Fabricated lightweight wall construction system and method based on AI and BIM

Abstract

The invention relates to the technical field of intelligent construction, and discloses an assembly type light wall construction system and method based on AI and BIM, and the system comprises a data collection unit, a CAC-GAN model unit, a DPT model unit, a reinforcement learning unit, an ICAS model unit, a DBMF model unit, an SHMPS model unit, a decision support model unit, and an interaction and updating unit. The data acquisition unit integrates BIM model data through a standardized interface, unifies timestamps and then stores the data into a central database. The DPT model unit optimizes a construction plan by using Transform, the SHMPS model unit predicts faults and evaluates structural health, and the decision support model unit fuses multi-source data to recommend an optimal decision. According to the invention, the data acquisition unit is combined with the CAC-GAN model unit to generate an accurate construction model and a dynamically adjusted construction plan, and real-time consistency of the model and a site construction state is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction technology, specifically to a prefabricated lightweight wall construction system and method based on AI and BIM. Background Technology

[0002] In the field of modern construction engineering, the widespread application of Building Information Modeling (BIM) technology has significantly improved the digitalization, visualization, and informatization of project management. BIM technology provides accurate three-dimensional digital representations for the design, construction, and operation phases of building projects, including multi-dimensional information such as structure, material properties, time, and cost. This technology greatly optimizes the architectural design process, reduces the time and cost required for design modifications, and improves construction efficiency and quality. However, despite the many advantages brought by BIM technology, its application in actual construction management still faces some key challenges.

[0003] First, existing BIM-based construction management systems often fail to fully realize real-time data updates and dynamic process optimization. In traditional applications, once the BIM model is established, its synchronization with the actual construction progress is not timely enough, leading to deviations between the model and the actual site conditions. This static information processing mechanism limits the ability to respond quickly to emergencies during construction, affecting the overall project schedule and quality control. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a high-speed data isolation method and application method based on DPDK, aiming to improve the lack of timely synchronization updates of actual construction progress in the prior art, which leads to deviations between the model and the actual site conditions.

[0005] In a first aspect, the present invention provides the following technical solution: a prefabricated lightweight wall construction system based on AI and BIM, comprising: a data acquisition unit, a CAC-GAN model unit, a DPT model unit, a reinforcement learning unit, an ICAS model unit, a DBMF model unit, a SHMPS model unit, a decision support model unit, and an interaction and update unit.

[0006] The data acquisition unit integrates BIM model data, environmental sensor data, and real-time progress reports through a standardized interface, and stores them in the central database after unifying the timestamps.

[0007] The CAC-GAN model unit combines environmental factors and time characteristics to predict the construction model, and dynamically adjusts the input through a loss function.

[0008] The DPT model unit uses Transformer to analyze the output of the CAC-GAN model unit and integrates BIM static data and BIM dynamic data to optimize the construction plan through supervised learning.

[0009] The reinforcement learning unit generates decision vectors using BIM model data and real-time monitoring data as states, and dynamically optimizes the strategy based on reward functions of completion, cost, and safety indicators.

[0010] The ICAS model unit integrates the output of the preceding module and automatically executes control commands through a rule engine and stream processing technology.

[0011] The DBMF model unit compares the deviation between the construction status and the plan in real time and triggers a feedback mechanism;

[0012] The SHMPS model unit predicts faults using sensor data and anomaly detection algorithms, and assesses the structural health level by combining BIM model data.

[0013] The decision support model unit integrates multi-source data and dynamically recommends the optimal decision through a utility function.

[0014] The interaction and update unit provides real-time data feedback through a visual interface and continuously optimizes parameters through online learning.

[0015] Through the above technical solutions: the data acquisition unit integrates BIM model data, environmental sensor data, and real-time progress reports in real time via RESTAPI and OPCUA protocols, and stores them in the central database after unifying the timestamps to ensure the synchronization of the model with the actual site conditions; the CAC-GAN model unit and DPT model unit achieve accurate prediction of the construction model and dynamic optimization of the construction plan through generative adversarial networks and Transformer mechanisms; the reinforcement learning unit, ICAS model unit, and DBMF model unit work together to quickly respond to emergencies in the construction process through decision vector generation, rule engine execution, and deviation feedback mechanisms; the SHMPS model unit combines anomaly detection algorithms and BIM data to assess structural health, the decision support model unit integrates multi-source data to recommend optimal decisions, and the interaction and update unit continuously optimizes system parameters through real-time feedback and online learning, thereby improving the progress control, quality management, and responsiveness of prefabricated lightweight wall construction.

[0016] Preferably, the specific process by which the data acquisition unit integrates BIM model data, environmental sensor data, and real-time progress reports through a standardized interface and stores them in the central database after unifying the timestamps includes: using RESTAPI to process BIM static data and BIM dynamic data, using OPCUA to process environmental sensor data, cleaning, normalizing, and unifying the timestamps of the collected data, and storing it in the central database in the form of a structured table.

[0017] Preferably, the CAC-GAN model unit combines environmental factors and time features to predict the construction model, and dynamically adjusts the input through a loss function. The specific process includes: the generator inputs a noise vector, environmental factors, and time features to output a predicted construction model; the discriminator inputs real data and generated data to output a discrimination probability; and the generator and discriminator are alternately updated through a mini-maximum game framework.

[0018] Preferably, the DPT model unit uses Transformer to analyze the output of the CAC-GAN model unit and integrates BIM static data and BIM dynamic data. The specific process of optimizing the construction plan through supervised learning includes: inputting the output of the CAC-GAN model unit, BIM static data and BIM dynamic data, extracting feature sequences using the multi-head attention mechanism of Transformer, and optimizing using supervised learning based on the loss function of mean square error.

[0019] Preferably, the reinforcement learning unit generates decision vectors using BIM model data and real-time monitoring data as states, and dynamically optimizes the strategy based on reward functions of completion, cost, and safety indicators. The specific process includes: using BIM model data and real-time monitoring data as states, generating decision vectors as actions, and updating the state-action value function by combining the multi-indicator weighted sum of the reward function with the Bellman equation.

[0020] Preferably, the specific process by which the ICAS model unit integrates the output of the preceding module and automatically executes control commands through a rule engine and stream processing technology includes: inputting the outputs of the CAC-GAN model unit, the DPT model unit, and the reinforcement learning unit; using the rule engine to match conditions to generate a command stream; and triggering the execution of control commands when the deviation exceeds a threshold.

[0021] Preferably, the specific process of the DBMF model unit comparing the deviation between the construction status and the plan in real time and triggering the feedback mechanism includes: inputting the current construction status and the plan status, calculating the deviation vector, and triggering the resource reallocation feedback mechanism when the deviation vector exceeds the feedback threshold.

[0022] Preferably, the specific process by which the SHMPS model unit predicts faults using sensor data and anomaly detection algorithms, and assesses the structural health level in conjunction with BIM model data, includes: inputting sensor data sequences, calculating isolation scores using the isolation forest algorithm to predict faults, and assessing the structural health level in conjunction with BIM model data.

[0023] Preferably, the decision support model unit integrates multi-source data and dynamically recommends the optimal decision through a utility function. The specific process includes: inputting the outputs of all preceding modules, calculating the utility value through multi-factor weighted sum of the utility function, and selecting the decision option with the highest utility value as the optimal decision. The interaction and update unit provides real-time data feedback through a visual interface and continuously optimizes parameters through online learning. The specific process includes: displaying data indicators on a web-based dashboard and continuously optimizing model parameters through gradient updates.

[0024] Secondly, this invention provides the following technical solution: a prefabricated lightweight wall construction method based on AI and BIM, comprising the following steps:

[0025] S1. Data is integrated through the data acquisition unit and stored in the central database;

[0026] S2. Predict the construction model using CAC-GAN model units; optimize the construction plan using DPT model units;

[0027] S3. Dynamically optimize the strategy through reinforcement learning units; execute control commands through ICAS model units; trigger the feedback mechanism through DBMF model units;

[0028] S4. Assess the structural health level using SHMPS model units; recommend the optimal decision using decision support model units.

[0029] S5. Optimize parameters by interacting with and updating unit feedback data.

[0030] This invention provides a prefabricated lightweight wall construction system and method based on AI and BIM. It has the following beneficial effects:

[0031] 1. This invention integrates BIM model data, environmental sensor data, and real-time progress reports in real time through the data acquisition unit using RESTAPI and OPCUA protocols. After unifying the timestamps, the data is stored in a central database. Combined with the generative adversarial network prediction of CAC-GAN model units and the Transformer optimization of DPT model units, an accurate construction model and dynamically adjusted construction plan are generated, achieving real-time consistency between the model and the on-site construction status.

[0032] 2. This invention generates decision vectors using BIM model data and real-time monitoring data as states through reinforcement learning units. It dynamically optimizes strategies using reward functions based on completion, cost, and safety indicators. Combined with the rule engine and stream processing technology of the ICAS model unit, it automatically executes control commands. The invention also incorporates the deviation comparison and resource reallocation mechanism of the DBMF model unit to achieve automated and efficient optimization of resource allocation.

[0033] 3. This invention uses the SHMPS model unit to analyze sensor data sequences using the isolated forest algorithm to predict faults and combines BIM model data to assess the structural health level. The decision support model unit integrates multi-source data and recommends the optimal decision through a utility function. The interaction and update unit provides real-time data feedback and optimizes parameters online through a Web dashboard, thereby improving construction safety and the level of intelligent decision-making. Attached Figure Description

[0034] Figure 1 This is a diagram of the prefabricated lightweight wall construction system based on AI and BIM proposed in this invention.

[0035] Figure 2 This is a flowchart of the data acquisition unit of the prefabricated lightweight wall construction system based on AI and BIM proposed in this invention.

[0036] Figure 3 The flowchart of the CAC-GAN model unit of the prefabricated lightweight wall construction system based on AI and BIM proposed in this invention is shown.

[0037] Figure 4 This is a flowchart of the DPT model unit architecture of the prefabricated lightweight wall construction system based on AI and BIM proposed in this invention.

[0038] Figure 5 This is a flowchart of the reinforcement learning unit for the prefabricated lightweight wall construction system based on AI and BIM proposed in this invention.

[0039] Figure 6 This is a flowchart of the prefabricated lightweight wall construction method based on AI and BIM proposed in this invention. Detailed Implementation

[0040] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1

[0042] Reference Figures 1-5In the first embodiment of the present invention, the present invention provides an AI and BIM-based prefabricated lightweight wall construction system, including: a data acquisition unit, a CAC-GAN model unit, a DPT model unit, a reinforcement learning unit, an ICAS model unit, a DBMF model unit, a SHMPS model unit, a decision support model unit, and an interaction and update unit.

[0043] The data acquisition unit integrates BIM model data, environmental sensor data, and real-time progress reports through standardized interfaces, and stores them in the central database after unifying the timestamps;

[0044] The CAC-GAN model unit combines environmental factors and time characteristics to predict the construction model, and dynamically adjusts the input through a loss function.

[0045] The DPT model unit utilizes Transformer to analyze the output of the CAC-GAN model unit and integrates BIM static and BIM dynamic data to optimize the construction plan through supervised learning.

[0046] The reinforcement learning unit generates decision vectors using BIM model data and real-time monitoring data as states, and dynamically optimizes the strategy based on reward functions for completion, cost, and safety indicators.

[0047] The ICAS model unit integrates the output of the preceding module and automatically executes control commands through the rule engine and stream processing technology.

[0048] DBMF model units compare the deviation between construction status and plan in real time and trigger the feedback mechanism;

[0049] The SHMPS model unit predicts faults using sensor data and anomaly detection algorithms, and assesses the structural health level by combining BIM model data.

[0050] The decision support model unit integrates multi-source data and dynamically recommends the optimal decision through a utility function.

[0051] The interaction and update unit provides real-time data feedback through a visual interface and continuously optimizes parameters through online learning.

[0052] Specifically, the system collects BIM model data, environmental sensor data, and real-time progress reports through the data acquisition unit, and then integrates these data through a standardized interface and stores them in a central database with a unified timestamp. This database uses PostgreSQL structured storage and includes source fields, value fields, and timestamp fields.

[0053] CAC-GAN model units receive environmental factors and time characteristics generator Input noise vector , and Output predictive construction model Discriminator Input real data and Output the discrimination probability; the generator loss function is defined as:

[0054] (1);

[0055] in, , This is the matrix of wall construction parameters. The vector is a standard normal distribution. For the sequence of construction days, For temperature and humidity vectors, For generator loss, As the expected value, For noise vectors, For noise distribution, For the output of the discriminator, The generator output is used; the discriminator loss function is defined as:

[0056] (2);

[0057] in, For discriminator loss, For real data, The model is based on real data distribution; the Adam optimizer alternately updates G and D to generate a predictive construction model.

[0058] The DPT model unit inputs the output P of the CAC-GAN model unit, BIM static data BS, and BIM dynamic data Bd, where BS is the wall design parameter matrix and Bd is the real-time monitoring parameter matrix. The sequence is processed using the Transformer's multi-head attention mechanism, with the attention formula defined as:

[0059] (3);

[0060] in, For querying the matrix, The key matrix, For value matrices, The dimension of the key vector. The function is softmax; the loss function is defined as:

[0061] (4);

[0062] in, For DPT loss, For the sample size, To truly optimize plan tags, To predict labels; through gradient descent training, output an optimized construction plan.

[0063] The reinforcement learning unit combines BIM model data and real-time monitoring data into a state st, where st is a multi-dimensional state vector, and generates a decision vector at, where at is a resource allocation vector. The reward function r is defined as:

[0064] (5);

[0065] in, For completion level, For cost, For safety indicators, , , The weights are 1; the Q-value update formula is defined as:

[0066] (6);

[0067] in, For state-action value function, For learning rate, As a discount factor, For the next state, For the next action; adopt Explore and optimize strategies.

[0068] The ICAS model unit takes the outputs of the CAC-GAN model unit, the DPT model unit, and the reinforcement learning unit as inputs, uses a rule engine to match conditions to generate an instruction stream, and triggers the execution of control instructions when the deviation exceeds a threshold th. The threshold judgment formula is defined as follows:

[0069] (7);

[0070] in, For planned output, For the actual output, Real-time execution is achieved through Apache Flink stream processing, based on a preset threshold.

[0071] The DBMF model element takes the current construction state St and the planned state Sp as inputs, and calculates the deviation vector Dv as defined by:

[0072] (8);

[0073] Where St is the real-time state vector, Sp is the target state vector, and in > Resource reallocation is triggered at certain times, among which, The feedback threshold is used. The SHMPS model unit takes the sensor data sequence Sd as input and calculates the isolation score Is using the isolated forest algorithm, defined as: ;in, For the expected path length, This represents the average path length. The sample size is given; the structural health level L(t) is defined as: Where k is the slope parameter and m is the median, which is evaluated in conjunction with BIM model data.

[0074] The decision support model unit takes all preceding outputs as input, and its utility function U is defined as follows: Where M is the number of factors, As weight, The factor values ​​are used; the optimal decision R is defined as: Where d represents the decision option, enabling dynamic recommendation.

[0075] The interaction and update unit displays data metric di based on a web dashboard and updates it via a gradient update formula: ;in, For model parameters, For learning rate, To achieve online optimization, the loss gradient is used.

[0076] The system achieves management of prefabricated lightweight wall construction through the sequential collaboration of the above modules, forming a closed loop from data collection to decision recommendation.

[0077] The data acquisition unit integrates BIM model data, environmental sensor data, and real-time progress reports through standardized interfaces, and stores them in the central database after unifying the timestamps. The specific process includes: using RESTAPI to process BIM static data and BIM dynamic data, using OPCUA to process environmental sensor data, cleaning, normalizing, and unifying the timestamps of the collected data, and storing it in the central database in the form of structured tables.

[0078] Specifically, the data acquisition unit includes an interface adapter, a preprocessor, and a storage controller. The interface adapter connects the BIM system server and the field sensor devices. It obtains BIM static and dynamic data from the BIM system server via RESTAPI, using the HTTP GET method and carrying an authentication token. It also obtains temperature, humidity, and vibration data from the environmental sensor devices via the OPCUA protocol, using a client-server model with node IDs of ns=2 and s=Sensor.Temp.

[0079] The preprocessor receives the data stream from the interface adapter, first performing a cleaning operation, including removing missing values ​​and filtering outliers, then performing normalization to map the numerical range to the [0,1] interval using the formula (x-min) / (max-min), where x is the original value, and min and max are the minimum and maximum values ​​of the dataset. Timestamps are uniformly in UTC format, converting local time to Unix timestamps. The storage controller inserts the processed data into the central database in a structured table format, supporting transaction rollback to ensure data integrity. The data acquisition unit sequentially performs acquisition, processing, and storage through the above components, forming a data input pipeline that supports real-time access by subsequent modules of the system.

[0080] The CAC-GAN model unit combines environmental factors and time features to predict the construction model, and dynamically adjusts the input through a loss function. The specific process includes: the generator takes noise vector, environmental factors and time features as input to output the predicted construction model, the discriminator takes real data and generated data as input to output the discrimination probability, and the generator and discriminator are alternately updated through a mini-maximum game framework.

[0081] Specifically, the CAC-GAN model unit includes a generator network, a discriminator network, a training controller, and a deployment interface. The generator network adopts a multilayer perceptron structure, with the input layer dimensions being the noise vector z, environmental factors e, and temporal features. The system processes data through three fully connected layers, and the output layer generates a predicted construction model P. The discriminator network adopts a convolutional neural network structure. The input layer receives real data R and generated data P, and outputs the discrimination probability through two convolutional layers and two fully connected layers.

[0082] The training controller manages the min-maximum game process. First, it loads a real data sample set R from the central database. The generator loss function is defined by formula (1), and the discriminator loss function is defined by formula (2). In each iteration of the training loop, the discriminator is updated k times (k=1), and the calculation... The discriminator parameters are adjusted via backpropagation using the Adam optimizer, and then the generator is updated once, calculating... It then uses backpropagation to adjust the generator parameters. The deployment interface saves the trained model parameters as an HDF5 file, loads it via the cloud platform API, supports synchronization of model weights on edge computing nodes, and inputs real-time e and e during the inference process. The output P is transmitted to the DPT model unit to realize construction model prediction.

[0083] The DPT model unit utilizes Transformer to analyze the output of the CAC-GAN model unit and integrates BIM static and dynamic data. The specific process of optimizing the construction plan through supervised learning includes: inputting the output of the CAC-GAN model unit, BIM static data, and BIM dynamic data; extracting feature sequences using the multi-head attention mechanism of Transformer; and optimizing using supervised learning based on the loss function of mean squared error.

[0084] Specifically, the DPT model unit includes an input integration layer, a Transformer encoder, an output decoding layer, and an optimization trainer. The input integration layer receives the output P, ​​BIM static data BS, and BIM dynamic data Bd from the CAC-GAN model unit. P, Bs, and Bd are connected along the feature dimension through a concatenation operation to form an input sequence X, and a position code PE is added. The Transformer encoder adopts a 6-layer stacked structure, with each layer containing a multi-head self-attention sublayer and a feedforward network sublayer. The multi-head attention mechanism divides the input into h heads, and each head calculates the attention formula (3). The multi-head results are connected through a linear transformation WO. The feedforward network sublayer is a two-layer fully connected layer, and LayerNorm and residual connections are added after each sublayer.

[0085] The output decoding layer extracts feature sequences from the final hidden state of the encoder, maps them to the optimized construction plan Y through a fully connected layer, and labels them with historical optimized plan data. The optimization trainer uses supervised learning, loads the training dataset from the central database, and defines the loss function as Equation (4). Gradient descent is performed through the AdamW optimizer, and the loss is monitored and verified to prevent overfitting. The optimized construction plan is then output to the reinforcement learning unit.

[0086] The reinforcement learning unit generates decision vectors using BIM model data and real-time monitoring data as states, and dynamically optimizes the strategy based on reward functions of completion, cost, and safety indicators. The specific process includes: using BIM model data and real-time monitoring data as states, generating decision vectors as actions, and updating the state-action value function by combining the multi-indicator weighted sum of the reward function with the Bellman equation.

[0087] Specifically, the reinforcement learning unit includes a state encoder, a Q-network, a reward calculator, an experience replay buffer, and an update optimizer. The state encoder extracts BIM model data and real-time monitoring data from the central database, and concatenates the two into a state st through an embedding layer. It then uses a fully connected layer for feature extraction. The Q-network adopts a deep neural network structure. The input layer receives st, and the output layer generates the Q-value vector of the action space. The network architecture is a three-layer fully connected structure, and the output dimension is equal to the number of actions. The reward calculator calculates the reward r at each time step, defined by formula (5). The experience replay buffer stores the transition tuple (st,at,r,st'). The sampling strategy is priority experience replay, and the priority is based on the absolute value of the TD error.

[0088] The update optimizer executes a training loop, samples mini-batches from the buffer, calculates the target Q-value using the target network, and uses the Bellman equation update formula (6). The MSE loss is minimized through backpropagation by the Adam optimizer, the exploration policy adopts ε-greedy, and the output optimization decision vector is transmitted to the ICAS model unit to realize dynamic policy adjustment.

[0089] The specific process by which the ICAS model unit integrates the output of the preceding modules and automatically executes control commands through a rule engine and stream processing technology includes: inputting the outputs of the CAC-GAN model unit, the DPT model unit, and the reinforcement learning unit; using the rule engine to match conditions to generate a command stream; and triggering the execution of control commands when the deviation exceeds a threshold.

[0090] Specifically, the ICAS model unit includes an input aggregator, a rule engine processor, a stream processing engine, and an instruction executor. The input aggregator receives the output P of the CAC-GAN model unit, the output Y of the DPT model unit, and the output at of the reinforcement learning unit from the upstream module, and aggregates them asynchronously through a message queue to form a unified input data packet.

[0091] The rule engine processor uses the Drools rule engine, loads predefined rule files, defines conditional expressions and actions for each rule, parses data packets, applies rule matching, and outputs a command stream. The stream processing engine uses the Apache Flink framework, configures data stream jobs, takes the input instruction stream as the source stream, calculates the aggregation deviation through window operations, and determines the deviation using formula (7). If the conditions are met, the engine filters and forwards the execution instructions. The instruction executor receives the filtered instructions, calls the field device controller through the REST API, performs actions such as resource reallocation, records the execution log to the central database, and outputs an acknowledgment signal to the DBMF model unit to achieve control closed loop.

[0092] The specific process of DBMF model units comparing the deviation between construction status and plan in real time and triggering the feedback mechanism includes: inputting the current construction status and plan status, calculating the deviation vector, and triggering the resource reallocation feedback mechanism when the deviation vector exceeds the feedback threshold.

[0093] Specifically, the DBMF model unit includes a state input interface, a vector calculation module, a threshold judgment module, and a feedback execution module. The state input interface pulls the current construction state (St) and planned state (Sp) from the central database in real time and receives updates via a WebSocket connection. The vector calculation module performs element-wise subtraction to calculate the deviation vector. The formula is (8), and the threshold judgment module compares... With feedback threshold If it exceeds the limit, a feedback signal is generated. The feedback execution module receives The resource allocation table is updated through database transactions, and downstream modules such as ICAS model units are notified to re-execute instructions, thus realizing a closed loop for deviation correction.

[0094] The specific process by which the SHMPS model unit predicts faults using sensor data and anomaly detection algorithms, and assesses the structural health level in conjunction with BIM model data, includes: inputting sensor data sequences, using the isolation forest algorithm to calculate isolation scores to predict faults, and combining BIM model data to assess the structural health level.

[0095] Specifically, the SHMPS model unit includes a sensor data acquisition unit, an anomaly detection processor, a health assessment calculator, and an output integrator. The sensor data acquisition unit subscribes to a data sequence Sd from environmental sensor devices. Sd is a time-series vector transmitted via the MQTT protocol, and CRC32 hashing is used to ensure data integrity verification. The anomaly detection processor loads an isolated forest model, constructs a forest for Sd, randomly splits each tree until an isolation point is reached, and calculates the expected path length E(p) and the average path length c(n), where c(n) is 2H(n-1) - 2(n-1) / n, where H is the harmonic number, n is the number of samples, and the isolation score Is is defined as... .

[0096] Health assessment calculator results Using BIM model data, the structural health level L(t) is calculated and defined as: The output integrator will combine L(t) and Packaged into a JSON report, it is sent to the decision support model unit via the API endpoint to achieve health monitoring integration.

[0097] The decision support model unit integrates multi-source data and dynamically recommends the optimal decision through a utility function. The specific process includes: inputting the outputs of all preceding modules, calculating the utility value through multi-factor weighted sum of the utility function, and selecting the decision option with the highest utility value as the optimal decision. The interaction and update unit provides real-time data feedback through a visual interface and continuously optimizes parameters through online learning. The specific process includes: displaying data indicators on a web-based dashboard and continuously optimizing model parameters through gradient updates.

[0098] Specifically, the decision support model unit includes a data fusion unit, a utility calculation module, a decision selector, and an output interface. The data fusion unit collects all preceding outputs from upstream modules, including those from the CAC-GAN model unit, the DPT model unit (Y), the reinforcement learning unit (at), and the ICAS model unit. The Dv of the DBMF model unit and the L(t) of the SHMPS model unit are concatenated into a fusion vector through feature normalization and stored in a memory buffer to support batch processing. The utility calculation module calculates the utility value for a predefined decision option set D_set. For each d, the utility function U is defined as: .

[0099] The policy selector iterates through the list of U values ​​and selects the d corresponding to the largest U as the optimal decision R, defined as follows: If there are tied maximum values, the option with the lowest cost is selected first; the output interface encapsulates R as a JSON response and sends it to the interaction and update unit via HTTP POST to achieve dynamic recommendation.

[0100] The interaction and update unit includes a visualization server, a data feedback processor, an online learning engine, and a parameter storage. The visualization server uses the Flask framework to deploy web dashboards, queries data metrics di from a central database, and renders charts using the Chart.js library.

[0101] The data feedback processor parses the upstream decision R and system logs, generates feedback data packets, and broadcasts them to the client via a WebSocket channel for real-time display. The online learning engine monitors the system loss L, calculates the gradient ∇L(θ), and the update formula is defined as... It uses the SGD optimizer to process small batches of data. The parameter store serializes and saves the updated θ to the Redis cache and persistent database, supporting rollback to the previous version for continuous optimization.

[0102] Example 2:

[0103] Reference Figure 6 In a second embodiment of the present invention, the present invention provides a prefabricated lightweight wall construction method based on AI and BIM, comprising the following steps:

[0104] S1. Data is integrated through the data acquisition unit and stored in the central database;

[0105] S2. Predict the construction model using CAC-GAN model units; optimize the construction plan using DPT model units;

[0106] S3. The strategy is dynamically optimized through reinforcement learning units; control commands are executed through ICAS model units; and a feedback mechanism is triggered through DBMF model units.

[0107] S4. Assess the structural health level using SHMPS model units; recommend the optimal decision using decision support model units.

[0108] S5. Optimize parameters by interacting with and updating unit feedback data.

[0109] Specifically, firstly, the data acquisition unit obtains wall design parameters and construction progress data from the BIM system server via RESTAPI, and at the same time collects temperature, humidity and vibration data from sensor devices via OPCUA protocol. After data cleaning, normalization and timestamp unification are completed, the data is stored in a PostgreSQL database in the form of a structured table, which contains source field, value field and timestamp field.

[0110] Subsequently, the CAC-GAN model unit extracts environmental factors and time feature data from the database, combines them with random noise vectors, and generates a predictive construction model matrix through a multilayer perceptron. The discriminator network receives the predictive model and historical construction data, outputs the discrimination probability, and generates a reliable construction model prediction by alternately optimizing the generator and discriminator parameters, which is then transmitted to the DPT model unit. The DPT model unit integrates the predictive construction model, wall design parameters, and real-time monitoring data, uses the Transformer's multi-head attention mechanism to extract feature sequences, and generates an optimized construction plan through supervised learning, including resource scheduling and timetable, which is then output to the reinforcement learning unit.

[0111] The reinforcement learning unit uses BIM data and real-time monitoring data as state inputs to generate decision vectors for resource allocation. It evaluates the effectiveness of decisions based on a weighted calculation of completion, cost, and safety indicators, and iteratively optimizes the strategy through a deep Q-network before outputting the decision vectors to the ICAS model unit. The ICAS model unit receives the predicted construction model, optimized construction plan, and decision vectors. It uses a rule engine to match conditions and generate an instruction stream. It analyzes deviations in real-time using Apache Flink stream processing technology and triggers control instructions, which are sent to field equipment for execution via REST API. The DBMF model unit compares the current construction status with the planned status in real-time, calculates the deviation vector, and triggers resource reallocation when the deviation exceeds a preset threshold, updating the database resource table and recording logs.

[0112] The SHMPS model unit acquires time-series data from sensors, applies the isolated forest algorithm to detect anomalies, marks potential fault locations, and calculates the health level of the walls by combining structural parameters in the BIM data, generating a health report. The decision support model unit integrates the outputs of all modules, calculates the utility value of each decision option, selects the option with the highest utility as the recommended decision, and transmits it to the interaction and update unit via API.

[0113] The interaction and update unit displays indicators such as construction progress, cost, and health level through a web dashboard, using Chart.js to render charts. At the same time, it calculates global loss based on user feedback and system logs, updates model parameters using gradient descent, and saves parameters to Redis cache and database to ensure continuous system optimization.

[0114] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A prefabricated lightweight wall construction system based on AI and BIM, characterized in that, include: The system includes a data acquisition unit, a CAC-GAN model unit, a DPT model unit, a reinforcement learning unit, an ICAS model unit, a DBMF model unit, a SHMPS model unit, a decision support model unit, and an interaction and update unit. The data acquisition unit integrates BIM model data, environmental sensor data, and real-time progress reports through a standardized interface, and stores them in the central database after unifying the timestamps. The CAC-GAN model unit combines environmental factors and time characteristics to predict the construction model, and dynamically adjusts the input through a loss function. The DPT model unit uses Transformer to analyze the output of the CAC-GAN model unit and integrates BIM static data and BIM dynamic data to optimize the construction plan through supervised learning. The reinforcement learning unit generates decision vectors using BIM model data and real-time monitoring data as states, and dynamically optimizes the strategy based on reward functions of completion, cost, and safety indicators. The ICAS model unit integrates the output of the preceding module and automatically executes control commands through a rule engine and stream processing technology. The DBMF model unit compares the deviation between the construction status and the plan in real time and triggers a feedback mechanism; The SHMPS model unit predicts faults using sensor data and anomaly detection algorithms, and assesses the structural health level by combining BIM model data. The decision support model unit integrates multi-source data and dynamically recommends the optimal decision through a utility function. The interaction and update unit provides real-time data feedback through a visual interface and continuously optimizes parameters through online learning.

2. The prefabricated lightweight wall construction system based on AI and BIM according to claim 1, characterized in that, The specific process by which the data acquisition unit integrates BIM model data, environmental sensor data, and real-time progress reports through a standardized interface and stores them in the central database after unifying the timestamps includes: using RESTAPI to process BIM static data and BIM dynamic data, using OPCUA to process environmental sensor data, cleaning, normalizing, and unifying the timestamps of the collected data, and storing it in the central database in the form of structured tables.

3. The prefabricated lightweight wall construction system based on AI and BIM according to claim 1, characterized in that, The CAC-GAN model unit combines environmental factors and time features to predict the construction model, and dynamically adjusts the input through a loss function. The specific process includes: the generator inputs a noise vector, environmental factors, and time features to output a predicted construction model; the discriminator inputs real data and generated data to output a discrimination probability; and the generator and discriminator are alternately updated through a mini-maximum game framework.

4. The prefabricated lightweight wall construction system based on AI and BIM according to claim 1, characterized in that, The DPT model unit uses Transformer to analyze the output of the CAC-GAN model unit and integrates BIM static and BIM dynamic data. The specific process of optimizing the construction plan through supervised learning includes: inputting the output of the CAC-GAN model unit, BIM static data, and BIM dynamic data; extracting feature sequences using the multi-head attention mechanism of Transformer; and optimizing using supervised learning based on the loss function of mean square error.

5. The prefabricated lightweight wall construction system based on AI and BIM according to claim 1, characterized in that, The reinforcement learning unit generates decision vectors using BIM model data and real-time monitoring data as states, and dynamically optimizes the strategy based on reward functions of completion, cost, and safety indicators. The specific process includes: using BIM model data and real-time monitoring data as states, generating decision vectors as actions, and updating the state-action value function by combining the multi-indicator weighted sum of the reward function with the Bellman equation.

6. The prefabricated lightweight wall construction system based on AI and BIM according to claim 1, characterized in that, The specific process by which the ICAS model unit integrates the output of the preceding module and automatically executes control commands through a rule engine and stream processing technology includes: inputting the outputs of the CAC-GAN model unit, the DPT model unit, and the reinforcement learning unit; using the rule engine to match conditions to generate a command stream; and triggering the execution of control commands when the deviation exceeds a threshold.

7. The prefabricated lightweight wall construction system based on AI and BIM according to claim 1, characterized in that, The specific process of the DBMF model unit comparing the deviation between the construction status and the plan in real time and triggering the feedback mechanism includes: inputting the current construction status and the planned status, calculating the deviation vector, and triggering the resource reallocation feedback mechanism when the deviation vector exceeds the feedback threshold.

8. The prefabricated lightweight wall construction system based on AI and BIM according to claim 1, characterized in that, The specific process by which the SHMPS model unit predicts faults using sensor data and anomaly detection algorithms, and assesses the structural health level in conjunction with BIM model data, includes: inputting sensor data sequences, calculating isolation scores using the isolation forest algorithm to predict faults, and assessing the structural health level in conjunction with BIM model data.

9. The prefabricated lightweight wall construction system based on AI and BIM according to claim 1, characterized in that, The decision support model unit integrates multi-source data and dynamically recommends the optimal decision through a utility function. The specific process includes: inputting the outputs of all preceding modules, calculating the utility value through multi-factor weighted sum of the utility function, and selecting the decision option with the highest utility value as the optimal decision. The interaction and update unit provides real-time data feedback through a visual interface and continuously optimizes parameters through online learning. The specific process includes: displaying data indicators on a web-based dashboard and continuously optimizing model parameters through gradient updates.

10. A prefabricated lightweight wall construction method based on AI and BIM, characterized in that, The prefabricated lightweight wall construction system based on AI and BIM as described in claim 1 includes the following steps: S1. Data is integrated through the data acquisition unit and stored in the central database; S2. Predict the construction model using CAC-GAN model units; optimize the construction plan using DPT model units; S3. Dynamically optimize the strategy through reinforcement learning units; execute control commands through ICAS model units; trigger the feedback mechanism through DBMF model units; S4. Assess the structural health level using SHMPS model units; recommend the optimal decision using decision support model units. S5. Optimize parameters by interacting with and updating unit feedback data.

Images