Intelligent Monitoring Method and Monitoring System for Construction Environment Based on Internet of Things

Abstract

The present application provides an intelligent monitoring method and monitoring system for construction environment based on the Internet of Things, which includes obtaining an environmental sensing data set to be analyzed; when the environmental sensing data set includes a first event, obtaining a first local relationship network in the event relationship network, where the first local relationship network includes a first component item, component items within a preset path length range, and connection lines between the component items; based on the first local relationship network, obtaining a first relationship embedding array based on a first feature extraction component in the construction monitoring neural network; based on the environmental sensing data set, obtaining a sensing embedding array based on a second feature extraction component in the construction monitoring neural network; generating an integrated embedding array according to the first relationship embedding array and the sensing embedding array; and based on the integrated embedding array, obtaining a construction monitoring result based on a decision-making component in the construction monitoring neural network. The present application overcomes the problems of missing context and inaccurate recognition caused by the short content of the data stream detected in real time.

Description

Technical Field

[0001] This application relates to the field of data processing. Specifically, it relates to an intelligent monitoring method and monitoring system for construction environment based on the Internet of Things. Background Art

[0002] In various construction fields such as buildings, roads, and bridges, accurately monitoring the construction site environment is a key link to ensure construction safety, project quality, and construction efficiency. Traditional construction environment monitoring methods mainly rely on manual inspections and simple sensor data collection. Manual inspections have problems such as low monitoring frequency, strong subjectivity, and difficulty in obtaining comprehensive information in real time, and cannot timely detect potential risks and abnormal situations that occur during the construction process. Although simple sensor data collection can provide certain real-time data, due to the often short data streams obtained, lack of sufficient context information, the understanding and judgment of the construction environment are limited. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent monitoring method and monitoring system for construction environment based on the Internet of Things. This application is implemented as follows:

[0004] In a first aspect, this application provides an intelligent monitoring method for construction environment based on the Internet of Things, including: obtaining an environmental sensing data set to be analyzed; when the environmental sensing data set includes a first event, obtaining a first local relationship network in an event relationship network, where each component item in the event relationship network represents an event, each connection line in the event relationship network represents an event relationship, the first event belongs to the event represented by the first component item in the event relationship network, and the first local relationship network includes the first component item, component items within a preset path length range obtained based on the first component item, and the connection lines between the component items; based on the first local relationship network, obtaining a first relationship embedding array based on a first feature extraction component in a construction monitoring neural network; based on the environmental sensing data set, obtaining a sensing embedding array based on a second feature extraction component in the construction monitoring neural network; generating an integrated embedding array based on the first relationship embedding array and the sensing embedding array; and obtaining a construction monitoring result corresponding to the environmental sensing data set based on the integrated embedding array and based on a decision-making component in the construction monitoring neural network.

[0005] In a second aspect, this application provides a monitoring system, including: one or more processors; a memory; one or more computer programs; where the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the above-mentioned method is implemented.

[0006] In the intelligent construction environment monitoring method based on the Internet of Things provided by the present application, an environmental sensing data set to be analyzed is obtained. If it is recognized that the environmental sensing data set includes a first event, a first local relationship network is obtained in the event relationship network. Each component item in the event relationship network represents an event, and each connection line in the event relationship network represents an event relationship. The first event belongs to the event represented by the first component item in the event relationship network. The first local relationship network includes the first component item, the component items within a preset path length range obtained based on the first component item, and the connection lines between the component items. Based on this, the present application obtains a first relationship embedding array based on the first local relationship network and the first feature extraction component in the construction monitoring neural network. At the same time, a sensing embedding array is obtained based on the environmental sensing data set and the second feature extraction component in the construction monitoring neural network. Then, an integrated embedding array is generated according to the first relationship embedding array and the sensing embedding array. Finally, based on the integrated embedding array, a construction monitoring result corresponding to the environmental sensing data set is obtained based on the decision-making component in the construction monitoring neural network. Based on the above process, the solution of the present application combines the events and event relationships in the event relationship network to generate a local relationship network, and improves the context of the environmental sensing data set through the local relationship network, thus overcoming the problems of missing context and inaccurate recognition caused by the short content of the real-time detection data stream.

[0007] In the following description, other features will be partially stated. When examining the following content and the drawings, those skilled in the art will partially discover these features, or may learn about these features through production or application. Through practicing or using various aspects of the methods, tools, and combinations listed in the detailed examples described later, the features in the current application can be implemented and obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for describing the embodiments of the present application will be briefly introduced below.

[0009] Figure 1 is a flowchart of an intelligent construction environment monitoring method based on the Internet of Things provided by an embodiment of the present application.

[0010] Figure 2 is a schematic diagram of the composition of a monitoring system provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] The following describes the embodiments of the present application in combination with the drawings in the embodiments of the present application. The terms used in the implementation part of the embodiments of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0012] In the embodiments of this application, the execution subject of the intelligent monitoring method for the construction environment based on the Internet of Things is a monitoring system, including but not limited to servers, personal computers, laptops, tablets, smartphones, etc. The monitoring system can run independently to implement this application, or can be connected to a network and implement this application through interaction with other monitoring systems in the network. Among them, the network where the monitoring system is located includes but is not limited to the Internet, wide area network, metropolitan area network, local area network, VPN network, etc. The monitoring system can be used to store the environmental sensing data collected by the sensor cluster at the construction site.

[0013] The intelligent monitoring method for the construction environment based on the Internet of Things provided by the embodiments of this application, as Figure 1 shown, includes the following steps:

[0014] Step S100: Obtain the environmental sensing data set to be analyzed.

[0015] In step S100, the environmental sensing data set to be analyzed contains various data collected by various environmental sensors at the construction site, and these data can reflect the environmental status and related event information at the construction site. At the construction site, the environmental sensing data set is collected by different types of sensors. For example, a temperature sensor will collect the temperature data at the construction site, a humidity sensor will obtain the air humidity information, a light sensor will record the light intensity, a noise sensor will monitor the noise level generated during construction, a vibration sensor can detect the vibration of equipment, and a distance sensor can detect positions with distance requirements in the construction scene, such as the gap between walls, the distance between personnel and high-risk equipment, etc.

[0016] For wired-connected sensors, they can be directly connected to the sensors through a data interface, and receive the data sent by the sensors according to a preset communication protocol. For example, using the RS-485 bus protocol, the sensors encode the collected data according to the format of this protocol, and then transmit the data through the bus, and then parse the data according to the corresponding decoding rules. For a wireless sensor network, wireless communication technologies such as Wi-Fi and ZigBee can be used to receive sensor data. Taking Wi-Fi as an example, the sensors encapsulate the collected data into data packets conforming to the Wi-Fi protocol and send them out through wireless signals. After receiving these data packets, they are unpacked and processed to obtain the environmental sensing data therein.

[0017] Step S200: When the environmental sensing data set includes a first event, obtain a first local relationship network in the event relationship network, where each component item in the event relationship network represents an event, each connection line in the event relationship network represents an event relationship, the first event belongs to the event represented by the first component item in the event relationship network, and the first local relationship network includes the first component item, component items within a preset path length obtained based on the first component item, and the connection lines between the component items.

[0018] In step S200, when the environmental sensing data set includes a first event, a first local relationship network is obtained in the event relationship network. The core of this step is to use the information of the event relationship network to improve the context of the environmental sensing data set, thereby solving the problems of missing context caused by short real-time detection data streams and inaccurate construction site monitoring. The event relationship network is a structured graph structure, where each component item, i.e., a graph vertex, represents an event, and each connection line represents an event relationship. These events and event relationships form a complex network structure, reflecting the internal connections between various events at the construction site.

[0019] In the actual scenario of the construction site, the first event may be a specific construction activity, such as "concrete pouring". First, it is necessary to determine the corresponding component item of this event in the event relationship network, that is, to find the first component item representing the event of "concrete pouring". The events in the event relationship network can cover all aspects of the construction process, such as equipment operation, material transportation, personnel activities, etc. Each event may be associated with multiple other events.

[0020] When obtaining the first local relationship network, other component items related to the first component item are determined according to the preset path length. The preset path length refers to the number of connection lines passed from one component item to another, which limits the scope of the local relationship network. Taking the "concrete pouring" event as an example, when the preset path length is 2, the events directly connected to "concrete pouring" and the events connected to these directly connected events are searched. The directly connected events may include "concrete mixing" and "formwork installation", etc. The events connected to "concrete mixing" may include "cement transportation" and "sand and gravel supply", etc. These events and the connection lines between them together constitute the first local relationship network.

[0021] To obtain the first local relationship network, a graph traversal algorithm can be used. A graph traversal algorithm is an algorithm for accessing all nodes and edges in a graph, such as breadth-first search (BFS) and depth-first search (DFS). Taking the breadth-first search algorithm as an example, first, the first set of components is used as the starting node, which is marked as visited and added to the queue. Then, a node is taken out from the queue, and all its adjacent nodes are checked. If an adjacent node has not been visited and the path length from it to the starting node is not greater than the preset path length, it is marked as visited and added to the queue, and at the same time, the connection line information between the nodes is recorded. This process is repeated until the queue is empty. In this way, all components and connection lines that meet the conditions can be found to generate the first local relationship network.

[0022] The generation of the first local relationship network can provide richer context information to help the system better understand the first event and other related events.

[0023] When detecting an event, taking the event of detecting concrete pouring from an environmental sensing dataset as an example, different types of environmental sensors can be comprehensively used to obtain relevant data, and the event can be identified through the analysis and processing of these data. First, a vibration sensor can be used to detect the vibration situation during concrete pouring. When concrete is poured, operations such as transporting concrete from a mixer truck to the pouring site and vibrating the concrete with vibrating equipment will generate vibrations. The vibration sensor can convert these vibration signals into electrical signals and record them. For example, vibration sensors are installed at parts such as formwork and support structures at the construction site. When concrete pouring starts and the vibrating equipment is turned on, the vibration sensor will detect vibration signals with specific frequencies and intensities. By performing spectral analysis on these vibration signals, the main frequency components and intensity ranges are determined. Generally, the vibration frequency during concrete pouring will be concentrated in a specific interval, and the intensity will also have a certain variation law as the vibration progresses. A vibration intensity threshold can be set. When the detected vibration intensity exceeds this threshold and the vibration frequency is within the preset frequency interval, it can be preliminarily judged that concrete pouring may be in progress.

[0024] Secondly, the pressure sensor can also provide important detection basis. During the concrete pouring process, as the concrete accumulates continuously, it will cause pressure changes to the formwork, support structure, etc. Install pressure sensors at key positions such as the bottom of the formwork and support columns to monitor the pressure changes in real time. When the concrete starts to be poured, the pressure will gradually increase, and the increasing rate is related to factors such as the pouring speed and quantity of the concrete. A pressure change model can be established. According to the concrete pouring process and design requirements, calculate the pressure change curve over time under normal pouring conditions. Compare the actually detected pressure change data with this model. If the change trends of the two are similar and the pressure values are within a reasonable range, the occurrence of the concrete pouring event can be further confirmed. For example, for a formwork with a specific size and shape, according to parameters such as the density and pouring speed of the concrete, the theoretical change value of the pressure at the bottom of the formwork during pouring can be calculated. When the actually detected pressure value is close to this theoretical value, it can be judged that the concrete is being poured as expected.

[0025] Temperature sensors can also be used to detect the concrete pouring event. During the concrete pouring process, the hydration reaction of cement will release heat, resulting in an increase in the temperature of the concrete. Arrange temperature sensors inside the concrete and in the surrounding environment to monitor the temperature changes in real time. After the concrete starts to be poured, the internal temperature will gradually rise, and the rising amplitude and speed are related to factors such as the type and dosage of cement and the ambient temperature. A temperature change threshold can be set. When it is detected that the internal temperature of the concrete rises beyond this threshold within a short period of time, it can be inferred that concrete pouring may be in progress. For example, in a normal temperature environment, the internal temperature of concrete made of ordinary Portland cement may rise by 5 - 10 degrees Celsius within a few hours after pouring. If it is detected that the internal temperature of the concrete has a similar rising amplitude during this period, it can be used as a basis for judging concrete pouring.

[0026] Finally, fuse and comprehensively analyze the data collected by the above various sensors. Multisensor data fusion technology can be used to process different types of data such as vibration, pressure, and temperature, remove noise and interference, and extract effective information related to the concrete pouring event. By establishing a comprehensive judgment model, according to the weights and combination rules of different sensor data, make a final judgment on the concrete pouring event. For example, the fuzzy logic algorithm can be used to assign corresponding weights according to the credibility and importance of different sensor data, perform fuzzy reasoning on these data, and obtain a comprehensive judgment result. When this result exceeds a preset threshold, it can be determined that the concrete pouring event is occurring.

[0027] Step S300: Based on the first local relationship network, obtain the first relationship embedding array by means of the first feature extraction component in the construction monitoring neural network.

[0028] In step S300, the events and their relationship information included in the first local relationship network are converted into a numerical form that can be processed and analyzed for more in-depth calculations and decision-making. The first local relationship network is composed of the components related to the first event in the event relationship network and the connecting lines between them, which reflects the associations and context information among various events at the construction site. Graph embedding technology can be used to obtain the first relationship embedding array by mapping the components and connecting lines in the graph structure (such as the first local relationship network) into a low-dimensional vector space, so that nodes with similar structures and relationships in the graph also have similar representations in the vector space. Graph embedding algorithms include, for example, DeepWalk, Node2Vec, etc.

[0029] Taking the DeepWalk algorithm as an example, it generates a node sequence in the first local relationship network through random walks, and then uses these node sequences as text sequences to learn the vector representations of the components using a word embedding algorithm (such as the Skip-Gram model). The specific steps are as follows: First, start random walks from each component in the first local relationship network to generate a series of component sequences. The process of random walks is similar to randomly selecting a path in the graph, walking from one component to an adjacent component, and repeating this process until the preset walk length is reached. For example, starting from the "tower crane operation" component, it may randomly walk to the "material hoisting" node, and then from the "material hoisting" component to the "equipment inspection" component, forming a component sequence "tower crane operation - material hoisting - equipment inspection". Then, use these component sequences as sensing data sequences and use the Skip-Gram model to learn the vector representations of the components. The Skip-Gram model predicts the context nodes of a component and learns the vector representation of the component by maximizing the probability of this prediction. Specifically, for each component in a sensing data sequence, the model predicts the nodes within a certain range before and after the component. By continuously optimizing the parameters of the model to maximize the prediction probability, the vector representation of each component is finally obtained.

[0030] After obtaining the vector representations of each component, these vectors are integrated to obtain the first relationship embedding array, for example, by averaging or weighted averaging the vectors of all components.

[0031] The first relationship embedding array contains important information about the events and their relationships in the first local relationship network. It converts the graph structure information into a numerical vector for subsequent calculations and analysis.

[0032] Step S400: Based on the environmental sensing data set, obtain a sensing embedding array based on the second feature extraction component in the construction monitoring neural network.

[0033] In step S400, valuable feature information is extracted from the environmental sensing dataset and converted into a vector form suitable for processing and analysis, so as to be integrated with the first relational embedding array subsequently, thereby providing a basis for accurate construction monitoring results. The environmental sensing dataset is composed of data collected by various environmental sensors at the construction site, and these data contain various environmental information at the construction site, such as temperature, humidity, light, noise, vibration, and so on.

[0034] The second feature extraction component is part of the construction monitoring neural network, and it can be a convolutional neural network (CNN) or a recurrent neural network (RNN) in deep learning. When using a convolutional neural network for feature extraction, the environmental sensing dataset is preprocessed and converted into a format suitable for the input of the convolutional neural network. For example, if the environmental sensing dataset is a series of time series data, it can be organized into a two-dimensional matrix form, where each row represents the sensor data at a time point and each column represents the measurement value of a sensor. Then, the convolutional neural network will perform a convolution operation on the input data through the convolutional layer, and the convolutional kernels in the convolutional layer will slide on the input data to extract features at different scales and positions. Each convolutional kernel generates a feature map, and these feature maps reflect the responses of the input data in different feature dimensions. After the convolutional layer, a pooling layer can be cascaded to downsample the feature maps, reducing the data dimension while retaining important feature information.

[0035] After being processed by the convolutional layer and the pooling layer, a series of feature maps are obtained, and these feature maps contain important feature information in the environmental sensing dataset. Finally, these feature maps are converted into a vector through the fully connected layer, and this vector is the sensing embedding array. The fully connected layer will connect all the neurons in each feature map to the neurons in the next layer, and integrate the information of the feature maps into a vector through linear transformation and non-linear activation functions.

[0036] As described above, in addition to the convolutional neural network, the recurrent neural network (RNN) can also be used to extract feature information from the environmental sensing dataset. The principle can refer to the existing technology and will not be elaborated here.

[0037] The sensing embedding array contains the key feature information in the environmental sensing dataset, which converts the environmental sensing data into a numerical vector, facilitating subsequent calculations and analyses.

[0038] Step S500: Generate an integrated embedding array based on the first relational embedding array and the sensing embedding array.

[0039] Step S500 aims to effectively integrate the event association information contained in the first relation embedding array extracted from the event relation network with the construction site environment information contained in the sensing embedding array extracted from the environmental sensing dataset, thereby laying a foundation for obtaining more accurate construction monitoring results in the subsequent stage. The first relation embedding array reflects the relationships and context information among the construction site events, while the sensing embedding array embodies various environmental state data of the construction site. By generating the integrated embedding array, the information from these two aspects can be comprehensively utilized to improve the accuracy and reliability of the construction environment monitoring.

[0040] One implementation method is to generate the integrated embedding array through matrix operations. Multiply the sensing embedding array with the first weight array to obtain the first multiplication result. The first weight array is a two-dimensional array (matrix), and its role is to weight each element in the sensing embedding array to highlight the importance of different environmental features. For example, at the construction site, if the temperature has a greater impact on the construction quality, then the weight corresponding to the temperature feature in the first weight array may be set higher. Then, flip the result of multiplying the first relation embedding array with the second weight array to obtain the flipped result. The second weight array is also used to weight the elements in the first relation embedding array to reflect the importance of different event relationships. Next, multiply the first multiplication result with the flipped result to obtain the second multiplication result. Finally, add the bias array to the second multiplication result to obtain the integrated embedding array. Among them, the first weight array and the second weight array are obtained by extracting from a third-order tensor. The dimension of the third-order tensor is i×s×t, the dimension of the first weight array is i×1×t, the dimension of the second weight array is 1×s×t, i is the dimension of the sensing embedding array, s is the dimension of the first relation embedding array, and t is the dimension of the integrated embedding array.

[0041] Another implementation method is to combine the first relation embedding array and the sensing embedding array to obtain the integrated embedding array. This method is simple and direct, connecting the two arrays in a certain order to form a new array. For example, if the dimension of the sensing embedding array is i and the dimension of the first relation embedding array is s, then the dimension of the integrated embedding array is i + s. This method can retain the original information of the two arrays, but may lead to a relatively high dimension of the integrated embedding array, increasing the complexity of subsequent calculations.

[0042] Another implementation is to obtain the Hadamard product of the first relational embedding array and the sensing embedding array to get the integrated embedding array. The Hadamard product is an operation of multiplying corresponding elements, and requires that the dimensions of the two arrays be equal. Through the Hadamard product, the corresponding elements in the first relational embedding array and the sensing embedding array can be multiplied element by element to obtain a new array with the same dimension as the original arrays. This method can effectively fuse the information in the two arrays and highlight the interaction of the corresponding elements in the two arrays.

[0043] When the environmental sensing data set also includes a second event, it is necessary to obtain a second local relational network in the event relational network and obtain a second relational embedding array based on it. At this time, the process of generating the integrated embedding array will be more complex. It is necessary to comprehensively consider the information of the first relational embedding array, the second relational embedding array, and the sensing embedding array. For example, the first relational embedding array and the second relational embedding array can be processed first, such as performing Hadamard products with a preset parameter array respectively to obtain a first intermediate result and a second intermediate result, and then fusing these two intermediate results to obtain a target relational embedding array. Then, methods such as the matrix operations, combinations, or Hadamard products mentioned above can be used to integrate the target relational embedding array and the sensing embedding array to generate the final integrated embedding array.

[0044] Step S600: Based on the integrated embedding array, obtain the construction monitoring result corresponding to the environmental sensing data set based on the decision-making component in the construction monitoring neural network.

[0045] The integrated embedding array is generated by fusing the first relational embedding array and the sensing embedding array. It synthesizes the event association information in the event relational network and the on-site environmental information in the environmental sensing data set, providing comprehensive and effective input data for the decision-making component.

[0046] When executing step S600, a fully connected neural network can be used as the decision-making component. The fully connected neural network consists of an input layer, a hidden layer, and an output layer. The input layer receives the integrated embedding array as input data, the hidden layer processes and transforms the input data through a series of neurons, and the output layer outputs the final construction monitoring result.

[0047] In the fully connected neural network, each neuron will perform a weighted sum on the input data and perform a non-linear transformation through an activation function. Assume that the input layer has n neurons corresponding to n elements of the integrated embedding array. The input of the j-th neuron in the hidden layer can be expressed as: ; where, is the weight from the i-th neuron in the input layer to the j-th neuron in the hidden layer, is the input value of the i-th neuron in the input layer (i.e., the i-th element of the integrated embedding array), is the bias term of the j-th neuron in the hidden layer.

[0048] Then, the output of the j-th neuron in the hidden layer can be calculated through the activation function f: , and the activation function is, for example, the Sigmoid function, the ReLU function, etc.

[0049] After being processed by the hidden layer, the neurons in the output layer will perform a weighted sum on the output of the hidden layer and output the final construction monitoring result.

[0050] The final construction monitoring result can be judged according to the output value of the neurons in the output layer. For example, if there is only one neuron in the output layer and its output value represents the probability of potential safety hazards at the construction site, then when the output value is greater than a preset threshold (such as 0.5), it can be judged that there are potential safety hazards at the construction site; when the output value is less than the threshold, it is judged that the construction site is safe.

[0051] The construction monitoring result can include various types of information, such as whether the construction progress is normal, whether there are potential safety hazards in construction, whether the construction environment meets the requirements, etc. For example, through the analysis of the integrated embedding array, the decision-making component can judge whether the current construction progress lags behind the planned progress. If environmental parameters such as temperature and humidity exceed the range required by the construction process, the decision-making component can issue an alarm for environmental anomalies; if some key events in the event relationship network do not occur in the normal order, the decision-making component can prompt possible construction risks.

[0052] Through the analysis and judgment of the integrated embedding array by the decision-making component of the construction monitoring neural network, it provides an important basis for the management and decision-making of the construction site, helps to timely discover problems and risks in the construction process, and ensures the smooth progress of construction and the safety of personnel.

[0053] As an implementation manner, after obtaining the environmental sensing data set to be analyzed, the method further includes a process of event detection, specifically including:

[0054] Step S101: Perform event detection on the environmental sensing data set to obtain an event detection result;

[0055] Step S102: When the event detection result indicates that the environmental sensing data set includes events, compare the events in the environmental sensing data set with the events in the event relationship network to obtain an event comparison result;

[0056] Step S103: When the event comparison result indicates that the events in the environmental sensing data set correspond to the events in the event relationship network, regard the corresponding events as the first events.

[0057] In step S101, event detection is performed on the environmental sensing data set to obtain an event detection result. The environmental sensing data set is composed of multi-source data collected by various sensors at the construction site. These data contain various physical information of the construction site, such as temperature, humidity, noise, vibration, etc. The purpose of event detection is to identify events with specific meanings from these complex data, such as the startup of construction equipment, the transportation of materials, the activities of personnel, etc.

[0058] One method for event detection is the threshold-based detection method. For some types of sensor data, such as temperature, noise, etc., a reasonable threshold range can be set. When the data collected by the sensor exceeds this threshold range, it is considered that an event may have occurred. For example, at the construction site, the normal environmental temperature range may be between 20°C and 30°C. If the temperature detected by the temperature sensor suddenly rises to 40°C, it can be judged that events such as equipment overheating or fire may have occurred. The setting of the threshold can be obtained through statistical analysis of historical data or can be set manually according to construction technology and safety standards.

[0059] Another method is the event detection method based on pattern recognition. The sensor data patterns corresponding to various events can be learned in advance, and then the real-time collected environmental sensing data can be matched with these patterns. For example, when a construction equipment starts up, the vibration sensor will detect a vibration signal with specific frequency and amplitude. By analyzing the characteristics of the vibration signal, such as frequency distribution, amplitude change, etc., it can be judged whether the equipment has started. Pattern recognition can use machine learning algorithms, such as support vector machine (SVM), decision tree, etc., or can also use deep learning algorithms, such as convolutional neural network (CNN), recurrent neural network (RNN), etc.

[0060] Taking the support vector machine as an example, its basic idea is to find an optimal hyperplane in the feature space to separate event data of different categories. Assuming that the environmental sensing data can be represented as a feature vector x and the event category can be represented as a label y (y = ±1), the goal of the support vector machine is to find a weight vector w and a bias b to maximize the classification margin. The specific optimization problem can be expressed as:

[0061] ;

[0062] where n is the number of samples.

[0063] In step S102, when the event detection result indicates that the environmental sensing data set includes an event, the event in the environmental sensing data set is compared with the events in the event relationship network to obtain an event comparison result. The event relationship network is a pre-constructed network structure, where each component represents an event and the connection lines represent the relationships between events.

[0064] When conducting event comparison, a semantic matching - based method can be adopted. For each detected event, its key information can be extracted, such as the name of the event, the occurrence time, relevant devices, etc. Then, this information is compared with the event information in the event relationship network. For example, if the detected event is "Tower crane starts", relevant components related to "Tower crane starts" can be searched in the event relationship network. The comparison process can use natural language processing techniques, such as lexical analysis, syntactic analysis, etc., to convert the event description into a form that can be processed by a computer, and then calculate the similarity. Multiple methods can be used for similarity calculation, such as edit distance, cosine similarity, etc. Edit distance refers to the minimum number of operations required to convert one string into another through insertion, deletion, and replacement operations. Another comparison method is rule - based comparison, where some comparison rules can be predefined. For example, if the detected event involves a certain specific device, then search for events related to that device in the event relationship network.

[0065] In step S103, when the event comparison result indicates that the event in the environmental sensing data set corresponds to the event in the event relationship network, the corresponding event is taken as the first event.

[0066] As an implementation manner, in step S102, when the event detection result indicates that the environmental sensing data set includes events, the events in the environmental sensing data set are compared with the events in the event relationship network to obtain an event comparison result, including:

[0067] Step S1021: When the event detection result indicates that the environmental sensing data set includes multiple events, obtain the conditional probability corresponding to each event in the multiple events;

[0068] Step S1022: Take the event corresponding to the maximum conditional probability among the multiple events as the event in the environmental sensing data set;

[0069] Step S1023: Compare the event in the environmental sensing data set with the events in the event relationship network to obtain an event comparison result.

[0070] In step S1021, when the event detection result indicates that the environmental sensing data set includes multiple events, obtain the conditional probability corresponding to each event in the multiple events. Conditional probability refers to the probability of an event occurring under certain known conditions. In the context of the environmental sensing data set, these conditions can be data collected by sensors, time information, the context of the event, etc. By calculating the conditional probability, the possibility of each event occurring can be quantified, providing a basis for subsequent event screening.

[0071] One way to calculate the conditional probability of an event is based on Bayes' theorem. Bayes' theorem describes the probability that a certain hypothesis holds given some evidence. Suppose event A represents a specific event, and event B represents the data collected by a sensor or other relevant conditions. The formula for Bayes' theorem is: ; where P(A|B) is the conditional probability that event A occurs given condition B, P(B|A) is the probability that condition B appears given that event A has occurred, P(A) is the prior probability that event A occurs, and P(B) is the probability that condition B appears. The prior probability P(A) can be obtained through statistical analysis of historical data. For example, by analyzing the frequency of a certain event occurring over a period of time to estimate its prior probability. P(B|A) can be obtained through statistical analysis of the relevant condition B when the known event A occurs. P(B) can be calculated using the law of total probability: ; where A i represents all possible events. In addition to Bayes' theorem, machine learning algorithms can also be used to calculate the conditional probability of an event. For example, using the Naive Bayes classifier, which is a classification algorithm based on Bayes' theorem and the assumption of feature conditional independence. The Naive Bayes classifier can learn from the training data to obtain the probability distribution of each event under different feature conditions. During prediction, according to the input feature conditions, the probability of each event occurring is calculated.

[0072] In step S1022, the event corresponding to the maximum conditional probability among multiple events is taken as the event in the environmental sensing dataset. The purpose of this step is to select the most likely event from multiple possible events to improve the accuracy of subsequent event comparison and monitoring. After calculating the conditional probability of each event, by comparing the magnitudes of these probability values, the event corresponding to the maximum conditional probability is found.

[0073] In step S1023, the event in the environmental sensing dataset is compared with the events in the event relationship network to obtain an event comparison result. The event relationship network is a pre-constructed network structure, where each component represents an event and the connecting lines represent the relationships between events.

[0074] When performing event comparison, a semantic matching-based method can be adopted. For the events in the determined environmental sensing dataset, their key information can be extracted, such as the name of the event, the occurrence time, related devices, etc., and then this information is compared with the event information in the event relationship network. For example, if the determined event is "tower crane startup", the components related to "tower crane startup" can be searched in the event relationship network. The comparison process can use natural language processing technologies, such as lexical analysis, syntactic analysis, etc., to convert the event description into a form that can be processed by a computer, and then perform similarity calculation. Multiple methods can be used for similarity calculation, such as edit distance, cosine similarity, etc. Another comparison method is rule-based comparison. Some comparison rules can be predefined. For example, if the determined event involves a certain specific device, then search for the events related to this device in the event relationship network. The definition of the rules can be customized according to the characteristics and requirements of the construction process.

[0075] As an implementation manner, in step S200, when the environmental sensing dataset includes the first event, obtaining a first local relationship network in the event relationship network includes:

[0076] Step S210: When the environmental sensing dataset includes the first event, taking the components corresponding to the first event in the event relationship network as the first components;

[0077] Step S220: Based on the event relationship network, obtaining a set of components involved with the first components, and the path length between each component in the set of components and the first components is not greater than a preset path length;

[0078] Step S230: Generating a first local relationship network based on the first components, the set of components, and the connection lines between the components.

[0079] Steps S210 - S230 in an implementation manner of step S200 are important processes for obtaining the first local relationship network from the event relationship network when the environmental sensing dataset includes the first event. This local relationship network can provide rich context information for subsequent intelligent monitoring of the construction environment, and helps to solve the problems of context loss and inaccurate monitoring caused by the short real-time detection data stream.

[0080] In step S210, when the environmental sensing dataset includes the first event, taking the components corresponding to the first event in the event relationship network as the first components.

[0081] In step S220, based on the event relationship network, obtain a set of components that are involved (i.e., have an association) with the first component set. The path length between each component in this component set and the first component is not greater than a preset path length. The preset path length is a pre-set parameter that limits the scope of the local relationship network and reflects the tightness of the association between events. The path length refers to the number of connection lines passed from one component to another. A graph traversal algorithm can be used to obtain the component set related to the first component. Graph traversal algorithms include breadth-first search (BFS) and depth-first search (DFS). As described above, it will not be elaborated here.

[0082] In step S230, based on the first component set, the component set, and the connection lines between the components, generate a first local relationship network. The first local relationship network is a subgraph of the event relationship network. It contains important information related to the first event and can reflect the association and context relationship between events.

[0083] When generating the first local relationship network, extract the first component set, the component set, and the connection lines between them from the event relationship network. The connection lines represent the relationships between events, such as causal relationships, sequential relationships, collaboration relationships, etc. A new graph structure, i.e., the first local relationship network, can be constructed according to the identification of the components and the information of the connection lines.

[0084] In practical applications, the implementation process of steps S210 - S230 may be affected by factors such as the complexity of the event relationship network and the accuracy of the data. To improve the accuracy and efficiency of obtaining the first local relationship network, the event relationship network can be optimized and sorted out to remove redundant information and improve the connectivity of the graph. At the same time, for the problem of data accuracy, a data verification and error correction mechanism can be adopted to ensure that the information in the event relationship network is accurate.

[0085] By determining the first component set in the event relationship network, obtaining the relevant component set, and generating the first local relationship network in the event relationship network, important context information is provided for subsequent intelligent monitoring of the construction environment. In the complex environment of the construction site, the first local relationship network can help better understand the association between events, accurately judge the construction status, timely discover potential problems and risks, and provide strong support for construction management and decision-making.

[0086] As a second implementation method, step S200, when the environmental sensing data set includes the first event, obtaining the first local relationship network in the event relationship network includes:

[0087] Step S201: When the environmental sensing data set includes the first event, use the component corresponding to the first event in the event relationship network as the first component set;

[0088] Step S202: Based on the event relation network, obtain the set of components involved with the first component item, where the path length between each component item in the set of component items and the first component item is not greater than the preset path length;

[0089] Step S203: When the environmental sensing data set includes the first event relation, based on the event relation network, obtain the first local set of component items in the set of component items that have the first event relation with the first component item;

[0090] Step S204: Generate the first local relation network based on the first component item, the first local set of component items, and the connection lines between the component items.

[0091] Steps S201 and S202 can refer to the aforementioned steps S210~S220.

[0092] In step S203, when the environmental sensing data set includes the first event relation, based on the event relation network, obtain the first local set of component items in the set of component items that have the first event relation with the first component item. The first event relation refers to a certain specific association between events, such as a causal relationship, a sequential relationship, a collaborative relationship, etc. When executing this step, check the relationship between each component item in the set of component items and the first component item to determine whether there is a first event relation. For example, if the first event relation is a causal relationship, for the first component item of "concrete pouring", search in the set of component items for the component item that has a causal relationship with "concrete pouring". For example, "concrete mixing" is a pre - causal event of "concrete pouring", then "concrete mixing" will be included in the first local set of component items. The relationship between events can be judged through the marking information of the connection lines in the event relation network. Each connection line can carry a label indicating the type of event relation it represents, and by reading this label information, filter out the component items that meet the first event relation.

[0093] In step S204, based on the first component item, the first local component item set, and the connection lines between the component items, a first local relationship network is generated. The first local relationship network is a subgraph of the event relationship network. It contains information that is closely related to the first event and has specific event relationships, and can more accurately reflect the associations and contextual relationships between events. When generating the first local relationship network, it is necessary to extract the first component item, the first local component item set, and the connection lines between them from the event relationship network. The connection lines represent various relationships between events. According to the identifiers of the component items and the information of the connection lines, a new graph structure, that is, the first local relationship network, is constructed. For example, for the first component item of "concrete pouring", the first local component item set may include "concrete mixing", "formwork installation", etc. These component items and the connection lines representing causal, sequential, and other relationships between them will be extracted to generate a first local relationship network centered on "concrete pouring". This local relationship network can intuitively display a series of events related to "concrete pouring" and having specific relationships and their mutual relationships, helping to better understand the situation at the construction site.

[0094] Steps S201 - S204 provide more targeted and relevant contextual information for subsequent intelligent monitoring of the construction environment by determining the first component item in the event relationship network, obtaining the relevant component item set, screening out the first local component item set with specific event relationships, and generating the first local relationship network. In the complex environment of the construction site, the first local relationship network can help better understand the associations between events, accurately judge the construction status, timely discover potential problems and risks, and provide strong support for construction management and decision-making. In addition, based on the above method, the coverage range of the local relationship network is controlled according to the preset path length. This can not only avoid the problem of data volume explosion of the local relationship network caused by too long path length, but also make the information aggregation effect better and better represent the events in the environmental sensor data set.

[0095] As a third implementation manner, in step S200, when the environmental sensor data set includes the first event, obtaining the first local relationship network in the event relationship network includes:

[0096] Step S2001: When the environmental sensor data set includes the first event, use the component item corresponding to the first event in the event relationship network as the first component item;

[0097] Step S2002: Based on the event relationship network, obtain the set of component items involved with the first component item, and the path length between each component item in the component item set and the first component item is not greater than the preset path length;

[0098] Step S2003: When the environmental sensing data set includes a first event relationship and a second event relationship, based on the event relationship network, obtain a second local component set in the component set that has either one or both of the first event relationship and the second event relationship with the first component;

[0099] Step S2004: Generate a first local relationship network based on the first component, the second local component set, and the connection lines between the components.

[0100] Steps S2001 and S2002 can refer to the aforementioned steps S210~S220.

[0101] In step S2003, when the environmental sensing data set includes a first event relationship and a second event relationship, based on the event relationship network, obtain a second local component set in the component set that has either one or both of the first event relationship and the second event relationship with the first component. The first event relationship and the second event relationship refer to specific associations between events, such as causal relationships, sequential relationships, collaborative relationships, dependency relationships, etc. When performing this step, it is necessary to carefully check the relationship between each component in the component set and the first component to determine whether there is a first event relationship or a second event relationship. For example, if the first event relationship is a causal relationship and the second event relationship is a collaborative relationship, for the first component of "the tower crane hoists large components", the components in the component set that have a causal relationship with "the tower crane hoists large components" will be searched, such as "the components are ready" being the precondition causal event of "the tower crane hoists large components"; at the same time, the components with a collaborative relationship will be searched, such as "the ground command personnel cooperate" having a collaborative relationship with "the tower crane hoists large components", and these components will all be included in the second local component set. The relationship between events can be judged through the marking information of the connection lines in the event relationship network. Each connection line can carry a label indicating the type of event relationship it represents, and the components that meet the first or second event relationship are screened out by reading this label information.

[0102] In step S2004, based on the first component item, the second local component item set, and the connection lines between the component items, a first local relationship network is generated. The first local relationship network is a subgraph of the event relationship network. It contains information that is closely related to the first event and has specific event relationships, and can more accurately reflect the associations and context relationships between events. When generating the first local relationship network, it is necessary to extract the first component item, the second local component item set, and the connection lines between them from the event relationship network. The connection lines represent various relationships between events. According to the identifiers of the component items and the information of the connection lines, a new graph structure, namely the first local relationship network, is constructed. For example, for the first component item of "the tower crane hoists large components", the second local component item set may include "component fixation", "tower crane operator operation", "component ready", "ground commander cooperation", etc. These component items and the connection lines representing causal, collaborative, etc. relationships between them will be extracted to generate a first local relationship network with "the tower crane hoists large components" as the core. This local relationship network can visually display a series of events related to "the tower crane hoists large components" and having specific relationships and their mutual relationships, helping to better understand the situation at the construction site.

[0103] Steps S2001 - S2004 provide more targeted and comprehensive context information for subsequent intelligent monitoring of the construction environment by determining the first component item in the event relationship network, obtaining the relevant component item set, screening out the second local component item set with specific event relationships, and generating the first local relationship network. In the complex environment of the construction site, the first local relationship network can help better understand the associations between events, accurately judge the construction status, and timely discover potential problems and risks, providing strong support for construction management and decision-making. In addition, based on the above method, the coverage range of the local relationship network is controlled according to the preset path length. This can not only avoid the problem of data volume explosion of the local relationship network caused by too long path length, but also make the information aggregation effect better and better represent the events in the environmental sensor data set.

[0104] As an implementation method, in step S500, based on the first relationship embedding array and the sensing embedding array, an integrated embedding array is generated, including:

[0105] Step S510: Multiply the sensing embedding array by the first weight array to obtain a first multiplication result. The dimension of the sensing embedding array is i;

[0106] Step S520: Flip the result of multiplying the first relationship embedding array by the second weight array to obtain a flipped result. The dimension of the first relationship embedding array is s;

[0107] Step S530: Multiply the first multiplication result by the flipped result to obtain a second multiplication result;

[0108] Step S540: Add the second product result to the bias array to obtain an integrated embedding array; wherein, the first weight array and the second weight array are obtained by extracting from a third-order tensor, the dimension of the third-order tensor is i×s×t, the dimension of the first weight array is i×1×t, the dimension of the second weight array is 1×s×t, i is the dimension of the sensing embedding array, s is the dimension of the first relational embedding array, and t is the dimension of the integrated embedding array.

[0109] In step S510, the sensing embedding array is multiplied by the first weight array to obtain a first product result. The dimension of the sensing embedding array is i. The sensing embedding array is a numerical representation of the feature information extracted from the environmental sensing dataset, which contains various environmental state data at the construction site, such as data collected by sensors for temperature, humidity, noise, vibration, etc. The first weight array is a two-dimensional array (matrix), and its role is to weight each element in the sensing embedding array to highlight the importance of different environmental features. When performing this step, matrix multiplication operations will be carried out according to the dimensions of the sensing embedding array and the first weight array. Suppose the sensing embedding array is a, with a dimension of i×1, and the first weight array is W1, with a dimension of i×t (t is the dimension of the integrated embedding array), then the calculation method of the first product result b is: , where the rule of matrix multiplication is: the j-th row element b of j is equal to the sum of the products of the j-th row of W1 and the corresponding elements of a, that is , where represents the element in the j-th row and k-th column of k and a

[0110] For example, if the sensing embedding array is , and the first weight array is , then the first product result is . Different weight values can be set according to the importance of different environmental features for construction monitoring. For example, if at the construction site, temperature has a greater impact on construction quality, then the weight corresponding to the temperature feature in the first weight array may be set higher, so as to highlight the role of temperature information during the multiplication process.

[0111] In step S520, the result of multiplying the first relationship embedding array by the second weight array is flipped to obtain a flipped result. The dimension of the first relationship embedding array is s. The first relationship embedding array is a numerical representation of the feature information extracted from the first local relationship network of the event relationship network, which reflects the relationships and context information among the construction site events. The second weight array is also used to weight the elements in the first relationship embedding array to reflect the importance of different event relationships. First, perform a matrix multiplication operation on the first relationship embedding array c (with dimension s×1) and the second weight array W2 (with dimension s×t) to obtain an intermediate result d: , and its calculation method is similar to that in step S510. The j-th element d j of d is equal to the sum of the products of the j-th row of W2 and the corresponding elements of c. Then, perform a flipping operation on d. The flipping operation can be to reverse the order of the vector elements. For example, if , the flipped result . Through the flipping operation, the corresponding relationship when multiplying the first relationship embedding array by the sensing embedding array can be changed, increasing the diversity of information fusion.

[0112] In step S530, multiply the first product result by the flipped result to obtain a second product result. In this step, perform an element-wise multiplication on the first product result b (with dimension t×1) and the flipped result e (with dimension t×1) to obtain the second product result f. The rule for element-wise multiplication is: the j-th element f j of f is equal to the product of the j-th element of b and the j-th element of e, that is . For example, if , then . Through this product operation, the information of the sensing embedding array and the first relationship embedding array is fused, so that the second product result contains both the environmental information of the construction site and the relationship information among the events.

[0113] In step S540, add the bias array to the second product result to obtain an integrated embedding array. Among them, the first weight array and the second weight array are obtained by extracting from a third-order tensor. The dimension of the third-order tensor is i×s×t, the dimension of the first weight array is i×t, the dimension of the second weight array is s×t, i is the dimension of the sensing embedding array, s is the dimension of the first relationship embedding array, and t is the dimension of the integrated embedding array. The bias array b (the b here has a different meaning from the b in the previous steps) is a vector with dimension t×1, and its function is to translate the second product result, adjust the overall numerical level of the integrated embedding array, and increase the flexibility of the model. Add the second product result f and the bias array b element-wise to obtain the integrated embedding array g: , that is, the j-th element g jequal to the sum of the j-th element of f and the j-th element of b, g j = f j + b j .

[0114] The values of the first weight array and the second weight array can be determined through the training process. A large number of environmental sensing dataset samples and corresponding prior construction monitoring result labels can be used, and an optimization algorithm (such as the stochastic gradient descent method) can be adopted to adjust the parameters of the first weight array, the second weight array, and the bias array, so that the generated integrated embedding array can better reflect the real situation of the construction environment, thereby improving the accuracy of subsequent construction monitoring. Steps S510 - S540 effectively fuse the information of the sensing embedding array and the first relational embedding array through steps such as matrix operations, flipping operations, and element-wise addition. The generated integrated embedding array synthesizes the environmental information and event relationship information of the construction site, providing an important data basis for the decision-making component based on the construction monitoring neural network to obtain accurate construction monitoring results.

[0115] As a second implementation manner, step S500, based on the first relational embedding array and the sensing embedding array, generates an integrated embedding array, including:

[0116] Step S501: Combine the first relational embedding array and the sensing embedding array to obtain an integrated embedding array; wherein, the dimension of the sensing embedding array is i, the dimension of the first relational embedding array is s, and the dimension of the integrated embedding array is i + s, where i ≥ 2 and s ≥ 2.

[0117] In step S501 of the second implementation manner of step S500, the first relational embedding array and the sensing embedding array are combined to obtain an integrated embedding array. This step simply and directly fuses the event association information contained in the first relational embedding array extracted from the event relationship network with the construction site environmental information contained in the sensing embedding array extracted from the environmental sensing dataset, thereby providing more comprehensive data support for subsequent construction monitoring.

[0118] The first relational embedding array is obtained by extracting features from the first local relationship network in the event relationship network, which reflects the relationships and context information between events at the construction site.

[0119] When performing step S501, the combination method adopted is to splice the first relationship embedding array and the sensing embedding array in a certain order to form a new array, that is, the integrated embedding array. Assume that the dimension of the sensing embedding array is i and the dimension of the first relationship embedding array is s, then the dimension of the integrated embedding array is i + s. The elements of the sensing embedding array can be arranged in sequence in the front, and the elements of the first relationship embedding array can be arranged in sequence in the back to form a new and longer vector. Of course, the arrangement order can also be adjusted according to specific requirements. For example, the elements of the first relationship embedding array can be arranged in the front and the elements of the sensing embedding array can be arranged in the back.

[0120] As a third implementation manner, step S500, generating an integrated embedding array based on the first relationship embedding array and the sensing embedding array, includes:

[0121] Step S5001: Obtain the Hadamard product of the first relationship embedding array and the sensing embedding array to obtain the integrated embedding array; the dimensions of the first relationship embedding array, the sensing embedding array, and the integrated embedding array are equal.

[0122] The Hadamard product, also known as the element-wise product, is an operation of multiplying corresponding elements, and requires that the dimensions of the two arrays participating in the operation are equal. Through the Hadamard product operation, the corresponding elements in the first relationship embedding array and the sensing embedding array are multiplied point by point to obtain a new array. This operation can highlight the interaction between the corresponding elements in the two arrays. For example, in the construction monitoring scenario, a certain element in the first relationship embedding array may represent the correlation strength between "tower crane operation" and "material hoisting", while the corresponding element in the sensing embedding array may represent the current wind speed. By the Hadamard product, multiplying these two elements, the obtained result can reflect the actual influence degree of the correlation between "tower crane operation" and "material hoisting" under the current wind speed condition. If the wind speed is relatively high and the correlation strength between "tower crane operation" and "material hoisting" is relatively high, then the result after multiplication may be relatively large, which may mean that in the current environment, the coordinated operation of these two events needs to be more cautious and there may be certain safety risks.

[0123] In one implementation manner, the method provided by the embodiments of the present application may further include:

[0124] Step S200a: When the environmental sensing data set further includes a second event, obtain a second local relationship network in the event relationship network, where the second event belongs to the event represented by the second constituent item in the event relationship network, and the second local relationship network includes the second constituent item, the constituent items within a preset path length range obtained based on the second constituent item, and the connection lines between the constituent items;

[0125] Step S200b: Based on the second local relationship network, obtain the second relationship embedding array based on the first feature extraction component in the construction monitoring neural network.

[0126] At this time, as a fourth implementation, step S500, generate an integrated embedding array based on the first relationship embedding array and the sensing embedding array, including:

[0127] Step S500a: Generate an integrated embedding array based on the first relationship embedding array, the second relationship embedding array, and the sensing embedding array.

[0128] In step S200a, when the environmental sensing data set further includes a second event, obtain the second local relationship network in the event relationship network, where the second event belongs to the event represented by the second component item in the event relationship network, and the second local relationship network includes the second component item, the component items within the preset path length obtained based on the second component item, and the connection lines between the component items. The second event is another representative event detected from the environmental sensing data set other than the first event. When performing this step, find the second component item corresponding to the second event in the event relationship network. For example, if the first event is "tower crane operation" and the second event is "concrete pouring", find the component item representing "concrete pouring" in the event relationship network as the second component item. Then, centered on the second component item, according to the preset path length, use a graph traversal algorithm (such as breadth-first search or depth-first search) to obtain other component items related to the second component item and the connection lines between them, thereby generating the second local relationship network. The preset path length limits the scope of the local relationship network and reflects the tightness of the association between events. Taking the breadth-first search algorithm as an example, starting from the second component item, mark it as visited and add it to the queue. Then, take out a node from the queue and check all its adjacent nodes. If an adjacent node has not been visited and the path length to the second component item is not greater than the preset path length, mark it as visited and add it to the queue, and record the connection line information between the nodes. Repeat this process until the queue is empty, and finally obtain the second local relationship network.

[0129] In step S200b, based on the second local relationship network, obtain the second relationship embedding array based on the first feature extraction component in the construction monitoring neural network. The first feature extraction component is a part of the construction monitoring neural network, and its role is to extract representative feature information from the second local relationship network and convert it into a numerical vector, that is, the second relationship embedding array. Graph embedding technology can be used to achieve this goal, and graph embedding algorithms include DeepWalk, Node2Vec, etc., and specific references can be made to the relevant content introduction above.

[0130] Based on steps S200a - S200b, in step S500a of the fourth implementation manner of step S500, an integrated embedding array is generated based on the first relationship embedding array, the second relationship embedding array, and the sensing embedding array. The purpose of this step is to fuse the relationship information of multiple events with the environmental sensing information to provide more comprehensive information for construction monitoring.

[0131] When executing step S500a, one method is to first process the first relationship embedding array and the second relationship embedding array, and then integrate them with the sensing embedding array. For example, first obtain the Hadamard product of the first relationship embedding array and a preset parameter array (such as a weight vector) to get a first intermediate result; obtain the Hadamard product of the second relationship embedding array and the preset parameter array to get a second intermediate result. The role of the preset parameter array is to weight the elements in the first relationship embedding array and the second relationship embedding array to highlight the importance of different event relationships. Next, multiply the sensing embedding array by the first weight array to get a first product result. The first weight array is a two-dimensional array (matrix), and its role is to weight each element in the sensing embedding array to highlight the importance of different environmental features. Then, multiply the target relationship embedding array by the second weight array and then flip it to get a flipped result. The second weight array is also used to weight the elements in the target relationship embedding array to reflect the importance of different event relationships. Finally, multiply the first product result by the flipped result to get a second product result, and then add the bias array to the second product result to obtain the integrated embedding array. Another method to implement step S500a is to combine the target relationship embedding array and the sensing embedding array to obtain the integrated embedding array. For example, connect the target relationship embedding array and the sensing embedding array in a certain order to form a new array. Suppose the dimension of the sensing embedding array is i and the dimension of the target relationship embedding array is s, then the dimension of the integrated embedding array is i + s. There is also a method to fuse the target relationship embedding array and the sensing embedding array to obtain the integrated embedding array. The fusion method can be to take the average value, weighted average value, etc. of the two, and it is required that the dimensions of the target relationship embedding array, the sensing embedding array, and the integrated embedding array are the same.

[0132] As an implementation manner, after obtaining the environmental sensing data set to be analyzed, the method provided by the embodiments of the present application may further include the determination process of the above second event:

[0133] Step S1001: Perform event detection on the environmental sensing data set to obtain an event detection result;

[0134] Step S1002: When the event detection result indicates that the environmental sensing data set includes two events, compare the two events in the environmental sensing data set with the events in the event relationship network respectively to obtain the event comparison result of each event;

[0135] Step S1003: When an event comparison result indicates that an event in the environmental sensing data set corresponds to an event in the event relationship network, regard the corresponding event as the first event;

[0136] Step S1004: When the remaining event comparison result indicates that the remaining event in the environmental sensing data set corresponds to an event in the event relationship network, regard the corresponding remaining event as the second event.

[0137] In step S1001, event detection is performed on the environmental sensing data set to obtain an event detection result. The environmental sensing data set is composed of multi-source data collected by various sensors at the construction site. These data contain various physical information of the construction site, such as temperature, humidity, noise, vibration, etc. The purpose of event detection is to identify events with specific meanings from these complex data, such as the startup of construction equipment, the transportation of materials, the activities of personnel, etc.

[0138] Multiple technical means can be used for event detection. One method is the threshold-based detection method. For some types of sensor data, such as temperature, noise, etc., a reasonable threshold range can be set. When the data collected by the sensor exceeds this threshold range, it is considered that an event may have occurred. For example, at the construction site, the normal environmental temperature range may be between 20°C and 30°C. If the temperature detected by the temperature sensor suddenly rises to 40°C, it can be judged that events such as equipment overheating or fire may have occurred. The setting of the threshold can be obtained through statistical analysis of historical data or can be set artificially according to construction technology and safety standards.

[0139] Another method is the event detection method based on pattern recognition. The sensor data patterns corresponding to various events can be learned in advance, and then the real-time collected environmental sensing data can be matched with these patterns. For example, when a construction equipment starts up, the vibration sensor will detect vibration signals with specific frequencies and amplitudes. By analyzing the characteristics of the vibration signals, such as frequency distribution, amplitude change, etc., it can be judged whether the equipment has started. Pattern recognition can use machine learning algorithms, such as support vector machine (SVM), decision tree, etc., or can also use deep learning algorithms, such as convolutional neural network (CNN), recurrent neural network (RNN), etc.

[0140] In addition to the above methods, an event detection method based on time series analysis can also be adopted. Since environmental sensing data usually changes over time, time series analysis can capture the dynamic change characteristics of the data. For example, by analyzing the trends, periodicity, seasonality, etc. of the data, potential events can be discovered. Time series analysis methods include autoregressive integrated moving average model (ARIMA), seasonal decomposition, etc.

[0141] In step S1002, when the event detection result indicates that the environmental sensing data set includes two events, the two events in the environmental sensing data set are respectively compared with the events in the event relationship network to obtain the event comparison result of each event. The event relationship network is a pre-constructed network structure, where each component represents an event and the connection lines represent the relationships between events. Multiple methods can be used for similarity calculation, such as edit distance, cosine similarity, etc. Edit distance refers to the minimum number of operations required to convert one string into another through insertion, deletion, and replacement operations between two strings. Cosine similarity measures the similarity between two vectors by calculating the cosine value of the angle between them. Another comparison method is rule-based comparison. Some comparison rules can be predefined. For example, if the detected event involves a certain specific device, then search for the events related to that device in the event relationship network.

[0142] In step S1003, when an event comparison result indicates that an event in the environmental sensing data set corresponds to an event in the event relationship network, the corresponding event is taken as the first event.

[0143] In step S1004, when the remaining event comparison result indicates that the remaining event in the environmental sensing data set corresponds to an event in the event relationship network, the corresponding remaining event is taken as the second event. The determination of the second event further enriches the event information in the environmental sensing data set and provides a basis for more comprehensive analysis of the construction environment in the follow-up.

[0144] For example, if "tower crane startup" is determined as the first event and the "concrete mixing" event also corresponds to a component in the event relationship network, then "concrete mixing" is taken as the second event. Subsequently, the second local relationship network can be obtained based on the second event in the event relationship network, and the relevant information of the first event and the second event can be integrated to more accurately understand the situation at the construction site.

[0145] Steps S1001 - S1004 provide an important basis for subsequent intelligent monitoring of the construction environment through event detection of the environmental sensing data set, comparison with the event relationship network, and determination of the first event and the second event. Relevant context information can be obtained in the event relationship network using the first event and the second event, thereby better understanding the situation at the construction site and improving the accuracy and reliability of monitoring. In the complex environment of the construction site, accurate event detection and determination of the first and second events contribute to timely discovery of potential problems and risks, providing strong support for construction management and decision-making. For example, by determining the two events of "tower crane start" and "concrete mixing", their associations can be analyzed, such as whether the tower crane is used to lift concrete-related materials, etc., so as to make coordination and safety prevention measures in advance to ensure the smooth progress of the construction.

[0146] As an implementation manner, step S500a, based on the first relationship embedding array, the second relationship embedding array, and the sensing embedding array, generates an integrated embedding array, including:

[0147] Step S500a1: Obtain the Hadamard product of the first relationship embedding array and the preset parameter array to get the first intermediate result;

[0148] Step S500a2: Obtain the Hadamard product of the second relationship embedding array and the preset parameter array to get the second intermediate result;

[0149] Step S500a3: Fuse the first intermediate result and the second intermediate result to get the target relationship embedding array;

[0150] Step S500a4: Multiply the sensing embedding array by the first weight array to get the first multiplication result, where the dimension of the sensing embedding array is i;

[0151] Step S500a5: Multiply the target relationship embedding array by the second weight array and then flip it to get the flipped result, where the dimension of the target relationship embedding array is s;

[0152] Step S500a6: Multiply the first multiplication result by the flipped result to get the second multiplication result;

[0153] Step S500a7: Add the bias array to the second multiplication result to get the integrated embedding array; where the first weight array and the second weight array are obtained by extracting from a third-order tensor, the dimension of the third-order tensor is i×s×t, the dimension of the first weight array is i×1×t, the dimension of the second weight array is 1×s×t, and the dimension of the integrated embedding array is t.

[0154] In step S500a1, obtain the Hadamard product of the first relational embedding array and the preset parameter array to get the first intermediate result. The preset parameter array can be a preset weight vector, which is used to weight each element in the first relational embedding array to highlight the importance of different event relationships. Let the first relational embedding array be a and the preset parameter array be w, and they have the same dimension s. The Hadamard product is an operation of multiplying corresponding elements, so the calculation formula for the first intermediate result c is: , where represents the Hadamard product. For example, if the first relational embedding array , and the preset parameter array , then the first intermediate result . Through the Hadamard product operation, the elements in the first relational embedding array can be adjusted according to the preset parameter array, so that the importance of different event relationships can be reflected in subsequent calculations.

[0155] In step S500a2, obtain the Hadamard product of the second relational embedding array and the preset parameter array to get the second intermediate result. The second relational embedding array is a numerical representation of the feature information extracted from the second local relational network of the event relational network, which also reflects the relationships and context information between the events related to the construction site. The process of obtaining the second intermediate result can refer to the example in step S500a1 and will not be elaborated here.

[0156] In step S500a3, fuse the first intermediate result and the second intermediate result to obtain the target relational embedding array. There are various ways of fusion, such as dot product calculation, summation calculation, averaging, etc., which are not specifically limited.

[0157] In step S500a4, multiply the sensing embedding array by the first weight array to get the first product result, where the dimension of the sensing embedding array is i. The sensing embedding array is a numerical representation of the feature information extracted from the environmental sensing dataset, which contains various environmental state data of the construction site, such as data collected by sensors for temperature, humidity, noise, vibration, etc. The first weight array is a two-dimensional array (matrix), which is used to weight each element in the sensing embedding array to highlight the importance of different environmental features.

[0158] In step S500a5, multiply the target relational embedding array by the second weight array and then flip it to get the flipped result, where the dimension of the target relational embedding array is s. The second weight array is also used to weight the elements in the target relational embedding array to reflect the importance of different event relationships. Through this step, the target relational embedding array is weighted and its order is adjusted, increasing the diversity of information fusion.

[0159] In step S500a6, the first product result and the flipped result are multiplied to obtain the second product result. In this step, the first product result g (with dimension t×1) and the flipped result i (with dimension t×1) are multiplied element by element to obtain the second product result j. Through this multiplication operation, the information of the sensing embedding array and the target relationship embedding array is fused, so that the second product result contains both the environmental information of the construction site and the relationship information between events.

[0160] In step S500a7, the bias array is added to the second product result to obtain the integrated embedding array. The bias array b is a vector with dimension t×1, and its function is to translate the second product result, adjust the overall numerical level of the integrated embedding array, and increase the flexibility of the model. In practical applications, the values of the first weight array, the second weight array, and the preset parameter array can be determined through the training process. A large number of environmental sensing data set samples and corresponding prior construction monitoring result labels can be used, and an optimization algorithm (such as the stochastic gradient descent method) can be adopted to adjust these parameters, so that the generated integrated embedding array can better reflect the real situation of the construction environment, thereby improving the accuracy of subsequent construction monitoring.

[0161] Steps S500a1 - S500a7 effectively fuse the information of the first relationship embedding array, the second relationship embedding array, and the sensing embedding array through a series of operations. The generated integrated embedding array combines the environmental information of the construction site and the relationship information of multiple events, providing an important data basis for the decision-making component based on the construction monitoring neural network to obtain accurate construction monitoring results.

[0162] As a second implementation manner, step S500a, based on the first relationship embedding array, the second relationship embedding array, and the sensing embedding array, generates an integrated embedding array, including:

[0163] Step S500a01: Obtain the Hadamard product of the first relationship embedding array and the preset parameter array to get the first intermediate result;

[0164] Step S500a02: Obtain the Hadamard product of the second relationship embedding array and the preset parameter array to get the second intermediate result;

[0165] Step S500a03: Fuse the first intermediate result and the second intermediate result to obtain the target relationship embedding array;

[0166] Step S500a04: Combine the target relationship embedding array and the sensing embedding array to obtain the integrated embedding array; where, the dimension of the sensing embedding array is i, the dimension of the target relationship embedding array is s, and the dimension of the integrated embedding array is i + s.

[0167] In step S500a01, obtain the Hadamard product of the first relational embedding array and the preset parameter array to get the first intermediate result. In step S500a02, obtain the Hadamard product of the second relational embedding array and the preset parameter array to get the second intermediate result.

[0168] In step S500a03, fuse the first intermediate result and the second intermediate result to obtain the target relational embedding array. The fusion operation can be carried out in various ways, such as dot product calculation, summation calculation, averaging, etc., and is not specifically limited.

[0169] In step S500a04, combine the target relational embedding array and the sensing embedding array to obtain the integrated embedding array. The combination operation is to concatenate the target relational embedding array and the sensing embedding array in sequence. When it is detected that the dataset contains the first event and the second event, through the above steps, the relational information and sensing information related to these two events can be integrated into one array. The first relational embedding array and the second relational embedding array respectively represent the local relational information in the event relational network related to these two events, and the sensing embedding array represents the actual data collected by each sensor. Through the processing and combination of these arrays, the obtained integrated embedding array can more comprehensively reflect the situation at the construction site. The decision-making component of the construction monitoring neural network can make a more accurate judgment on the construction monitoring results based on this integrated embedding array, such as judging whether there are construction safety hazards and whether the construction progress is affected.

[0170] As a third implementation manner, step S500a, based on the first relational embedding array, the second relational embedding array, and the sensing embedding array, generate the integrated embedding array, including:

[0171] Step S500a001: Obtain the Hadamard product of the first relational embedding array and the preset parameter array to get the first intermediate result;

[0172] Step S500a002: Obtain the Hadamard product of the second relational embedding array and the preset parameter array to get the second intermediate result;

[0173] Step S500a003: Fuse the first intermediate result and the second intermediate result to obtain the target relational embedding array;

[0174] Step S500a004: Fuse the target relational embedding array and the sensing embedding array to obtain the integrated embedding array; wherein, the dimensions of the target relational embedding array, the sensing embedding array, and the integrated embedding array are the same.

[0175] In step S500a001, obtain the Hadamard product of the first relational embedding array and the preset parameter array to get the first intermediate result. In step S500a002, obtain the Hadamard product of the second relational embedding array and the preset parameter array to get the second intermediate result. In step S500a003, fuse the first intermediate result and the second intermediate result to obtain the target relational embedding array. The fusion operation can be performed in various ways, such as summation, averaging, etc. In step S500a004, fuse the target relational embedding array and the sensing embedding array to obtain the integrated embedding array, and the dimensions of the target relational embedding array, the sensing embedding array, and the integrated embedding array are the same. The sensing embedding array is obtained through the second feature extraction component in the construction monitoring neural network based on the environmental sensing data set, and it is a numerical representation of the environmental sensing data. There are also various fusion methods, such as weighted summation, taking the average value, etc.

[0176] Through the above steps, the relational information and sensing information related to these two events are fused. The first relational embedding array and the second relational embedding array respectively reflect the local relational information in the event relational network related to these two events, and the sensing embedding array is the actual data collected by each sensor. By processing and fusing these arrays, the obtained integrated embedding array can more comprehensively reflect the actual situation of bridge construction. It helps to effectively integrate and utilize multi-source data. The information in the event relational network and the environmental sensing data itself have different characteristics and sources. Through the Hadamard product and fusion operations, they can be organically combined. The role of the preset parameter array is to adjust the importance of each relational information according to different application scenarios and requirements. In different construction stages, the importance of different events may change. By adjusting the preset parameter array, the integrated embedding array can more accurately reflect the current construction state.

[0177] As an implementation manner, the training process of the construction monitoring neural network provided by this application is as follows:

[0178] Step S10: Obtain an environmental sensing data set sample library, where each environmental sensing training data set in the environmental sensing data set sample library includes a prior event label, the prior event label in each environmental sensing training data set belongs to the event represented by the target component in the event relational network, and each environmental sensing training data set carries a corresponding prior construction monitoring result label;

[0179] Step S20: For each environmental sensing training data set, extract the training local relational network corresponding to the environmental sensing training data set in the event relational network, where the training local relational network includes the target component, the components within the preset path length range obtained based on the target component, and the connection lines between the components;

[0180] Step S30: For each environmental sensing training data set, obtain a training relationship embedding array based on the training local relationship network and the first feature extraction component in the construction monitoring neural network;

[0181] Step S40: For each environmental sensing training data set, obtain a training sensing embedding array based on the environmental sensing training data set and the second feature extraction component in the construction monitoring neural network;

[0182] Step S50: For each environmental sensing training data set, generate a training integrated embedding array based on the training relationship embedding array and the training sensing embedding array;

[0183] Step S60: For each environmental sensing training data set, obtain the construction monitoring prediction result corresponding to the environmental sensing training data set based on the training integrated embedding array and the decision-making component in the construction monitoring neural network;

[0184] Step S70: Train the construction monitoring neural network according to the construction monitoring prediction result corresponding to each environmental sensing training data set and the error marked by the prior construction monitoring result.

[0185] Steps S10 - S70 in the training process of the construction monitoring neural network are a series of operations performed to enable the construction monitoring neural network to accurately monitor the construction site. Its core purpose is to enable the neural network to learn the internal relationship between environmental sensing data and construction monitoring results through a large amount of training data and a reasonable training method, thereby improving the accuracy of construction site monitoring and overcoming the problems of missing context and inaccurate monitoring caused by the short real-time detection data stream.

[0186] In step S10, an environmental sensing data set sample library is obtained. Each environmental sensing training data set in the environmental sensing data set sample library includes a prior event label, and the prior event label in each environmental sensing training data set belongs to the event represented by the target component in the event relationship network. Each environmental sensing training data set also carries a corresponding prior construction monitoring result label. The environmental sensing data set sample library is a collection containing a large amount of training data, which is collected from the actual construction site and contains information collected by various environmental sensors, such as temperature, humidity, pressure, vibration, etc. The prior event label is a pre-label of the events occurring in the data set, which corresponds to the target component in the event relationship network. The event relationship network is a network structure describing the relationships between events, where the components represent events and the connection lines represent event relationships. The prior construction monitoring result label is the label of the actual construction monitoring result corresponding to each environmental sensing training data set, such as whether the construction is safe, whether there are faults, etc.

[0187] In step S20, for each environmental sensing training data set, a training local relationship network corresponding to the environmental sensing training data set is extracted from the event relationship network, where the training local relationship network includes target components, components within a preset path length obtained based on the target components, and connection lines between the components. The preset path length refers to the number limit of connection lines passed from one component to another. By extracting the training local relationship network, the local event relationship information related to the prior event markers in the current environmental sensing training data set can be focused on, thereby providing more targeted information for subsequent feature extraction. For example, if the target component corresponding to the prior event marker is "construction equipment failure" and the preset path length is 2, then the training local relationship network will include the component "construction equipment failure" and other components connected to it through no more than 2 connection lines, such as the components "inadequate equipment maintenance" and "equipment part aging" and the connection lines between them. A graph traversal algorithm, such as the breadth-first search algorithm, can be used to start from the target component and traverse the event relationship network according to the limit of the preset path length to extract the relevant components and connection lines and construct the training local relationship network.

[0188] In step S30, for each environmental sensing training data set, based on the training local relationship network, a training relationship embedding array is obtained by the first feature extraction component in the construction monitoring neural network. The first feature extraction component is an important part of the construction monitoring neural network, and its role is to perform feature extraction and numerical representation on the information in the training local relationship network to obtain the training relationship embedding array. The training relationship embedding array is a numerical encoding of the event relationship information in the training local relationship network, and it can reflect the degree of association and features between events. For example, the training local relationship network contains events such as "construction equipment failure", "inadequate equipment maintenance", and "equipment part aging". The first feature extraction component may convert the relationships between these events into a vector, such as [0.2, 0.3, 0.5], where each element represents the feature value of a different event relationship. A graph neural network (GNN) in deep learning can be used to implement the first feature extraction component, and the graph neural network can effectively process graph-structured data and extract the features of nodes and edges in the graph.

[0189] In step S40, for each environmental sensing training data set, based on the environmental sensing training data set, a training sensing embedding array is obtained based on the second feature extraction component in the construction monitoring neural network. The role of the second feature extraction component is to extract features and numerically represent the sensing information in the environmental sensing training data set to obtain the training sensing embedding array. The training sensing embedding array is a numerical encoding of the environmental sensing data, which can reflect the characteristics of various information collected by the environmental sensors. For example, the environmental sensing training data set contains readings from sensors such as temperature, humidity, and pressure. The second feature extraction component may convert these readings into a vector, such as [25.5, 60, 101.3], where each element represents the feature value of a different sensor. A multi-layer perceptron (MLP) can be used to implement the second feature extraction component, and the multi-layer perceptron can perform non-linear transformations on the input sensor data to extract useful features.

[0190] In step S50, for each environmental sensing training data set, a training integrated embedding array is generated based on the training relationship embedding array and the training sensing embedding array. The purpose of generating the training integrated embedding array is to fuse the event relationship information and the sensing information so that the subsequent decision-making component can comprehensively consider these two aspects of information to judge the construction monitoring results. There are various methods for generating the training integrated embedding array. For example, the training sensing embedding array can be multiplied by the first weight array, the training relationship embedding array can be multiplied by the second weight array and then flipped, and then the two results are multiplied and added with the bias array to obtain the training integrated embedding array; or the training relationship embedding array and the training sensing embedding array can be combined, or their Hadamard product can be obtained, etc. Taking the combination method as an example, if the training sensing embedding array is [25.5, 60, 101.3] and the training relationship embedding array is [0.2, 0.3, 0.5], then the training integrated embedding array is [25.5, 60, 101.3, 0.2, 0.3, 0.5]. The appropriate method for generating the training integrated embedding array can be selected according to specific application requirements and experimental results.

[0191] In step S60, for each environmental sensing training data set, according to the training integration embedding array, the construction monitoring prediction result corresponding to the environmental sensing training data set is obtained based on the decision-making component in the construction monitoring neural network. The decision-making component is the last component of the construction monitoring neural network, and its role is to make judgments and predictions according to the information in the training integration embedding array and output the construction monitoring prediction result. The decision-making component can be implemented using a fully connected layer and an activation function. For example, the softmax activation function is used to convert the output into a probability distribution, so as to obtain the probabilities of different construction monitoring results. For example, the probability of the construction monitoring prediction result being "construction safety" may be 0.8, and the probability of "hidden dangers in construction" may be 0.2. The forward propagation algorithm can be used to input the training integration embedding array into the decision-making component, and through a series of calculations and transformations, the construction monitoring prediction result is obtained.

[0192] In step S70, the construction monitoring neural network is trained according to the construction monitoring prediction result corresponding to each environmental sensing training data set and the error marked by the prior construction monitoring result. The error refers to the difference between the construction monitoring prediction result and the prior construction monitoring result mark. The parameters of the neural network need to be adjusted through an optimization algorithm to minimize the error. The optimization algorithm can be Stochastic Gradient Descent (SGD), Adaptive Moment Estimation (Adam), etc. For example, the prior construction monitoring result mark is "construction safety", while the probability of "construction safety" in the construction monitoring prediction result is 0.6, and the probability of "hidden dangers in construction" is 0.4, then there is a certain error. The backpropagation algorithm can be used to calculate the gradient of each parameter according to the error, and then update the parameters according to the rules of the optimization algorithm. Taking the stochastic gradient descent algorithm as an example, the parameter update formula is , where is the parameter of the neural network, is the learning rate, is the gradient of the error function with respect to the parameter.

[0193] During the training process, the batch training method can be adopted. The environmental sensing training data set is divided into multiple batches for training, and each batch of data is trained at a time. This can reduce the use of memory and improve the training efficiency. In addition, in order to prevent the occurrence of overfitting, regularization methods can be adopted. For example, L1 and L2 regularization, by adding a regularization term to the error function, limit the size of the neural network parameters and avoid the model from being too complex. The error function of L2 regularization can be expressed as , where is the original error function, is the regularization coefficient, is the parameter of the neural network.

[0194] The embodiment of the present application provides a monitoring system, such asFigure 2 As shown in Figure 2 , the monitoring system 100 includes: a processor 101 and a memory 103. Among them, the processor 101 and the memory 103 are connected, such as connected by a bus 102. Optionally, the monitoring system 100 may further include a transceiver 104. It should be noted that in practical applications, the number of transceivers 104 is not limited to one, and the structure of the monitoring system 100 does not constitute a limitation to the embodiments of the present application.

[0195] The processor 101 can be a CPU, a general-purpose processor, a GPU, a DSP, an ASIC, an FPGA or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It can implement or execute various exemplary logic blocks, modules and circuits described in combination with the disclosure of the present application. The processor 101 can also be a combination that realizes computing functions, such as a combination including one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0196] The bus 102 may include a path for transmitting information between the above components. The bus 102 can be a PCI bus or an EISA bus, etc. The bus 102 can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, Figure 2 only a thick line is used to represent it in the figure, but it does not mean that there is only one bus or one type of bus.

[0197] The memory 103 can be a ROM or other types of static storage devices that can store static information and instructions, a RAM or other types of dynamic storage devices that can store information and instructions, or it can also be an EEPROM, a CD-ROM or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto.

[0198] The memory 103 is used to store the application program code for executing the solution of the present application, and is controlled by the processor 101 to execute. The processor 101 is used to execute the application program code stored in the memory 103 to implement the content shown in any of the foregoing method embodiments.

[0199] The embodiments of the present application provide a monitoring system. The monitoring system in the embodiments of the present application includes: one or more processors; a memory; one or more computer programs, where one or more computer programs are stored in the memory and are configured to be executed by one or more processors. When the one or more programs are executed by the processor, the intelligent monitoring method for the construction environment based on the Internet of Things provided by the embodiments of the present application is implemented.

[0200] An embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program runs on a processor, the processor can execute the corresponding content in the foregoing method embodiment.

[0201] It should be understood that although the steps in the flowchart of the accompanying drawings are shown in sequence according to the indication of the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, the execution of these steps has no strict order limitation, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

[0202] The above are only some embodiments of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. An intelligent monitoring method for construction environment based on the Internet of Things, characterized in that, Including: Obtain an environmental sensing data set to be analyzed; the environmental sensing data set is composed of multi-source data collected by various sensors at a construction site, and this data contains various physical information of the construction site; Perform event detection on the environmental sensing data set to obtain an event detection result; the purpose of event detection is to identify events with specific meanings from this complex data; When the environmental sensing data set includes a first event, obtain a first local relationship network in the event relationship network, where each component item in the event relationship network represents an event, each connection line in the event relationship network represents an event relationship, the first event belongs to the event represented by the first component item in the event relationship network, and the first local relationship network includes the first component item, component items within a preset path length range obtained based on the first component item, and the connection lines between the component items; Based on the first local relationship network, obtain a first relationship embedding array based on the first feature extraction component in the construction monitoring neural network; Based on the environmental sensing data set, obtain a sensing embedding array based on the second feature extraction component in the construction monitoring neural network; Generate an integrated embedding array based on the first relationship embedding array and the sensing embedding array; Based on the integrated embedding array, obtain a construction monitoring result corresponding to the environmental sensing data set based on the decision-making component in the construction monitoring neural network.

2. The method according to claim 1, characterized in that, After obtaining the environmental sensing data set to be analyzed, the method further includes: When the event detection result indicates that the environmental sensing data set includes an event, compare the event in the environmental sensing data set with the events in the event relationship network to obtain an event comparison result; When the event comparison result indicates that the event in the environmental sensing data set corresponds to the event in the event relationship network, use the corresponding event as the first event.

3. The method according to claim 2, wherein The step of, when the event detection result indicates that the environmental sensing data set includes an event, comparing the event in the environmental sensing data set with the events in the event relationship network to obtain an event comparison result includes: When the event detection result indicates that the environmental sensing data set includes multiple events, obtain the conditional probability corresponding to each event in the multiple events; Use the event corresponding to the maximum conditional probability among the multiple events as the event in the environmental sensing data set; Compare the event in the environmental sensing data set with the events in the event relationship network to obtain the event comparison result.

4. The method according to claim 1, wherein The step of, when the environmental sensing data set includes a first event, obtaining a first local relationship network in the event relationship network includes: Use the component item corresponding to the first event in the event relationship network as the first component item; Based on the event relationship network, obtain a set of component items involved with the first component item, and the path length between each component item in the set of component items and the first component item is not greater than the preset path length; Generate the first local relationship network based on the first component item, the set of component items, and the connection lines between the component items; Or; Use the components corresponding to the first event in the event relationship network as the first components; Based on the event relationship network, obtain a set of components involved with the first components, where the path length between each component in the set of components and the first components is not greater than the preset path length; When the environmental sensing data set includes a first event relationship, based on the event relationship network, obtain a first local set of components in the set of components that have the first event relationship with the first components; Generate the first local relationship network based on the first components, the first local set of components, and the connection lines between the components; Or; Use the components corresponding to the first event in the event relationship network as the first components; Based on the event relationship network, obtain a set of components involved with the first components, where the path length between each component in the set of components and the first components is not greater than the preset path length; When the environmental sensing data set includes a first event relationship and a second event relationship, based on the event relationship network, obtain a second local set of components in the set of components that have either or both of the first event relationship and the second event relationship with the first components; Generate the first local relationship network based on the first components, the second local set of components, and the connection lines between the components.

5. The method according to claim 1, characterized in that, The generating the integrated embedding array based on the first relationship embedding array and the sensing embedding array includes: Multiply the sensing embedding array by a first weight array to obtain a first multiplication result, where the dimension of the sensing embedding array is i; Flip the result of multiplying the first relationship embedding array by a second weight array to obtain a flipped result, where the dimension of the first relationship embedding array is s; Multiply the first multiplication result by the flipped result to obtain a second multiplication result; Add the bias array to the second multiplication result to obtain the integrated embedding array; Wherein, the first weight array and the second weight array are obtained by extracting a third-order tensor, the dimension of the third-order tensor is i×s×t, the dimension of the first weight array is i×1×t, the dimension of the second weight array is 1×s×t, i is the dimension of the sensing embedding array, s is the dimension of the first relationship embedding array, and t is the dimension of the integrated embedding array; Or; The generating the integrated embedding array based on the first relationship embedding array and the sensing embedding array includes: Combine the first relationship embedding array and the sensing embedding array to obtain the integrated embedding array; wherein, the dimension of the sensing embedding array is i, the dimension of the first relationship embedding array is s, and the dimension of the integrated embedding array is i + s, i≥2, s≥2; Or; The generating the integrated embedding array based on the first relationship embedding array and the sensing embedding array includes: Obtain the Hadamard product of the first relational embedding array and the sensing embedding array to obtain the integrated embedding array; the first relational embedding array, the sensing embedding array, and the integrated embedding array have the same dimension.

6. The method according to claim 1, characterized in that, The method further includes: When the environmental sensing data set further includes a second event, obtain a second local relational network in the event relational network, where the second event belongs to the event represented by the second component item in the event relational network, and the second local relational network includes the second component item, the component items within the preset path length range obtained based on the second component item, and the connection lines between the component items; Based on the second local relational network, obtain a second relational embedding array based on the first feature extraction component in the construction monitoring neural network; Generating an integrated embedding array based on the first relational embedding array and the sensing embedding array includes: Generate the integrated embedding array based on the first relational embedding array, the second relational embedding array, and the sensing embedding array.

7. The method according to claim 6, characterized in that, After obtaining the environmental sensing data set to be analyzed, the method further includes: Perform event detection on the environmental sensing data set to obtain an event detection result; When the event detection result indicates that the environmental sensing data set includes two events, compare the two events in the environmental sensing data set with the events in the event relational network respectively to obtain an event comparison result for each event; When an event comparison result indicates that an event in the environmental sensing data set corresponds to an event in the event relational network, use the corresponding event as the first event; When the remaining event comparison result indicates that the remaining event in the environmental sensing data set corresponds to an event in the event relational network, use the corresponding remaining event as the second event.

8. The method according to claim 6, characterized in that The generating the integrated embedding array based on the first relational embedding array, the second relational embedding array, and the sensing embedding array includes: Obtain the Hadamard product of the first relational embedding array and a preset parameter array to obtain a first intermediate result; Obtain the Hadamard product of the second relational embedding array and the preset parameter array to obtain a second intermediate result; Fuse the first intermediate result and the second intermediate result to obtain a target relational embedding array; Multiply the sensing embedding array by a first weight array to obtain a first multiplication result, where the dimension of the sensing embedding array is i; Multiply the target relational embedding array by a second weight array and then flip it to obtain a flipped result, where the dimension of the target relational embedding array is s; Multiply the first multiplication result by the flipped result to obtain a second multiplication result; Add a bias array to the second multiplication result to obtain the integrated embedding array; Among them, the first weight array and the second weight array are obtained by extracting a third-order tensor. The dimension of the third-order tensor is i×s×t, the dimension of the first weight array is i×1×t, the dimension of the second weight array is 1×s×t, and the dimension of the integrated embedding array is t; Or; generating the integrated embedding array based on the first relational embedding array, the second relational embedding array, and the sensing embedding array includes: Obtaining the Hadamard product of the first relational embedding array and a preset parameter array to obtain a first intermediate result; Obtaining the Hadamard product of the second relational embedding array and the preset parameter array to obtain a second intermediate result; Fusing the first intermediate result and the second intermediate result to obtain a target relational embedding array; Combining the target relational embedding array and the sensing embedding array to obtain the integrated embedding array; Among them, the dimension of the sensing embedding array is i, the dimension of the target relational embedding array is s, and the dimension of the integrated embedding array is i + s; Or; generating the integrated embedding array based on the first relational embedding array, the second relational embedding array, and the sensing embedding array includes: Obtaining the Hadamard product of the first relational embedding array and a preset parameter array to obtain a first intermediate result; Obtaining the Hadamard product of the second relational embedding array and the preset parameter array to obtain a second intermediate result; Fusing the first intermediate result and the second intermediate result to obtain a target relational embedding array; Fusing the target relational embedding array and the sensing embedding array to obtain the integrated embedding array; Among them, the dimensions of the target relational embedding array, the sensing embedding array, and the integrated embedding array are the same.

9. The method according to claim 1, wherein The method further includes: Obtaining an environmental sensing dataset sample library, where each environmental sensing training dataset in the environmental sensing dataset sample library includes a prior event label. The prior event label in each environmental sensing training dataset belongs to the event represented by the target component in the event relationship network, and each environmental sensing training dataset carries a corresponding prior construction monitoring result label; For each environmental sensing training dataset, extracting a training local relationship network corresponding to the environmental sensing training dataset in the event relationship network, where the training local relationship network includes the target component, the components within the preset path length range obtained based on the target component, and the connection lines between the components; For each environmental sensing training dataset, obtaining a training relational embedding array based on the training local relationship network and the first feature extraction component in the construction monitoring neural network; For each environmental sensing training dataset, obtaining a training sensing embedding array based on the environmental sensing training dataset and the second feature extraction component in the construction monitoring neural network; For each environmental sensing training dataset, generating a training integrated embedding array based on the training relational embedding array and the training sensing embedding array; For each of the environmental sensing training data sets, based on the training integration embedding array, obtain the construction monitoring prediction result corresponding to the environmental sensing training data set based on the decision-making component in the construction monitoring neural network; Train the construction monitoring neural network according to the error marked by the construction monitoring prediction result corresponding to each environmental sensing training data set and the prior construction monitoring result.

10. A monitoring system, characterized in that, Comprising: One or more processors; A memory; One or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processor, the method according to any one of claims 1 to 9 is implemented.

Images