HPC curtain wall installation safety protection intelligent monitoring system

Abstract

The invention discloses an HPC curtain wall installation safety protection intelligent monitoring system, and relates to the technical field of intelligent monitoring, and the system comprises a risk rehearsal module which constructs a digital neural rehearsal model, compares installation attitude data with historical accident features through a reinforcement learning algorithm, generates a safety risk thermodynamic diagram, and sends the safety risk thermodynamic diagram to an intelligent monitoring module; performing spatial superposition on the safety risk thermodynamic diagram and a preset installation scene to generate a safety risk index; the risk early warning module is used for generating an instruction type after the voice instruction signal is subjected to natural language analysis, performing contradiction value calculation on the instruction type and a safety risk index, and activating a risk protection protocol of a digital neural rehearsal model and sending out an alarm signal when the contradiction value exceeds a preset safety threshold value; and the communication module transmits the alarm signal and the safety risk index to a remote monitoring terminal. According to the invention, through dynamic visual presentation of the digital nerve rehearsal model and the security risk thermodynamic diagram, intelligent quantitative analysis and spatial visual expression of the security risk are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring, in particular to an HPC curtain wall installation safety protection intelligent monitoring system. BACKGROUND

[0002] With the continuous development of the construction industry, the popularity of high-rise and super high-rise buildings makes HPC curtain walls widely used due to their excellent durability and decorative properties. As the core link of high-altitude operation, curtain wall installation has extremely high safety risks, so the HPC curtain wall installation safety protection intelligent monitoring becomes particularly important. At present, the HPC curtain wall installation safety protection intelligent monitoring method in the construction industry mainly relies on manual inspection as the main method, supplemented by simple sensor alarms, which can to some extent find safety hazards and reduce accidents.

[0003] However, the core of the conventional safety protection and management method is still based on the timeliness and reliability of personnel observation, and there are inherent limitations in active prevention and risk warning. A prominent challenge in the industry is the difficulty in achieving no-blind-angle monitoring of personnel, equipment and key connection points on the high-altitude curtain wall installation operation surface. There are monitoring blind spots and response lags, and safety officers cannot cover all high-risk points and dynamic changes in the operating environment at all times, making it difficult to discover and warn in a timely manner, resulting in an unavoidable time difference between risk identification and emergency disposal. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an HPC curtain wall installation safety protection intelligent monitoring system to solve the problems of difficult no-blind-angle monitoring of high-altitude curtain wall operation surface and untimely risk disposal.

[0006] To solve the above technical problems, the present application provides the following technical solutions: The present application provides an HPC curtain wall installation safety protection intelligent monitoring system, which comprises, a data acquisition module, which acquires installation posture data of the HPC curtain wall and voice instruction signals of safety officers in real time, and extracts historical accident features from a historical accident database; a risk pre-performance module, which constructs a digital neural pre-performance model, compares the installation posture data with the historical accident features through a reinforcement learning algorithm, generates a safety risk heat map, and superimposes the safety risk heat map on a preset installation scene to generate a safety risk index; a risk warning module, which generates an instruction type after natural language analysis of the voice instruction signals, calculates a contradiction value of the instruction type and the safety risk index, activates a risk protection protocol of the digital neural pre-performance model and issues an alarm signal when the contradiction value exceeds a preset safety threshold; A communication module transmits the alarm signal and the safety risk index to a remote monitoring terminal.

[0007] As a preferred scheme of the HPC curtain wall installation safety protection intelligent monitoring system, the installation posture data includes displacement, inclination angle and vibration frequency parameters. The voice instruction signal of the safety officer includes operation instructions, safety warnings, state confirmations and emergency instructions.

[0008] As a preferred scheme of the HPC curtain wall installation safety protection intelligent monitoring system, the historical accident features are extracted by the following specific steps, High-frequency features are screened from the historical accident database by the association rule mining method to generate a basic accident association feature set. The basic accident association feature set is verified by a causal chain and enhanced by time sequence interaction to output historical accident features.

[0009] As a preferred scheme of the HPC curtain wall installation safety protection intelligent monitoring system, the digital neural pre-play model is constructed by the following specific steps, The convolutional neural network and the long short-term memory network are called and initialized to build posture extraction layers and accident matching layers. The posture extraction layers and the accident matching layers are executed by layer topology search and interaction mechanism optimization using the NAS algorithm to construct the digital neural pre-play model.

[0010] As a preferred scheme of the HPC curtain wall installation safety protection intelligent monitoring system, the safety risk heat map is generated by the following specific steps, The installation posture data and the historical accident features are input into the digital neural pre-play model, the posture extraction layer extracts fluctuation features by a gated recurrent network to generate a time sequence feature vector. The accident matching layer mines causal patterns by a self-attention mechanism to generate an accident association feature matrix. The time sequence feature vector and the accident association feature matrix are integrated by a feature channel to generate a safety risk heat map.

[0011] As a preferred scheme of the HPC curtain wall installation safety protection intelligent monitoring system, the preset installation scene is generated by the following specific steps, The scene structure coordinates are extracted from the equipment monitoring records and matched with the safety risk heat map to generate a space mapping table. The space mapping table is segmented by space semantics and valued by dynamic risk weights to generate a preset installation scene.

[0012] As a preferred scheme of the HPC curtain wall installation safety protection intelligent monitoring system, the generated safety risk index refers to extracting the structure type and dynamic risk weight in the preset installation scene based on the safety risk thermal map, and quantitatively fusing the structure type and the dynamic risk weight to generate the safety risk index.

[0013] As a preferred scheme of the HPC curtain wall installation safety protection intelligent monitoring system, the generated safety risk index refers to extracting the structure type and dynamic risk weight in the preset installation scene based on the safety risk thermal map, and quantitatively fusing the structure type and the dynamic risk weight to generate the safety risk index. The spectral features of the safety officer voice instruction are extracted, a text string is generated, and the text string is subjected to lexical analysis and semantic association to generate an instruction type. The instruction type and the safety risk index are subjected to operation risk matching and difference calculation to obtain a contradiction value.

[0014] As a preferred scheme of the HPC curtain wall installation safety protection intelligent monitoring system, the generated safety risk index refers to extracting the structure type and dynamic risk weight in the preset installation scene based on the safety risk thermal map, and quantitatively fusing the structure type and the dynamic risk weight to generate the safety risk index. The highest safety threshold value causing the accident in the historical accident database is extracted, and the highest safety threshold value is associated and corrected to obtain a safety threshold value. The contradiction value of the current operation area is monitored in real time, and if the contradiction value exceeds the safety threshold value, the risk protection protocol of the digital neural pre-play model is triggered, and an alarm signal is sent.

[0015] As a preferred scheme of the HPC curtain wall installation safety protection intelligent monitoring system, the generated safety risk index refers to extracting the structure type and dynamic risk weight in the preset installation scene based on the safety risk thermal map, and quantitatively fusing the structure type and the dynamic risk weight to generate the safety risk index.

[0016] The present application has the advantages that: through the digital neural pre-play model, the installation posture data and the historical accident characteristics are compared, the intelligent quantitative analysis and spatial visual expression of the safety risk are realized, the potential risk of the construction personnel is reduced by fusing the monitoring data and the historical accident data, the dynamic visual presentation of the safety risk thermal map provides the on-site and remote monitoring personnel with the intuitive, global and real-time situation awareness of the high-risk area position, the severity and the evolution trend, the safety supervision focus can be accurately pre-positioned to the prevention link, the output of the safety risk index provides a quantitative evaluation scale, the scientificity and timeliness of the safety risk early warning are enhanced, the key decision basis for risk disposal is provided, and the occurrence probability of the major installation safety accidents is greatly reduced. BRIEF DESCRIPTION OF DRAWINGS

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

[0018] Fig. 1 A schematic diagram of installing a safety protection intelligent monitoring system for an HPC curtain wall.

[0019] Fig. 2 A flowchart for generating a safety risk index.

[0020] Fig. 3 A flowchart for generating an alarm signal.

[0021] Fig. 4 A flowchart for safety data transmission. DETAILED DESCRIPTION

[0022] In order to make the above objectives, features and advantages of the present application more apparent, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0023] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0024] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0025] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides a safety protection intelligent monitoring system for an HPC curtain wall installation, comprising the following steps: A data acquisition module acquires real-time installation posture data of the HPC curtain wall and voice instruction signals of the safety officer, and extracts historical accident features from a historical accident database; The installation posture data includes displacement, inclination angle and vibration frequency parameters; It should be noted that the displacement amount is obtained by measuring the displacement change of the support by a high-precision laser ranging sensor; the inclination degree is obtained by monitoring the inclination angle offset of the panel relative to the horizontal plane by a MEMS inclination sensor; and the vibration frequency parameter is obtained by collecting vibration signals by a three-axis acceleration sensor and analyzing the vibration frequency parameter by fast Fourier transform. The voice instruction signal of the safety officer includes operation instruction, safety warning, state confirmation and emergency instruction. It should be noted that the operation instruction is naturally captured by the omnidirectional microphone array covering the operation area; the safety warning is accurately picked up by the noise reduction type directional microphone pointing to the operation area; the state confirmation focuses on the safety officer's mouth by using the near-field sound collector to ensure that the instruction is clear and there is no reverberation; and the emergency instruction is quickly responded by the high-sensitivity goose neck microphone.

[0026] Collect and preprocess multi-source accident data; It should be noted that the collection method of multi-source accident data is as follows: accident reports are obtained from the regular archiving of enterprise and institutional accident archives; safety logs are entered by organizing on-site operation records; device monitoring records are derived from historical operation data of sensors, instruments and other device supporting management settings; third-party investigation reports are obtained through public information platforms; and then the collected multi-source accident data including accident reports, safety logs, device monitoring records and third-party investigation reports are data cleaned to remove incomplete, incorrect or irrelevant data records. Format conversion is performed to ensure that all data sources comply with uniform format standards for subsequent processing. De-duplication operation is performed to eliminate duplicate data entries and ensure the uniqueness of the data set. The normalization step is used to map data values from different sources to the same scale to avoid undue influence on the results due to large differences in the order of magnitude of certain features. Finally, through the abnormal value processing, the data points deviating from the normal range are identified and corrected to ensure the accuracy and reliability of the data set.

[0027] The preprocessed multi-source accident data is classified by keyword matching to construct a historical accident database. It should be noted that, based on the basic dimensions of accident analysis, the first dimension is the type of accident, the second dimension is the cause mechanism, and the third dimension is the severity, and the accident classification framework is constructed; then the data labeling is carried out by using the "information extraction" process: for text data such as accident reports, the preliminary labeling is completed by regular expression matching and named entity recognition; for ambiguous expressions (such as "support deformation"), the type label is completed in combination with the context association (such as the subsequent "overall collapse"), and the direct cause (equipment abnormality) and the indirect cause (maintenance deficiency) are labeled. When the historical accident database is constructed, a hierarchical storage structure is used: the main table stores the unique identification, timestamp, geographical location, type label, cause classification, severity level and other accident fields; the extension table stores the scene image path, involved equipment number, casualty information and other detailed data through foreign key association, to ensure logical association and query efficiency.

[0028] The high-frequency feature set of basic accident association is generated by using the association rule mining method to screen the historical accident database; It should be noted that, based on the core analysis items of accident type, direct cause, indirect cause and severity, all accident records in the historical accident database are traversed to identify high-frequency co-occurrence feature combinations (such as "collapse type" and "operation deviation" "general accident" often occur together), and a candidate frequent item set is formed; then, pruning optimization is performed to remove low-frequency subsets: check whether the support of all k-1 dimensional subsets of each candidate frequent item set is higher than the minimum support; if the support of any subset in the candidate frequent item set is less than the minimum support, the candidate frequent item set is removed; finally, only the candidate frequent item set whose subsets all meet the support requirement is reserved, then accidental association (such as "equipment aging-object impact" is removed because there is no stable causal relationship) is excluded; the screened candidate frequent item set is converted into the basic accident association feature set.

[0029] It should also be noted that the support is an index for measuring the frequency of feature combination in the data set in association rule mining, which is the ratio of the number of accident records containing the feature combination to the total number of accident records; The minimum support is a statistical threshold value defined based on the proportion of the frequency of the feature combination in the data set to the total number of records, and the exemplary value range is 1%~5%.

[0030] The basic accident association feature set is verified for causal chain and enhanced for time sequence interaction, and the historical accident features are output; It should be noted that the time sequence and statistical dependence between the basic accident correlation feature set are verified by Granger causality test (such as whether "operation deviation" often occurs before "collapse"), and the strong causal chain with real causal relationship is reserved by backtracking the accident case to exclude false correlation; then, time sequence interaction enhancement is carried out: the dynamic change information of the basic accident correlation feature set in the time dimension is extracted, and the time sequence dependence of the basic accident correlation feature set is captured through sliding window analysis; the time interaction information between the features is quantified, such as "operation deviation" increasing by 1 time, and the subsequent "collapse risk" increasing by 20%, and the time interaction information is coded as time sequence interaction features, such as "operation deviation-vibration anomaly"; finally, the strong causal chain after causality verification and the time sequence interaction features are integrated, and the historical accident features containing causal logic and dynamic time sequence information are output.

[0031] The risk pre-performance module constructs a digital neural pre-performance model, compares the installation posture data with the historical accident features through the reinforcement learning algorithm, generates a safety risk heat map, and performs spatial superposition of the safety risk heat map and the preset installation scene to generate a safety risk index. The convolutional neural network and the long short-term memory network are called and initialized to build the posture extraction layer and the accident matching layer. It should be noted that in the TensorFlow framework, the posture extraction layer and the accident matching layer are built by sequentially calling the convolutional neural network and the long short-term memory network through the Keras API: the posture extraction layer takes the installation posture data as input, configures a 3x3x3 convolution kernel, stacks 4 layers of convolutional network, and after each layer of convolutional network, a BatchNormalization (batch normalization) layer is connected for standardization processing, then a ReLU activation function is used to enhance the non-linear expression, and finally a 256-dimensional posture feature vector is flattened and output; the accident matching layer takes the posture feature vector as input, calls the bidirectional long short-term memory network, and connects the multi-head attention mechanism at the output end of the long short-term memory network to capture the time sequence dependence of the posture feature vector before the accident occurs, and the initialization and building of the posture extraction layer and the accident matching layer are completed.

[0032] The NAS algorithm is used to perform layer topology search and interaction mechanism optimization on the posture extraction layer and the accident matching layer to construct a digital neural pre-performance model. It should be noted that the searchable layer topology space is defined: for the pose extraction layer, the search space covers the selectable types of convolution kernel size, the selectable range of convolution layer number, and the selectable types of cross-modal connection mode; for the accident matching layer, the search space covers the selectable range of long short-term memory network number, the selectable range of long short-term memory network layer number, and the selectable type of time interaction mode; then, a gradient-based NAS (Network Attached Storage) algorithm is used, taking the accident matching accuracy and the reasoning delay in the historical accident features as the joint optimization target, performing probability sampling through the relaxation continuous relaxation strategy, iteratively evaluating the performance of different topology structures, and generating a set of layer topology structure parameters: during the optimization process, the inter-layer interaction mechanism is dynamically adjusted: based on the set of layer topology structure parameters, the high-dimensional pose feature vector is mapped to a low-dimensional space compatible with the historical accident features; a gating interaction network is added before the accident matching layer, with the input being the concatenation of the low-dimensional pose feature vector and the historical accident features, and the information fusion ratio is fused through the gating function; finally, when the accident matching accuracy and the reasoning delay no longer improve, the search is terminated, the optimal set of layer topology structure parameters is determined, and the digital neural rehearsal model is output.

[0033] The installation pose data and the historical accident features are divided into sample set, training set and validation set in the ratio of 7:2:1. The missing time step installation pose data in the sample set is filled by linear interpolation, and is standardized by a batch normalization layer to form an enhanced standard sample. The Adam optimizer is used to dynamically adjust the parameters in the training set, and the cosine annealing learning rate scheduling is applied synchronously, and the early stopping method is used to monitor the validation loss, and if the validation loss does not decrease for 10 consecutive rounds, the training is terminated. The loss function uses Focal Loss to focus on improving the accuracy of accident matching. Finally, when the validation set loss fluctuates less than the convergence threshold for 5 consecutive rounds, it is determined that the digital neural rehearsal model converges, and the trained digital neural rehearsal model is output.

[0034] It should also be noted that the accident matching accuracy is obtained by matching the basic accident correlation feature set with the historical accident features; The reasoning delay is derived from the actual time measurement of the digital neural rehearsal model processing the historical accident features; The convergence threshold is defined based on the error change rate of the prediction error on the validation set, with a value range of 0.0001 to 0.01.

[0035] The installation pose data and the historical accident features are input into the digital neural rehearsal model, and the pose extraction layer extracts fluctuation features through the gating recurrent network to generate a time series feature vector; It should be noted that the attitude extraction layer time aligns and pre-processes the installation attitude data and historical accident features: the installation attitude data is divided according to a fixed time window, and attitude numerical features such as displacement deviation, inclination offset, and vibration frequency average of each time step are extracted; the operation instructions, safety warnings and other discrete features in the historical accident features corresponding to the time of the attitude window are encoded into accident discrete features at the same time; then, the attitude numerical features and accident discrete features within a fixed time step are spliced in the channel dimension to form an input matrix input to the gated recurrent network. The gated recurrent network dynamically captures the time sequence association of attitude numerical features and accident discrete features through input gate, forget gate and output gate. Finally, the hidden state of the last time step is extracted and mapped into a time sequence feature vector.

[0036] It should also be noted that the hidden state is an internal information carrier maintained by the recurrent neural network when processing sequence data, which integrates key information of historical steps and current input; The specific steps of the gated recurrent network capturing the time sequence association of attitude numerical features and accident discrete features are as follows: after the input gate receives the spliced input matrix, it generates an input gating coefficient through a sigmoid function, and filters out key information important for association analysis (such as the combination of “incline sudden increase 0.5° + emergency braking instruction” is retained, and stable low-frequency vibration is suppressed); then, the forget gate generates a forget gate coefficient through a sigmoid function based on the association mode recorded in the previous time step, such as “no abnormal vibration of the device in the past 10 seconds”, and the input matrix, to determine how much historical information to retain—if the association mode is strongly related to the current input matrix (such as similar vibration before past accidents), more association mode is retained, otherwise, redundant historical information is reduced to interfere; then, the key information filtered by the input gate and the historical information retained by the forget gate are weighted and fused to generate the current cell state, completing the fusion of historical information and current key information; finally, the output gate generates an output gating coefficient through a sigmoid function, which acts on the current cell state to generate the hidden state of the current time step, representing the time sequence association of the current attitude numerical features and accident discrete features.

[0037] The accident matching layer uses a self-attention mechanism to mine causal patterns and generate an accident association feature matrix; It should be noted that the temporal feature vector output by the attitude extraction layer is linearly transformed to generate the Q (query) matrix, K (key) matrix, and V (value) matrix: The temporal feature vector output by the attitude extraction layer is input into three independent linear layers, and Q-space mapping, K-space mapping, and V-space mapping are performed respectively to generate the corresponding Q matrix, K matrix, and V matrix. The Q matrix is ​​used to capture the "query requirement" of the current time step (such as "whether the current vibration anomaly is related to historical accidents"), the K matrix is ​​used to encode the "key information" of each time step (such as "vibration pattern characteristics before historical accidents"), and the V matrix is ​​used to store the "value content" of each time step (such as "tilt offset associated with the accident"). Next, the correlation scores between each time step are obtained by the dot product of the Q matrix and the K matrix. Then, the correlation scores are normalized by the softmax function to generate the attention weight matrix. Then, the V matrix is ​​weighted and aggregated using causal attention weights: the attention weight matrix and the V matrix are summed element-wise to obtain the context-aware features of each time step. Finally, the context-aware features are mapped to the target dimension through a non-linear activation function and a fully connected layer to generate the accident correlation feature matrix.

[0038] The time-series feature vector and the accident correlation feature matrix are integrated through feature channels to generate a safety risk heat map; It should be noted that in the feature channel, a timestamp interpolation algorithm is used to align the time-series feature vector with the accident-related feature matrix in the time dimension to eliminate differences in time steps. Then, feature fusion is performed through a multi-head attention gating layer: the aligned time-series feature vector uses a one-dimensional convolutional kernel with a sliding window covering the current and adjacent time steps to capture vibration amplitude fluctuations in continuous time steps and output local perceptual features. Simultaneously, the sliding window method is used to segment the time-series data contained in the accident-related feature matrix at fixed time intervals, forming a continuous sequence of feature segments. Subsequently, dynamic change features of displacement, tilt angle, and vibration frequency parameters are extracted within each feature segment sequence. The similarity scores of the dynamic change features at different time intervals are obtained and sorted. Based on the similarity score sorting results, the dynamic change features at different time intervals are aggregated to extract the long-range dependency of the accident-related feature matrix. Then, the local perceptual features of the time-series feature vector and the long-range dependency of the accident-related feature matrix are weighted and fused through a feature splicing channel to generate a multi-dimensional feature tensor. Finally, the multi-dimensional feature tensor is used for feature enhancement and spatial mapping using a non-linear activation function to generate a safety risk heatmap.

[0039] Extract scene structure coordinates from equipment monitoring records and match them with the spatial coordinates of the safety risk heat map to generate a spatial mapping table; It should be noted that the spatial coordinates of the scene structure are parsed from the installation log, three-dimensional scanning point cloud, sensor position record and other spatial coordinates monitored by the device, and the three-dimensional coordinates and dimensions, orientation and other geometric features of the device components such as transmission devices and sensor interfaces are extracted, while filtering noise data and calibrating the spatial coordinates to the same spatial reference system as the scene structure coordinates; then, the spatial analysis of the safety risk heat map is performed: the spatial coordinate information of the risk distribution in the safety risk heat map is extracted and mapped to the spatial reference system through affine transformation to ensure the uniformity of the spatial reference; then, the spatial coordinate matching is performed: the coordinate points of the scene structure are matched with the high-risk area coordinate points in the safety risk heat map, and the risk intensity value of the safety risk heat map is mapped to each coordinate point of the scene structure by combining the interpolation algorithm to obtain the matched structure coordinates; finally, the matched structure coordinates and corresponding risk intensity value, risk type and other information are arranged into a structured table to generate a spatial mapping table containing the fields of "structure coordinates-risk value-risk type".

[0040] It should also be noted that the risk intensity value of the safety risk heat map is directly determined by the output value of the multi-dimensional feature tensor after processing by the nonlinear activation function. The high-risk area in the safety risk heat map is derived from the continuous area in the safety risk heat map where the risk intensity value continuously sets the risk threshold, wherein the risk threshold is defined based on the accident matching accuracy of the accident-related features in the historical accident database, and the exemplary value range is 0.7-0.9.

[0041] The spatial mapping table is subjected to spatial semantic segmentation and dynamic risk weight assignment to generate a preset installation scene. It should be noted that the DBSCAN clustering algorithm is used to group all coordinate points in the spatial mapping table according to the risk type (such as collision and tipping) and spatial proximity, aggregate coordinate points of the same risk type and low spatial proximity (such as a coordinate difference of ≤0.5 meters) into independent semantic areas, and label the area type and the coordinate range contained. Subsequently, a dynamic risk weight is assigned to each independent semantic area: the dynamic risk weight is determined by three types of data, including the number of accidents in the independent semantic area in the historical accident database (the more the number of accidents in the independent semantic area, the higher the dynamic risk weight), the frequency of daily maintenance of the structure component in the installation log (the higher the maintenance frequency of the component, the higher the dynamic risk weight of the independent semantic area where the component is located), and the risk level of the independent semantic area in the spatial mapping table (the higher the risk level of the independent semantic area, the higher the dynamic risk weight). The dynamic risk weight ensures that the dynamic risk weight directly reflects the potential risk and protection priority of the independent semantic area. Finally, the position information, type and dynamic risk weight of the independent semantic area are integrated to generate a preset installation scene.

[0042] Based on the safety risk heat map, the structure type and dynamic risk weight of the corresponding position in the preset installation scene are extracted, the structure type and dynamic risk weight are quantitatively fused, and the safety risk index is generated; It should be noted that the positions of each coordinate point in the safety risk heat map are corresponded to the preset installation scene, and the structure type and dynamic risk weight of the area where each coordinate point is located are extracted. Then, the structure type is quantitatively assigned: according to the functional attributes of the components in the installation log (such as the transmission device being a "core component" and the sensor interface being an "auxiliary component"), a basic value is assigned (such as a core component basic value of 5 and an auxiliary component basic value of 3); the product of the basic value and the dynamic risk weight is taken as the initial risk value of the coordinate point; finally, the initial risk values of all coordinate points are linearly scaled and normalized to generate the safety risk index of each coordinate point, which is finally integrated into the safety risk index covering the whole scene.

[0043] The risk warning module generates an instruction type after natural language analysis of the voice instruction signal, and calculates a contradiction value for the instruction type and the safety risk index. When the contradiction value exceeds a preset safety threshold, the risk protection protocol of the digital neural rehearsal model is activated and an alarm signal is issued. The spectral features of the safety officer's voice instruction are extracted, a text string is generated, and the text string is subjected to lexical analysis and semantic association to generate an instruction type. It should be noted that the safety officer's voice instruction is subjected to frame processing, each frame is smoothed by a Hamming window, and then converted to a frequency domain representation using short-time Fourier transform, and the Mel frequency cepstral coefficient reflecting human ear perception characteristics is extracted as the spectral feature; then the spectral feature is subjected to speech recognition and converted to a corresponding text string through pattern matching; after obtaining the text string, the text string is subjected to word segmentation processing, and the continuous text in the text string is cut into independent words; then the part-of-speech tagging is performed to identify the grammatical roles of each independent word, such as verbs, nouns, adverbs, etc. Then, the semantic association analysis is performed to extract key action words (such as "hoist", "stop", "check"), safety state words (such as "loose", "in position", "abnormal"), and emergency degree adverbs (such as "immediately", "soon", "carefully"), etc. Finally, the instruction type is determined according to the extracted key words and corresponding grammatical roles: if it contains strong warning words such as "stop", "evacuate", "danger", it is determined as an emergency instruction; if it contains reminder words such as "carefully", "cautiously", "observe", it is determined as a safety warning; if it starts with a verb such as "execute", "start", "move", it is associated with specific operation actions and determined as an operation instruction; if it contains state description words such as "confirm", "already", "complete", it is determined as a state confirmation. Finally, the clear instruction type is generated according to the judgment result.

[0044] The operation risk matching and difference calculation are performed on the instruction type and the safety risk index to obtain a contradiction value; The expression for calculating the contradiction value is: ; wherein, represents the contradiction value, represents the safety risk index, represents the instruction type risk expectation function, represents the weight coefficient; It should be noted that the weight coefficient is defined based on the risk sensitivity, and is used to adjust the sensitivity of the contradiction value to the actual risk deviation. The exemplary value range is 0.5-2.

[0045] The instruction type risk expectation function is defined based on the correlation between the instruction type and the potential risk in historical experience.

[0046] The highest safety threshold value causing the accident in the historical accident database is extracted, and the highest safety threshold value is associated and corrected to obtain the safety threshold value; It should be noted that the monitoring parameters (such as equipment temperature, pressure, displacement, etc.) directly causing the accident are selected from the historical accident database, and the effective data in the monitoring parameters before and at the moment of the accident are extracted through the fields of accident report and equipment monitoring record, invalid and interference data are filtered, and an accident data set is formed. Secondly, based on the association rule mining, the key cause parameters (such as "when the panel inclination angle exceeds 30°, 90% of the accidents occur within 10 minutes") that have the greatest impact on the occurrence of the accident are identified, then the maximum measured value of the locked key cause parameters in all accidents (such as the highest record of the panel inclination angle in the historical accident is 32°) is counted, and the key cause parameters causing the accident are determined according to the severity of the accident consequences. The highest safety threshold value; subsequently, the associated correction is carried out: based on the actual situation of the current scene, such as the aging degree of the equipment, the change of the environmental conditions, the maintenance state, etc., the correction factor (such as the aging of the equipment makes the material strength decrease by 15%, so the threshold value needs to be reduced by 15%) affecting the highest safety threshold value is associated and analyzed through correlation analysis, and the highest threshold value is adjusted to obtain the safety threshold value adapted to the current scene.

[0047] The contradiction value of the current operation area is monitored in real time, and if the contradiction value exceeds the safety threshold value, the risk protection protocol of the digital neural pre-play model is triggered, and an alarm signal is issued; It should be noted that the steps of monitoring the contradiction value of the current operation area in real time and triggering the protection process are as follows: First, the safety risk index of the current operation area and the type of currently executed instructions are collected in real time, and the contradiction value of the current operation area is calculated. Then, the calculated contradiction value is compared with the preset safety threshold: if the contradiction value exceeds the safety threshold, the risk protection protocol of the digital neural pre-play model is triggered immediately; after triggering, the digital neural pre-play model quickly executes the protection action: dynamically adjusts the current operation parameters, such as reducing the equipment running speed, starts the associated protection equipment, such as automatically reinforcing the support structure, and generates specific protection scheme suggestions (such as "the current contradiction value exceeds the safety threshold, the panel adjustment needs to be stopped immediately and the support stability needs to be checked") and sends an alarm signal through the sound and light alarm device at the same time.

[0048] The communication module transmits the alarm signal and the safety risk index to the remote monitoring terminal.

[0049] The alarm signal and the safety risk index are data encapsulated, and the MQTT protocol is used for encrypted transmission to the remote monitoring terminal.

[0050] It should be noted that the transmission data is arranged: the alarm signal and the safety risk index are extracted from the current operation area and structuredly encapsulated to ensure that the data fields are consistent with the parsing rules of the remote monitoring terminal; then, the encapsulated transmission data is encrypted: the transmission data is encrypted using the TLS 1.3 protocol to generate an encrypted binary data packet, ensuring that the data content cannot be tampered with or stolen during transmission. Next, establish an MQTT connection: configure the local MQTT client parameters, maintain a long connection through the heartbeat mechanism, and set an automatic reconnection strategy to ensure the stability of the communication link. Finally, publish the encrypted data to the remote terminal: send the encrypted binary data packet to the pre-defined topic through the MQTT PUBLISH message, set the quality of service level to ensure that the message is delivered at least once. After the remote monitoring terminal receives the topic, it decrypts and parses the encrypted binary data packet, completing the remote transmission and reception of the alarm signal and the safety risk index.

[0051] It should also be noted that the steps of configuring the local MQTT client parameters and maintaining a stable communication link are as follows: By setting the connection ground and port of the MQTT agent, generating a unique client ID and configuring authentication information, the configuration of the client basic parameters is completed, and it is ensured that the client can be connected to the agent legally; then, the heartbeat mechanism is enabled: by setting the connection parameters, the local MQTT client sends a heartbeat packet to the agent periodically, and the agent returns a response after receiving it; if the local MQTT client does not receive a response within the response time of the connection parameters, it is determined that the connection is abnormal, and the disconnection process is triggered. Finally, the automatic reconnection strategy is set: the initial waiting time (such as 1 second), the maximum waiting time (such as 30 seconds) and the incremental rule (such as doubling the waiting time each time) are configured, and the subscribed topics are automatically restored after successful reconnection, ensuring that the communication link can be quickly rebuilt after interruption, maintaining the continuity of data transmission; The pre-defined topic is defined based on the security risk classification, the alarm type and the security risk index, and is used to distinguish different types of security data transmission requirements; The risk protection protocol of the digital neural pre-play model is an emergency response mechanism triggered when the contradiction value exceeds the safety threshold in real-time monitoring, which dynamically adjusts the operation parameters, starts the associated protection equipment and generates specific protection suggestions.

[0052] In summary, the present application realizes intelligent quantitative analysis and spatial visualization expression of safety risks by comparing the installation attitude data with historical accident characteristics through the digital neural pre-play model, reduces the potential risks of construction personnel by fusing monitoring data and historical accident data, and provides intuitive, global real-time situational awareness of high-risk area location, severity and evolution trend for on-site and remote monitoring personnel through dynamic visualization presentation of safety risk heat maps, so that the safety supervision focus can be accurately pre-positioned to the prevention link, the output of the safety risk index provides a quantitative evaluation scale, enhances the scientificity and timeliness of safety risk warning, provides a key decision basis for risk disposal, and greatly reduces the probability of occurrence of major installation safety accidents.

[0053] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. An intelligent monitoring system for safety protection during HPC curtain wall installation, characterized by: include, The data acquisition module collects real-time installation posture data of the HPC curtain wall and voice command signals from safety officers, and extracts historical accident features from the historical accident database. The risk simulation module constructs a digital neural simulation model, compares installation posture data with historical accident characteristics through reinforcement learning algorithms, generates a safety risk heat map, and spatially overlays the safety risk heat map with a preset installation scenario to generate a safety risk index. The risk warning module parses the voice command signal into natural language to generate command type, calculates the contradiction value between command type and security risk index, and activates the risk protection protocol of digital neural pre-simulation model and issues an alarm signal when the contradiction value exceeds the preset security threshold. The communication module transmits alarm signals and safety risk indices to the remote monitoring terminal.

2. The intelligent monitoring system for safety protection of HPC curtain wall installation as described in claim 1, characterized in that: The installation posture data includes displacement, tilt angle, and vibration frequency parameters. The safety officer's voice command signals include operation instructions, safety warnings, status confirmations, and emergency instructions.

3. The intelligent monitoring system for safety protection of HPC curtain wall installation as described in claim 2, characterized in that: The specific steps for extracting historical accident features are as follows: High-frequency features are filtered from the historical accident database using association rule mining to generate a basic accident association feature set. The causal chain verification and temporal interaction enhancement are performed on the basic accident association feature set to output historical accident features.

4. The intelligent monitoring system for safety protection of HPC curtain wall installation as described in claim 3, characterized in that: The specific steps for constructing the digital neural pre-model are as follows. Call and initialize the convolutional neural network and the long short-term memory network, and build the pose extraction layer and the accident matching layer; A digital neural pre-model is constructed by optimizing the topology search and interaction mechanism of the attitude extraction layer and the accident matching layer execution layer using the NAS algorithm.

5. The intelligent monitoring system for safety protection of HPC curtain wall installation as described in claim 4, characterized in that: The specific steps for generating the security risk heatmap are as follows: The installation attitude data and historical accident features are input into the digital neural pre-model. The attitude extraction layer extracts the fluctuation features through a gated recurrent network to generate a time-series feature vector. The accident matching layer uses a self-attention mechanism to mine causal patterns and generate an accident association feature matrix. The time-series feature vector and the accident-related feature matrix are integrated through feature channels to generate a safety risk heat map.

6. The intelligent monitoring system for safety protection of HPC curtain wall installation as described in claim 5, characterized in that: The specific steps for obtaining the preset installation scenario are as follows. Extract scene structure coordinates from equipment monitoring records and match them with the spatial coordinates of the safety risk heat map to generate a spatial mapping table; Spatial semantic segmentation and dynamic risk weight assignment are performed on the spatial mapping table to generate preset installation scenarios.

7. The intelligent monitoring system for safety protection of HPC curtain wall installation as described in claim 6, characterized in that: The generation of the safety risk index refers to the process of extracting the structural types and dynamic risk weights in a preset installation scenario based on a safety risk heat map, and then quantifying and integrating the structural types and dynamic risk weights to generate the safety risk index.

8. The intelligent monitoring system for safety protection of HPC curtain wall installation as described in claim 7, characterized in that: The specific steps for obtaining the contradictory value are as follows. Extract the spectral features of the safety officer's voice commands, generate a text string, and perform lexical analysis and semantic association on the text string to generate the command type; Operational risk matching and difference calculation are performed on instruction types and security risk indices to obtain contradictory values.

9. The intelligent monitoring system for safety protection of HPC curtain wall installation as described in claim 8, characterized in that: The specific steps for issuing the alarm signal are as follows: Extract the highest safety threshold value that led to the occurrence of an accident from the historical accident database, and perform correlation correction on the highest safety threshold value to obtain the safety threshold; The system monitors the conflict values ​​in the current operating area in real time. If the conflict value exceeds the safety threshold, it triggers the risk protection protocol of the digital neural pre-simulation model and issues an alarm signal.

10. The intelligent monitoring system for safety protection of HPC curtain wall installation as described in claim 9, characterized in that: The transmission of alarm signals and security risk indices to remote monitoring terminals refers to encapsulating the alarm signals and security risk indices into data and transmitting them to remote monitoring terminals using the MQTT protocol for encryption.

Images