Building safety monitoring system integrating natural language interaction and self-iterative optimization

Abstract

The invention provides a building safety monitoring system integrating natural language interaction and self-iterative optimization. Comprising a natural language interaction module, an authority management module, a task sequence scheduling module, an API data input module, a data preprocessing module, a picture and video real-time monitoring module, a data analysis and image display module, an energy consumption optimization module, a block chain data storage module, a GIS and BIM model interaction module, a real-time monitoring and early warning module and an emergency response module. And a model iterative optimization module. According to the method, the field vacancy in the aspects of data acquisition timeliness and convenience, multi-source data comprehensive processing, emergency real-time response, energy efficiency management and the like is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of building structure health monitoring, and in particular to a building safety monitoring system integrating natural language interaction and self-iterative optimization. Background Art

[0002] With the rapid development of contemporary society, the urbanization process is accelerating, the number and scale of buildings are continuously increasing, and the building structures are becoming increasingly complex and diverse. This phenomenon reflects the growing demand for building space by the economy and population, and reflects the innovation of building technology and design concepts. From skyscrapers to integrated commercial buildings, from multifunctional industrial facilities to residential communities, various types of buildings are constantly emerging. However, with the increase in the number of buildings, natural disasters, structural aging and human damage pose a major challenge to the safety of buildings. Global climate change has brought about the frequent occurrence of extreme weather events, including strong winds, heavy rains, earthquakes, floods, etc., which directly threaten the safety of building structures. Over time, the aging of building materials and the reduction of structural bearing capacity are particularly prominent in old buildings. In addition, human factors such as improper construction, illegal renovation and accidental collisions may also damage buildings. Therefore, ensuring the health and safety of building structures has become a problem that needs to be solved urgently.

[0003] In this context, the importance of building structure health monitoring technology has become increasingly prominent. By collecting and analyzing various performance data of buildings in real time, this technology can accurately assess the health status of building structures, provide a scientific basis for building maintenance and renovation, prevent structural collapse and ensure public safety. However, the existing building structure health monitoring system has many defects, including manual operation of data analysis, cumbersome and inefficient interaction process. The data collection coverage of traditional monitoring systems is incomplete, false alarms and misjudgments are frequent, and the system flexibility is poor. In addition, when processing video image data, traditional systems require manual processing, resulting in slow processing speed and low accuracy of processing results, affecting the response timeliness of the monitoring system. BIM and GIS technologies are inefficient in information integration, and the three-dimensional visualization effect is limited, making it difficult to intuitively present the health status of building structures. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a building safety monitoring system that integrates natural language interaction and self-iterative optimization. Based on natural language interaction, it integrates API data input and preprocessing, image recognition, energy consumption optimization, blockchain data storage, GIS and BIM model interaction and other functions to achieve real-time monitoring, abnormal warning and intelligent management of building structures.

[0005] The present invention proposes an innovative building safety monitoring system that integrates natural language interaction and self-optimization modules. The system functional modules include natural language interaction, self-iterative optimization mechanism, API data input and preprocessing, image recognition, energy efficiency optimization, blockchain data storage, and BIM / GIS model interaction. Among them, the natural language interaction and iterative optimization module is the core innovation of the system.

[0006] The natural language interaction module provides an efficient and convenient user interaction method through advanced natural language processing technology, allowing users to query the real-time status of building structures through natural language, and obtain health data analysis reports, status images and warning information. Compared with traditional systems that require manual input of instructions and operations, the new system uses the LangChain framework to achieve dynamic matching of instructions, enhancing the system's ability to quickly parse and execute instructions. After the natural language input by the user is analyzed through sentence segmentation, word segmentation and dependency relationships, the system automatically generates an API request and obtains the required data from the backend, presenting it in an easy-to-read Markdown format. This interactive mode greatly improves the efficiency of information acquisition, allowing users to more quickly and accurately grasp the health status of building structures, and optimizes the building safety monitoring process.

[0007] The self-optimization module realizes the intelligent self-iteration and optimization of the system by integrating the scoring mechanism and online learning function. The system can continuously adjust the model parameters and strategies according to user feedback and execution results, reduce false alarms and misjudgments, and improve the accuracy of data analysis and system response speed. This function ensures that the system maintains high accuracy and reliability in dynamic and complex building safety scenarios, providing strong technical support for the long-term safe operation of buildings.

[0008] In summary, the system of the present invention solves the shortcomings of current building monitoring technology in data acquisition, processing, storage, image recognition and energy consumption management by integrating advanced technologies such as natural language interaction, self-optimization and API data processing. The system significantly improves the efficiency and reliability of building structure health monitoring, provides intelligent support for building maintenance and management, and has broad application prospects and far-reaching technical value.

[0009] The specific technical solution is:

[0010] A building safety monitoring system integrating natural language interaction and self-iterative optimization, comprising:

[0011] The natural language interaction module is used to obtain building structure health monitoring data through natural language processing technology, provide data analysis and corresponding image display, and allow users to use natural language to query the real-time status of the building structure. The returned results include analysis reports of building health data, real-time status images and early warning notifications;

[0012] The authority management module is used to authenticate users, manage user rights, verify the rationality of user instructions, and apply for authorization from users with higher authority when necessary. The system can record and audit user operations to ensure security and traceability;

[0013] The task sequence scheduling module is used to mobilize different model modules to perform tasks, ensure the efficient collaboration between modules, and coordinate the execution order of each model through preset task sequence rules and priority sorting rules to ensure that the system can still effectively complete key tasks under high load conditions;

[0014] The API data input module is used to input data, data source and data description, request URL construction, and obtain data through the API interface. It further includes a data verification unit to perform integrity verification on the input data and generate a verification report on the data source and description. The data is transmitted in JSON format, and the image data is transmitted in base64 encoding format to ensure data consistency and transmission security;

[0015] The data preprocessing module is used to clean and mark the data received from the API data input module according to the data description, and supports users to establish an API database on the back end. The database is used to store information such as building name, API address, description, required parameters and examples to ensure data integrity and consistency, and provide high-quality data input for subsequent data analysis;

[0016] The image and video real-time monitoring module uses a deep learning model to monitor and identify input images or videos in real time. It is used to detect structural abnormalities such as cracks, deformation, and water leakage in buildings, and generate corresponding alarms and data reports. The detection results include location coordinates, confidence levels, and category labels.

[0017] The data analysis and image display module includes a multi-dimensional data analysis unit, which is used to perform multi-dimensional analysis on the cleaned data and generate a comprehensive health assessment report of the building structure. The report includes charts, data image displays and related analysis conclusions. Users can use visualization tools through the natural language interaction module to display the analysis results in the form of line charts, bar charts, scatter plots, etc.;

[0018] The energy consumption optimization module is used to dynamically adjust the system's energy consumption configuration according to the needs of the task sequence, dynamically mobilize the model to optimize energy consumption management, and use reinforcement learning to optimize the energy consumption configuration during the model learning process;

[0019] The blockchain data storage module is used to store key data and logs generated in the system on the blockchain to ensure the data is tamper-proof and traceable, and use hash algorithms to ensure data integrity and security;

[0020] GIS and BIM model interaction module, which is used to display and interact with buildings and their environment in 3D through GIS and BIM technology, and realize intuitive monitoring and management of the health status of building structures. Users can view real-time monitoring data, health assessment results and early warning information in the 3D model;

[0021] The real-time monitoring and early warning module is used to perform real-time monitoring based on input data, detect abnormal data and issue alarms, and send tasks to the emergency response module when necessary. It includes an abnormality detection unit and a real-time response unit, which can identify various abnormal conditions in the building structure and send real-time early warning notifications to users through communication protocols;

[0022] The emergency response module is used to actively intervene after the real-time monitoring and early warning module detects an abnormality. According to the abnormal state in the system, it automatically calls the preset emergency treatment plan and adjusts the emergency response strategy in real time;

[0023] The model iteration optimization module performs self-training and optimization based on usage feedback and task execution results, and uses online learning methods to continuously improve the accuracy and response speed of the model.

[0024] Furthermore, the natural language interaction module is displayed in Markdown format, allowing users to issue commands to control corresponding modules or devices. The system adopts the LangChain framework, which replaces the traditional method of using the cosine similarity algorithm to calculate the similarity between user commands and predefined commands. The framework dynamically matches user needs through the model, finds the appropriate API interface from the backend API database, builds the API request function and returns data.

[0025] Furthermore, the task sequence scheduling module adopts an adaptive scheduling algorithm and a reinforcement learning strategy to dynamically adjust the task execution order according to task priority, resource conditions and real-time system load to ensure that the system can still effectively complete key tasks under high load conditions.

[0026] Furthermore, the data preprocessing module includes a multi-level data cleaning and labeling unit, which can parse the descriptive information of the input data, identify data features and abnormal patterns, and automatically format the data according to the characteristics of the data source. In data missing supplementation, for numerical data, statistical methods are used to fill in missing values; for categorical data, the nearest neighbor filling method is used to fill in missing values.

[0027] Furthermore, during the model recognition process, the image and video real-time monitoring module receives input images or video frames and outputs detection results through forward propagation, the results including the detection box position, detection confidence and category probability.

[0028] Furthermore, the energy consumption optimization module includes an energy consumption monitoring unit and a dynamic adjustment unit. The energy consumption monitoring unit is used to monitor the energy consumption status of each model module in real time. The dynamic adjustment unit dynamically adjusts the energy consumption configuration of each module according to the requirements of the task sequence to achieve the purpose of optimizing system performance and reducing energy consumption.

[0029] Furthermore, the blockchain data storage module records and verifies building monitoring data through smart contracts and uses a hash algorithm to ensure the integrity and security of the data.

[0030] Furthermore, the emergency response module's emergency handling plan includes structural reinforcement suggestions, personnel evacuation instructions and alarm notifications. The system can adjust strategies based on real-time data to deal with different emergency situations.

[0031] Furthermore, the iterative optimization module continuously optimizes the performance of each module through online learning based on feedback data and task execution results. It can dynamically adjust model parameters and strategies, update the knowledge base in real time, improve the accuracy, response speed and resource utilization efficiency of the system, and ensure the efficient operation of the system in different scenarios.

[0032] The present invention obtains structural health monitoring data through natural language interaction, and provides analysis of the data and images that may need to be displayed corresponding to the data; uses API to input data, data sources and data descriptions, and pre-processes the input data; monitors in real time in the background based on the data input to the system, issues timely warnings for abnormal data, and performs emergency responses in some cases; receives natural language instructions from users, executes corresponding tasks, and determines the rationality of the instructions; monitors and warns image data input to the system in real time; and mobilizes different models through task sequences so that the models can interact quickly and directly.

[0033] The building safety monitoring system combined with the iteratively optimized large language model of the present invention has the following beneficial effects:

[0034] (1) Intelligent interaction: Through the natural language interaction module, users can interact with the system in natural language, improving the user experience.

[0035] (2) Multi-source data fusion: Supports multiple data sources such as API data input, real-time image and video monitoring, and provides comprehensive building safety monitoring.

[0036] (3) Efficient task scheduling: The task sequence scheduling module is used to ensure that each module works efficiently and collaboratively, thus improving system performance.

[0037] (4) Energy consumption optimization: Through the energy consumption optimization module, the system energy consumption configuration is dynamically adjusted to reduce energy consumption.

[0038] (5) Data security: Blockchain technology is used to achieve data immutability and traceability, thereby ensuring data security.

[0039] (6) Three-dimensional visualization: Through the interaction module of GIS and BIM model, the three-dimensional visualization of the building can be realized, which is convenient for intuitive monitoring and management.

[0040] (7) Real-time warning and emergency response: It has real-time monitoring and warning functions, and can detect abnormalities in time and take emergency measures.

[0041] (8) Iterative optimization: The system can perform self-training and optimization based on usage feedback, continuously improving the accuracy and response speed of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the overall architecture of the building safety monitoring system of the present invention;

[0043] Figure 2 This is a usage diagram of the natural language interaction module of the present invention.

[0044] Figure 3 It is a flow chart of the API data input and preprocessing module of the present invention;

[0045] Figure 4 It is a processing flow chart of the picture and video real-time monitoring module of the present invention;

[0046] Figure 5 It is a schematic diagram of the processing result of the picture and video real-time monitoring module of the present invention;

[0047] Figure 6 It is a task scheduling flow chart of the task sequence scheduling module of the present invention;

[0048] Figure 7 This is a schematic diagram of the data storage structure of the blockchain data storage module of the present invention. DETAILED DESCRIPTION

[0049] The specific technical solution of the present invention is described in conjunction with the accompanying drawings.

[0050] like Figure 1 As shown, the overall architecture of the system of the present invention is shown, including the relationship between each functional module and the data flow. The process of implementing each part is as follows:

[0051] 1. Implementation of the natural language interaction module (see Figure 2 )

[0052] 1. Module Function

[0053] Natural language processing: The natural language interaction module is designed to provide a humanized user interaction method, allowing users to use natural language to query the real-time status, historical data and analysis reports of building structures, or issue control instructions. At the same time, the LangChain framework is introduced to optimize the model calling and task execution process.

[0054] 2. Technical Implementation

[0055] The LangChain framework is used to replace the traditional cosine similarity method for direct instruction matching. LangChain supports the decomposition, model interaction and tool integration of complex instructions through its chain call mechanism, thereby achieving efficient matching between user input and predefined instructions. Its chain structure can combine multiple tools such as pre-trained language models and external API calls to achieve comprehensive semantic analysis and intent parsing.

[0056] 3. Interaction process

[0057] User input: The user enters natural language text in the interactive interface, such as "query the real-time structural status of Building A".

[0058] Semantic parsing: The LangChain framework receives user input, calls the NLP model for semantic parsing, and extracts key intents and entities.

[0059] Instruction matching: The system directly matches the user's intention through the chain structure of LangChain, matches the most suitable API through the model, calls the encoding tool to generate the request parameter encoding required by the API, executes the obtained results and passes them to the data analysis and image display module to facilitate the model to analyze the results.

[0060] Result return: Display query results in Markdown format, including analysis reports of building health data, real-time status images and early warning notifications.

[0061] 4. Example process:

[0062] Input: Query the real-time status of Bridge No. 1 on January 1, 2024.

[0063] Parsing process:

[0064] Text reception: The LangChain framework receives input and calls the pre-trained language model for pre-processing to identify key entities (such as "Bridge No. 1", "January 1, 2024") and intent ("query real-time status").

[0065] Instruction chain matching: Directly match inputs with predefined system instructions through the model's contextual understanding capabilities.

[0066] Model and tool call: LangChain's Tool Call module triggers the relevant API interface to retrieve the real-time status data of Bridge No. 1 from the database.

[0067] Result generation and return: After the system calls the API to obtain data, it generates and returns a report in Markdown format through the built-in model of the data analysis and image presentation module.

[0068] Output:

[0069] ###Query Results

[0070] **Building Name**: Bridge No. 1

[0071] **Query date**: January 1, 2024

[0072] **Real-time status**:

[0073] - **Structural status**: Stable

[0074] - **Sensor Data**:

[0075] - Temperature: 15℃

[0076] -Vibration: within normal range

[0077] **Note**: No anomalies detected.

[0078] 2. Implementation of API data input and preprocessing module (see Figure 3 )

[0079] 1.API data input module

[0080] Function description: By constructing the address of the request data and using the API RequestBuilder, the required data can be obtained from various data sources.

[0081] Data acquisition method: The system automatically generates a request URL based on the predefined API address and parameters, and sends an HTTP request to obtain data.

[0082] Data format: Data is transmitted in JSON format, and image data is transmitted in base64 encoding.

[0083] Data verification: including integrity verification, format verification and content verification, generating a verification report on the data source and description.

[0084] 2. API Database

[0085] Table 1 Database structure

[0086]

[0087] Field Description:

[0088] Building Name: Building name.

[0089] API URL: data source address.

[0090] Description: The detailed description of the API.

[0091] Required Parameters: Parameters required by the API.

[0092] Example: An example of the parameter.

[0093] Function: Improve the efficiency of model retrieval of API interfaces in the database.

[0094] 3. Data preprocessing module

[0095] Multi-level data cleaning:

[0096] Missing value handling:

[0097] Numerical data: use mean or median to fill in.

[0098] Categorical data: Use KNN to fill in the gaps, calculation formula:

[0099]

[0100] Data labeling: Automatically label data types, ranges, units, and other information based on data descriptions. Anomaly detection: Identify outliers and abnormal patterns in data and perform preliminary processing.

[0101] 3. Data acquisition process

[0102] Step 1: The system retrieves the corresponding API URL and required request parameters from the API database. Step 2: Construct request parameters params, including date, time, channel and other information.

[0103] Step 3: Use a request library (such as requests) to send an HTTP GET request to obtain data:

[0104] Step 4: Receive the JSON data returned by the server and perform data verification and preprocessing.

[0105] Step 5: Pass the preprocessed data to the subsequent analysis module.

[0106] 3. Implementation of the real-time monitoring module for images and videos (see Figure 4 , Figure 5 )

[0107] 1. Module Function

[0108] The input images and videos are monitored in real time to detect structural anomalies of buildings, such as cracks, deformation, and water leakage.

[0109] 2. Technical Implementation

[0110] Model selection: The image recognition module can use a deep learning model (such as the YOLO series model) to achieve efficient target detection.

[0111] Model training: Use public datasets and self-made datasets for training to improve the recognition accuracy of the model.

[0112] Testing process:

[0113] Input: Receives images or video frames.

[0114] Processing: Output the detection results through forward propagation.

[0115] Output: detection box position (x, y, w, h), detection confidence C and category probability p(c).

[0116] Calculation formula:

[0117] Output=YOLO_Model(Input Image)

[0118] 4. Implementation of Energy Consumption Optimization Module

[0119] 1. Module Function

[0120] According to the needs of the task sequence, the system's energy consumption configuration is dynamically adjusted to achieve optimization of energy consumption management.

[0121] 2. Technical Implementation

[0122] Energy consumption monitoring unit: monitor the energy consumption status of each model module in real time.

[0123] Dynamic adjustment unit: Dynamically adjust the operating status of the model according to task requirements and system load, such as turning on or off some functional modules.

[0124] Reinforcement Learning Strategy:

[0125] Objective function:

[0126]

[0127] Parameter Description:

[0128] J(π): Expected return of strategy π.

[0129] The mathematical expectation of all possible paths under policy π.

[0130] r(s t , a t ): In state s t Take action a t Timely rewards received.

[0131] γ: discount factor, γ∈(0,1).

[0132] α, β, γ: weight coefficients (the weight coefficients are based on the urgency of the task, and the urgency is based on the control).

[0133] *Explanation of discount factor:

[0134]

[0135] Strategy optimization: choose the action that maximizes future benefits t ,Right now:

[0136]

[0137] Parameter Description:

[0138] Q(s t , a): State-action value function.

[0139] 5. Implementation of the task sequence scheduling module (see Figure 6 )

[0140] 1. Module Function

[0141] Coordinate the execution order of each model to ensure that the system can still effectively complete key tasks under high load conditions.

[0142] 2. Technical Implementation

[0143] Adaptive Scheduling Algorithm:

[0144]

[0145] Parameter Description:

[0146] P new : The priority of the task after update.

[0147] P base : Basic task priority.

[0148] t wait : Task waiting time.

[0149] L current : Current system load.

[0150] α, β, γ: weight coefficients (the weight coefficients are based on the urgency of the task, and the urgency is based on the control).

[0151] 6. Implementation of blockchain data storage module (see Figure 7 )

[0152] 1. Module Function

[0153] Using blockchain technology, key data and logs are stored on the blockchain to ensure the data is tamper-proof and traceable.

[0154] 2. Technical Implementation

[0155] Smart contracts: define rules for data storage and access, and automate data verification and recording.

[0156] Hash algorithm: Use the SHA-256 algorithm to calculate the hash value of the data. The formula is:

[0157] Hash=SHA-256(D)

[0158] Parameter Description:

[0159] D: data block.

[0160] SHA-256: Hash encryption algorithm.

[0161] Distributed ledger: Data is stored on multiple nodes to prevent single point failure and data tampering.

[0162] 7. Implementation of GIS and BIM model interaction module

[0163] 1. Module Function

[0164] Through GIS and BIM technology, three-dimensional visualization and interaction of buildings and their environment can be achieved.

[0165] 2. Technical Implementation

[0166] Data integration: Combine GIS geographic information and BIM building information model to build a complete three-dimensional model.

[0167] Visual display: Use 3D rendering technology to present the structural details and environmental information of the building.

[0168] Data overlay: Overlay real-time monitoring data and early warning information on the 3D model to intuitively display the health status of the building.

[0169] 8. Implementation of the Model Self-Optimization Module

[0170] 1. Module Function

[0171] Through user feedback and task execution results, the model supports self-training and optimization.

[0172] 2. Technical Implementation

[0173] Self-training and optimization: Using RLHF (reinforcement learning based on human feedback), the model can self-optimize parameters through user ratings after interaction to meet user preference answers.

[0174] Implementation steps:

[0175] Step 1: Collect supervised data and train an initial supervised learning model

[0176] Collect descriptive data: During operation, the system will receive command requests from different modules, such as "output and analyze the vibration data of Bridge No. 1 for the past 3 minutes."

[0177] Generate standard examples: Professionals provide standard example outputs for these instructions to form a preliminary data set. For example, operators may construct data such as building name: {building_name} time: {time} feature data: data name: {data_name} data value: {data_value} data analysis: {data_analysis}, etc.

[0178] Training supervised learning model: Use these descriptive data to train an initial supervised learning model so that the system can give reasonable and expected monitoring results. This stage helps the system master basic task execution capabilities.

[0179] Step 2: Collect feedback data and train the reward model

[0180] Collect comparative data: In actual system operation, the model will generate multiple potential outputs for different instructions. For example, when the system performs real-time analysis on the vibration data of Bridge No. 1, it may output different risk assessment or detection results.

[0181] Compare and score: Professionals can score the accuracy, timeliness, and clarity of these outputs. For example, output 1 scores high because it locates the risk point; output 2 scores low because its description is unclear.

[0182] Training the reward model: The expert’s rating data is used to train the reward model so that it learns to predict the “preference score” of different outputs, that is, whether the output meets expectations and human preferences.

[0183] Step 3: PPO reinforcement learning optimization model

[0184] Initialize the PPO model: Based on the supervised learning model, use the PPO algorithm to initialize a reinforcement learning model.

[0185] Generate answers and feedback scores: The PPO model generates data analysis results according to instructions. The reward model provides a score for each result, and results with lower scores represent room for improvement.

[0186] PPO optimization: The PPO algorithm uses the feedback provided by the reward model to update the model strategy and optimize the results generated by the system to make them more in line with human preferences (accurate, clear, and fast response). Through continuous sampling and optimization, the system gradually improves its performance in building safety monitoring tasks.

[0187] Reward model training steps:

[0188] Data Collection: Obtaining Comparative Data During system operation, the model generates multiple different outputs for a specific instruction (such as "Analyze the vibration data of Bridge 1"). Each output may include a detailed description of the risk detection, risk level, location, recommended actions, etc.

[0189] Professionals will score or rank these outputs, evaluating them based on accuracy, timeliness, clarity, etc. For example, if there are three different outputs, the experts will rank them from high to low based on preference, or directly score each output.

[0190] Example data:

[0191] Command x input: Analyze the vibration data of Bridge No. 1.

[0192] Output result set {y1, y2, y3}: three different outputs generated by the model for the same instruction.

[0193] Human ratings {r1, r2, r3}: The rating or preference order of each output, where the higher the rating, the higher the preference for that output.

[0194] The training reward model is implemented using a neural network, which accepts input (x, y) and outputs a score prediction value. in are the parameters of reward modeling (weights and biases in the network).

[0195] For the text input instruction x, a pre-trained language model is used to encode it into h x , extract the vector representing the instruction content. For the text output instruction y, the pre-trained language model is used to encode it into h y , including textual features such as the accuracy of the analysis report (which can be scored by professionals), the clarity of wording (which can be scored by professionals or another model) and numerical features such as statistical information of vibration data.

[0196] Concatenate the input and output vectors to generate the complete feature vector h[h x ,h y ].

[0197]

[0198] The eigenvector h[h x ,h y ] is input to the fully connected layer, the activation function is used to extract features, and the corresponding preference score is output.

[0199] In this process, different evaluation indicators are used to evaluate the effect of the reward model and the deviation between the prediction and the expectation:

[0200] For continuous scores (0-10 points), the mean square error loss function is used:

[0201]

[0202] in:

[0203] The expected value is the average of all samples. Each sample consists of instruction x, model output y, and human rating r, and the expected value is obtained by r.

[0204] R φ (x, y): represents the reward model score corresponding to input x and output y.

[0205] For sorting, the sorting loss function is used:

[0206]

[0207] in:

[0208] y i and j They represent different outputs of the model for the same problem.

[0209] R φ (x,y i ) and R φ (x,y j ): indicates that for input x and output y i (y j ) corresponding reward model score

[0210] 1-(R φ (x,y j )-R φ (x,y j )): The goal of scoring difference, if R φ (x,y i )-R φ (x,yj ) is less than 1, there is a loss; if the difference is greater than 1, the loss is 0.

[0211] max(0,·): only when R φ (x,y i )≤R φ (x,y j )+1, the loss is greater than 0. That is, the model's score for the more preferred output should be at least 1 point higher than the score for the suboptimal output, to ensure that the output that is more in line with human preferences gets a higher score.

[0212] The present invention provides a building safety monitoring system that integrates multiple advanced technologies and has the characteristics of intelligent interaction, efficient data processing, real-time warning, energy consumption optimization, data security and three-dimensional visualization. The system can comprehensively improve the efficiency and reliability of building safety monitoring and has broad application prospects.

Claims

1. A building safety monitoring system integrating natural language interaction and self-iterative optimization, characterized in that: include: The natural language interaction module is used to obtain building structure health monitoring data through natural language processing technology, provide data analysis and corresponding image display, and allow users to use natural language to query the real-time status of the building structure. The returned results include analysis reports of building health data, real-time status images and early warning notifications; The authority management module is used to authenticate user identities, manage user permissions, verify the rationality of user instructions, and apply for authorization from users with higher permissions when necessary. The system can record and audit user operations. The task sequence scheduling module is used to mobilize different model modules to perform tasks, ensure efficient collaboration among modules, and coordinate the execution order of each model through preset task sequence rules and priority sorting rules; API data input module, used to input data, data source and data description, request URL construction, and obtain data through API interface, including data verification unit, which verifies the integrity of input data and generates verification report of data source and description. Data is transmitted in JSON format, and image data is transmitted in base64 encoding format; The data preprocessing module is used to clean and mark the data received from the API data input module according to the data description, and supports users to establish an API database on the back end; The image and video real-time monitoring module uses a deep learning model to monitor and identify input images or videos in real time to detect abnormal problems in buildings and generate corresponding alarms and data reports. The detection results include location coordinates, confidence levels, and category labels. The data analysis and image display module includes a multi-dimensional data analysis unit, which is used to perform multi-dimensional analysis on the cleaned data and generate a comprehensive health assessment report of the building structure. Users can use visualization tools through the natural language interaction module to display the analysis results; The energy consumption optimization module is used to dynamically adjust the system's energy consumption configuration according to the needs of the task sequence, dynamically mobilize the model to optimize energy consumption management, and use reinforcement learning to optimize the energy consumption configuration during the model learning process; Blockchain data storage module, used to store key data and logs generated in the system on the blockchain; GIS and BIM model interaction module, which is used to display and interact with buildings and their environment in three dimensions through GIS and BIM technology, and realize intuitive monitoring and management of the health status of building structures. Users can view real-time monitoring data, health assessment results and early warning information in the three-dimensional model; The real-time monitoring and early warning module is used to perform real-time monitoring based on input data, detect abnormal data and issue alarms, and send tasks to the emergency response module when necessary. It includes an abnormality detection unit and a real-time response unit, which can identify various abnormal conditions in the building structure and send real-time early warning notifications to users through communication protocols; The emergency response module is used to actively intervene after the real-time monitoring and early warning module detects an abnormality. According to the abnormal state in the system, it automatically calls the preset emergency treatment plan and adjusts the emergency response strategy in real time; The model iteration optimization module performs self-training and optimization based on usage feedback and task execution results, and adopts online learning methods to continuously improve the accuracy and response speed of the model.

2. A building safety monitoring system integrating natural language interaction and self-iterative optimization according to claim 1, characterized in that: The natural language interaction module is displayed in Markdown format, allowing users to issue instructions to control corresponding modules or devices; the system adopts the LangChain architecture to replace the traditional method of directly using the cosine similarity algorithm to match user instructions with predefined instructions; the model in the system dynamically matches the appropriate API interface in the backend API database according to the user instructions, constructs the API request function and returns data.

3. The building safety monitoring system integrating natural language interaction and self-iterative optimization according to claim 1 is characterized in that: The task sequence scheduling module adopts an adaptive scheduling algorithm and a reinforcement learning strategy to dynamically adjust the task execution order according to task priority, resource conditions and real-time system load to ensure that the system can still effectively complete key tasks under high load conditions.

4. A building safety monitoring system integrating natural language interaction and self-iterative optimization according to claim 1, characterized in that: The data preprocessing module includes a multi-level data cleaning and labeling unit, which can parse the descriptive information of the input data, identify data features and abnormal patterns, and automatically format the data according to the characteristics of the data source. In data missing supplementation, for numerical data, statistical methods are used to fill in missing values; for categorical data, the nearest neighbor filling method is used to fill in missing values.

5. The building safety monitoring system integrating natural language interaction and self-iterative optimization according to claim 1 is characterized in that: During the model recognition process, the picture and video real-time monitoring module receives input images or video frames and outputs detection results through forward propagation. The results include detection box position, detection confidence and category probability.

6. The building safety monitoring system integrating natural language interaction and self-iterative optimization according to claim 1 is characterized in that: The energy consumption optimization module includes an energy consumption monitoring unit and a dynamic adjustment unit. The energy consumption monitoring unit is used to monitor the energy consumption status of each model module in real time. The dynamic adjustment unit dynamically adjusts the energy consumption configuration of each module according to the requirements of the task sequence to achieve the purpose of optimizing system performance and reducing energy consumption.

7. The building safety monitoring system integrating natural language interaction and self-iterative optimization according to claim 1 is characterized in that: The blockchain data storage module records and verifies building monitoring data through smart contracts and uses hash algorithms to ensure the integrity and security of the data.

8. The building safety monitoring system integrating natural language interaction and self-iterative optimization according to claim 1 is characterized in that: The emergency response module's emergency handling plan includes structural reinforcement suggestions, personnel evacuation instructions and alarm notifications. The system can adjust strategies based on real-time data to respond to different emergency situations.

9. The building safety monitoring system integrating natural language interaction and self-iterative optimization according to claim 1, characterized in that: The iterative optimization module continuously optimizes according to user rating feedback and task execution results, dynamically adjusts model parameters and strategies through online learning algorithms, updates the knowledge base in real time, and continuously improves the accuracy and response speed of the system.

Images