected layer performs abnormal behavior classification calculation. Finally, the high-risk judgment result of not wearing safety ropes and the medium-risk judgment result of illegal operation are output. The entire inference process is completed on the edge device, and the inference delay is controlled within three milliseconds.

[0069] In a specific embodiment, the process of performing step S105 can specifically include the following steps:

[0070] The construction anomaly probability value in the inference result is compared with the preset three-level warning threshold value, the warning level is preliminarily divided according to the threshold interval in which the anomaly probability is located, and a warning level classification result is obtained;

[0071] Based on the warning level classification result, the probability confidence of the first-level warning, the second-level warning and the third-level warning is calculated by the Bayesian decision theory, and the uncertainty factor is also included in the warning decision process for comprehensive evaluation, and a warning level confidence value is obtained;

[0072] According to the early warning level confidence value, the abnormal probability change trend in the continuous time window is dynamically tracked through early warning timing correlation analysis, and the early warning threshold is adaptively adjusted when the abnormal probability shows an upward trend, to obtain a dynamic early warning decision result.

[0073] The dynamic early warning decision result is converted into a corresponding early warning signal format, and the sound prompt, sound and light alarm and forced shutdown instruction are triggered in stages through the construction safety communication protocol to obtain a construction safety early warning signal.

[0074] Specifically, when the construction abnormal probability value in the inference result is compared with the preset three-level early warning threshold, the comparison algorithm uses an interval determination method to determine the threshold range to which the probability value belongs. The three-level early warning threshold includes a first-level early warning threshold, a second-level early warning threshold and a third-level early warning threshold, each threshold corresponding to a different risk level. The first-level early warning corresponds to a low-risk anomaly, the second-level early warning corresponds to a medium-risk anomaly, and the third-level early warning corresponds to a high-risk anomaly. The numerical comparison process determines in which threshold interval the construction abnormal probability value falls through the size relationship judgment. The comparison result is directly mapped to the corresponding early warning level, and the threshold interval in which the abnormal probability is located determines the preliminary early warning level classification. The early warning level classification result includes an early warning type identifier and an initial risk level assessment. The Bayesian decision theory evaluates the reliability of the early warning level through the calculation of prior probability, likelihood probability and posterior probability. The prior probabilities of the first-level early warning, the second-level early warning and the third-level early warning are derived based on historical construction safety statistical data, reflecting the frequency of occurrence of different early warning levels on the construction site. The likelihood probability reflects the possibility of the current observation data under a specific early warning level, which is calculated by analyzing the correlation between the construction anomaly characteristics and the early warning level. The posterior probability combines the prior probability and the likelihood probability through the Bayesian formula to calculate the true probability of each early warning level under the current observation condition. The uncertainty factor is quantified through the information entropy theory, reflecting the degree of uncertainty in the early warning decision process, which is related to factors such as model prediction confidence, environmental interference intensity and data quality. The comprehensive evaluation process weights and fuses the posterior probability and the uncertainty factor, and calculates the early warning level confidence value through weighted average or weighted product.

[0075] The early warning timing correlation analysis establishes a dynamic monitoring mechanism for the abnormal probability in the time window, and continuously collects abnormal probability values in a certain time range through a sliding window technique. The continuous time window contains abnormal probability observation values at multiple time points, and the length of the time window is determined according to the characteristics of the construction operation and the duration of the abnormal behavior. The trend of the abnormal probability is identified through a trend analysis algorithm, which includes linear regression fitting, moving average calculation, and difference sequence analysis. Dynamic tracking processes monitor the change direction and amplitude of the abnormal probability in real time, and triggers the adaptive adjustment of the early warning threshold when a sustained upward trend in the abnormal probability is detected. The adaptive adjustment process dynamically modifies the early warning threshold according to the trend strength and change rate, and reduces the early warning threshold to increase the early warning sensitivity when the upward trend is strong, and appropriately increases the early warning threshold to reduce false positives when the downward trend is strong. The dynamic early warning decision result considers the current abnormal probability, historical trend and adjusted early warning threshold to form an early warning decision.

[0076] The early warning signal format conversion converts the digital dynamic early warning decision result into specific control instructions and communication messages. The construction safety communication protocol defines the signal format and transmission method corresponding to different early warning levels, and the protocol includes structured information such as message header, data segment, and check code. The sound prompt corresponds to a first-level early warning signal, and uses a gentle buzzer sound or voice prompt, with signal frequency and volume set to a level that does not interfere with normal construction operations. The sound and light alarm corresponds to a second-level early warning signal, combining a flashing warning light and a loud alarm sound, the warning light uses orange or yellow light source, and the alarm sound uses intermittent high pitch. The forced shutdown instruction corresponds to a third-level early warning signal, directly sending a device control instruction to shut down the power of related devices or stop dangerous operations, and the shutdown instruction is implemented through digital signals or relay control. The hierarchical triggering process automatically selects the corresponding signal type and intensity according to the early warning level, ensuring that the early warning signal matches the risk level, and the construction safety early warning signal is simultaneously published through various means such as on-site public address equipment, LED display screen, and device controller.

[0077] In a specific embodiment, the process of performing step S106 can specifically include the following steps:

[0078] The first-level early warning, the second-level early warning, and the third-level early warning in the early warning level classification result are taken as Bayesian prior probabilities, and the occurrence frequency of each level of early warning is calculated and processed through probability distribution based on historical construction abnormal event statistical data, to obtain the prior probability distribution of the early warning level;

[0079] Based on the prior probability distribution of the early warning level and the current construction abnormal probability value, the posterior probability of each early warning level is updated and calculated through the Bayesian formula, and the conditional probability is modified and processed by introducing the construction site environment factor and the operation type factor, to obtain the posterior probability value of the early warning level;

[0080] According to the early warning level posterior probability value, the uncertainty degree of early warning decision is quantitatively calculated and processed through information entropy theory, the model prediction confidence is weighted and fused with the environmental interference factor to obtain an early warning uncertainty factor;

[0081] The early warning level posterior probability value and the early warning uncertainty factor are comprehensively evaluated and processed, the early warning levels meeting the reliability requirements are confirmed through confidence threshold screening to obtain early warning level confidence values.

[0082] Specifically, when the first-level early warning, the second-level early warning and the third-level early warning in the early warning level classification result are the Bayesian prior probability, the prior probability reflects the basic occurrence frequency of each early warning level in the construction site. The historical construction abnormal event statistical data are obtained by collecting construction safety records, accident reports, violation behavior records and other data sources in a certain period of time in the past, and the statistical data contain detailed information such as the type, severity, occurrence time and processing result of abnormal events. The occurrence frequency of each level of early warning is obtained by calculating the ratio of the number of occurrences of each early warning level to the total number of early warnings in a specific time range. The frequency calculation adopts the relative frequency statistical method, and the historical data are classified and counted according to the early warning level. The probability distribution calculation process converts the statistical frequency value into a probability form. The specific calculation method is to divide the number of occurrences of each early warning level by the sum of the number of occurrences of all early warning levels, and then take the quotient as the prior probability of the early warning level. The sum of the prior probabilities of all early warning levels is ensured to be equal to one, and the prior probability distribution of the early warning level establishes a basic probability framework for subsequent Bayesian inference calculation. The Bayesian formula calculates the posterior probability by combining the prior probability and the likelihood probability. The specific calculation form of the Bayesian theorem is that the posterior probability is equal to the likelihood probability multiplied by the prior probability and then divided by the marginal probability. The likelihood probability reflects the possibility of each early warning level under the condition of the current construction abnormal probability value. The calculation of the likelihood probability is determined by analyzing the similarity between the current abnormal probability value and the historical abnormal probability value corresponding to each early warning level. The current construction abnormal probability value is input into the Bayesian formula as the observation evidence, the marginal probability is obtained by multiplying the likelihood probability of all early warning levels by the corresponding prior probability and then summing, and the posterior probability calculation result needs to be normalized to ensure the effectiveness of the probability distribution. The construction site environmental factors include illumination conditions, weather conditions, noise levels, equipment interference and other environmental variables that affect monitoring accuracy. The work type factor includes characteristic parameters of different construction activities such as high-altitude work, lifting work, welding work and ground work. The conditional probability correction process adjusts the posterior probability according to the current environmental factors and work type factors. The correction calculation adopts the multiplication formula of conditional probability, quantifies the influence degree of the environmental factors and the work type factors as correction coefficients, multiplies the correction coefficients with the original posterior probability to obtain the corrected probability value, and the early warning level posterior probability value after correction more accurately reflects the real risk level under the current construction scene.

[0083] The information entropy theory quantifies the degree of uncertainty by calculating the entropy value of the probability distribution. The calculation method of Shannon entropy is to multiply the posterior probability value of each warning level by the logarithm of the probability value, and then sum the calculation results of all warning levels and take the negative value. The degree of uncertainty of the warning decision reflects the reliability of the warning result. When the entropy value is close to zero, it indicates that the probability distribution is highly concentrated on a certain warning level, and the degree of uncertainty is low. When the entropy value is large, it indicates that the probability distribution is relatively uniform, and the degree of uncertainty is high. The quantitative calculation process converts the abstract concept of uncertainty into a specific numerical indicator. The numerical range is usually between zero and the logarithm base number multiplied by the number of warning levels. The calculation result provides a quantitative evaluation of the reliability of the warning decision. The model prediction confidence reflects the credibility of the AI model for the current prediction result. The confidence calculation is determined by analyzing factors such as the sharpness of the probability distribution of the model output layer, the activation intensity of the feature extraction layer, and the stability of the network weights. The environmental interference factor quantifies the influence of environmental conditions on monitoring accuracy. The interference factor calculation considers specific environmental quality indicators such as image contrast, signal-to-noise ratio, and occlusion area proportion. The weighted fusion process linearly combines the model prediction confidence and the environmental interference factor according to the preset weight ratio. The fusion calculation uses the weighted average method. The confidence is multiplied by its weight coefficient, the interference factor is multiplied by its weight coefficient, and then the two weighted results are added to obtain the final uncertainty factor. The warning uncertainty factor comprehensively reflects the comprehensive influence of various sources of uncertainty in the warning decision process.

[0084] The comprehensive evaluation process performs mathematical operations on the posterior probability value of the warning level and the warning uncertainty factor. The evaluation calculation uses the confidence calculation formula. The posterior probability value is multiplied by one minus the uncertainty factor to obtain the final confidence value considering the influence of uncertainty. The confidence threshold screening sets the minimum acceptable confidence level. The threshold is usually determined according to construction safety requirements and false alarm tolerance. Only the warning level with a confidence value exceeding the threshold is considered to meet the reliability requirements. The reliability requirement confirmation process determines the size relationship between the calculated confidence value and the preset threshold. The comparison operation uses a simple numerical size comparison. The warning level that meets the requirements enters the final decision process, and the warning level that does not meet the requirements is filtered out or marked as a low-confidence result. The warning level confidence value is the final output, which contains both the classification information of the warning level and the quantitative indicator of the confidence of the classification result. The confidence value range is usually between zero and one. The higher the value, the more reliable the warning result.

[0085] The above describes the construction monitoring method based on artificial intelligence in the embodiments of the present application. The construction monitoring system based on artificial intelligence in the embodiments of the present application is described below. Please refer to Figure 2 An embodiment of the construction monitoring system based on artificial intelligence in the embodiments of the present application includes:

[0086] The acquisition module is used for collecting the construction personnel operation behavior in real time through a high-speed camera arranged on the construction operation surface, performing frequency filtering processing on the original construction site data according to the construction machinery vibration frequency spectrum, and obtaining standardized construction behavior data.

[0087] The optimization module is used for inputting the standardized construction behavior data into a construction behavior feature genetic algorithm, performing optimization processing on the identification network through behavior feature coding of high-altitude operation and hoisting operation, and obtaining a lightweight construction abnormality identification model.

[0088] The calculation module is used for performing probability calculation processing on the construction abnormal behavior of not wearing a safety helmet and illegally entering a dangerous area according to the lightweight construction abnormality identification model, and obtaining a construction abnormality probability value.

[0089] The compression module is used for transmitting the construction abnormality probability value to a construction site edge computing device, performing compression processing on the model parameters through weight pruning, and obtaining an inference result.

[0090] The grading module is used for performing grading determination processing on the construction abnormality probability through a three-level early warning threshold according to the inference result, and obtaining a construction safety early warning signal.

[0091] The above Figure 2 The construction monitoring system based on artificial intelligence in the embodiment of the application is described in detail from the perspective of a modular functional entity, and the construction monitoring device based on artificial intelligence in the embodiment of the application is described in detail from the perspective of hardware processing.

[0092] Referring Figure 3 , the embodiment of the application also provides a construction monitoring device based on artificial intelligence. The construction monitoring device based on artificial intelligence can be a server, and its internal structure can be as shown in Figure 3 The construction monitoring device based on artificial intelligence includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. The processor of the computer is used to provide calculation and control capabilities. The memory of the construction monitoring device based on artificial intelligence includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the construction monitoring device based on artificial intelligence is used to store the corresponding data in the embodiment. The network interface of the construction monitoring device based on artificial intelligence is used to communicate with the external terminal through network connection. The computer program is executed by the processor to realize the above method.

[0093] Those skilled in the art can understand Figure 3The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the artificial intelligence-based construction monitoring device to which the scheme of the present application is applied.

[0094] The present application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium, and can also be a volatile computer readable storage medium, and the computer readable storage medium stores instructions, and when the instructions run on a computer, the computer executes the steps of the artificial intelligence-based construction monitoring method.

[0095] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, system and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0096] The integrated unit, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical scheme of the present application or the part that essentially contributes to the prior art or the whole or part of the technical scheme can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing an artificial intelligence-based construction monitoring device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0097] The above embodiments are only used to illustrate the technical scheme of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical scheme recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical scheme deviate from the spirit and scope of the technical scheme of each embodiment of the present application.

Claims

1. A construction monitoring method based on artificial intelligence, characterized in that: The method comprises: High-speed cameras placed on the construction site collect real-time data on construction workers' work behavior. The original construction site data is processed by frequency division and filtering based on the vibration spectrum of the construction machinery to obtain standardized construction behavior data. The standardized construction behavior data is input into the construction behavior feature genetic algorithm, and the recognition network is optimized by encoding the behavior features of high-altitude work and lifting operations to obtain a lightweight construction anomaly recognition model; The lightweight construction abnormality recognition model is used to calculate the probability of abnormal construction behaviors such as not wearing a safety helmet and illegally entering a dangerous area to obtain a construction abnormality probability value; The construction abnormality probability value is transmitted to the edge computing device at the construction site, and the model parameters are compressed through weight pruning to obtain the inference result; According to the inference results, the probability of construction anomaly is graded and determined using three-level warning thresholds to obtain a construction safety warning signal.

2. The construction monitoring method based on artificial intelligence according to claim 1, characterized in that: The high-speed cameras placed on the construction site are used to collect the construction workers' working behaviors in real time, and the original construction site data is processed by frequency division and filtering according to the vibration spectrum of the construction machinery to obtain standardized construction behavior data, including: The high-speed camera is set to high-frequency sampling mode to continuously capture images of the construction work surface. At the same time, the acceleration sensor is used to obtain the vibration signal of the construction site to obtain multi-modal raw construction site data. Based on the construction machinery vibration spectrum characteristics, the multimodal original construction site data is divided into a mechanical vibration frequency band and a human activity frequency band, and the signals of different frequency bands are separated and processed by an adaptive bandpass filter to obtain a construction signal combination after frequency division; Performing denoising and illumination compensation processing on the image data in the frequency-divided construction signal combination, and converting the vibration data into an amplitude-frequency feature vector to obtain a preprocessed construction data set; The pre-processed construction data set is normalized and standardized according to the construction operation type label, and the data range is unified into a standard interval through a standardization method to obtain standardized construction behavior data.

3. The construction monitoring method based on artificial intelligence according to claim 1, characterized in that: The standardized construction behavior data is input into the construction behavior feature genetic algorithm, and the recognition network is optimized by coding the behavior features of aerial work and lifting operations to obtain a lightweight construction anomaly recognition model, including: The action sequences of high-altitude workers and the operation trajectories of lifting equipment in the standardized construction behavior data are used as feature samples, and the feature samples are annotated with behavior types through construction behavior semantic analysis to obtain a labeled construction behavior feature sample set; Based on the helmet wearing status characteristics and dangerous area location characteristics in the construction behavior feature sample set, the number of layers and neurons of the neural network are encoded using the construction behavior feature genetic algorithm to obtain the chromosome structure of the construction anomaly recognition network; According to the chromosome structure of the construction anomaly identification network, chromosome gene segments are recombined through a crossover operation guided by construction behavior semantics, and chromosomes are mutated based on construction safety constraints to obtain an optimized network structure encoding; The optimized network structure is encoded into a specific neural network parameter configuration, and the network performance is evaluated using a construction abnormality behavior fitness function to obtain a construction abnormality recognition performance score; Based on the construction anomaly recognition performance score, multiple generations of evolutionary selection are performed, and iterative optimization processing is performed by retaining high-fitness network structures and eliminating low-performance structures to obtain a lightweight construction anomaly recognition model.

4. The construction monitoring method based on artificial intelligence according to claim 1, characterized in that: The lightweight construction abnormality recognition model is used to perform probability calculation processing on abnormal construction behaviors such as not wearing a safety helmet and illegally entering a dangerous area to obtain a construction abnormality probability value, including: The input layer of the lightweight construction anomaly recognition model receives standardized construction behavior data, and the convolution layer extracts features of the construction worker's head area and body posture to obtain a construction worker behavior feature map; Based on the construction worker behavior feature map, the pooling layer is used to reduce the dimension of the helmet wearing status features and the personnel position coordinate features. At the same time, the spatial relationship between the dangerous area boundary coordinates and the personnel position coordinates is calculated to obtain the construction abnormal behavior feature vector; The abnormal construction behavior feature vector is input into the fully connected layer for nonlinear transformation processing, and the probability of not wearing a helmet and the probability of illegally entering a dangerous area are mapped through an activation function to obtain the probability distribution of abnormal construction behavior; The probability distribution of abnormal construction behavior is normalized by a Softmax function, and the multi-category abnormal behavior probabilities are converted into a single comprehensive probability value of construction abnormality to obtain a numerical value of the construction abnormality probability.

5. The construction monitoring method based on artificial intelligence according to claim 1, characterized in that: The construction abnormality probability value is transmitted to the edge computing device at the construction site, and the model parameters are compressed by weight pruning to obtain the inference result, including: Transmitting the construction anomaly probability value to the edge computing device via the construction site wireless communication network, and loading the network parameters of the lightweight construction anomaly recognition model into the edge device memory to obtain the edge device model configuration; Based on the importance weight distribution of construction behaviors in the edge device model configuration, a weight pruning algorithm is used to remove low-contribution network connections while retaining network paths that are critical for identifying construction anomalies, thereby obtaining a compressed model parameter set. Converting the floating-point parameters in the compressed model parameter set into integer parameters through a quantization algorithm, and performing parallel reconstruction processing on the model structure to obtain an edge optimization model; Based on the edge optimization model, local reasoning and calculation processing is performed on the construction abnormality probability values ​​received in real time, and the abnormality judgment result is output through forward propagation of the model to obtain the reasoning result.

6. The construction monitoring method based on artificial intelligence according to claim 1, characterized in that: According to the inference result, the construction abnormality probability is graded and determined by using three-level warning thresholds to obtain a construction safety warning signal, including: Compare the construction abnormality probability value in the inference result with the preset three-level warning threshold, and preliminarily divide the warning level according to the threshold interval of the abnormality probability to obtain the warning level classification result; Based on the warning level classification results, the probability confidence of the first-level warning, second-level warning and third-level warning is calculated and processed through Bayesian decision theory. At the same time, the uncertainty factor is incorporated into the warning decision process for comprehensive evaluation and processing to obtain the warning level confidence value; According to the warning level confidence value, the abnormal probability change trend in the continuous time window is dynamically tracked through warning time series correlation analysis. When the abnormal probability shows an upward trend, the warning threshold is adaptively adjusted to obtain a dynamic warning judgment result; The dynamic early warning determination result is converted into a corresponding early warning signal format, and the sound prompt, sound and light alarm and forced shutdown instruction are triggered in a hierarchical manner through the construction safety communication protocol to obtain a construction safety early warning signal.

7. The artificial intelligence-based construction monitoring method according to claim 6, characterized in that: Based on the warning level classification results, the probability confidence of the first-level warning, second-level warning and third-level warning is calculated and processed through Bayesian decision theory. At the same time, uncertainty factors are incorporated into the warning decision process for comprehensive evaluation and processing to obtain the warning level confidence value, including: The first-level warning, second-level warning, and third-level warning in the warning level classification results are used as Bayesian prior probabilities, and the probability distribution of the occurrence frequency of each level of warning is calculated and processed based on the historical construction abnormal event statistical data to obtain the warning level prior probability distribution; Based on the prior probability distribution of the warning level and the current construction abnormality probability value, the posterior probability of each warning level is updated and calculated using the Bayesian formula. At the same time, the construction site environmental factors and the operation type factors are introduced to perform conditional probability correction processing to obtain the posterior probability value of the warning level; According to the posterior probability value of the warning level, the uncertainty degree of the warning decision is quantitatively calculated through the information entropy theory, and the model prediction confidence and the environmental interference factor are weighted and integrated to obtain the warning uncertainty factor; The posterior probability value of the warning level and the warning uncertainty factor are comprehensively evaluated, and the warning level that meets the reliability requirements is confirmed through confidence threshold screening to obtain a warning level confidence value.

8. A construction monitoring system based on artificial intelligence, characterized in that: For implementing the artificial intelligence-based construction monitoring method according to any one of claims 1 to 7, the artificial intelligence-based construction monitoring system comprises: The acquisition module is used to collect the construction workers' working behavior in real time through high-speed cameras placed on the construction work surface, and perform frequency division and filtering on the original construction site data according to the vibration spectrum of the construction machinery to obtain standardized construction behavior data; An optimization module is used to input the standardized construction behavior data into a construction behavior feature genetic algorithm, optimize the recognition network through the behavioral feature coding of high-altitude work and lifting operations, and obtain a lightweight construction anomaly recognition model; A calculation module is used to perform probability calculation processing on abnormal construction behaviors such as not wearing a safety helmet and illegally entering a dangerous area based on the lightweight construction abnormality recognition model to obtain a construction abnormality probability value; A compression module is used to transmit the construction abnormality probability value to the edge computing device at the construction site, compress the model parameters through weight pruning, and obtain the inference result; The grading module is used to perform graded judgment processing on the construction abnormality probability through three-level warning thresholds based on the inference results to obtain a construction safety warning signal.

9. A construction monitoring device based on artificial intelligence, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the artificial intelligence-based construction monitoring method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is caused to execute the artificial intelligence-based construction monitoring method according to any one of claims 1 to 7.

Images