Deep learning based building health monitoring sensor data anomaly detection method

Abstract

The application discloses a building health monitoring sensor data anomaly detection method based on deep learning, and relates to the technical field of building health monitoring; the application effectively captures the spatial correlation features and time sequence evolution features of a building sensor from the whole period of new deployment to aging and scrapping by constructing a whole-life-cycle adaptive deep learning detection model, combining a convolutional neural network and a long short-term memory network fusion architecture, and especially by designing a dynamic adaptive weight adjustment mechanism, so that the model can capture performance changes such as sensor sensitivity attenuation and response delay in real time, and dynamically optimize the detection threshold and feature extraction weight, thereby greatly improving the accuracy of anomaly detection and the adaptability to sensor performance in different stages.

Description

Technical Field

[0001] This invention relates to the field of building health monitoring technology, specifically a method for detecting abnormal data from building health monitoring sensors based on deep learning. Background Technology

[0002] With the acceleration of urbanization and the continuous development of building technology, large and complex building structures are becoming increasingly common, and their safety and durability have become the focus of social attention. Building health monitoring, as an important means to ensure the safety of building structures, can detect and deal with potential structural anomalies in a timely manner by monitoring key parameters such as strain, acceleration, and stress of building structures in real time. This is of great significance for preventing catastrophic accidents and extending the service life of buildings.

[0003] However, traditional methods for detecting anomalies in building health monitoring sensor data have many limitations. First, traditional methods often rely on fixed detection thresholds and feature extraction methods, making it difficult to adapt to performance changes throughout the entire lifecycle of sensors, from deployment to aging and decommissioning, resulting in a significant decrease in detection accuracy over time. Second, traditional methods lack sufficient flexibility and robustness when dealing with complex and changing building environmental conditions, making them prone to false alarms and missed alarms. Furthermore, traditional methods often fail to independently and accurately identify sensor failures, instead conflating sensor failure with structural anomalies, increasing the complexity and cost of operation and maintenance management. Finally, traditional methods mostly require frequent manual calibration of model parameters, which not only increases the burden of operation and maintenance but also fails to meet the long-term needs of building health monitoring throughout its entire lifecycle.

[0004] In view of the many problems existing in traditional methods for detecting anomalies in building health monitoring sensor data, this invention proposes a deep learning-based method for detecting anomalies in building health monitoring sensor data, which is of particular importance. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based method for detecting anomalies in building health monitoring sensor data. This method can effectively capture the spatial correlation features and temporal evolution features of building sensors throughout their entire lifecycle by constructing a full lifecycle adaptive deep learning detection model and combining a fusion architecture of convolutional neural networks and long short-term memory networks. At the same time, a dynamic adaptive weight adjustment mechanism is introduced, enabling the model to capture changes in sensor performance in real time and dynamically optimize the detection threshold and feature extraction weights, thereby significantly improving the accuracy and adaptability of anomaly detection.

[0006] To solve the above-mentioned technical problems, this invention provides the following technical solution: a method for detecting anomalies in building health monitoring sensor data based on deep learning, the specific steps of which are as follows: S1. Build a building health monitoring sensor data acquisition terminal to collect core monitoring data of building structure strain, acceleration, and stress, as well as sensor operating status data, and construct a full life cycle data sample library. S2. Construct a full lifecycle adaptive deep learning detection model, which adapts to the full lifecycle performance changes of building sensors from new deployment to aging and scrapping; S3. Design a dynamic adaptive weight adjustment mechanism in the deep learning detection model to capture real-time performance change data such as sensor sensitivity decay and response delay, and dynamically optimize the model detection threshold and feature extraction weights. S4. Embed a sensor fault self-diagnosis branch in the model to independently identify sensor failure-type anomalies. S5. Input the collected data into the optimized model to complete anomaly detection, and output the classification results of real structural anomalies and sensor failure anomalies; the method does not require frequent manual calibration of model parameters, and the long-term stability of the model is improved by 30% compared with traditional methods.

[0007] Furthermore, the specific implementation steps for constructing the full life cycle data sample library in S1 are as follows: First, clarify the core monitoring points and sensor deployment locations for building health monitoring, prioritizing the deployment of corresponding sensors on the main beams, columns, shear walls, and key load-bearing components of the building. Then, start the data acquisition terminal to collect data in stages. First, collect baseline data without attenuation during the newly deployed sensor phase, with a collection period of 30 consecutive days and a daily collection frequency of no less than 288 times. Next, collect normal data during the stable operation phase, with a collection period of the entire time when the sensor is working normally and its performance is not attenuated. Then, collect drift data during the performance attenuation phase, with this phase focusing on sensor sensitivity. The initial fluctuation in temperature was used as the starting point for continuous data collection. Finally, data on failure before aging and scrapping was collected. The starting point for the sensor response delay exceeding the initial value by 50% was used as the starting point until the sensor could no longer output valid data. During the collection process, the real-time operating condition data of the building at the corresponding stage was synchronously associated. After data classification and labeling, redundancy removal, missing value removal, and data standardization preprocessing were performed. Then, the training set, validation set, and test set were divided in an 8:1:1 ratio to form a complete and usable full life cycle data sample library. This provides accurate data support for subsequent model building and training, effectively solving the problem of insufficient model adaptability caused by the incomplete coverage of scenarios in traditional sample libraries.

[0008] Furthermore, the specific implementation steps for constructing the full lifecycle adaptive deep learning detection model in S2 are as follows: First, the basic architecture of the model is determined to be a fusion architecture of convolutional neural network and long short-term memory network. First, a convolutional neural network feature extraction layer is built. First, an input layer is set to adapt to the multi-dimensional monitoring data input of the sensor. Then, convolutional layer, pooling layer, and activation layer are built in sequence. The spatial correlation features of the sensor data are extracted through the convolutional layer. The feature dimension is compressed to reduce the computational power consumption through the pooling layer. The nonlinear fitting ability of the model is enhanced through the activation layer. Then, a long short-term memory network temporal capture layer is built, with three core structures of input gate, forget gate, and output gate. This captures the temporal evolution features of the sensor's full lifecycle data and adapts to the dynamic change law of sensor performance from stability to decay to failure. Next, a model fusion layer is built to deeply fuse spatial features and temporal features. Then, a dynamic weight adjustment interface and a fault self-diagnosis branch embedding interface are reserved to ensure that the model can adapt to the access of subsequent mechanisms and branches. Finally, a dual output layer is built, corresponding to the structural anomaly detection output and the sensor failure diagnosis output, respectively. This completes the construction of the full lifecycle adaptive deep learning detection model, ensuring that the model has both feature extraction capability and full lifecycle adaptation capability.

[0009] Furthermore, the specific steps for training the full lifecycle adaptive deep learning detection model in S2 are as follows: Retrieve the full lifecycle data sample library, input the training set, validation set, and test set into the model according to their corresponding proportions, initialize the model parameters, determine the initial learning rate and number of iterations, then start iterative training of the model. First, extract spatial features of the data through a convolutional neural network, then capture temporal features of the data through a long short-term memory network, then complete feature fusion through a fusion layer, then output preliminary detection results through a dual output layer, then calculate the error between the detection results and the sample labeling results, optimize the parameters of each layer of the model through error backpropagation, and then repeat the iterative training process. After each iteration, verify the model accuracy through a validation set. Stop iterating when the model accuracy does not improve for several consecutive iterations. Then, perform performance testing on the trained model through a test set to ensure the model's adaptability and detection accuracy to data at different stages of the full lifecycle. Finally, perform lightweight optimization on the model, compressing the model size while retaining core performance to ensure that the model can adapt to the computing power requirements of building monitoring scenarios, thus completing the complete training process of the model.

[0010] Furthermore, the core of the dynamic adaptive weight adjustment mechanism in S3 is the calculation of the sensor performance degradation coefficient, the formula of which is: ,in The sensor performance degradation coefficient is the core basis for weight adjustment. , , The weighting coefficients are derived from a large amount of sample data in the full lifecycle data sample library and determined through iterative optimization using the gradient descent method, and satisfy the following conditions: , The value is determined based on the weighting of the sensor's operating time on performance degradation. The value is determined based on the weighting of the impact of sensor sensitivity changes on performance degradation. The value is determined based on the weight of the impact of sensor response delay on performance degradation. For the real-time running time of the sensor, The rated total operating time of the sensor, This represents the initial sensitivity of the sensor. For the sensor's real-time sensitivity, For the initial response delay of the sensor, To address the real-time response delay of sensors, this formula breaks through the limitations of traditional single-parameter attenuation determination, achieving multi-dimensional parameter fusion calculation. It can accurately quantify the degree of sensor performance attenuation, providing a precise basis for subsequent weight adjustment and avoiding the problem of weight adjustment failure caused by inaccurate attenuation determination in traditional methods.

[0011] Furthermore, the formula for calculating the detection threshold of the dynamic optimization model in S3 is as follows: ,in This is the threshold for real-time anomaly detection by the sensor. This is the initial anomaly detection threshold for the sensor. The sensor performance degradation coefficient, The threshold adjustment correction coefficient is derived from iterative training based on false anomaly misjudgment case data in the full lifecycle data sample library. It is used to correct the adaptation relationship between the attenuation coefficient and the threshold adjustment. This formula breaks through the limitations of traditional fixed threshold detection and realizes dynamic optimization of the threshold as the sensor performance degrades. When the sensor performance degrades to a low degree, the threshold is finely adjusted to avoid false anomaly misjudgment. When the sensor performance degrades to a high degree, the threshold is moderately increased to effectively avoid the problem of data drift caused by sensor performance degrade being misjudged as structural anomalies. The formula parameters and weights have clear sources and can greatly improve the accuracy of anomaly detection, providing core support for the dynamic adaptive weight adjustment mechanism.

[0012] Furthermore, the specific implementation steps of the sensor fault self-diagnosis branch embedded in S4 are as follows: First, build an independent architecture for the fault self-diagnosis branch to ensure that it runs in parallel with the structural anomaly detection branch without interfering with each other. Second, build a failure feature extraction module, which extracts three core failure features of the sensor. First, extract the data disconnection feature by monitoring the data transmission link in real time to determine whether there is a continuous lack of data transmission. Third, extract the data fluctuation feature by monitoring the data change amplitude to determine whether it exceeds the normal fluctuation range. Fourth, extract the response lag feature by monitoring the time difference between data acquisition and transmission to determine whether there is a response delay exceeding the standard. Fifth, build a failure judgment module, first setting independent judgment logic for the three types of failure features, and then setting comprehensive judgment rules. If any core failure feature is met and the duration reaches the set standard, it is judged as a sensor failure. Sixth, build a result output module to independently label the sensor failure judgment results. Seventh, build a linkage feedback module to synchronously feed back the failure results to the structural anomaly detection branch. The data corresponding to the failed sensor is directly marked as invalid data and does not participate in the structural anomaly judgment calculation. This completes the embedding and operation of the sensor fault self-diagnosis branch, realizing independent and accurate identification of sensor failure anomalies.

[0013] Furthermore, the specific implementation steps of model self-updating in S5 are as follows: First, a self-updating trigger module is built in the model to monitor changes in sensor performance and model detection accuracy in real time. When a sudden change in sensor performance or a continuous decline in model detection accuracy is detected, the model self-updating program is triggered. Then, a data retrieval module is built to automatically retrieve sample data from the full lifecycle data sample library that matches the current sensor performance state, without retrieving the full sample data. Next, a local fine-tuning module is built to fine-tune the local parameters of the feature extraction layer and weight allocation layer related to performance changes in the model, without retraining the entire model. Then, a fine-tuning verification module is built to retrieve the corresponding verification set data to verify the performance of the fine-tuned model, ensuring that the detection accuracy of the fine-tuned model meets the standard and is adapted to the current sensor performance state. Finally, an update and solidification module is built to solidify the fine-tuned parameters after verification, completing the model self-updating. Then, the model anomaly detection process is restarted. The entire self-updating process does not require manual intervention and is completed automatically, effectively solving the problems of high maintenance costs and untimely adaptation caused by the need for manual retraining of traditional models.

[0014] Furthermore, the specific implementation steps for the anomaly detection result output and subsequent processing in S5 are as follows: First, receive the classification results of structural anomalies and sensor failure anomalies output by the model; then, build a result grading module to classify structural anomalies into three levels according to severity: mild, moderate, and severe; and classify sensor failure anomalies into three levels according to failure degree: mild failure, moderate failure, and complete failure; next, build a result labeling module to label the anomaly results of different levels, clarifying the anomaly type, anomaly level, and corresponding monitoring point; finally, build an early warning triggering module to set differentiated settings for anomalies of different levels. The early warning mechanism triggers alerts for minor anomalies, warnings for moderate anomalies, and emergency warnings for severe anomalies. A data archiving module is then built to uniformly archive anomaly detection results, corresponding sensor data, and early warning records, forming a complete monitoring archive. An operation and maintenance linkage module is further built to simultaneously push anomaly results and early warning information to the building operation and maintenance management terminal, providing operation and maintenance personnel with accurate decision-making basis. Simultaneously, replacement reminders are pushed for completely failed sensors, and emergency response reminders are pushed for severe structural anomalies, achieving seamless integration of anomaly detection and operation and maintenance management, and improving the closed-loop management capability of building health monitoring.

[0015] Furthermore, the method also includes specific implementation steps for full-process quality control: First, in the S1 data acquisition stage, a data acquisition quality control module is built to monitor the completeness, accuracy, and continuity of data acquisition in real time. If data is missing or distorted, a supplementary acquisition or correction procedure is immediately triggered to ensure the quality of the sample library data. Then, in the S2 model building stage, an architecture control module is built to verify the rationality of the architecture at each level of the model and the adaptability of the interfaces, so as to avoid architectural defects that prevent subsequent mechanisms from being accessed. Then, in the S3 dynamic adjustment stage, an adjustment control module is built to monitor the rationality of weight and threshold adjustments in real time, verify the adaptability of the adjusted model, and avoid over-adjustment that leads to detection failure. Then, in the S4 fault diagnosis stage, a diagnosis control module is built to verify the accuracy of the failure judgment results and avoid missed or false judgments. Then, in the S5 anomaly detection stage, a detection control module is built to verify the accuracy of the classification results, compare the anomalies with historical monitoring data to verify their authenticity, and regularly verify the long-term operational stability of the model to ensure that the model maintains efficient detection capabilities throughout the entire life cycle. Full-process control requires no manual intervention, further ensuring the practicality and reliability of the method and meeting the long-term needs of building life cycle health monitoring.

[0016] Compared with existing technologies, this deep learning-based method for detecting anomalies in building health monitoring sensor data has the following advantages: I. This method constructs a full lifecycle adaptive deep learning detection model, combining a convolutional neural network and a long short-term memory network architecture. It effectively captures the spatial correlation and temporal evolution features of building sensors throughout their entire lifecycle, from deployment to aging and disposal. In particular, the design of the dynamic adaptive weight adjustment mechanism enables the model to capture sensor performance changes such as sensitivity decay and response delay in real time, and dynamically optimize the detection threshold and feature extraction weights. This significantly improves the accuracy of anomaly detection and the adaptability to sensor performance at different stages. Compared with traditional methods, this method does not require frequent manual calibration of model parameters, and the long-term stability of the model is improved by 30%.

[0017] Second, this method embeds a sensor fault self-diagnosis branch into the model. This branch operates independently of the structural anomaly detection branch. By extracting three core failure characteristics of the sensor—data disconnection, data fluctuation, and response lag—it achieves independent and accurate identification of sensor failure anomalies. Simultaneously, combined with subsequent processing steps such as anomaly detection result classification, labeling, early warning triggering, data archiving, and operation and maintenance linkage, a complete closed-loop management system for monitoring and operation and maintenance is formed. This not only provides operation and maintenance personnel with accurate decision-making basis but also pushes replacement reminders and emergency response reminders for completely failed sensors and severe structural anomalies, respectively, effectively improving the efficiency and safety of building health monitoring.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 This is a flowchart of a deep learning-based method for detecting anomalies in building health monitoring sensor data. Figure 2 A flowchart for constructing a full lifecycle data sample library for a deep learning-based method for detecting anomalies in building health monitoring sensor data; Figure 3 This is a flowchart illustrating the sensor fault self-diagnosis branch embedding and operation process of a deep learning-based building health monitoring sensor data anomaly detection method. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] Example

[0023] A data acquisition terminal for building health monitoring sensors was established, clearly defining the core monitoring points and sensor deployment locations. Strain sensors, acceleration sensors, and stress sensors were prioritized for key load-bearing components such as main beams, columns, shear walls, and joints in the office building, along with a sensor operation status monitoring module. The acquisition terminal was then activated to collect data in stages: first, baseline data without degradation was collected during the initial sensor deployment phase, with a collection period of 30 consecutive days and a daily collection frequency of 300 times; next, normal data during the stable operation phase was collected, covering the entire period when the sensor was operating normally and its performance was not degraded; then, starting with the first fluctuation in sensor sensitivity, drift data during the performance degradation phase was continuously collected; finally, starting with the sensor response delay exceeding 50% of the initial value, failure data before aging and eventual scrapping was collected until the sensor could no longer output valid data. During the data collection process, real-time operational data of the office building is synchronously linked, including daily pedestrian flow distribution, air conditioning operating power, and external environmental temperature and humidity. After data classification and labeling, redundancy removal, missing value removal, and data standardization preprocessing are performed sequentially. Then, the data is divided into training, validation, and test sets in an 8:1:1 ratio to construct a complete full-lifecycle data sample library. Figure 2 As shown.

[0024] The overall process of this method is as follows: Figure 1 As shown, the next step is to construct a full lifecycle adaptive deep learning detection model. This model employs a fusion architecture of convolutional neural networks and long short-term memory networks. First, a convolutional neural network feature extraction layer is built, with an input layer adapted to the multi-dimensional monitoring data input from the sensor. Convolutional layers, pooling layers, and activation layers are then constructed sequentially. The convolutional layers extract spatial correlation features from the sensor data, the pooling layers compress feature dimensions to reduce computational consumption, and the activation layers enhance the model's nonlinear fitting capability. Next, a long short-term memory network temporal capture layer is built, with a three-layer core structure consisting of an input gate, a forget gate, and an output gate. This captures the temporal evolution characteristics of the sensor's full lifecycle data, adapting to the dynamic changes in sensor performance from stability to degradation and then to failure.

[0025] Next, a model fusion layer is built to deeply fuse spatial and temporal features, reserving interfaces for dynamic weight adjustment and fault self-diagnosis branch embedding. Finally, a dual-output layer is built, corresponding to structural anomaly detection output and sensor failure diagnosis output, respectively, completing the model construction. During model training, a full lifecycle data sample library is retrieved, and the training set, validation set, and test set are input proportionally to initialize model parameters. After determining the initial learning rate and number of iterations, iterative training is started. Spatial features of the data are extracted through a convolutional neural network, and temporal features are captured by a long short-term memory network. The fusion layer completes feature fusion, and the dual-output layer outputs preliminary detection results. The error between the detection results and the sample labeling results is calculated, and the parameters of each layer of the model are optimized through error backpropagation. The iterative training process is repeated. After each iteration, the model accuracy is verified using a validation set. When the model accuracy does not improve for several consecutive iterations, the iteration stops. The performance of the trained model is then tested using a test set. Finally, the model is optimized for lightweighting, compressing the model size while retaining core performance.

[0026] A dynamic adaptive weight adjustment mechanism is designed in the deep learning detection model. The core of this mechanism is the calculation of the sensor performance attenuation coefficient. This calculation is used to capture real-time performance changes such as sensor sensitivity attenuation and response delay. The formula is as follows: ,in The sensor performance degradation coefficient, , , These are the weighting coefficients. For the real-time running time of the sensor, The rated total operating time of the sensor, This represents the initial sensitivity of the sensor. For the sensor's real-time sensitivity, For the initial response delay of the sensor, To account for the sensor's real-time response delay, and in conjunction with the real-time anomaly detection threshold calculation, the detection threshold and feature extraction weights of the model are dynamically optimized. The formula is as follows: ,in This is the threshold for real-time anomaly detection by the sensor. This is the initial anomaly detection threshold for the sensor. The sensor performance degradation coefficient, Adjust the correction coefficients for the threshold to ensure that the model can adapt to the performance changes of the sensor at different stages of its life cycle in real time, thereby improving the accuracy of anomaly detection.

[0027] Embed a sensor fault self-diagnosis branch in the model, such as Figure 3As shown, this branch adopts an independent architecture, running in parallel with the structural anomaly detection branch without interfering with each other. It extracts three core failure features through a failure feature extraction module: data disconnection features, data fluctuation features, and response lag features. The failure judgment module sets independent judgment logic and comprehensive judgment rules for these three types of failure features; if any core failure feature is met and its duration reaches a set standard, the sensor is judged as having failed. The result output module independently labels the sensor failure judgment results. The linkage feedback module synchronously feeds the failure results back to the structural anomaly detection branch; the data corresponding to the failed sensor is directly marked as invalid data and does not participate in the structural anomaly judgment calculation.

[0028] The optimized model is input with real-time data collected from the office building structure monitoring terminal and sensor operating status data to complete anomaly detection and output classification results of actual structural anomalies and sensor failure anomalies. During the model self-update process, the self-update trigger module monitors sensor performance changes and model detection accuracy in real time. When a sudden change in sensor performance or a continuous decline in model detection accuracy is detected, the model self-update program is triggered. The data retrieval module automatically retrieves sample data from the full lifecycle data sample library that matches the current sensor performance status, without retrieving the full sample. The local fine-tuning module performs local parameter fine-tuning on the feature extraction layer and weight allocation layer related to performance mutations in the model, without retraining the entire model. The fine-tuning verification module retrieves the corresponding verification set data to verify the performance of the fine-tuned model. The update and solidification module solidifies the fine-tuned parameters after successful verification, completes the model self-update, and restarts the anomaly detection process, all without manual intervention.

[0029] When processing anomaly detection results, the result grading module classifies actual structural anomalies into three levels: mild, moderate, and severe, and sensor failure anomalies into three levels: mild failure, moderate failure, and complete failure. The result labeling module labels anomalies of different levels, clearly indicating the anomaly type, anomaly level, and corresponding monitoring point. The early warning triggering module sets differentiated early warning mechanisms for different levels of anomalies: mild anomalies trigger suggestive warnings, moderate anomalies trigger cautionary warnings, and severe anomalies trigger emergency warnings. The data archiving module archives anomaly detection results, corresponding sensor data, and early warning records in a unified manner, forming a complete monitoring archive. The operation and maintenance linkage module pushes anomaly results and early warning information to the office building operation and maintenance management terminal simultaneously, providing operation and maintenance personnel with accurate decision-making basis. At the same time, it pushes replacement reminders for completely failed sensors and emergency response reminders for severe structural anomalies.

[0030] In terms of end-to-end quality control, the S1 data acquisition phase uses a data acquisition quality control module to monitor the completeness, accuracy, and continuity of data acquisition in real time, and immediately triggers supplementary acquisition or correction procedures if data is missing or distorted; the S2 model building phase uses an architecture control module to verify the rationality of the model's architecture at each level and the compatibility of its interfaces; the S3 dynamic adjustment phase uses an adjustment control module to monitor the rationality of weight and threshold adjustments in real time and verify the adaptability of the adjusted model; the S4 fault diagnosis phase uses a fault diagnosis control module to verify the accuracy of failure judgment results; and the S5 anomaly detection phase uses a fault detection control module to verify the accuracy of classification results, compare historical monitoring data to verify the authenticity of anomalies, and periodically verify the long-term operational stability of the model.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for detecting anomalies in building health monitoring sensor data based on deep learning, characterized in that, The specific steps of this method are as follows: S1. Build a building health monitoring sensor data acquisition terminal to collect core monitoring data of building structure strain, acceleration, and stress, as well as sensor operating status data, and construct a full life cycle data sample library. S2. Construct a full lifecycle adaptive deep learning detection model, which adapts to the full lifecycle performance changes of building sensors from new deployment to aging and scrapping; S3. Design a dynamic adaptive weight adjustment mechanism in the deep learning detection model to capture real-time performance change data such as sensor sensitivity decay and response delay, and dynamically optimize the model detection threshold and feature extraction weights. S4. Embed a sensor fault self-diagnosis branch in the model to independently identify sensor failure-type anomalies. S5. Input the collected data into the optimized model to complete the anomaly detection, and output the classification results of the actual structural anomalies and sensor failure anomalies.

2. The method for detecting abnormal data from building health monitoring sensors based on deep learning according to claim 1, characterized in that, The specific implementation steps for constructing the full life cycle data sample library in S1 are as follows: First, clarify the core monitoring points and sensor deployment locations for building health monitoring. Prioritize the deployment of corresponding sensors on the main beams, columns, shear walls, and key load-bearing components of the building. Then, start the acquisition terminal to collect data in stages. First, collect baseline data without attenuation during the newly deployed sensor stage, with a collection period of 30 consecutive days and a daily collection frequency of no less than 288 times. Next, collect normal data during the stable operation stage, with a collection period of the entire time when the sensor is working normally and its performance is not attenuated. Then, collect drift data during the performance attenuation stage, with the first fluctuation in sensor sensitivity as the starting point for continuous collection. Finally, collect failure data before aging and scrapping, with the sensor response delay exceeding the initial value by 50% as the starting point until the sensor can no longer output valid data. During the collection process, synchronously associate the real-time operating condition data of the building in the corresponding stage. After completing the data classification and labeling, perform redundancy removal, missing value removal, and data standardization preprocessing. Then, divide the training set, validation set, and test set in an 8:1:1 ratio to form a complete and usable full life cycle data sample library.

3. The method for detecting abnormal data from building health monitoring sensors based on deep learning according to claim 1, characterized in that, The specific implementation steps for constructing the full lifecycle adaptive deep learning detection model in S2 are as follows: First, the basic architecture of the model is determined to be a fusion architecture of convolutional neural network and long short-term memory network. First, a convolutional neural network feature extraction layer is built. First, an input layer is set to adapt to the multi-dimensional monitoring data input of the sensor. Then, convolutional layer, pooling layer, and activation layer are built in sequence. The spatial correlation features of the sensor data are extracted through the convolutional layer. The feature dimension is compressed to reduce the computational power consumption through the pooling layer. The nonlinear fitting ability of the model is enhanced through the activation layer. Then, a long short-term memory network temporal capture layer is built, with three core structures of input gate, forget gate, and output gate. This captures the temporal evolution features of the sensor's full lifecycle data and adapts to the dynamic change law of sensor performance from stability to decay to failure. Next, a model fusion layer is built to deeply fuse spatial features and temporal features. Then, a dynamic weight adjustment interface and a fault self-diagnosis branch embedding interface are reserved. Finally, a dual output layer is built, corresponding to the structural anomaly detection output and the sensor failure diagnosis output, respectively, to complete the construction of the full lifecycle adaptive deep learning detection model.

4. The method for detecting abnormal data from building health monitoring sensors based on deep learning according to claim 1, characterized in that, The specific steps for training the full lifecycle adaptive deep learning detection model in S2 are as follows: Retrieve the full lifecycle data sample library, input the training set, validation set, and test set into the model according to their corresponding proportions, initialize the model parameters, determine the initial learning rate and number of iterations, and then start iterative training. First, extract spatial features of the data through a convolutional neural network, then capture temporal features of the data through a long short-term memory network, then complete feature fusion through a fusion layer, and then output preliminary detection results through a dual-output layer. Next, calculate the error between the detection results and the sample labeling results, optimize the parameters of each layer of the model through error backpropagation, and then repeat the iterative training process. After each iteration, verify the model accuracy through the validation set. Stop iterating when the model accuracy does not improve for several consecutive iterations. Then, perform performance testing on the trained model through the test set. Finally, perform lightweight optimization on the model, compressing the model size while retaining core performance.

5. The method for detecting abnormal data from building health monitoring sensors based on deep learning according to claim 1, characterized in that, The core of the dynamic adaptive weight adjustment mechanism in S3 is the calculation of the sensor performance degradation coefficient, the formula of which is: ,in The sensor performance degradation coefficient, , , These are the weighting coefficients. For the real-time running time of the sensor, The rated total operating time of the sensor, This represents the initial sensitivity of the sensor. For the sensor's real-time sensitivity, For the initial response delay of the sensor, This refers to the real-time response delay of the sensor.

6. The method for detecting abnormal data from building health monitoring sensors based on deep learning according to claim 1, characterized in that, The formula for calculating the detection threshold of the dynamic optimization model in S3 is as follows: ,in This is the threshold for real-time anomaly detection by the sensor. This is the initial anomaly detection threshold for the sensor. The sensor performance degradation coefficient, Adjust the correction coefficient for the threshold.

7. The method for detecting abnormal data from building health monitoring sensors based on deep learning according to claim 1, characterized in that, The specific implementation steps of the sensor fault self-diagnosis branch embedded in S4 are as follows: First, build an independent architecture for the fault self-diagnosis branch to ensure that it runs in parallel with the structural anomaly detection branch without interfering with each other. Second, build a failure feature extraction module, which extracts three core failure features of the sensor. First, extract the data disconnection feature by monitoring the data transmission link in real time to determine whether there is a continuous lack of data transmission. Then, extract the data fluctuation feature by monitoring the data change amplitude to determine whether it exceeds the normal fluctuation range. Third, extract the response lag feature by monitoring the time difference between data acquisition and transmission to determine whether there is a response delay exceeding the standard. Fourth, build a failure judgment module, first setting independent judgment logic for the three types of failure features, and then setting comprehensive judgment rules. If any core failure feature is met and the duration reaches the set standard, the sensor is judged to be faulty. Fifth, build a result output module to independently label the sensor failure judgment results. Sixth, build a linkage feedback module to synchronously feed back the failure results to the structural anomaly detection branch. The data corresponding to the faulty sensor is directly marked as invalid data and does not participate in the structural anomaly judgment calculation. This completes the embedding and operation of the sensor fault self-diagnosis branch.

8. The method for detecting abnormal data from building health monitoring sensors based on deep learning according to claim 1, characterized in that, The specific implementation steps of model self-updating in S5 are as follows: First, a self-updating trigger module is built in the model to monitor changes in sensor performance and model detection accuracy in real time. When a sudden change in sensor performance or a continuous decline in model detection accuracy is detected, the model self-updating program is triggered. Then, a data retrieval module is built to automatically retrieve sample data from the full lifecycle data sample library that matches the current sensor performance state, without retrieving the full sample data. Next, a local fine-tuning module is built to fine-tune the local parameters of the feature extraction layer and weight allocation layer related to performance changes in the model, without retraining the entire model. Then, a fine-tuning verification module is built to retrieve the corresponding verification set data to verify the performance of the fine-tuned model. Finally, an update and solidification module is built to solidify the fine-tuned parameters after verification, completing the model self-updating. Then, the model anomaly detection process is restarted. The entire self-updating process is completed automatically without manual intervention.

9. The method for detecting abnormal data from building health monitoring sensors based on deep learning according to claim 1, characterized in that, The specific implementation steps for anomaly detection result output and subsequent processing in S5 are as follows: First, receive the classification results of actual structural anomalies and sensor failure anomalies output by the model. Then, build a result grading module to classify actual structural anomalies into three levels: mild, moderate, and severe, and sensor failure anomalies into three levels: mild failure, moderate failure, and complete failure. Next, build a result labeling module to label anomaly results of different levels, clarifying the anomaly type, anomaly level, and corresponding monitoring point. Then, build an early warning triggering module to set differentiated early warning mechanisms for different levels of anomalies: mild anomalies trigger suggestive warnings, moderate anomalies trigger cautionary warnings, and severe anomalies trigger emergency warnings. Next, build a data archiving module to uniformly archive anomaly detection results, corresponding sensor data, and early warning records to form a complete monitoring archive. Finally, build an operation and maintenance linkage module to synchronously push anomaly results and early warning information to the building operation and maintenance management terminal, providing accurate decision-making basis for operation and maintenance personnel. Simultaneously, push replacement reminders for completely failed sensors and emergency response reminders for severe structural anomalies.

10. The method for detecting anomalies in building health monitoring sensor data based on deep learning according to claim 1, characterized in that, The method also includes the following specific implementation steps for full-process quality control: First, in the S1 data acquisition stage, a data acquisition quality control module is built to monitor the completeness, accuracy, and continuity of data acquisition in real time. If data is missing or distorted, a supplementary acquisition or correction procedure is immediately triggered. Then, in the S2 model building stage, an architecture control module is built to verify the rationality of the architecture at each level of the model and the adaptability of the interfaces. Then, in the S3 dynamic adjustment stage, an adjustment control module is built to monitor the rationality of weight and threshold adjustments in real time and verify the adaptability of the adjusted model. Then, in the S4 fault diagnosis stage, a diagnosis control module is built to verify the accuracy of the failure judgment results. Then, in the S5 anomaly detection stage, a detection control module is built to verify the accuracy of the classification results, compare the anomalies with historical monitoring data to verify their authenticity, and periodically verify the long-term operational stability of the model.

Images