Data analysis method and system applied to smart construction site platform

Abstract

The present invention provides a data analysis method and system applied to a smart construction site platform, which relates to the field of smart construction site technology. By acquiring the construction interaction video data stream and sample influence coefficient of the construction site personnel, neural network parameter learning is continuously performed to generate a global construction risk decision network composed of multiple construction risk decision member networks. By clustering the construction risk decision distribution of the member network and dynamically updating the sample influence coefficient, the global construction risk decision network is continuously optimized until the training convergence requirement is met. The target global construction risk decision network finally generated can accurately make risk decisions on any input target construction interaction data, and generate construction risk prompt information for the construction site personnel, thereby effectively improving the intelligent level and risk warning capability of the construction safety management of the construction site, reducing the occurrence of construction accidents, and ensuring the safety of the construction site personnel.

Description

Technical Field

[0001] The present invention relates to the technical field of smart construction sites, and in particular to a data analysis method and system applied to a smart construction site platform. Background Art

[0002] With the rapid development of smart construction site platforms, intelligent management and monitoring of construction sites has become an industry trend. Smart construction site platforms integrate various advanced technologies such as the Internet of Things, big data analysis, and artificial intelligence to improve construction efficiency, reduce costs, and ensure site safety. However, in actual applications, smart construction site platforms still face many challenges, especially in terms of efficient and accurate risk analysis of construction interaction data.

[0003] Traditional construction risk analysis methods mainly rely on manual monitoring and empirical judgment, which is not only time-consuming and labor-intensive, but also easily affected by subjective factors, resulting in limited accuracy and reliability of risk analysis results. In recent years, although some machine learning-based methods have been proposed for construction risk analysis, most of these methods use a single neural network model, which is difficult to capture the complex features and diversity in construction interaction data, thus limiting the accuracy and generalization ability of risk analysis. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a data analysis method applied to a smart construction site platform, the method comprising:

[0005] Acquire a construction interaction video data stream of on-site personnel as a sample learning data sequence and a sample influence coefficient of each sample construction interaction video data in the sample learning data sequence;

[0006] Based on the sample learning data sequence and each sample influence coefficient, continuously performing neural network parameter learning to generate a global construction risk decision network, wherein the global construction risk decision network is composed of multiple construction risk decision member networks;

[0007] Obtaining the member network construction risk decision distribution of each of the construction risk decision member networks for the sample learning data sequence, clustering the construction risk decision member networks based on the construction risk decision distribution of each of the member networks, and generating at least one member network cluster;

[0008] Based on the construction risk decision distribution of each member network and the clustering of each member network, each sample influence coefficient is updated to enhance the sample learning data sequence, and then the step of returning to the sample learning data sequence and the sample influence coefficient to continue neural network parameter learning to generate a global construction risk decision network is performed until the global construction risk decision network meets the training convergence requirements, and a target global construction risk decision network is generated. After making a construction risk decision on any input target construction interaction data based on the target global construction risk decision network, construction risk warning information for the on-site personnel of the construction site is generated based on the construction risk decision result.

[0009] On the other hand, an embodiment of the present invention also provides a data analysis system applied to a smart construction site platform, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0010] Based on the above aspects, the embodiment of the present application obtains the construction interaction video data stream and sample influence coefficient of the construction site personnel, continuously performs neural network parameter learning, and generates a global construction risk decision network composed of multiple construction risk decision member networks. By clustering the construction risk decision distribution of the member network and dynamically updating the sample influence coefficient, the global construction risk decision network is continuously optimized until the training convergence requirements are met. The target global construction risk decision network finally generated can accurately make risk decisions for any input target construction interaction data, and generate construction risk warning information for construction site personnel, thereby effectively improving the intelligent level and risk warning capabilities of construction safety management on the construction site, reducing the occurrence of construction accidents, and ensuring the safety of construction site personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the data analysis method applied to the smart construction site platform provided in an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of the hardware architecture of a data analysis system applied to a smart construction site platform provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a data analysis method applied to a smart construction site platform provided by an embodiment of the present invention. The data analysis method applied to a smart construction site platform is introduced in detail below.

[0014] Step S110 , obtaining a construction interaction video data stream of on-site personnel as a sample learning data sequence and a sample influence coefficient of each sample construction interaction video data in the sample learning data sequence.

[0015] In this embodiment, a large construction site is taken as an example, and the server is responsible for collecting and processing data. In this construction site, there are multiple cameras distributed, which can capture various construction interaction scenes of on-site personnel, such as workers carrying construction materials, operating construction machinery, and collaborating with other workers when performing high-altitude operations.

[0016] The server obtains construction interaction video data streams from these cameras. For example, video stream 1 may show a group of workers building scaffolding, including how they pass steel pipes and fix fasteners; video stream 2 may be a scene of workers operating concrete mixers and vibrators at a concrete pouring site. These video streams constitute the sample learning data sequence.

[0017] To determine the sample influence coefficient, first obtain the annotated interest preference distribution corresponding to each sample construction interaction video data. For example, for the video stream 1 of scaffolding construction, according to previous safety standards and experience, the importance weights of different operation links are annotated. For example, the correct way of holding the steel pipe with both hands when passing it has a higher weight because it is related to the safety of the workers and the stability of the scaffolding construction. Assume that there are 5 construction risk decision-making member networks. In the scaffolding construction video, after analysis, it was found that 3 member networks made correct decisions on key operations (such as correctly holding the steel pipe for passing).

[0018] Set the initial parameter a = 0.5 (this parameter is set in advance based on experience to balance the impact of the number of member networks that make correct decisions in the calculation), and the smoothing factor b = 0.8 (used to control data stability and avoid excessive coefficient fluctuations). For this video stream 1 (assuming it is the sample construction interaction video data i = 1), the number of member networks k_1 = 3, n = 5.

[0019] Calculate the intermediate variable x_1 = k_1 / n + a * (n - k_1) / n = 3 / 5 + 0.5 * (5 - 3) / 5 = 0.6 + 0.2 = 0.8.

[0020] Calculate the sample influence coefficient c_1 = (x_1 * b) / (1+(x_1 * b - 1) * e ^ - (x_1 * b)) = (0.8 * 0.8) / (1+(0.8 * 0.8 - 1) * e ^ - (0.8 * 0.8))≈0.64 / (1+(0.64 - 1)* e ^ - 0.64)≈0.69.

[0021] Similarly, for the video stream 2 of concrete pouring, its sample influence coefficient is determined using the same method.

[0022] Step S120, based on the sample learning data sequence and each sample influence coefficient, continuously perform neural network parameter learning to generate a global construction risk decision network, wherein the global construction risk decision network is composed of multiple construction risk decision member networks.

[0023] In this embodiment, the server starts to perform neural network parameter learning based on the acquired sample learning data sequence and sample influence coefficient.

[0024] Parse sample construction interaction video data from the sample learning data sequence, such as the scaffolding construction video and concrete pouring video mentioned above.

[0025] For the scaffolding video (sample construction interaction video data), the sample influence coefficient is 0.69, and candidate sample learning data is determined based on this video data and the sample influence coefficient. This may involve extracting features from the video, such as extracting features such as the worker's posture and the way the tool is used, and weighting these features according to the sample influence coefficient.

[0026] Assume that there are three construction risk decision-making member networks. Load the processed candidate sample learning data into the three construction risk decision-making member networks respectively.

[0027] In the construction risk decision member network 1, the internal neural network calculation generates member network decision data. For example, based on the input weighted feature data, this network determines that the risk probability of a worker not wearing a safety belt at a certain moment in the scaffolding construction process is 0.3.

[0028] Similarly, construction risk decision-making member networks 2 and 3 also generate their own member network decision data.

[0029] Assume that the network participation coefficient is initially set to 0.33 (indicating that each member network has the same contribution weight to the global decision in the initial stage).

[0030] Based on the network participation coefficient corresponding to each construction risk decision-making member network, the decision data of each member network is fused. For example, the decision data of the three member networks are weighted and summed according to their respective network participation coefficients to obtain the global network decision data. If the decision data of the construction risk decision-making member network 1 is 0.3 (the risk probability of not wearing a seat belt), the network participation coefficient is 0.33; the decision data of the construction risk decision-making member network 2 is 0.2 (also the risk probability of not wearing a seat belt), and the network participation coefficient is 0.33; the decision data of the construction risk decision-making member network 3 is 0.4, and the network participation coefficient is 0.34 (due to calculation errors and other reasons, the sum of the coefficients is 1).

[0031] Global network decision data = 0.3 * 0.33 + 0.2 * 0.33 + 0.4 * 0.34 = 0.299.

[0032] Then, based on this global network decision data, the participation coefficients of each network and each construction risk decision member network are updated. For example, if the global network decision data shows that the risk probability is high, the network participation coefficient may be adjusted to increase the coefficient of the member network that performs better in this case, and at the same time adjust the parameters within the member network (such as the weight of the neuron, etc.).

[0033] Repeat this process, parsing new sample construction interaction video data from the sample learning data sequence each time, such as parsing concrete pouring video data and repeating the above operation. Until the participation coefficient of each network meets the set participation coefficient requirements (for example, the fluctuation range of a certain network participation coefficient is less than 0.01) and the member network parameter information of each construction risk decision member network meets the set convergence requirements (for example, the fluctuation of the decision result output by the member network is less than 0.05), a global construction risk decision network is generated.

[0034] Step S130, obtaining the member network construction risk decision distribution of each of the construction risk decision member networks for the sample learning data sequence, clustering the construction risk decision member networks based on the construction risk decision distribution of each of the member networks, and generating at least one member network cluster.

[0035] In this embodiment, the server obtains the member network construction risk decision distribution of each construction risk decision member network for the sample learning data sequence.

[0036] Taking the previous three construction risk decision-making member networks as an example, for each operation link in the scaffolding construction video, such as the workers' positioning, tool transfer, use of safety protection equipment, etc., each member network will generate a construction risk decision distribution.

[0037] Member network 1 may make a low risk probability (0.1) decision for worker position risk, a medium risk probability (0.3) decision for tool transfer risk, and a high risk probability (0.6) decision for safety protection equipment use risk; member network 2's risk decision results for the same operation links are 0.2, 0.4, and 0.5 respectively; member network 3's results are 0.15, 0.35, and 0.45.

[0038] Obtain the characteristic distance between the construction risk decision distributions of each member network. For example, the characteristic distance between member network 1 and member network 2 can be calculated by calculating the sum of the squares of the differences in the risk decision results of each operation link and then taking the square root. The difference in the risk decision results for worker position is (0.1 - 0.2)² = 0.01, the difference in the risk decision results for tool transfer is (0.3 - 0.4)² = 0.01, and the difference in the risk decision results for the use of safety protection equipment is (0.6 - 0.5)² = 0.01. The characteristic distance = √(0.01 + 0.01+0.01)≈0.173.

[0039] This characteristic distance is used as the clustering characteristic index between each construction risk decision-making member network. Based on this clustering characteristic index, each construction risk decision-making member network is clustered. Assuming that a distance threshold is set to 0.2, if the characteristic distance between two member networks is less than 0.2, they are grouped in the same cluster. According to this rule, if the characteristic distance between member network 1 and member network 3 is less than 0.2 after calculation, they can be grouped in one member network cluster, and member network 2 is a separate member network cluster.

[0040] Step S140, based on the construction risk decision distribution of each member network and the clustering of each member network, update the influence coefficient of each sample to enhance the sample learning data sequence, and then return to step S120 until the global construction risk decision network meets the training convergence requirements, generate a target global construction risk decision network, and make a construction risk decision on any input target construction interaction data based on the target global construction risk decision network, and generate construction risk warning information for the construction site personnel based on the construction risk decision result.

[0041] In this embodiment, the sample influence coefficient is updated based on the member network construction risk decision distribution and member network clustering obtained previously.

[0042] Analyze the target member network clusters that meet the target requirements in the member network clusters. Assuming the correlation requirement is considered, obtain the global construction risk decision distribution of the global construction risk decision network for the sample learning data sequence. For example, the overall risk decision distribution of the global construction risk decision network for the scaffolding video is that the risk probability of not wearing a safety belt is 0.3, the risk probability of improper standing position is 0.2, etc.

[0043] For the member network cluster containing member network 1 and member network 3 obtained by the previous clustering, calculate the correlation between this cluster and the global construction risk decision network. The correlation can be measured by calculating the sum of the similarities between the member network decision results in the cluster and the global decision results in each operation link. If this correlation is less than the set correlation value (assuming it is 0.5), this cluster is the target member network cluster.

[0044] Based on the target member network clustering and the construction risk decision distribution of each member network, the influence coefficient of each sample is updated.

[0045] Analyze the construction risk decision distribution of member network 1 and member network 3 in the target member network cluster, and determine the decision tendency of each construction risk decision member network for different examples of construction interaction video data (such as scaffolding video). For example, member network 1 tends to have a high risk probability for the risk decision of not wearing a seat belt, and member network 3 tends to have a medium risk probability. Record the distribution parameters of the decision tendency within the target member network cluster, and obtain the internal decision association parameters of the target member network cluster.

[0046] According to the internal decision-related parameters and the construction risk decision distribution of each member network, the sample construction interaction video data is statistically analyzed for correct decision-making. For example, in this cluster, for the risk decision of not wearing a safety belt in the scaffolding video, member network 1 made a correct decision (assuming that its decision result is consistent with a more accurate reference result), and member network 3 made an incorrect decision. Then the number of member networks with correct decisions in this cluster accounts for 1 / 2 = 0.5 of the total number of member networks in the target member network cluster, which is used as the preliminary correct decision ratio of the sample construction interaction video data in the target member network cluster.

[0047] The initial correct decision ratio of each sample construction interaction video data in the target member network cluster is correlated with the overall construction risk decision distribution of each member network. For example, for the risk of not wearing a seat belt, if most member networks in the global construction risk decision network consider it a high risk, and the initial correct decision ratio in this target member network cluster is low, this indicates that the decision of this sample construction interaction video data in the cluster is different from the overall decision.

[0048] According to the results of the association analysis, the sample influence coefficient adjustment direction of each sample construction interaction video data (such as the situation of not wearing a safety belt in the scaffolding video) is preliminarily determined. Since there are differences between the decision-making within the target member network cluster and the overall decision-making and the correct decision-making ratio within the cluster is low, it is preliminarily determined that the sample influence coefficient of this sample construction interaction video data needs to be increased.

[0049] Combined with the internal decision-making correlation parameters, the preliminary judgment results are optimized. For example, considering that the decision tendency of member network 1 in the cluster is a high risk probability and the decision accuracy is high in other similar scenarios, the magnitude of the sample influence coefficient that needs to be increased is further determined.

[0050] The original sample influence coefficient is updated according to the sample influence coefficient adjustment range.

[0051] Then, return to step S120 to continue neural network parameter learning, and repeat this process until the global construction risk decision network meets the training convergence requirements, such as the network participation coefficient no longer changes and the decision results output by the member network are stable.

[0052] After the target global construction risk decision network is generated, any input target construction interaction data, such as a new construction scene video, is input into the target global construction risk decision network for construction risk decision. Assuming that the new video is about steel structure welding, the target global construction risk decision network analyzes and concludes that the risk probability of a welder not wearing a protective mask is 0.8.

[0053] Based on the construction risk decision result, the server generates construction risk warning information for the construction site personnel. For example, it sends information to the construction site management system to remind workers in the steel structure welding area that there are workers who are not wearing protective masks, and the risk is high. Please remind the workers to wear protective masks in time.

[0054] Based on the above steps, the embodiment of the present application obtains the construction interaction video data stream and sample influence coefficient of the construction site personnel, continuously performs neural network parameter learning, and generates a global construction risk decision network composed of multiple construction risk decision member networks. By clustering the construction risk decision distribution of the member network and dynamically updating the sample influence coefficient, the global construction risk decision network is continuously optimized until the training convergence requirements are met. The target global construction risk decision network finally generated can accurately make risk decisions for any input target construction interaction data, and generate construction risk warning information for construction site personnel, thereby effectively improving the intelligent level and risk warning capabilities of construction safety management on the construction site, reducing the occurrence of construction accidents, and ensuring the safety of construction site personnel.

[0055] In a possible implementation, the step of continuously performing neural network parameter learning based on the sample learning data sequence and each sample influence coefficient to generate a global construction risk decision network, wherein the global construction risk decision network is composed of multiple construction risk decision member networks, includes:

[0056] Step A110, parsing sample construction interaction video data from the sample learning data sequence, and determining candidate sample learning data based on the sample construction interaction video data and the sample influence coefficient corresponding to the sample construction interaction video data.

[0057] Step A120, loading the candidate sample learning data into each of the construction risk decision-making member networks respectively to generate decision data for each member network.

[0058] Step A130, based on the network participation coefficient corresponding to each of the construction risk decision-making member networks, the decision data of each member network are integrated to generate global network decision data.

[0059] Step A140, based on the global network decision data, updates the network participation coefficients and the construction risk decision member networks, and returns to execute step A110 until the network participation coefficients meet the set participation coefficient requirements and the member network parameter information of each construction risk decision member network meets the set convergence requirements, thereby generating the global construction risk decision network.

[0060] In this embodiment, continuing to take the construction site as an example, the server has acquired a sample learning data sequence containing various construction scenes, including the scaffolding construction video and concrete pouring video mentioned earlier.

[0061] For the scaffolding video (sample construction interaction video data), the sample impact coefficient is 0.69. The server parses the video, and this parsing process involves in-depth processing of the video content. First, the video is decomposed into frames, and then features are extracted for each frame. For example, for the workers' movements, their hand gestures (such as the way of holding the steel pipe, the direction of the arm when passing the steel pipe, etc.), the center of gravity of the body (which is related to the stability of the workers on the scaffolding), and the direction of the face (to determine whether the workers are paying attention to the safety conditions around them) are identified. For construction tools, features such as the placement of the steel pipe and the connection status of the fasteners are identified.

[0062] After extracting these features, they are weighted according to the sample influence coefficient of 0.69. For example, since hand gestures are crucial to construction safety, their original feature values ​​may be multiplied by a relatively large weight (determined by the sample influence coefficient), while some relatively minor features (such as non-critical objects in the surrounding environment) are multiplied by a smaller weight. After such weighted processing, the data containing weighted features becomes the candidate sample learning data. The same method is also applied to other sample construction interaction video data such as concrete pouring videos.

[0063] Assume that there are three construction risk decision-making member networks in the server, namely member network A, member network B and member network C.

[0064] The candidate sample learning data of the scaffolding video is loaded into the member network A. The internal structure of the member network A is a complex neural network, which contains multiple neuron layers. When the input data enters the network, it first passes through the input layer, and then the hidden layer for further processing and analysis of the features. For example, the neurons in the hidden layer will perform complex calculations on the weighted features such as the worker's posture and tool status to determine the construction risk status represented by the combination of these features. After being processed by multiple hidden layers, the member network decision data about this scaffolding scene is finally obtained in the output layer. For example, the member network A determines that at a certain moment, the risk probability of the local structure of the scaffolding being unstable due to improper operation of the workers during the scaffolding process is 0.35.

[0065] Similarly, the same candidate sample learning data is loaded into member network B and member network C. Member network B may use its unique neural network structure and algorithm to conclude that the risk probability of instability of the local structure of the scaffold at this moment is 0.4; member network C concludes that the risk probability is 0.3. These are the member network decision data generated by member network B and member network C for the scaffolding construction video.

[0066] For other candidate sample learning data such as concrete pouring videos, they are also loaded into the three construction risk decision-making member networks in the same way to generate their corresponding member network decision data.

[0067] Initially, the network participation coefficients of member network A, member network B, and member network C are set to 0.33, 0.33, and 0.34, respectively (the sum of these coefficients is 1, indicating the relative contribution of each member network to the global decision in the initial stage).

[0068] For the risk decision data fusion of the scaffolding construction video, the calculation is based on the above network participation coefficient. The decision data of member network A is 0.35, multiplied by its network participation coefficient 0.33 to get 0.1155; the decision data of member network B is 0.4 multiplied by 0.33 to get 0.132; the decision data of member network C is 0.3 multiplied by 0.34 to get 0.102. Add these three results, that is, 0.1155 + 0.132 + 0.102 = 0.3495, which is the global network decision data for the scaffolding construction video, indicating the comprehensive risk probability of the local structure of the scaffolding being unstable due to improper operation of workers during the scaffolding construction process at this moment.

[0069] The same method is applied to the member network decision data fusion of other sample construction interaction video data such as concrete pouring video to obtain the corresponding global network decision data.

[0070] Since the global network decision data obtained for the scaffolding construction video is 0.3495, the server will update each network participation coefficient and each construction risk decision member network based on this result.

[0071] If it is found in subsequent evaluations that the decision result (0.4) of member network B is closer to the actual risk status than other member networks (perhaps obtained by comparing with historical data or manually annotated data), then the network participation coefficient of member network B will be appropriately increased, while the network participation coefficients of member networks A and C will be correspondingly reduced. For example, the network participation coefficient of member network B is updated to 0.35, member network A is updated to 0.32, and member network C is updated to 0.33.

[0072] At the same time, the server will also update the internal parameters of each construction risk decision member network based on the global network decision data. Taking member network A as an example, since the global network decision data shows that the risk probability is 0.3495, while the decision data of member network A is 0.35, this indicates that there may be a certain deviation in the decision of member network A. The server will adjust the weights, biases and other parameters of the neurons inside member network A so that the next time it processes similar candidate sample learning data, it can get a result closer to the global network decision data.

[0073] After completing these updates, the server will return to the steps of parsing the sample construction interaction video data from the sample learning data sequence, such as continuing to parse the next concrete pouring video, and repeating the above process of determining candidate sample learning data, loading it into each member network, fusing decision data, and updating coefficients and network parameters.

[0074] This cycle will continue until the participation coefficient of each network meets the set participation coefficient requirements (for example, the fluctuation range of each network participation coefficient is less than 0.01 in multiple consecutive iterations) and the member network parameter information of each construction risk decision member network meets the set convergence requirements (for example, the fluctuation of the decision results output by the member network is less than 0.05 in multiple consecutive iterations). When these conditions are met, the global construction risk decision network is successfully generated. This global construction risk decision network can comprehensively consider the advantages of each construction risk decision member network and accurately make risk decisions for various construction scenarios.

[0075] In a possible implementation, step S130 includes:

[0076] Step S131, obtaining the characteristic distance between the construction risk decision distributions of each member network, and using the characteristic distance as a clustering characteristic indicator between each construction risk decision member network.

[0077] In this embodiment, it is further assumed that there are three construction risk decision member networks: member network 1, member network 2, and member network 3. For each operation link in the previous scaffolding construction video, such as worker positioning, tool delivery, and use of safety protection equipment, each member network has its own construction risk decision distribution.

[0078] The decision results of member network 1 for worker position risk are low risk probability (0.1), tool transfer risk is medium risk probability (0.3), and safety protection equipment use risk is high risk probability (0.6); the risk decision results of member network 2 for these operation links are 0.2, 0.4, and 0.5 respectively; the results of member network 3 are 0.15, 0.35, and 0.45.

[0079] The server calculates the characteristic distance between member network 1 and member network 2. The calculation method is to calculate the sum of the squares of the differences in the risk decision results of each operation link and then take the square root. The difference in the risk decision results for worker position is (0.1 - 0.2)^2 = 0.01, the difference in the risk decision results for tool transfer is (0.3 - 0.4)^2 = 0.01, and the difference in the risk decision results for safety protection equipment use is (0.6 - 0.5)^2 = 0.01. The characteristic distance is about 0.173.

[0080] Similarly, the server calculates the characteristic distances between member network 1 and member network 3, and between member network 2 and member network 3. These calculated characteristic distances are used as clustering characteristic indicators between each construction risk decision-making member network.

[0081] Step S132: clustering each of the construction risk decision-making member networks based on the clustering characteristic index to generate at least one member network cluster.

[0082] Assume that a distance threshold is set to 0.2. The server clusters the member networks according to the feature distance calculated previously.

[0083] The characteristic distance between member network 1 and member network 3 is less than 0.2 after calculation, so they can be classified into one member network cluster. The characteristic distance between member network 2 and member network 1 and member network 3 is greater than 0.2, so member network 2 is a separate member network cluster. In this way, the clustering of the construction risk decision-making member network is completed based on the clustering characteristic index, and at least one member network cluster is generated, which provides a basis for further analysis and processing.

[0084] In a possible implementation, step S140 includes:

[0085] Step S141 : parsing target member network clusters that meet target requirements in each of the member network clusters, wherein the target requirements include at least one of a correlation requirement and a weight requirement.

[0086] Step S142: updating the influence coefficient of each sample based on the target member network clustering and the construction risk decision distribution of each member network.

[0087] In a possible implementation, step S142 includes:

[0088] Step S1421, analyze the construction risk decision distribution of each construction risk decision member network contained in each of the target member network clusters to determine the decision tendency of each construction risk decision member network for different sample construction interaction video data, and record the distribution parameters of the decision tendency within the target member network cluster to obtain the internal decision association parameters of the target member network cluster, wherein the distribution parameters include the number of member networks with different decision tendencies and the type of sample construction interaction video data targeted by the decision tendency.

[0089] Step S1422, based on the internal decision association parameters and the construction risk decision distribution of each member network, the sample construction interaction video data is statistically analyzed for the correct decision situation, and the ratio of the number of member networks with correct decisions in the target member network cluster to the total number of member networks in the target member network cluster is calculated as the preliminary correct decision ratio of the sample construction interaction video data in the target member network cluster.

[0090] Step S1423, performing correlation analysis on the initial correct decision ratio of each sample construction interaction video data in the target member network cluster and the overall member network construction risk decision distribution to obtain correlation analysis results, and based on the correlation analysis results, performing a preliminary determination on the adjustment direction of the sample influence coefficient of each sample construction interaction video data to obtain a preliminary determination result.

[0091] Step S1424, optimizes the preliminary determination result in combination with the internal decision-related parameters to obtain an adjustment direction optimization result of the sample influence coefficient of each sample construction interaction video data, and determines the adjustment amplitude of the sample influence coefficient of each sample construction interaction video data based on the adjustment direction optimization result and the internal decision-related parameters.

[0092] Step S1425 , updating the original sample influence coefficient according to the sample influence coefficient adjustment range of each sample construction interaction video data.

[0093] In a possible implementation, step S1423 includes:

[0094] Step S1423-1, determine the distribution pattern of decision tendencies of different construction risk decision-making member networks for each sample construction interaction video data, and generate a distribution pattern feature description.

[0095] Step S1423-2, extracting corresponding distribution features based on the initial correct decision ratio of each sample construction interaction video data in the target member network cluster.

[0096] Step S1423-3, based on the distribution pattern feature description and the distribution features of the preliminary decision error-free ratio, establish a basic framework for association comparison, and perform association analysis on the overall member network construction risk decision distribution and the preliminary decision error-free ratio according to the basic framework for association comparison to obtain preliminary association analysis results. The preliminary association analysis results are used to characterize the dimension of difference or consistency between the decision within the target member network cluster and the overall decision in the sample construction interaction video data.

[0097] Step S1423-4, based on the preliminary correlation analysis results, a preliminary determination is made on the adjustment direction of the sample influence coefficient of each sample construction interaction video data. If the difference between the decision of the sample construction interaction video data within the target member network cluster and the overall decision is greater than the set difference, it is preliminarily determined that the sample influence coefficient of the sample construction interaction video data needs to be increased; otherwise, it is preliminarily determined that the sample influence coefficient of the sample construction interaction video data needs to be reduced.

[0098] In this embodiment, the server first obtains the global construction risk decision distribution of the global construction risk decision network for the sample learning data sequence (such as the scaffolding video, concrete pouring video, etc. mentioned above). For example, for the risk of not wearing a safety belt in the scaffolding video, the global construction risk decision network determines that its risk probability is 0.3, and the risk probability of improper standing position is 0.2, etc.

[0099] For the member network clusters obtained by the previous clustering (assuming that member network 1 and member network 3 are in one cluster, and member network 2 is a separate cluster), calculate the correlation between the cluster and the global construction risk decision network. Taking the cluster containing member network 1 and member network 3 as an example, the calculation method is to calculate the sum of the similarities between the member network decision results in the cluster and the global decision results in each operation link. For example, for the risk of not wearing a seat belt, the decision result of member network 1 is 0.35, the decision result of member network 3 is 0.3, and the global decision result is 0.3. Calculate their similarity (which can be measured by simple difference absolute value, etc.), perform such calculations on multiple operation links and sum them up. If this correlation is less than the set correlation value (assuming it is 0.5), this cluster is the target member network cluster.

[0100] Next, the server obtains the network participation coefficient of each construction risk decision-making member network. For example, the network participation coefficient of member network 1 is 0.32, and the network participation coefficient of member network 3 is 0.33.

[0101] Based on these network participation coefficients, the weight coefficients of each member network cluster for the global construction risk decision network are determined. For the cluster containing member network 1 and member network 3, the weight coefficient is 0.32 + 0.33 = 0.65. If this weight coefficient is less than the set weight coefficient (assuming it is 0.7), then this cluster is the target member network cluster.

[0102] The relevance requirement and the weight requirement may also be considered simultaneously, for example, a cluster having a relevance less than a set relevance value and a weight coefficient less than a set weight coefficient among the member network clusters is determined as the target member network cluster.

[0103] Then, taking the target member network clustering including member network 1 and member network 3 as an example, we analyze their construction risk decision distribution for different example construction interaction video data (such as scaffolding construction video). For the risk of not wearing a safety belt in the scaffolding construction video, the decision result of member network 1 is 0.35, which tends to be high risk; the decision result of member network 3 is 0.3, which tends to be medium risk.

[0104] Record the distribution parameters of decision-making tendencies within the target member network cluster. For the risk of not wearing a seat belt, the number of member networks with different decision-making tendencies is 2 (one tends to be high risk and one tends to be medium risk), and the sample construction interaction video data type targeted by the decision-making tendencies is the situation of not wearing a seat belt in the scaffolding construction video. In this way, the internal decision-making association parameters of the target member network cluster are obtained.

[0105] Next, according to the internal decision-related parameters and the construction risk decision distribution of each member network, the sample construction interaction video data is statistically analyzed for the correct decision. Assume that for the risk of not wearing a safety belt in the scaffolding video, there is a more accurate reference risk value (which can be based on manual annotation or historical accurate data) of 0.32. Compared with the reference value of 0.32, the decision result of member network 1, 0.35, is considered to be correct within a certain range (assuming that this range is ±0.03), and the decision result of member network 3, 0.3, is also considered to be correct within this range. The number of member networks with correct decisions in this cluster is 2, and the total number of member networks is 2, so the initial correct decision ratio of the sample construction interaction video data (not wearing a safety belt) in the target member network cluster is 2 / 2 = 1.

[0106] Furthermore, the distribution pattern of the decision-making tendency of different construction risk decision-making member networks for each sample construction interaction video data (such as not wearing a safety belt, improper standing position, etc. in the scaffolding video) is determined. For example, in the global scope, for the risk of not wearing a safety belt, most member networks (assuming more than 60%) tend to be high risk, which is a description of the distribution pattern characteristics.

[0107] The correct proportion of initial decisions for the risk of not wearing a seat belt in the target member network cluster is 1, and its distribution characteristics are extracted. This proportion of 1 means that all decisions in the cluster are correct, which is a relatively special distribution feature.

[0108] According to the description of the distribution pattern characteristics (the distribution of decision-making tendencies in the global scope) and the distribution characteristics of the initial decision-making correct ratio (the correctness of decisions within the cluster), a basic framework for association comparison is established. For example, if there are more member networks that tend to be high-risk in the global scope, and the correctness ratio of decisions within the target member network cluster is 1, this indicates that the decision within the cluster is consistent with the global decision. According to this basic framework for association comparison, the overall construction risk decision distribution of each member network and the initial decision-making correct ratio are analyzed to obtain preliminary association analysis results.

[0109] According to the preliminary correlation analysis results, if the difference between the decision of the sample construction interaction video data (no seat belt) in the target member network cluster and the overall decision is greater than the set difference (assuming that the set difference is a risk probability difference of 0.1), it is initially determined that the sample influence coefficient of the sample construction interaction video data needs to be increased; otherwise, it is initially determined that the sample influence coefficient needs to be reduced. In this example, since the decision within the cluster is consistent with the global decision, it is initially determined that the sample influence coefficient may not need to be adjusted or reduced (depending on more precise calculation and setting rules).

[0110] Combined with internal decision-making correlation parameters (such as the number of member networks with different decision-making tendencies, etc.), the preliminary judgment results are optimized. For example, if the proportion of correct decisions in the target member network cluster is high, but the number of member networks with different decision-making tendencies is very different (such as one tends to be high-risk and the other tends to be medium-risk, and the difference is obvious), this may affect the direction and magnitude of the adjustment.

[0111] The sample influence coefficient adjustment range of each sample construction interaction video data is determined based on the optimized results. Assuming that after comprehensive consideration, for the sample construction interaction video data without wearing a seat belt, the adjustment range is determined to be - 0.05 (indicating a reduction in the sample influence coefficient).

[0112] The original sample impact coefficient is updated according to this adjustment range. If the original sample impact coefficient is 0.69, the updated sample impact coefficient is 0.69 - 0.05 = 0.64.

[0113] For other sample construction interaction video data (such as various risk situations in concrete pouring videos, etc.), the server also processes them according to the same process, continuously updating the sample influence coefficient to enhance the sample learning data sequence, thereby improving the accuracy and effectiveness of the entire construction risk decision-making network.

[0114] In a possible implementation, step S141 includes:

[0115] Obtain the global construction risk decision distribution of the global construction risk decision network for the sample learning data sequence.

[0116] Based on the construction risk decision distribution of each member network and the global construction risk decision distribution, the correlation between each member network cluster and the global construction risk decision network is generated.

[0117] A target member network cluster whose correlation is less than a set correlation value is determined in each of the member network clusters.

[0118] In a possible implementation, step S141 further includes:

[0119] Obtain the network participation coefficient of each of the construction risk decision-making member networks.

[0120] Based on the network participation coefficient, a weight coefficient of each member network cluster for the global construction risk decision network is determined.

[0121] A target member network cluster whose weight coefficient is less than a set weight coefficient is determined in each of the member network clusters.

[0122] In a possible implementation, step S141 further includes:

[0123] A target member network cluster whose relevance is less than a set relevance value and whose weight coefficient is less than a set weight coefficient is determined in each of the member network clusters.

[0124] In this embodiment, it is assumed that the sample learning data sequence contains video data of multiple construction scenes, such as scaffolding erection videos, concrete pouring videos, height operation videos, etc. For different operation links and risk factors in each video, the global construction risk decision network has corresponding risk decision results.

[0125] Taking the scaffolding video as an example, for the workers' standing position on the scaffolding, the global construction risk decision network determines the risk probability to be 0.2 after the previous learning and decision-making process. For the operation of workers transferring construction materials (such as steel pipes), the risk probability is determined to be 0.3, because there may be a risk of materials falling and injuring people below, or the workers' improper posture during the transfer process may cause themselves to lose balance. For whether the workers wear safety protection equipment correctly (such as helmets, safety belts, etc.), the risk probability is determined to be 0.4, because not wearing safety protection equipment correctly is a serious safety hazard during the construction process.

[0126] For the concrete pouring video, during the concrete mixing process, the global construction risk decision network determines that the risk probability of uneven concrete mixing due to mechanical failure is 0.15, and the risk probability of concrete overflow or unstable formwork support during the pouring process is 0.25, etc.

[0127] These risk probability judgment results for different sample learning data (various risk factors in the video data) together constitute the global construction risk decision distribution of the global construction risk decision network for the sample learning data sequence.

[0128] Reviewing the member network construction risk decision distribution, it is assumed that there are three construction risk decision member networks, namely member network A, member network B and member network C, and they have been clustered before, for example, member network A and member network C are in one cluster, and member network B is in another cluster.

[0129] For the construction risk decision distribution of member network A in the scaffolding erection video, the risk judgment for worker position is 0.18 (different from 0.2 judged by the global construction risk decision network), the risk judgment for material transfer is 0.32, and the risk judgment for wearing safety protection equipment is 0.42.

[0130] Member network C judged the same risk factors as 0.22, 0.28, and 0.38, respectively.

[0131] Member network B assessed the risk of workers’ standing positions in the scaffolding construction video as 0.25, the risk of materials transfer as 0.35, and the risk of wearing safety protection equipment as 0.3.

[0132] For the concrete pouring video, member networks A, C, and B also have their own corresponding construction risk decision distributions. For example, member network A judges the risk of uneven concrete mixing as 0.13, and the risk of concrete overflow or unstable formwork support as 0.23.

[0133] For the clusters containing member networks A and C, calculate the correlation between them and the global construction risk decision network. Taking the scaffolding video as an example, the calculation method is as follows:

[0134] For the worker position risk, the sum of the absolute values ​​of the differences between the decision results of member network A and member network C and the global decision result is calculated, that is, |0.18 - 0.2|+|0.22 - 0.2| = 0.02 + 0.02 = 0.04.

[0135] For the material transfer risk, calculate |0.32 - 0.3|+|0.28 - 0.3| = 0.02 + 0.02 = 0.04.

[0136] For the risk of wearing safety protective equipment, calculate |0.42 - 0.4|+|0.38 - 0.4| = 0.02 + 0.02= 0.04.

[0137] Then the sum of the absolute values ​​of these differences is summed up for all risk factors (here we only take the scaffolding construction video as an example, the actual calculation should cover all risk factors in the sample learning data). Assuming that the sum of the calculation results for other risk factors is 0.1, the sum of the absolute values ​​of the total differences is 0.04+0.04 + 0.04+0.1 = 0.22.

[0138] This 0.22 is a correlation measure between this cluster (including member network A and member network C) and the global construction risk decision network (this is just a simple calculation method, and more complex weighting methods may be used in practice).

[0139] Similarly, the correlation between the cluster where member network B is located and the global construction risk decision network is calculated.

[0140] In the process of determining the target member network cluster whose correlation is less than the set correlation value in each member network cluster, it is assumed that the set correlation value is 0.3. For the correlation measurement result calculated previously, the correlation of the cluster containing member network A and member network C is 0.22, which is less than 0.3; the correlation calculation result of the cluster where member network B is located is 0.28, which is also less than 0.3 (the calculation result assumed here). Therefore, in this case, the clusters where member network A and member network C are located and the cluster where member network B is located can both be determined as target member network clusters that meet the correlation requirements.

[0141] Furthermore, in the previous process of building the global construction risk decision-making network, each construction risk decision-making member network has a network participation coefficient, which represents the relative importance or participation degree of the member network in the entire global decision-making process.

[0142] For example, the network participation coefficient of member network A is initially set to 0.3, the network participation coefficient of member network B is 0.35, and the network participation coefficient of member network C is 0.35. These coefficients are set based on some prior knowledge or the principle of average distribution at the beginning of network construction, but as the network learns and optimizes, these coefficients will continue to adjust.

[0143] When calculating the weight coefficient of the cluster containing member network A and member network C, since the network participation coefficient of member network A is 0.3 and the network participation coefficient of member network C is 0.35, the weight coefficient of this cluster is 0.3 + 0.35 = 0.65.

[0144] The network participation coefficient of member network B is 0.35, so the weight coefficient of its cluster is 0.35.

[0145] Assume that the weight coefficient is set to 0.7. For the weight coefficient calculated above, the weight coefficient of the cluster containing member networks A and C is 0.65, which is less than 0.7; the weight coefficient of the cluster containing member network B is 0.35, which is also less than 0.7. Therefore, in this case, the cluster containing member networks A and C and the cluster containing member network B can be determined as target member network clusters that meet the weight requirements.

[0146] According to the results of the previous calculation, the correlation of the clusters where member networks A and C are located is less than the set correlation value (0.3) and the weight coefficient is less than the set weight coefficient (0.7); the correlation of the cluster where member network B is located is less than the set correlation value (0.3) and the weight coefficient is less than the set weight coefficient (0.7). Therefore, these two clusters meet this comprehensive requirement and are determined as target member network clusters.

[0147] Thus, the process of parsing the target member network clusters that meet the target requirements (at least one of the correlation requirements and the weight requirements, and the comprehensive requirements of the two) in each member network cluster is completed, laying the foundation for subsequent operations such as updating the sample influence coefficient based on the target member network cluster.

[0148] In a possible implementation manner, the step S143 further includes:

[0149] Based on the target member network clustering and the construction risk decision distribution of each member network, the clustering decision distribution of the target member network clustering is determined.

[0150] Obtain the correct decision ratio of each of the sample construction interaction video data in the cluster decision distribution being correctly decided by each of the construction risk decision member networks, wherein the cluster decision distribution includes the member network construction risk decision distribution of each construction risk decision member network for the sample construction interaction video data.

[0151] It is determined whether the correct decision ratio is greater than a set ratio. If so, the sample influence coefficient corresponding to each of the sample construction interaction video data is reduced; otherwise, the sample influence coefficient corresponding to each of the sample construction interaction video data is increased.

[0152] In this embodiment, a target member network cluster has been determined, for example, a cluster including member network A and member network C. Scaffolding erection video and concrete pouring video are used as sample construction interaction video data for illustration.

[0153] For the scaffolding video, member network A judged the worker position risk as 0.18, the material transfer risk as 0.32, and the safety protection equipment wearing risk as 0.42; member network C judged the worker position risk as 0.22, the material transfer risk as 0.28, and the safety protection equipment wearing risk as 0.38. This is the member network construction risk decision distribution of member network A and member network C for the scaffolding video, which together constitutes the cluster decision distribution of the target member network cluster for the scaffolding video.

[0154] Similarly, for the concrete pouring video, member network A determines the risk of uneven concrete mixing as 0.13 and the risk of concrete overflow or unstable formwork support as 0.23; member network C determines the risk of uneven concrete mixing as 0.15 and the risk of concrete overflow or unstable formwork support as 0.25. These decision results together constitute the clustering decision distribution of the target member network clustering for the concrete pouring video.

[0155] Next, first of all, we need to determine a standard for making correct decisions. For example, for each risk factor, set a reasonable risk probability range as the standard for making correct decisions. Assume that for the worker position risk, if the determined risk probability is within the range of [0.15, 0.25], the decision is considered correct; for the material transfer risk, if it is within the range of [0.25, 0.35], the decision is considered correct; for the safety protection equipment wearing risk, if it is within the range of [0.35, 0.45], the decision is considered correct. For the uneven concrete mixing risk in the concrete pouring video, if it is within the range of [0.1, 0.2], the decision is considered correct; for the risk of concrete overflow or unstable formwork support, if it is within the range of [0.2, 0.3], the decision is considered correct.

[0156] For scaffolding video:

[0157] In terms of worker position risk, member network A is judged as 0.18 and member network C is judged as 0.22, both within the range of [0.15, 0.25], so the number of member networks that make correct decisions on this risk factor is 2. A total of 2 member networks (member network A and member network C) participate in the decision-making, so the correct decision ratio is 2 / 2 = 1.

[0158] In terms of material transfer risk, member network A was judged to be 0.32, which is within the range of [0.25, 0.35], and member network C was judged to be 0.28, which is also within the range of [0.25, 0.35]. The number of member networks with correct decisions is 2, and the correct decision ratio is 2 / 2= 1.

[0159] In terms of the risk of wearing safety protection equipment, member network A judged it to be 0.42, which is within the range of [0.35, 0.45], and member network C judged it to be 0.38, which is also within the range of [0.35, 0.45]. The correct decision ratio is 2 / 2 = 1.

[0160] For the concrete pouring video:

[0161] In terms of the risk of uneven concrete mixing, member network A judged that 0.13 is within the range of [0.1, 0.2], and member network C judged that 0.15 is also within the range of [0.1, 0.2]. The correct decision ratio is 2 / 2 = 1.

[0162] In terms of the risk of concrete overflow or unstable formwork support, member network A judged it to be 0.23, which is within the range of [0.2, 0.3], and member network C judged it to be 0.25, which is also within the range of [0.2, 0.3]. The correct decision ratio is 2 / 2 = 1.

[0163] Assume that the set ratio is 0.8. For each risk factor in the scaffolding video, the correct decision ratio is 1, which is greater than the set ratio of 0.8. Therefore, for the sample impact coefficient corresponding to the scaffolding video, the server will reduce its value. For example, if the sample impact coefficient of the scaffolding video was 0.69 before, the sample impact coefficient after reduction may be 0.6 (the specific reduction can be based on the pre-set rules, this is just an example).

[0164] For each risk factor in the concrete pouring video, the correct decision ratio is also 1, which is greater than the set ratio of 0.8. Similarly, the server will reduce the sample impact coefficient corresponding to the concrete pouring video. Assuming it was 0.7 before, it may be 0.65 after the reduction.

[0165] In this way, the sample influence coefficient is updated according to the clustering decision distribution and the correct decision ratio of the target member network clustering, which helps to optimize the sample learning data sequence and further improve the accuracy and effectiveness of the global construction risk decision network.

[0166] In a possible implementation manner, before step S110, the method further includes:

[0167] Step S101, obtaining the annotated interest preference distribution corresponding to each of the sample construction interaction video data.

[0168] Step S102, based on each of the sample construction interaction video data and each of the construction risk decision member networks, generating a member network construction risk decision distribution of each of the construction risk decision member networks for the sample learning data sequence.

[0169] Step S103, based on the annotated interest preference distribution and the member network construction risk decision distribution, determining the number of member networks whose sample construction interaction video data is correctly decided by each construction risk decision member network.

[0170] Step S104: generating a sample influence coefficient of each sample construction interaction video data based on the number of member networks, initial parameters and smoothing factors.

[0171] The calculation process of step S104 includes:

[0172] Define the following variables:

[0173] Suppose there are n construction risk decision-making member networks, m sample construction interaction video data, the number of member networks is k, k is related to n, and is used to represent the number of member networks for which each sample construction interaction video data is correctly decided, and the initial parameter is set to a, which is a pre-set basic parameter related to the calculation of the sample influence coefficient, the smoothing factor is set to b, which is used to control data stability, and the number of member networks for which sample construction interaction video data i is correctly decided is set to k_i, i=1, 2,..., m.

[0174] For the sample construction interaction video data i, calculate the intermediate variable x_i=k_i / n+a*(n-k_i) / n.

[0175] Calculate the sample influence coefficient c_i = (x_i*b) / (1+(x_i*b-1)*e^-(x_i*b)).

[0176] In this embodiment, the server first analyzes the scaffolding construction video (sample construction interaction video data) to obtain its annotation interest preference distribution. For the scaffolding construction scene, different operation links have different degrees of importance and focus. For example, in the link where workers pass steel pipes, the correct way of passing (such as holding with both hands, passing smoothly, etc.) is very critical, so in the annotation interest preference distribution, this operation link may be given a higher weight. Assume that the weight of the steel pipe passing operation is 0.3.

[0177] The position of workers on the scaffold is also very important. A proper position can ensure their own safety and the stability of the scaffold. The weight of this operation may be set to 0.25. Some non-critical items in the surrounding environment (such as some small tools that are temporarily placed and do not affect construction safety) are less concerned and may have a weight of only 0.05.

[0178] The weight of safety protection equipment (such as wearing a safety belt) is very high, set to 0.4, because it is directly related to the life safety of workers. The weights of these different operation links constitute the annotation interest preference distribution corresponding to the scaffolding construction video.

[0179] During the concrete pouring process, the uniformity of concrete mixing is very critical, and its weight in the annotation interest preference distribution may be 0.3. The height control during concrete pouring is also very important, with a weight of 0.25, because an inappropriate pouring height may cause concrete splashing or uneven formwork pressure.

[0180] The correct use of the vibrator has a weight of 0.2, while some auxiliary facilities on the construction site (such as temporary lighting equipment, etc.) have a weight of 0.05, and protective equipment (such as protective shoes, goggles, etc.) has a weight of 0.2. These weights constitute the annotation interest preference distribution corresponding to the concrete pouring video.

[0181] The three construction risk decision-making member networks are member network 1, member network 2, and member network 3. These member networks have their own internal parameters and structures at the beginning. For example, member network 1 may be a neural network with 3 layers of neurons, member network 2 is a 4-layer neural network with different neuron connections, and member network 3 is a 3-layer neural network with a special activation function.

[0182] Member network 1 analyzes the scaffolding construction video. For the link where workers pass the steel pipe, based on the image data in the video and the calculation logic within the network, the risk probability of this link is determined to be 0.25 (because it may be found that some workers' hand movements are slightly unstable when passing).

[0183] For the worker position risk, it was determined to be 0.18 (by analyzing factors such as the worker’s position and posture on the scaffold).

[0184] The risk of wearing safety protection equipment is judged as 0.38 (it may be detected that some workers do not wear their safety belts properly). The judgment results of these risk probabilities constitute the member network construction risk decision distribution of member network 1 for the scaffolding erection video.

[0185] Member network 2 analyzes the same scaffolding construction video and determines the risk probability of workers passing steel pipes to be 0.3 (possibly based on different feature extraction and analysis methods).

[0186] The worker position risk is determined to be 0.22, and the risk of wearing safety protection equipment is determined to be 0.4. This is the member network construction risk decision distribution of member network 2 for the scaffolding construction video.

[0187] Member network 3 analyzes the scaffolding construction video. For the worker passing the steel pipe, the risk probability is determined to be 0.28; for the worker standing risk, it is determined to be 0.2; for the risk of wearing safety protection equipment, it is determined to be 0.35. This constitutes the member network construction risk decision distribution of member network 3 for the scaffolding construction video.

[0188] Member network 1 analyzes the concrete pouring video and determines the risk of concrete mixing uniformity with a probability of 0.15 (probably based on the flow state of concrete and other characteristics).

[0189] The risk of controlling the height of concrete pouring is judged as 0.23; the risk of using the vibrator correctly is judged as 0.18; and the risk of wearing protective equipment is judged as 0.2. These constitute the member network construction risk decision distribution of member network 1 for the concrete pouring video.

[0190] Member network 2 analyzes the concrete pouring video. For the concrete mixing uniformity risk, the probability is 0.18; for the concrete pouring height control risk, it is 0.25; for the vibrator correct use risk, it is 0.22; for the protective equipment wearing risk, it is 0.18. This is the member network construction risk decision distribution of member network 2 for the concrete pouring video.

[0191] Member network 3 analyzes the concrete pouring video. For the concrete mixing uniformity risk, the judgment probability is 0.16; for the concrete pouring height control risk, it is judged to be 0.22; for the vibrator correct use risk, it is judged to be 0.2; for the protective equipment wearing risk, it is judged to be 0.19. This constitutes the member network construction risk decision distribution of member network 3 for the concrete pouring video.

[0192] For the worker passing the steel pipe, according to the distribution of the marked interest preference, the weight is 0.3, and a reasonable risk probability range is set to judge whether the decision is correct. Assume that for this link, the risk probability is in the range of [0.2, 0.3] and the decision is considered correct.

[0193] The risk probability determined by member network 1 is 0.25, which is within this range, so member network 1 makes the correct decision on the worker passing the steel pipe. Member network 2 determines the risk probability to be 0.3, which is also within this range, so the decision is correct. Member network 3 determines the risk probability to be 0.28, which is also within the range, so the decision is correct. Therefore, the number of member networks that make the correct decision on the worker passing the steel pipe is 3.

[0194] For worker position risk, the risk probability is set to be within the range of [0.15, 0.25], and the decision is correct. According to the distribution of labeled interest preferences, its weight is 0.25. Member network 1 is judged to be 0.18, which is within the range, and the decision is correct. Member network 2 is judged to be 0.22, which is within the range, and the decision is correct. Member network 3 is judged to be 0.2, which is within the range, and the decision is correct. Therefore, for worker position risk, the number of member networks with correct decisions is 3.

[0195] For the risk of wearing safety protection equipment, the risk probability is set in the range of [0.3, 0.45] to be a correct decision, and its weight is 0.4 according to the distribution of annotated interest preferences.

[0196] Member network 1 is judged as 0.38, which is within the range, and the decision is correct. Member network 2 is judged as 0.4, which is within the range, and the decision is correct. Member network 3 is judged as 0.35, which is within the range, and the decision is correct. Therefore, for the risk of wearing safety protection equipment, the number of member networks that make correct decisions is 3.

[0197] For the concrete mixing uniformity risk, according to the distribution of annotated interest preferences, its weight is 0.3, and the risk probability is set to be in the range of [0.1, 0.2] for a correct decision. Member network 1 is judged to be 0.15, which is within the range, and the decision is correct. Member network 2 is judged to be 0.18, which is within the range, and the decision is correct. Member network 3 is judged to be 0.16, which is within the range, and the decision is correct. Therefore, for the concrete mixing uniformity risk, the number of member networks with correct decisions is 3.

[0198] For the concrete pouring height control risk, the risk probability is set to be within the range of [0.2, 0.3], and the decision is correct. According to the distribution of the labeled interest preference, its weight is 0.25. Member network 2 is judged to be 0.25, which is within the range, and the decision is correct. Member network 3 is judged to be 0.22, which is within the range, and the decision is correct. Therefore, for the concrete pouring height control risk, the number of member networks with correct decisions is 3.

[0199] For the risk of vibrator correct use, the risk probability is set to be within the range of [0.15, 0.25], and the decision is correct. According to the distribution of labeled interest preferences, its weight is 0.2. Member network 1 is judged to be 0.18, which is within the range, and the decision is correct. Member network 2 is judged to be 0.22, which is within the range, and the decision is correct. Member network 3 is judged to be 0.2, which is within the range, and the decision is correct. Therefore, for the risk of vibrator correct use, the number of member networks with correct decisions is 3.

[0200] For the risk of wearing protective equipment, the risk probability is set to be within the range of [0.15, 0.25], and the decision is correct. According to the distribution of annotated interest preferences, its weight is 0.2. Member network 1 is judged to be 0.2, which is within the range, and the decision is correct. Member network 2 is judged to be 0.18, which is within the range, and the decision is correct. Member network 3 is judged to be 0.19, which is within the range, and the decision is correct. Therefore, for the risk of wearing protective equipment, the number of member networks with correct decisions is 3.

[0201] Furthermore, it is assumed that there are n = 3 construction risk decision-making member networks and m sample construction interaction video data (here the scaffolding construction video is taken as a sample, that is, m = 1, and there will be multiple samples in actual situations).

[0202] For the worker passing the steel pipe in the scaffolding video, the number of member networks that were correctly determined was k_1 = 3 (determined earlier). The initial parameter a = 0.5 (pre-set, basic parameter related to the calculation of the sample influence coefficient) and the smoothing factor b = 0.8 (used to control data stability) were set.

[0203] Calculate the intermediate variable x_1 = k_1 / n + a*(n - k_1) / n = 3 / 3+0.5*(3 - 3) / 3 = 1 + 0 =1.

[0204] Calculate the sample influence coefficient c_1=(x_1*b) / (1+(x_1*b - 1)*e^-(x_1*b))=(1*0.8) / (1+(1*0.8 - 1)*e^-(1*0.8)) = 0.8 / (1+(0.8 - 1)*e^-0.8)≈0.8 / (1 - 0.2*e^-0.8)≈0.86.

[0205] The same method can be applied to other operational links in the scaffolding construction video (such as worker positioning risks, safety protection equipment wearing risks, etc.) and various risk factors in the concrete pouring video to calculate their respective sample impact coefficients.

[0206] Figure 2 The hardware structure of the data analysis system 100 applied to the smart construction site platform for implementing the data analysis method applied to the smart construction site platform provided by the embodiment of the present invention is shown as follows: Figure 2 As shown, the data analysis system 100 applied to the smart construction site platform may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0207] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the data analysis system 100 applied to the smart construction site platform uses to execute or use to complete the exemplary method described in the present invention.

[0208] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the data analysis method applied to the smart construction site platform in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0209] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned data analysis system 100 applied to the smart construction site platform. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0210] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the data analysis method applied to the smart construction site platform as above is implemented.

[0211] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A data analysis method applied to a smart construction site platform, It is characterized in that The method comprises: Acquire a construction interaction video data stream of on-site personnel as a sample learning data sequence and a sample influence coefficient of each sample construction interaction video data in the sample learning data sequence; Based on the sample learning data sequence and each sample influence coefficient, continuously performing neural network parameter learning to generate a global construction risk decision network, wherein the global construction risk decision network is composed of multiple construction risk decision member networks; Obtaining the member network construction risk decision distribution of each of the construction risk decision member networks for the sample learning data sequence, clustering the construction risk decision member networks based on the construction risk decision distribution of each of the member networks, and generating at least one member network cluster; Based on the construction risk decision distribution of each member network and the clustering of each member network, the influence coefficient of each sample is updated to enhance the sample learning data sequence, and then the step of continuously performing neural network parameter learning based on the sample learning data sequence and the influence coefficient of each sample is returned to generate a global construction risk decision network until the global construction risk decision network meets the training convergence requirements, and a target global construction risk decision network is generated. After making a construction risk decision on any input target construction interaction data based on the target global construction risk decision network, construction risk warning information for the construction site personnel is generated based on the construction risk decision result; The step of continuously performing neural network parameter learning based on the sample learning data sequence and each sample influence coefficient to generate a global construction risk decision network, wherein the global construction risk decision network is composed of multiple construction risk decision member networks, comprises: Parsing the sample construction interaction video data from the sample learning data sequence, and determining candidate sample learning data based on the sample construction interaction video data and the sample influence coefficient corresponding to the sample construction interaction video data; Loading the candidate sample learning data into each of the construction risk decision-making member networks respectively to generate decision data for each member network; Based on the network participation coefficient corresponding to each of the construction risk decision-making member networks, the decision data of each of the member networks are integrated to generate global network decision data; Based on the global network decision data, update each of the network participation coefficients and each of the construction risk decision member networks, return to execute the steps of parsing the sample construction interaction video data from the sample learning data sequence, and determine the candidate sample learning data based on the sample construction interaction video data and the sample influence coefficient corresponding to the sample construction interaction video data, until each of the network participation coefficients meets the set participation coefficient requirements and the member network parameter information of each of the construction risk decision member networks meets the set convergence requirements, thereby generating the global construction risk decision network; The step of clustering the construction risk decision member networks based on the construction risk decision distribution of each member network to generate at least one member network cluster comprises: Obtaining characteristic distances between construction risk decision distributions of each of the member networks, and using the characteristic distances as clustering characteristic indicators between each of the construction risk decision member networks; Based on the clustering characteristic index, each of the construction risk decision-making member networks is clustered to generate at least one member network cluster.

2. According to the data analysis method applied to the smart construction site platform according to claim 1, It is characterized in that The step of updating the influence coefficient of each sample based on the construction risk decision distribution of each member network and the clustering of each member network to enhance the sample learning data sequence includes: parsing a target member network cluster that meets a target requirement in each of the member network clusters, wherein the target requirement includes at least one of a relevance requirement and a weight requirement; Based on the target member network clustering and the construction risk decision distribution of each member network, updating the impact coefficient of each sample; The step of updating the influence coefficient of each sample based on the target member network clustering and the construction risk decision distribution of each member network includes: Analyze the construction risk decision distribution of each construction risk decision member network contained in each of the target member network clusters to determine the decision tendency of each construction risk decision member network for different sample construction interaction video data, and record the distribution parameters of the decision tendency within the target member network cluster to obtain the internal decision association parameters of the target member network cluster, wherein the distribution parameters include the number of member networks with different decision tendencies and the type of sample construction interaction video data targeted by the decision tendency; According to the internal decision association parameters and the construction risk decision distribution of each member network, the sample construction interaction video data is statistically analyzed for the correctness of the decision, and the proportion of the number of member networks with correct decisions in the target member network cluster to the total number of member networks in the target member network cluster is calculated as the preliminary correctness ratio of the sample construction interaction video data in the target member network cluster; Perform correlation analysis on the initial correct decision ratio of each sample construction interaction video data in the target member network cluster and the overall construction risk decision distribution of each member network to obtain a correlation analysis result, and perform a preliminary determination on the sample influence coefficient adjustment direction of each sample construction interaction video data according to the correlation analysis result to obtain a preliminary determination result; In combination with the internal decision-making association parameters, the preliminary determination result is optimized to obtain an adjustment direction optimization result of the sample influence coefficient of each sample construction interaction video data, and according to the adjustment direction optimization result and the internal decision-making association parameters, an adjustment range of the sample influence coefficient of each sample construction interaction video data is determined; The original sample influence coefficient is updated according to the adjustment range of the sample influence coefficient of each sample construction interaction video data; The step of performing correlation analysis on the initial correct decision ratio of each sample construction interaction video data in the target member network cluster and the overall construction risk decision distribution of each member network to obtain the correlation analysis result, and performing preliminary determination on the sample influence coefficient adjustment direction of each sample construction interaction video data according to the correlation analysis result to obtain the preliminary determination result includes: Determine the distribution pattern of decision-making tendency of different construction risk decision-making member networks for each sample construction interaction video data, and generate a distribution pattern feature description; According to the proportion of correct initial decisions of each sample construction interaction video data in the target member network cluster, the corresponding distribution characteristics are extracted; According to the distribution pattern feature description and the distribution feature of the initial decision correct ratio, a basic framework for association comparison is established, and according to the basic framework for association comparison, an association analysis is performed on the overall construction risk decision distribution of each member network and the initial decision correct ratio to obtain a preliminary association analysis result, and the preliminary association analysis result is used to characterize the dimension of difference or consistency between the decision within the target member network cluster and the overall decision in the sample construction interaction video data; According to the preliminary correlation analysis results, a preliminary judgment is made on the adjustment direction of the sample influence coefficient of each sample construction interaction video data. If the difference between the decision of the sample construction interaction video data within the target member network cluster and the overall decision is greater than the set difference, it is preliminarily determined that the sample influence coefficient of the sample construction interaction video data needs to be increased; otherwise, it is preliminarily determined that the sample influence coefficient of the sample construction interaction video data needs to be reduced.

3. According to the data analysis method applied to the smart construction site platform according to claim 2, It is characterized in that The step of parsing the target member network clusters meeting the target requirements in each of the member network clusters, wherein the target requirements include at least one of a relevance requirement and a weight requirement, comprises: Obtaining the global construction risk decision distribution of the global construction risk decision network for the sample learning data sequence; Based on the construction risk decision distribution of each member network and the global construction risk decision distribution, generating the correlation between each member network cluster and the global construction risk decision network; A target member network cluster whose correlation is less than a set correlation value is determined in each of the member network clusters.

4. According to claim 2, the data analysis method applied to the smart construction site platform, It is characterized in that The step of parsing the target member network clusters meeting the target requirements in each of the member network clusters, wherein the target requirements include at least one of a relevance requirement and a weight requirement, further includes: Obtaining a network participation coefficient of each of the construction risk decision-making member networks; Based on the network participation coefficient, determining the weight coefficient of each of the member network clusters for the global construction risk decision network; A target member network cluster whose weight coefficient is less than a set weight coefficient is determined in each of the member network clusters.

5. The data analysis method applied to the smart construction site platform according to any one of claims 2 to 4, It is characterized in that The step of parsing the target member network clusters meeting the target requirements in each of the member network clusters, wherein the target requirements include at least one of a relevance requirement and a weight requirement, further includes: A target member network cluster whose relevance is less than a set relevance value and whose weight coefficient is less than a set weight coefficient is determined in each of the member network clusters.

6. The data analysis method applied to the smart construction site platform according to claim 2, It is characterized in that The step of updating the influence coefficient of each sample based on the target member network clustering and the construction risk decision distribution of each member network also includes: Determining the clustering decision distribution of the target member network cluster based on the target member network cluster and the construction risk decision distribution of each member network; Obtaining the correct decision ratio of each of the sample construction interaction video data in the clustering decision distribution being correctly decided by each of the construction risk decision member networks, wherein the clustering decision distribution includes the member network construction risk decision distribution of each construction risk decision member network for the sample construction interaction video data; It is determined whether the correct decision ratio is greater than a set ratio. If so, the sample influence coefficient corresponding to each of the sample construction interaction video data is reduced; otherwise, the sample influence coefficient corresponding to each of the sample construction interaction video data is increased.

7. The data analysis method applied to the smart construction site platform according to claim 1, It is characterized in that Before the step of obtaining the construction interaction video data stream of the construction site personnel as the sample learning data sequence and the sample influence coefficient of each sample construction interaction video data in the sample learning data sequence, the step further includes: Obtaining the annotated interest preference distribution corresponding to each of the sample construction interaction video data; Based on each of the sample construction interaction video data and each of the construction risk decision member networks, generating a member network construction risk decision distribution of each of the construction risk decision member networks for the sample learning data sequence; Based on the annotated interest preference distribution and the member network construction risk decision distribution, determining the number of member networks whose decisions on each of the sample construction interaction video data are correct by each of the construction risk decision member networks; Based on the number of member networks, initial parameters and smoothing factors, generating a sample influence coefficient of each sample construction interaction video data; The calculation process of generating the sample influence coefficient of each sample construction interaction video data based on the number of member networks, initial parameters and smoothing factors includes: Define the following variables: Assume that there are n construction risk decision-making member networks, m sample construction interaction video data, the number of member networks is k, k is related to n, and is used to represent the number of member networks for which each sample construction interaction video data is correctly decided, and the initial parameter is set to a, which is a pre-set basic parameter related to the calculation of the sample influence coefficient, the smoothing factor is set to b, which is used to control data stability, and the number of member networks for which sample construction interaction video data i is correctly decided is set to k_i, i=1, 2, ..., m; For the sample construction interaction video data i, calculate the intermediate variable x_i=k_i / n+a*(n-k_i) / n; Calculate the sample influence coefficient c_i = (x_i*b) / (1+(x_i*b-1)*e^-(x_i*b)).

8. A data analysis system applied to a smart construction site platform, It is characterized in that The data analysis system applied to the smart construction site platform includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the data analysis method applied to the smart construction site platform as described in any one of claims 1 to 7 above.

Images