A personnel behavior recognition method based on optical fiber sensing data driving

Abstract

The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on optical fiber sensing data driving and relates to the technical field of safety monitoring. The application discloses a personnel behavior recognition method based on opticalfiber sensing data driving and relates to the technical field of safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology, and in particular to a method for human behavior recognition based on fiber optic sensor data. Background Technology

[0002] With the increasing demands for the protection and utilization of traditional courtyards, age-friendly safety management, and smart security, multi-source fiber optic sensing and monitoring technologies for courtyard buildings such as siheyuan (traditional courtyard houses) and human behavior recognition technologies have received widespread attention. Existing ancient building safety monitoring solutions mainly rely on single FBG (fiber grating) fixed-point monitoring or single distributed fiber optic DTS / DAS environmental monitoring to sense structural, temperature, or vibration conditions. However, these solutions generally suffer from the following problems in practical applications: While single FBG fiber optic grating solutions can monitor strain and displacement changes at critical locations such as beams, columns, and walls with high precision, the monitoring points are scattered, making it difficult to cover continuous spaces such as courtyard boundaries, passageways, and kitchens, easily creating monitoring blind spots. Furthermore, they cannot determine the status of human behavior such as walking, falls, gatherings, or abnormal intrusions based on vibration behavior. While single distributed fiber optic DTS / DAS solutions can achieve continuous sensing of temperature or vibration, their ability to quantify minor structural deformations is insufficient, and they are prone to false alarms in complex courtyard environments such as construction disturbances, personnel passage, and wind and rain vibrations. At the same time, existing solutions are mostly geared towards monitoring bridges, tunnels, or large ancient buildings, lacking integrated designs for scenarios where the "non-intrusion and preservation of style" of cultural heritage courtyards coexist with the "gas, fall, and security" requirements of modern residences. This makes it difficult to achieve integrated identification and graded early warning of structural safety, environmental safety, personnel safety, and security.

[0003] Therefore, how to provide a method for human behavior recognition based on fiber optic sensing data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a personnel behavior recognition method driven by fiber optic sensor data. This invention employs a combination of DTS / DAS distributed fiber optic sensing and FBG fiber grating structural response monitoring. Through courtyard space mapping, functional zoning weight allocation, and spatiotemporal alignment processing, multi-source fiber optic sensor fusion data covering temperature, vibration, strain, and displacement information is constructed. Based on this, DAS vibration features are extracted using STFT and WT, and DTS temperature change features and FBG strain and displacement change features are fused to form a multi-parameter fusion feature vector. Simultaneously, a constraint mechanism based on DAS vibration propagation characteristics and FBG structural response characteristics is introduced to optimize the feature representation. Furthermore, a modified Res-GRU model is used for temporal behavior recognition, enabling the identification and graded early warning of behaviors such as normal walking, stillness, falls, gathering, and abnormal intrusion. This method offers advantages such as multi-risk fusion perception, low false alarm rate, strong real-time performance, and applicability to courtyard security monitoring scenarios.

[0005] A method for human behavior recognition based on fiber optic sensor data according to an embodiment of the present invention includes the following steps: S1. Synchronously acquire distributed temperature sensor (DTS) data, distributed acoustic sensor (DAS) data, and fiber optic grating (FBG) structure response data, and perform time synchronization to obtain multi-source fiber optic sensing raw data. S2. Based on the DTS / DAS composite optical cable deployment path and FBG sensor deployment location, the original multi-source fiber optic sensing data is spatially mapped, and after differentiating the sensing data of different areas by combining the courtyard functional zoning, the data is spatiotemporally aligned to generate multi-source fiber optic sensing fusion data. S3. Using the multi-source fiber optic sensing fusion data as the processing object, short-time Fourier transform (STFT) and wavelet transform (WT) are used to extract the vibration features of DAS, and combined with the temperature change features of DTS and the strain and displacement change features of FBG, a multi-parameter fusion feature vector is formed. S4. Input the multi-parameter fused feature vector into the feature representation model, and construct a consistency constraint mechanism based on the DAS vibration propagation characteristics and FBG structural response characteristics to constrain and optimize the feature representation, thereby obtaining the semantic representation of personnel behavior. S5. Construct a temporal behavior recognition model based on the improved Res-GRU, perform temporal modeling and analysis on the semantic representation of the personnel behavior, and output the personnel behavior category and confidence parameters. S6. Based on the personnel behavior category and confidence parameters, combined with DTS temperature anomaly information and FBG structural anomaly information, perform multi-source fusion judgment, and output early warning information based on hierarchical alarm rules.

[0006] Optionally, S1 specifically includes: Distributed acoustic sensing DAS composite fiber, distributed temperature sensing DTS fiber, and fiber Bragg grating (FBG) sensor array are deployed along the building structure path within the pre-defined area of ​​the courtyard. The DAS composite fiber is continuously laid along the perimeter wall, indoor passage and main path of the courtyard, the DTS fiber is continuously laid along the roof ridge, eaves, ring beam and temperature-sensitive area of ​​the kitchen, and the FBG sensor is set at the load-bearing column node, the mid-span of the main beam and key structural connection node. The Rayleigh scattering signal is continuously sampled by the DAS sensing unit based on the coherent optical time-domain reflectometry (C-OTDR) mechanism to obtain the vibration time-series signal at the corresponding spatial location. The echo signal is demodulated using the DTS sensing unit based on the Raman scattering effect to obtain the spatiotemporal temperature sequence data distributed along the optical fiber. The FBG sensing unit is used to detect the change in the wavelength of the grating reflection in real time, and the wavelength drift is converted into structural strain and displacement response data. The DTS / DAS demodulator and FBG demodulator are used to perform unified time reference calibration on various types of sensor data, and generate multi-source fiber optic sensor raw data with unified timestamps.

[0007] Optionally, the process of performing unified time reference calibration on various types of sensor data using a DTS / DAS demodulator and an FBG demodulator specifically involves: The original timestamp sequences output by the DTS sensing unit, DAS sensing unit and FBG sensing unit are obtained respectively, and a unified time reference time axis is constructed based on the internal clock of the DTS / DAS demodulator as the master reference clock. The time drift deviation sequence of FBG data relative to DAS / DTS data is obtained by calculating the point-by-point difference between the timestamp sequence output by the FBG demodulator and the time axis of the unified time reference. A time correction function is constructed based on the time drift deviation sequence, and the FBG structure response data is processed by segmented time axis mapping to align the FBG data in different time periods to the corresponding DAS / DTS time intervals. The DAS vibration time series data, DTS temperature time series data and FBG structural response data are divided into sliding time windows, and the window length and step size of the sliding window are uniformly set based on the preset sampling period. Within each sliding time window, time resampling processing is performed on the three types of sensor data to convert non-equal interval sampled data into an equal time interval sequence; The resampled DAS, DTS and FBG data are reconstructed and sorted in chronological order, and cross-sensor channel alignment mapping is performed on the time axis to generate multi-source fiber optic sensing raw data with a unified timestamp. In the process of time alignment, abnormal timestamp data is removed or interpolated and corrected, and the corrected timestamp sequence is arranged continuously.

[0008] Optionally, S2 specifically includes: Based on the actual deployment path of the DTS / DAS composite optical cable in the courtyard environment and the installation position of the FBG sensor at the key nodes of the structure, the collected DAS vibration data, DTS temperature data and FBG strain and displacement data are processed for spatial coordinate calibration. The continuous sampling position along the optical fiber is mapped to the spatial geometric position of the courtyard building structure, so that each segment of optical fiber sensing data corresponds to a specific courtyard space area. After completing the spatial calibration, the courtyard space is divided according to its functional attributes, including the perimeter protection area, indoor living area, personnel passage area and structural safety monitoring area. Based on the role of different areas in the monitoring task, corresponding initial weight coefficients are set for the three types of sensor data: DAS, DTS and FBG. The weight coefficients are preset according to the matching relationship between the sensor type and the spatial area. The weighted raw data of multi-source fiber optic sensing is processed by time slicing according to a unified time reference. DAS, DTS and FBG data are synchronously segmented using a fixed time window or a sliding time window. Data with different sampling frequencies are aligned within the same time segment. During time alignment, DAS vibration data is resampled according to the spatial position of the optical fiber to arrange the vibration signal in a spatially continuous manner. DTS temperature data is interpolated and reconstructed according to the distribution position of the optical fiber to form a continuous spatial sequence. FBG strain and displacement data are aligned in sequence according to the structural node number to make the discrete structural response data correspond one-to-one with the spatial position. Various types of sensor data are fused and stitched together according to both spatial location and time dimensions, so that DAS, DTS and FBG data in the same spatial area within the same time window are jointly organized and a unified data representation is generated. The final output is multi-source fiber optic sensor fusion data consistent with the courtyard spatial structure. This fusion data includes spatial location index, time window index, and corresponding DAS vibration data, DTS temperature data, and FBG structural response data sequences.

[0009] Optionally, the process of dually fusing and stitching together various sensor data according to both spatial location and temporal dimensions specifically involves: After spatial coordinate calibration and time alignment, the DAS vibration data, DTS temperature data, and FBG structural response data are classified and organized according to the spatial location index to form a multi-type sensor data set corresponding to each spatial location. Within the multi-type sensor data set, different types of sensor data are sorted by time dimension according to a unified time window, so that DAS vibration data, DTS temperature data and FBG structural response data are arranged in corresponding time segments. A two-dimensional index structure is constructed based on spatial location index and time window index, and multiple types of sensor data are combined and mapped based on the two-dimensional index structure, so that each spatial location corresponds to each time window to form a data combination unit. The data combination units are arranged sequentially according to the spatial location dimension and serialized and spliced ​​in the time dimension to form multi-source fiber optic sensing fusion data. The output includes multi-source fiber optic sensor fusion data containing spatial location index, time window index, and corresponding DAS vibration data, DTS temperature data, and FBG structural response data.

[0010] Optionally, S3 specifically includes: The DAS vibration signal in the multi-source fiber optic sensor fusion data after spatiotemporal alignment is segmented according to a preset time window, and the vibration data in each time window is processed by short-time Fourier transform (STFT) to calculate the time-frequency distribution matrix of the vibration signal in different time periods and extract the spectral amplitude features in the corresponding time window. Based on the STFT processing, the DAS vibration signal is further processed by continuous wavelet transform (WT). The vibration signal is decomposed into multiple scales according to different scale factors, the wavelet coefficients at each scale are calculated, and the transient change components and local abrupt change components of the vibration signal are extracted. The DTS temperature data is serialized according to spatial sampling points, and the temperature difference change sequence between adjacent sampling points is calculated. At the same time, the first-order difference is calculated on the temperature sequence to form a temperature change gradient feature sequence. The FBG strain data and displacement data are processed synchronously to calculate the strain change and displacement change of each structural monitoring node in adjacent time windows. The structural response data are then arranged in a time series to form a structural response change characteristic sequence. The DAS vibration time-frequency features obtained by STFT processing, the DAS wavelet coefficient features obtained by WT transform, the DTS temperature change features, and the FBG strain and displacement change features are spliced ​​together by feature dimensions, and the spliced ​​high-dimensional features are uniformly normalized to construct a multi-parameter fusion feature vector.

[0011] Optionally, S4 specifically includes: The multi-parameter fused feature vector is input into a preset feature representation model. The input feature vector is serialized and organized according to the time dimension, arranged into a time-series feature sequence according to a unified time window order, and an initial behavioral feature representation is established based on the sequence structure. In the feature representation process, based on the propagation attenuation characteristics of DAS vibration signals in optical fiber medium, the vibration characteristics corresponding to different spatial locations are constrained and modeled. Among them, the spatial attenuation constraint term is constructed according to the correspondence between optical fiber distance and vibration signal intensity, and the feature representation update process is introduced. Meanwhile, based on the physical consistency relationship between FBG sensors in structural strain and displacement response, the structural response characteristics are constrained and modeled. The synchronous change relationship between strain change and displacement change is embedded as a constraint condition in the feature optimization process, and the consistency correction of the structural response characteristics is performed. Under the constraints, the multi-parameter fusion feature vector is iteratively optimized. The input features are nonlinearly transformed by the feature mapping function, so that different types of sensing features are reconstructed and fused in a unified feature space to form an intermediate feature representation with structural constraints. Temporal consistency verification is performed on the intermediate feature representations after constraint optimization, the feature change trend is smoothed according to continuous time windows, and feature jumps between time windows are eliminated. The final output is a semantic representation of human behavior, which is a unified representation that integrates DAS vibration features, DTS temperature features, and FBG structural response features.

[0012] Optionally, S5 specifically includes: The semantic representation of the personnel behavior is arranged in chronological order to construct continuous time series input data, and the time series input data is subjected to sliding segmentation based on a preset time window; The segmented semantic feature sequence is input into the improved Res-GRU temporal behavior recognition model for temporal modeling processing to generate a global temporal behavior feature vector. Based on the global temporal behavior feature vector, classification mapping is performed. The temporal features are mapped to the personnel behavior category space through a preset behavior category mapping relationship, and the corresponding personnel behavior category and confidence parameters are output.

[0013] Optionally, the step of inputting the segmented semantic feature sequence into the improved Res-GRU temporal behavior recognition model for temporal modeling processing to generate a global temporal behavior feature vector specifically involves: The segmented semantic feature sequence is input into the improved Res-GRU temporal behavior recognition model for temporal modeling processing to generate a global temporal behavior feature vector. The improved Res-GRU temporal behavior recognition model includes an input processing module, a time-dependent recursive modeling module, a state update control module, a temporal enhancement propagation module, and a global feature aggregation module. The input processing module receives the segmented semantic feature sequence and arranges the semantic features corresponding to each time window in chronological order to form a standard time-series input vector sequence, while maintaining the semantic continuity index relationship between time windows; The time-dependent recursive modeling module recursively models the dependencies of the input vector sequence in the time dimension. At each time step, it jointly calculates the input vector of the current time step and the hidden state of the previous time step to generate the candidate hidden state of the current time step. The state update control module is based on the reset gate and update gate mechanism, introduces residual information path, fuses and updates candidate hidden states and historical hidden states, and constructs an adaptive residual adjustment factor based on the semantic feature sequence. The residual fusion weight is dynamically adjusted according to the semantic state change intensity of different time windows to generate the hidden state of the current time step. The step of constructing an adaptive residual adjustment factor based on the semantic feature sequence includes: extracting the change intensity of semantic features in the current time window; calculating the Euclidean distance between the semantic vector of the current time window and the semantic vectors of adjacent time windows; quantifying the intensity of temporal fluctuations in the current window; and inputting the quantized change intensity into a single-layer fully connected mapping layer and outputting it after passing through a Sigmoid activation function. , which serves as the adaptive residual adjustment factor for the current time step; The hidden state at the current time step is represented as: ,in, Indicates an update to the door. This indicates the hidden state at the previous time step. Indicates the candidate hidden state at the current time step; The temporal enhancement propagation module performs cross-temporal enhancement fusion propagation on the hidden state of the current time step and the historical hidden state based on the residual connection mechanism. The temporal enhancement propagation module receives the adaptive residual adjustment factor corresponding to the current time step and uses it as the dynamic weight benchmark for cross-temporal residual fusion. The cross-temporal enhancement fusion strength of the historical hidden state and the current hidden state is synchronously linked with the dynamic weight benchmark. The global feature aggregation module performs global aggregation processing on the hidden state sequences output at all time steps based on a combination of terminal hidden state extraction and time-dimensional attention-weighted aggregation, generating a global temporal behavior feature vector.

[0014] Optionally, S6 specifically includes: The personnel behavior categories and confidence parameters are synchronously input with DTS temperature data and FBG structural response data, and the three types of data are time-aligned based on a unified timestamp. The current behavior state is initially determined based on the personnel behavior categories, and personnel behavior category identification information within the corresponding time window is extracted. The personnel behavior category identification information includes normal walking behavior, stationary behavior, falling behavior, gathering behavior, and abnormal intrusion behavior. The DTS temperature data is divided into time windows, and the temperature change, temperature gradient change and temperature change characteristics within each time window are calculated. Based on the preset temperature threshold, DTS temperature anomaly information is generated. The FBG structural response data is divided into time synchronization segments to calculate the strain and displacement changes of each structural monitoring node, and FBG structural anomaly information is generated based on the structural response threshold. Spatial location mapping and time series alignment are performed on personnel behavior category identification information, DTS temperature anomaly information and FBG structure anomaly information to construct a multi-source fusion judgment data structure. Based on the multi-source fusion judgment data structure, hierarchical rule matching processing is performed to verify the rules for different categories of personnel behavior. Among them, fall behavior and abnormal intrusion behavior are given priority for high-level risk judgment, and cross-validation is performed by combining DTS temperature anomaly information and FBG structure anomaly information. The risk level output is generated based on the combination relationship between personnel behavior category, DTS temperature anomaly information and FBG structural anomaly information, and the first-level warning information, second-level warning information or third-level warning information is output according to the preset priority order. The warning information is sent to the courtyard safety monitoring terminal or remote management platform, and the time windows corresponding to fall behavior and abnormal intrusion behavior are marked and stored.

[0015] The beneficial effects of this invention are as follows: This invention addresses the problems of incomplete coverage, insufficient structural response accuracy, difficulty in identifying human behavior, and high false alarm rate in traditional courtyard monitoring by synergistically fusing DTS / DAS distributed fiber optic sensing and FBG fiber grating structural response monitoring. It employs unified time synchronization, spatial correspondence mapping, and courtyard functional zoning weight allocation to achieve unified organization of DAS vibration data, DTS temperature data, and FBG strain and displacement data in spatial location and time window. In the feature extraction stage, it combines STFT and WT to extract vibration time-frequency features and transient change features respectively, and fuses temperature change features and structural response change features to form a multi-parameter fused feature vector. To address the intertwining of vibration interference, temperature fluctuations, and structural micro-variations in the courtyard environment, this invention further constructs a consistency constraint mechanism based on DAS vibration propagation characteristics and FBG structural response characteristics to constrain and optimize the fused features, obtaining a more stable semantic representation of human behavior. Furthermore, it models behavioral changes within a continuous time window by improving the Res-GRU temporal behavior recognition model, outputting human behavior categories and confidence parameters. Ultimately, this invention combines DTS temperature anomaly information and FBG structural anomaly information for multi-source fusion judgment, realizing graded early warning of falls, gatherings, abnormal intrusions, temperature anomalies, and structural anomalies, thereby improving the integrated monitoring capabilities of personnel safety, environmental safety, and structural safety in courtyard settings. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a personnel behavior recognition method based on fiber optic sensing data proposed in this invention. Figure 2 This is a diagram of the courtyard fiber optic multi-parameter sensor deployment and data acquisition structure for a personnel behavior recognition method based on fiber optic sensor data driven by the present invention. Figure 3 This is a structural diagram of an improved Res-GRU time-series behavior recognition model based on fiber optic sensing data-driven human behavior recognition proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A method for human behavior recognition based on fiber optic sensor data includes the following steps: S1. Synchronously acquire distributed temperature sensor (DTS) data, distributed acoustic sensor (DAS) data, and fiber optic grating (FBG) structure response data, and perform time synchronization to obtain multi-source fiber optic sensing raw data. S2. Based on the DTS / DAS composite optical cable deployment path and FBG sensor deployment location, the original multi-source fiber optic sensing data is spatially mapped, and after differentiating the sensing data of different areas by combining the courtyard functional zoning, the data is spatiotemporally aligned to generate multi-source fiber optic sensing fusion data. S3. Using the multi-source fiber optic sensing fusion data as the processing object, short-time Fourier transform (STFT) and wavelet transform (WT) are used to extract the vibration features of DAS, and combined with the temperature change features of DTS and the strain and displacement change features of FBG, a multi-parameter fusion feature vector is formed. S4. Input the multi-parameter fused feature vector into the feature representation model, and construct a consistency constraint mechanism based on the DAS vibration propagation characteristics and FBG structural response characteristics to constrain and optimize the feature representation, thereby obtaining the semantic representation of personnel behavior. S5. Construct a temporal behavior recognition model based on the improved Res-GRU, perform temporal modeling and analysis on the semantic representation of the personnel behavior, and output the personnel behavior category and confidence parameters. S6. Based on the personnel behavior category and confidence parameters, combined with DTS temperature anomaly information and FBG structural anomaly information, perform multi-source fusion judgment, and output early warning information based on hierarchical alarm rules.

[0019] In this embodiment, S1 specifically refers to: Distributed acoustic sensing DAS composite fiber, distributed temperature sensing DTS fiber, and fiber Bragg grating (FBG) sensor array are deployed along the building structure path within the pre-defined area of ​​the courtyard. The DAS composite fiber is continuously laid along the perimeter wall, indoor passage and main path of the courtyard, the DTS fiber is continuously laid along the roof ridge, eaves, ring beam and temperature-sensitive area of ​​the kitchen, and the FBG sensor is set at the load-bearing column node, the mid-span of the main beam and key structural connection node. The Rayleigh scattering signal is continuously sampled by the DAS sensing unit based on the coherent optical time-domain reflectometry (C-OTDR) mechanism to obtain the vibration time-series signal at the corresponding spatial location. The echo signal is demodulated using the DTS sensing unit based on the Raman scattering effect to obtain the spatiotemporal temperature sequence data distributed along the optical fiber. The FBG sensing unit is used to detect the change in the wavelength of the grating reflection in real time, and the wavelength drift is converted into structural strain and displacement response data. The DTS / DAS demodulator and FBG demodulator are used to perform unified time reference calibration on various types of sensor data, and generate multi-source fiber optic sensor raw data with unified timestamps.

[0020] In the specific implementation process, a single courtyard was used as the monitoring object. The perimeter of the courtyard was set to 60m to 100m. DAS composite optical fiber was continuously laid along the inner side of the perimeter wall at a height of 0.3m to 0.5m and fixed with a combination of clips and structural adhesive. The spacing between adjacent fixing points was set to 0.5m. In the indoor passage and main path area of ​​the courtyard, DAS composite optical fiber was laid along the edge along the passage direction. The spacing between adjacent optical fiber segments in the passage area was set to 0.8m to 1.2m. The spatial sampling interval of the DAS sensing unit was set to 0.5m, the sampling frequency was set to 1000Hz, the pulse width was set to 50ns, and the working center wavelength was set to 1550nm. Vibration time sequence signals corresponding to each spatial sampling point were continuously collected. DTS optical fibers are continuously deployed along the roof ridge, eaves, ring beams, and temperature-sensitive areas of the kitchen. The fixed fiber spacing in the roof ridge and eaves areas is set to 0.6m, while the fiber spacing on the kitchen walls and near the gas pipes is set to 50mm to 100mm from the outer wall of the gas pipes. The spatial resolution of the DTS sensing unit is set to 1m, the temperature sampling period is set to 2s, the temperature measurement accuracy is set to ±0.5℃, and the temperature data are sequentially arranged along the fiber length to form a temperature spatiotemporal sequence. FBG sensors were installed at load-bearing column nodes, mid-span of main beams, and key structural connection nodes. At load-bearing column nodes, FBG sensors were installed at heights of 0.3m, 1.5m, and 2.7m above the ground, respectively. One FBG strain sensor was installed at the mid-span of the main beam, and one FBG displacement sensor was installed at each beam end connection. The center wavelength of the FBG sensors was set within the range of 1528nm to 1568nm, with a 2nm interval between the center wavelengths of adjacent sensors. The sampling frequency of the FBG demodulator was set to 100Hz, and the wavelength resolution was set to 1pm. The detected wavelength drift was converted into structural strain and displacement response data. The DTS / DAS demodulator and the FBG demodulator are synchronized via a unified timing module using PTP clock synchronization, with the time synchronization error controlled within 1ms. During data acquisition, DAS vibration timing signals, DTS temperature spatiotemporal sequence data, and FBG structural strain and displacement response data are all written to a buffer queue with a unified timestamp. The time window length of the buffer queue is set to 1 second, and the overlap rate of adjacent time windows is set to 50%. When DAS, DTS, and FBG data are received simultaneously within the same time window, the three types of data are aligned and encapsulated according to their timestamps to generate multi-source fiber optic sensing raw data with a unified timestamp.

[0021] In this embodiment, the unified time reference calibration processing of various sensor data using a DTS / DAS demodulator and an FBG demodulator is specifically as follows: The original timestamp sequences output by the DTS sensing unit, DAS sensing unit and FBG sensing unit are obtained respectively, and a unified time reference time axis is constructed based on the internal clock of the DTS / DAS demodulator as the master reference clock. The time drift deviation sequence of FBG data relative to DAS / DTS data is obtained by calculating the point-by-point difference between the timestamp sequence output by the FBG demodulator and the time axis of the unified time reference. A time correction function is constructed based on the time drift deviation sequence, and the FBG structure response data is processed by segmented time axis mapping to align the FBG data in different time periods to the corresponding DAS / DTS time intervals. The DAS vibration time series data, DTS temperature time series data and FBG structural response data are divided into sliding time windows, and the window length and step size of the sliding window are uniformly set based on the preset sampling period. Within each sliding time window, time resampling processing is performed on the three types of sensor data to convert non-equal interval sampled data into an equal time interval sequence; The resampled DAS, DTS and FBG data are reconstructed and sorted in chronological order, and cross-sensor channel alignment mapping is performed on the time axis to generate multi-source fiber optic sensing raw data with a unified timestamp. In the process of time alignment, abnormal timestamp data is removed or interpolated and corrected, and the corrected timestamp sequence is arranged continuously. In practical implementation, the DTS / DAS demodulator serves as the primary time reference device, while the FBG demodulator acts as the secondary time reference device. Both establish the same time reference through a unified timing signal or synchronous communication interface. When acquiring DTS temperature data and DAS vibration data, the DTS / DAS demodulator writes a primary reference timestamp to each set of sampled data. Similarly, when acquiring FBG structural response data, the FBG demodulator writes a local timestamp to each set of strain and displacement response data. The system reads the FBG demodulator's local timestamp according to a preset calibration cycle and compares it with the primary reference timestamp provided by the DTS / DAS demodulator to calculate the time deviation of the FBG demodulator relative to the primary time reference. When the time deviation is within the allowable range, the current acquisition status is maintained. When the time deviation exceeds a preset threshold, the FBG demodulator's timestamp is corrected based on the deviation value, ensuring that the FBG structural response data corresponds to the same acquisition time of the DTS / DAS data. Through the above processing, the DTS temperature data, DAS vibration data, and FBG structural response data all have a unified time reference, providing a time consistency basis for the synchronous generation of subsequent multi-source fiber optic sensing raw data.

[0022] In this embodiment, S2 specifically refers to: Based on the actual deployment path of the DTS / DAS composite optical cable in the courtyard environment and the installation position of the FBG sensor at the key nodes of the structure, the collected DAS vibration data, DTS temperature data and FBG strain and displacement data are processed for spatial coordinate calibration. The continuous sampling position along the optical fiber is mapped to the spatial geometric position of the courtyard building structure, so that each segment of optical fiber sensing data corresponds to a specific courtyard space area. After completing the spatial calibration, the courtyard space is divided according to its functional attributes, including the perimeter protection area, indoor living area, personnel passage area and structural safety monitoring area. Based on the role of different areas in the monitoring task, corresponding initial weight coefficients are set for the three types of sensor data: DAS, DTS and FBG. The weight coefficients are preset according to the matching relationship between the sensor type and the spatial area. The weighted raw data of multi-source fiber optic sensing is processed by time slicing according to a unified time reference. DAS, DTS and FBG data are synchronously segmented using a fixed time window or a sliding time window. Data with different sampling frequencies are aligned within the same time segment. During time alignment, DAS vibration data is resampled according to the spatial position of the optical fiber to arrange the vibration signal in a spatially continuous manner. DTS temperature data is interpolated and reconstructed according to the distribution position of the optical fiber to form a continuous spatial sequence. FBG strain and displacement data are aligned in sequence according to the structural node number to make the discrete structural response data correspond one-to-one with the spatial position. Various types of sensor data are fused and stitched together according to both spatial location and time dimensions, so that DAS, DTS and FBG data in the same spatial area within the same time window are jointly organized and a unified data representation is generated. The final output is multi-source fiber optic sensor fusion data consistent with the courtyard spatial structure. This fusion data includes spatial location index, time window index, and corresponding DAS vibration data, DTS temperature data, and FBG structural response data sequence. In the specific implementation process, the DAS vibration data acquisition adopts a distributed sampling unit with a spatial resolution of 0.5m. One sampling reference point is set every 10m in the perimeter protection area, one sampling reference point is set every 5m in the indoor passage area, and one high-density sampling point is set every 3m in the main personnel passage area. The DAS sampling frequency is set to 1000Hz, the single sampling time window length is set to 1s, and the sliding step size is set to 0.5s to ensure the continuous capture capability of dynamic behavior signals. DTS temperature data employs differentiated sampling strategies in different functional areas, with the sampling interval set at 1.5m for the perimeter area, 1m for the indoor area, and 0.5m for the kitchen and heat source sensitive areas; the temperature sampling period is set to 2s, and the temperature interpolation reconstruction uses the cubic spline interpolation method, with the spatial reconstruction error controlled within ±0.3℃. The FBG structural response data uses a fixed-point high-precision sampling method at key nodes. The sampling frequency for load-bearing column nodes, the mid-span position of the main beam, and connection nodes is set to 1Hz. The strain measurement resolution is set to 1με, and the displacement measurement resolution is set to 0.01mm to ensure the ability to identify minute changes in the structure. In time-slice processing, the uniform time window length is set to 1 second and the window overlap rate is set to 50%, that is, a fused data time slice is generated every 0.5 seconds; different modal data are aligned in the same time slice through the PTP time synchronization mechanism, and the time synchronization error is controlled within 1 ms; During the spatial coordinate calibration process, the building BIM model is used as the reference coordinate system, the spatial mapping accuracy is controlled within ±0.1m, and the least squares fitting method is used to reconstruct the optical fiber path spatially, so that DAS, DTS and FBG data can all correspond to a unified set of three-dimensional coordinate points. Finally, in the fusion stage, a unified representation structure is generated by vector-level splicing of multi-source data within the same spatial unit. The data dimensions of each spatial-temporal unit are unified as (DAS: 1000×1, DTS: 1×1, FBG: 1×1), and stored in JSON structured format to generate multi-source fiber optic sensing fusion data with a unified structure.

[0023] In this embodiment, the dual fusion and splicing of various sensor data according to the spatial location dimension and the time dimension is specifically as follows: After spatial coordinate calibration and time alignment, the DAS vibration data, DTS temperature data, and FBG structural response data are classified and organized according to the spatial location index to form a multi-type sensor data set corresponding to each spatial location. Within the multi-type sensor data set, different types of sensor data are sorted by time dimension according to a unified time window, so that DAS vibration data, DTS temperature data and FBG structural response data are arranged in corresponding time segments. A two-dimensional index structure is constructed based on spatial location index and time window index, and multiple types of sensor data are combined and mapped based on the two-dimensional index structure, so that each spatial location corresponds to each time window to form a data combination unit. The data combination units are arranged sequentially according to the spatial location dimension and serialized and spliced ​​in the time dimension to form multi-source fiber optic sensing fusion data; The output includes multi-source fiber optic sensor fusion data containing spatial location index, time window index, and corresponding DAS vibration data, DTS temperature data, and FBG structural response data; In the specific implementation process, the left corner of the courtyard entrance is used as the origin of the spatial coordinates. An X-axis is established along the east-west direction of the courtyard, a Y-axis is established along the north-south direction, and a Z-axis is established along the building height direction. The unit of spatial coordinates is set to m. A spatial location index is established for DAS vibration data at a spatial interval of 0.5m, a spatial location index is established for DTS temperature data at a spatial interval of 1m, and a spatial location index is established for FBG structural response data according to the actual installation node number of the sensor. The spatial coordinates of the load-bearing column node where the FBG sensor is located, the mid-span position of the main beam, and the key connection nodes of the structure are entered into the spatial index table. The time window length is set to 1 second, the sliding step size is set to 0.5 seconds, and a time window index is generated for each time window. The time window index is numbered sequentially according to the acquisition time. Within the same time window, 1000 sampling points are extracted from DAS vibration data, 1 temperature sampling value is extracted from DTS temperature data, and 100 strain and displacement sampling points are extracted from FBG structural response data. The three types of data are written into the corresponding data buffer units respectively. When constructing the two-dimensional index structure, the spatial location index is used as the row index and the time window index is used as the column index, forming a spatial-temporal two-dimensional index matrix; the minimum granularity of the spatial location index is set to 0.5m, the minimum granularity of the time window index is set to 0.5s, and each two-dimensional index position corresponds to a data combination unit; For each data combination unit, the DAS vibration data, DTS temperature data and FBG structural response data within the corresponding spatial location and time window are spliced ​​together in the order of DAS, DTS and FBG. The length of the DAS vibration data is set to 1000, the length of the DTS temperature data is set to 1, the length of the FBG structural response data is set to 1, and the length of the spliced ​​single data combination unit is set to 1002. During the serialization and splicing process, data combination units within the same time window are first arranged in ascending order of spatial location index, and then data combination units in different time windows are arranged in ascending order of time window index, forming multi-source fiber optic sensor fusion data with one-to-one correspondence between spatial location index, time window index, and sensor data.

[0024] In this embodiment, S3 specifically refers to: The DAS vibration signal in the multi-source fiber optic sensor fusion data after spatiotemporal alignment is segmented according to a preset time window, and the vibration data in each time window is processed by short-time Fourier transform (STFT) to calculate the time-frequency distribution matrix of the vibration signal in different time periods and extract the spectral amplitude features in the corresponding time window. Based on the STFT processing, the DAS vibration signal is further processed by continuous wavelet transform (WT). The vibration signal is decomposed into multiple scales according to different scale factors, the wavelet coefficients at each scale are calculated, and the transient change components and local abrupt change components of the vibration signal are extracted. The DTS temperature data is serialized according to spatial sampling points, and the temperature difference change sequence between adjacent sampling points is calculated. At the same time, the first-order difference is calculated on the temperature sequence to form a temperature change gradient feature sequence. The FBG strain data and displacement data are processed synchronously to calculate the strain change and displacement change of each structural monitoring node in adjacent time windows. The structural response data are then arranged in a time series to form a structural response change characteristic sequence. The DAS vibration time-frequency features obtained by STFT processing, the DAS wavelet coefficient features obtained by WT transform, the DTS temperature change features, and the FBG strain and displacement change features are spliced ​​together by feature dimensions, and the spliced ​​high-dimensional features are uniformly normalized to construct a multi-parameter fusion feature vector.

[0025] In the specific implementation process, the time window length of the DAS vibration signal was set to 1s, the sliding step size was set to 0.5s, and the sampling frequency was 1000Hz. The STFT processing adopted a 256-point FFT window with a 50% overlap rate to ensure the balance between time-frequency resolution and time continuity. The continuous wavelet transform (WT) adopted the db4 wavelet basis function and the number of decomposition layers was set to 4 to extract transient vibration features at different scales. The DTS temperature data was spatially sequenced and reconstructed at a sampling frequency of 1Hz. The FBG strain and displacement data were synchronously aligned at a sampling frequency of 100Hz and unified to a 1s time window scale through linear interpolation. After splicing, the multi-parameter features formed a three-dimensional tensor structure with dimensions of N×T×F, where N represents the number of spatial sampling points, T represents the number of time windows, and F represents the dimension of the fused features. Finally, the features were normalized by the Z-score normalization method, and their numerical range was constrained to improve the convergence stability of the model.

[0026] In this embodiment, S4 specifically refers to: The multi-parameter fused feature vector is input into a preset feature representation model. The input feature vector is serialized and organized according to the time dimension, arranged into a time-series feature sequence according to a unified time window order, and an initial behavioral feature representation is established based on the sequence structure. In the feature representation process, based on the propagation attenuation characteristics of DAS vibration signals in optical fiber medium, the vibration characteristics corresponding to different spatial locations are constrained and modeled. Among them, the spatial attenuation constraint term is constructed according to the correspondence between optical fiber distance and vibration signal intensity, and the feature representation update process is introduced. Meanwhile, based on the physical consistency relationship between FBG sensors in structural strain and displacement response, the structural response characteristics are constrained and modeled. The synchronous change relationship between strain change and displacement change is embedded as a constraint condition in the feature optimization process, and the consistency correction of the structural response characteristics is performed. Under the constraints, the multi-parameter fusion feature vector is iteratively optimized. The input features are nonlinearly transformed by the feature mapping function, so that different types of sensing features are reconstructed and fused in a unified feature space to form an intermediate feature representation with structural constraints. Temporal consistency verification is performed on the intermediate feature representations after constraint optimization, the feature change trend is smoothed according to continuous time windows, and feature jumps between time windows are eliminated. The final output is a semantic representation of human behavior, which is a unified representation after fusing DAS vibration features, DTS temperature features, and FBG structural response features. In the specific implementation process, the multi-parameter fusion feature vector is organized into a three-dimensional feature tensor according to the spatial location index and the time window index. The tensor dimension is set to N×T×F, where N is the number of spatial sampling points, T is the number of continuous time windows, and F is the fusion feature dimension. In the single courtyard example, N is set to 120, T is set to 20, and F is set to 64. The feature representation model includes an input mapping layer, a constraint modeling layer, a feature optimization layer, and a semantic output layer. The input mapping layer maps the multi-parameter fusion feature vector to a 128-dimensional hidden feature space and uses a linear mapping function and a ReLU activation function for nonlinear transformation.

[0027] In DAS vibration propagation constraint modeling, the spatial sampling point corresponding to the triggering location of personnel behavior is used as the reference point. The fiber optic path distance between the reference point and adjacent spatial sampling points is calculated, with a distance calculation interval of 0.5m. The spatial attenuation coefficient is set to 0.08, and the vibration intensity attenuation weight is calculated according to the fiber optic path distance. The vibration characteristic difference between adjacent spatial sampling points is combined with the corresponding attenuation weight to form a spatial attenuation constraint term. Spatial sampling points exceeding 15m are not included in the calculation of the same spatial attenuation constraint term.

[0028] In the FBG structural response consistency constraint modeling, the strain change and displacement change at the same structural node are paired according to the same time window, with the time window length set to 1s and the sliding step size set to 0.5s. The strain change is normalized to the 0-1 interval, and the displacement change is normalized to the 0-1 interval. The synchronous change difference between the two in 5 consecutive time windows is calculated. When the synchronous change difference is greater than 0.2, the corresponding structural response feature is input into the consistency correction branch.

[0029] During feature optimization, the number of iterations was set to 3, the weight coefficient of the spatial decay constraint term was set to 0.4, the weight coefficient of the structural response consistency constraint term was set to 0.3, and the weight coefficient of the original feature preservation term was set to 0.3. After each iteration, the intermediate feature representation was input into the 64-dimensional semantic output layer and smoothed according to the time window order. The smoothing window length was set to 3 time windows, and finally the 64-dimensional semantic representation of human behavior was output.

[0030] In this embodiment, S5 specifically refers to: The semantic representation of human behavior is arranged in chronological order to construct continuous time series input data, and the time series input data is processed by sliding segmentation based on a preset time window; The segmented semantic feature sequence is input into the improved Res-GRU temporal behavior recognition model for temporal modeling processing to generate a global temporal behavior feature vector. Classification and mapping processing is performed based on global temporal behavioral feature vectors. Temporal features are mapped to the personnel behavior category space through a preset behavior category mapping relationship, and the corresponding personnel behavior category and confidence parameters are output. In the specific implementation process, the semantic representation of personnel behavior is arranged from earliest to latest according to the timestamp. The dimension of the semantic representation of personnel behavior corresponding to a single time window is set to 64 dimensions, and the sequence length of the continuous time series input data is set to 20 time windows. The window length of the sliding segmentation processing is set to 20 seconds, the sliding step size is set to 5 seconds, and 15 overlapping data of time windows are retained between adjacent sliding windows. For sequences with fewer than 20 time windows, zero-value padding is used to make them complete. For sequences with more than 20 time windows, the 20 most recent time windows are truncated in chronological order to form a semantic feature sequence with a dimension of 20×64.

[0031] The improved Res-GRU temporal behavior recognition model is set as a two-layer recursive structure, with 128 hidden units in each layer, 128 residual connection mapping dimensions, and a Dropout parameter of 0.2. The semantic feature sequence is input into the first layer of Res-GRU to generate the first hidden state sequence, which is then input into the second layer of Res-GRU for temporal state updates. The terminal hidden state is then concatenated with the time-weighted convergence result to form a 256-dimensional global temporal behavior feature vector.

[0032] The classification mapping process is implemented using a fully connected classification layer. The input dimension of the fully connected classification layer is set to 256, and the output dimension is set to 5 categories according to the preset number of personnel behavior categories, corresponding to normal walking behavior, stationary behavior, falling behavior, gathering behavior, and abnormal intrusion behavior, respectively. The output results of the classification layer are calculated by the Softmax function to obtain the confidence parameter corresponding to each personnel behavior category, and the category with the largest confidence parameter is selected as the output personnel behavior category.

[0033] In this embodiment, the segmented semantic feature sequence is input into the improved Res-GRU temporal behavior recognition model for temporal modeling processing to generate a global temporal behavior feature vector, specifically as follows: The segmented semantic feature sequence is input into the improved Res-GRU temporal behavior recognition model for temporal modeling processing to generate a global temporal behavior feature vector. The improved Res-GRU temporal behavior recognition model includes an input processing module, a time-dependent recursive modeling module, a state update control module, a temporal enhancement propagation module, and a global feature aggregation module. The input processing module receives the segmented semantic feature sequence and arranges the semantic features corresponding to each time window in chronological order to form a standard time-series input vector sequence, while maintaining the semantic continuity index relationship between time windows; The time-dependent recursive modeling module recursively models the dependencies of the input vector sequence in the time dimension. At each time step, it jointly calculates the input vector of the current time step and the hidden state of the previous time step to generate the candidate hidden state of the current time step. The state update control module, based on the reset gate and update gate mechanism, introduces a residual information path to fuse and update candidate hidden states and historical hidden states. It also constructs an adaptive residual adjustment factor based on the semantic feature sequence, dynamically adjusting the residual fusion weights according to the intensity of semantic state changes in different time windows to generate the hidden state at the current time step. Specifically, constructing the adaptive residual adjustment factor based on the semantic feature sequence includes: extracting the intensity of semantic feature changes in the current time window; calculating the Euclidean distance between the semantic vector of the current time window and the semantic vectors of adjacent time windows to quantify the temporal fluctuation severity of the current window; and inputting the quantized intensity of change into a single-layer fully connected mapping layer and outputting it after passing through a Sigmoid activation function. , which serves as the adaptive residual adjustment factor for the current time step; The hidden state at the current time step is represented as: ,in, Indicates an update to the door. This indicates the hidden state at the previous time step. Indicates the candidate hidden state at the current time step; The temporal enhancement propagation module performs cross-temporal enhancement fusion propagation on the hidden state of the current time step and the historical hidden state based on the residual connection mechanism. The temporal enhancement propagation module receives the adaptive residual adjustment factor corresponding to the current time step and uses it as the dynamic weight benchmark for cross-temporal residual fusion. The cross-temporal enhancement fusion strength of the historical hidden state and the current hidden state is synchronously linked with the dynamic weight benchmark. The global feature aggregation module is based on a combination of terminal hidden state extraction and time-dimensional attention-weighted aggregation. It performs global aggregation processing on the hidden state sequences output at all time steps to generate a global temporal behavior feature vector. In its implementation, the improved Res-GRU temporal behavior recognition model inherits the residual recursive processing framework of the basic Res-GRU model in its overall structure, including input sequence encoding, gating state update, residual state propagation, and global feature aggregation. However, it further introduces an adaptive residual adjustment factor based on semantic feature sequences in the state update control module to replace the fixed residual weight state propagation method in the basic Res-GRU model. First, the segmented semantic feature sequences are arranged according to 20 time windows, each time window corresponding to a 64-dimensional semantic representation of human behavior, and a standard temporal input vector sequence is constructed in chronological order. Then, the correlation between the current time step input vector and the hidden state of the previous time step is calculated through the time-dependent recursive modeling module to generate candidate hidden states. Compared with ordinary GRU, this model retains historical semantic states through the residual information path, which can reduce feature decay in long-term sequence modeling; compared with LSTM, its gating structure is simpler and the computational cost is lower; compared with Transformer-type models, its modeling of short-window continuous behavior sequences is more suitable for deployment in courtyard edge monitoring equipment. By using an adaptive residual adjustment factor to differentiate the intensity of semantic state changes in different time windows, the temporal response features corresponding to sudden behaviors such as falls and abnormal intrusions can be enhanced, ultimately generating a global temporal behavior feature vector. Specifically, the adaptive residual factor can dynamically adjust the residual fusion strength based on the combined fluctuation intensity of the three types of sensor features within a single window: automatically increasing the strength when a sudden change in kitchen temperature occurs simultaneously with a gathering of people. Simultaneously amplify three types of abnormal characteristics: temperature, vibration, and structural deformation; when there is only slight wind vibration and no human activity, pressurize... Suppress noise residuals.

[0034] In this embodiment, S6 specifically refers to: The personnel behavior categories and confidence parameters are synchronously input with DTS temperature data and FBG structural response data, and the three types of data are time-aligned based on a unified timestamp. The current behavior status is initially determined based on the personnel behavior category, and the personnel behavior category identification information within the corresponding time window is extracted. The personnel behavior category identification information includes normal walking behavior, stationary behavior, falling behavior, gathering behavior, and abnormal intrusion behavior. The DTS temperature data is divided into time windows, and the temperature change, temperature gradient change and temperature change characteristics within each time window are calculated. Based on the preset temperature threshold, DTS temperature anomaly information is generated. The FBG structural response data is divided into time synchronization segments to calculate the strain and displacement changes of each structural monitoring node, and FBG structural anomaly information is generated based on the structural response threshold. Spatial location mapping and time series alignment are performed on personnel behavior category identification information, DTS temperature anomaly information and FBG structure anomaly information to construct a multi-source fusion judgment data structure. Based on the multi-source fusion judgment data structure, hierarchical rule matching processing is performed, and rule verification is performed for different personnel behavior categories. Among them, fall behavior and abnormal intrusion behavior are given priority for high-level risk judgment, and cross-validation is performed by combining DTS temperature anomaly information and FBG structure anomaly information. The risk level output is generated based on the combination relationship between personnel behavior category, DTS temperature anomaly information and FBG structural anomaly information, and the first-level warning information, second-level warning information or third-level warning information is output according to the preset priority order. The warning information is sent to the courtyard safety monitoring terminal or remote management platform, and the time windows corresponding to fall behavior and abnormal intrusion behavior are marked and stored. In the specific implementation process, personnel behavior categories and confidence parameters are output according to a 1-second time window, with the confidence parameter value ranged from 0 to 1. When the maximum confidence parameter corresponding to a personnel behavior category is greater than 0.80, that personnel behavior category is taken as the valid identification category for the current time window. When the maximum confidence parameter is less than or equal to 0.80, the current time window is marked as pending confirmation, and the identification results of the two preceding and following time windows are continuously compared. DTS temperature data is divided according to a 1-second time window, with the temperature change threshold set to 80℃ and the temperature gradient change threshold set to 5℃ / m. If the temperature at the same spatial location shows an upward trend within three consecutive time windows, DTS temperature anomaly information is generated. FBG structural response data is divided according to the same time window, with the strain threshold set to 2000με, the strain change threshold set to 500με / min, the displacement threshold set to 10mm, and the displacement change threshold set to 3mm / min. FBG structural anomaly information is generated when any structural monitoring node exceeds the corresponding threshold.

[0035] In the multi-source fusion judgment process, personnel behavior category identification information, DTS temperature anomaly information, and FBG structural anomaly information are matched according to spatial location index and time window index to form a multi-source fusion judgment data structure. The hierarchical alarm rules are set as follows: when the personnel behavior category is a fall and the confidence parameter is greater than 0.85, a level 1 warning is output; when the personnel behavior category is an abnormal intrusion and the confidence parameter is greater than 0.80, a level 1 warning is output; when the personnel behavior category is a gathering behavior or a stationary behavior, and DTS temperature anomaly information or FBG structural anomaly information exists within the same time window, a level 2 warning is output; when the personnel behavior category is a normal walking behavior and no DTS temperature anomaly information or FBG structural anomaly information is present, a level 3 warning is output. This invention, by cross-validating personnel behavior identification results with temperature anomalies and structural anomalies, can reduce false alarms caused by construction vibrations, short-term passage, or environmental temperature fluctuations compared to schemes that only rely on a single vibration signal for alarms, and can form a more explicit warning level in composite risk scenarios involving falls, abnormal intrusions, and accompanying temperature or structural anomalies. The warning information sending delay is set to no more than 3 seconds, and the corresponding time window, spatial location index, personnel behavior category and confidence parameters are stored synchronously.

[0036] Example 1: To verify the feasibility of this invention in practice, it was applied to a safety monitoring scenario of a traditional courtyard-style residential building. The courtyard is a brick-and-wood structure with a single-story building, covering an area of ​​approximately 486 m², with a perimeter wall length of approximately 82 m. It includes three bedrooms, a kitchen, a main north-south passageway, two side passageways, and a semi-open courtyard activity area. There are eight permanent residents in the courtyard, five of whom are over 60 years old. There is less activity at night, but during the day, there are instances of elderly people walking slowly, lingering briefly, groups of people gathering to chat, gas being used in the kitchen, and short-term entry by outsiders. The original monitoring methods in this scenario mainly relied on video cameras, ordinary smoke detectors, and manual patrols. However, due to issues such as wall obstruction, insufficient nighttime lighting, difficulty in promptly detecting falls by the elderly, and the difficulty in jointly judging abnormal kitchen temperatures and structural vibrations, traditional monitoring methods are prone to blind spots, false alarms, and response delays.

[0037] In this scenario, DAS composite optical fibers are laid along the perimeter wall, indoor passageways, and main courtyard paths at a height of 0.4m and a fixed point spacing of 0.5m. DTS optical fibers are laid along the roof ridge, eaves, ring beams, and areas adjacent to the kitchen gas pipeline, with the DTS fiber in the kitchen area approximately 80mm from the outer wall of the gas pipeline. A total of 18 FBG sensors are deployed at load-bearing column nodes, the mid-span of the main beam, and key structural connection nodes. The DAS sampling frequency is 1000Hz, with a spatial sampling interval of 0.5m; the DTS sampling period is 2s, with a spatial resolution of 1m; and the FBG demodulator sampling frequency is 100Hz. The system uses a 1s time window and a 0.5s sliding step for synchronous processing, and maps DAS vibration data, DTS temperature data, and FBG structural response data to a unified spatial location index and time window index to form multi-source fiber optic sensor fusion data.

[0038] The testing period lasted for two consecutive weeks, including both daytime and nighttime testing periods. During the testing period, approximately 12.09 million sets of valid windowed DAS vibration data, approximately 604,000 sets of DTS temperature data, and approximately 1.812 million sets of FBG structural response data were generated. On-site manual records and system logs served as the verification basis. Test behaviors included 312 normal walking maneuvers, 126 stationary stops, 32 simulated falls, 44 instances of multiple people gathering, 26 simulated abnormal intrusions, 18 simulated kitchen temperature anomalies, and 21 simulated minor structural vibrations. To ensure consistency in the comparison results, the comparison systems were evaluated using the same testing period, the same triggering events, and the same manually annotated records. The single DAS solution used only vibration data for behavior recognition, the video surveillance solution used ordinary fixed cameras for image recognition, and the traditional alarm solution used a combination of smoke detectors, manual inspections, and ordinary door magnetic alarms. The test results are shown in Table 1.

[0039] Table 1. Comparative Test Results of Field Application of Courtyard Personnel Behavior Recognition Methods As shown in Table 1, the method of the present invention outperforms both the single DAS vibration recognition scheme and the video surveillance recognition scheme in terms of overall behavior recognition accuracy, fall behavior recognition accuracy, abnormal intrusion recognition accuracy, early warning response time, and control of false alarms and missed alarms. Specifically, the overall behavior recognition accuracy of the present invention reaches 95.4%, which is 10.7 percentage points higher than the single DAS vibration recognition scheme and 15.6 percentage points higher than the video surveillance recognition scheme. This indicates that relying solely on a single vibration signal or image information is insufficient to comprehensively reflect changes in human behavior in a courtyard setting. The present invention, by integrating DAS vibration data, DTS temperature data, and FBG structural response data, can describe the courtyard state from multiple dimensions such as vibration, temperature, strain, and displacement, thereby improving the overall reliability of behavior recognition.

[0040] In fall detection, this invention achieves an accuracy rate of 96.9%, significantly higher than the 84.4% of a single DAS vibration detection scheme and the 81.3% of a video surveillance detection scheme. This is because falls not only cause sudden changes in local vibration but may also be accompanied by short-term structural response changes. This invention can model behavioral changes within a continuous time window through multi-parameter fusion of feature vectors and an improved Res-GRU temporal behavior detection model, thereby more accurately distinguishing falls from interference situations such as normal walking, brief pauses, or object collisions. In abnormal intrusion detection, this invention achieves an accuracy rate of 96.2%, indicating that even under conditions where video surveillance effectiveness is limited, such as at night, in obstructed lighting, or with insufficient illumination, it can still achieve stable identification based on fiber optic vibration response and spatial location mapping results.

[0041] Furthermore, this invention supports auxiliary judgment of temperature and structural anomalies, a function not found in single DAS vibration recognition schemes and video surveillance recognition schemes. This indicates that this invention can not only identify personnel behavior but also provide auxiliary analysis of kitchen temperature anomalies, gas-related risks, and courtyard structural response anomalies, achieving integrated monitoring of personnel safety, environmental safety, and structural safety. Regarding the average warning response time, this invention is 1.86s, shorter than the 3.42s of the single DAS vibration recognition scheme and the 4.75s of the video surveillance recognition scheme, demonstrating that this invention maintains good real-time response capabilities while completing multi-source fusion judgment. The number of false alarms and missed alarms is reduced to 5 and 3 respectively, indicating that the consistency constraint mechanism and multi-source fusion judgment rules can reduce misjudgments caused by construction vibrations, short-term personnel passage, changes in lighting, or local environmental disturbances. This invention is better suited to the practical application needs of courtyard scenarios with complex spatial structures, diverse personnel behaviors, and numerous environmental interferences, exhibiting high recognition accuracy, timely warning, and application stability.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for human behavior recognition based on fiber optic sensor data, characterized in that, include: S1. Synchronously acquire distributed temperature sensor (DTS) data, distributed acoustic sensor (DAS) data, and fiber optic grating (FBG) structure response data, and perform time synchronization to obtain multi-source fiber optic sensing raw data. S2. Based on the DTS / DAS composite optical cable deployment path and FBG sensor deployment location, the original multi-source fiber optic sensing data is spatially mapped, and after differentiating the sensing data of different areas by combining the courtyard functional zoning, the data is spatiotemporally aligned to generate multi-source fiber optic sensing fusion data. S3. Using the multi-source fiber optic sensing fusion data as the processing object, short-time Fourier transform (STFT) and wavelet transform (WT) are used to extract the vibration features of DAS, and combined with the temperature change features of DTS and the strain and displacement change features of FBG, a multi-parameter fusion feature vector is formed. S4. Input the multi-parameter fused feature vector into the feature representation model, and construct a consistency constraint mechanism based on the DAS vibration propagation characteristics and FBG structural response characteristics, perform constraint optimization, and obtain the semantic representation of personnel behavior. S5. Construct a temporal behavior recognition model based on the improved Res-GRU, perform temporal modeling and analysis on the semantic representation of the personnel behavior, and output the personnel behavior category and confidence parameters. S6. Based on the personnel behavior category and confidence parameters, combined with DTS temperature anomaly information and FBG structural anomaly information, perform multi-source fusion judgment, and output early warning information based on hierarchical alarm rules.

2. The method for human behavior recognition based on fiber optic sensor data as described in claim 1, characterized in that, Specifically, S1 is: Distributed acoustic sensing DAS composite fiber, distributed temperature sensing DTS fiber, and fiber Bragg grating (FBG) sensor array are deployed along the building structure path within the pre-defined area of ​​the courtyard. The DAS composite fiber is continuously laid along the perimeter wall, indoor passage and main path of the courtyard, the DTS fiber is continuously laid along the roof ridge, eaves, ring beam and temperature-sensitive area of ​​the kitchen, and the FBG sensor is set at the load-bearing column node, the mid-span of the main beam and key structural connection node. The Rayleigh scattering signal is continuously sampled by the DAS sensing unit based on the coherent optical time-domain reflectometry (C-OTDR) mechanism to obtain the vibration time-series signal at the corresponding spatial location. The echo signal is demodulated using the DTS sensing unit based on the Raman scattering effect to obtain the spatiotemporal temperature sequence data distributed along the optical fiber. The FBG sensing unit is used to detect the change in the wavelength of the grating reflection in real time, and the wavelength drift is converted into structural strain and displacement response data. The DTS / DAS demodulator and FBG demodulator are used to perform unified time reference calibration on various types of sensor data, and generate multi-source fiber optic sensor raw data with unified timestamps.

3. The method for human behavior recognition based on fiber optic sensor data as described in claim 2, characterized in that, The specific steps for performing unified time reference calibration on various types of sensor data using a DTS / DAS demodulator and an FBG demodulator are as follows: The original timestamp sequences output by the DTS sensing unit, DAS sensing unit and FBG sensing unit are obtained respectively, and a unified time reference time axis is constructed based on the internal clock of the DTS / DAS demodulator as the master reference clock. The time drift deviation sequence of FBG data relative to DAS / DTS data is obtained by calculating the point-by-point difference between the timestamp sequence output by the FBG demodulator and the time axis of the unified time reference. A time correction function is constructed based on the time drift deviation sequence, and the FBG structure response data is processed by segmented time axis mapping to align the FBG data in different time periods to the corresponding DAS / DTS time intervals. The DAS vibration time series data, DTS temperature time series data and FBG structural response data are divided into sliding time windows, and the window length and step size of the sliding window are uniformly set based on the preset sampling period. Within each sliding time window, time resampling processing is performed on the three types of sensor data to convert non-equal interval sampled data into an equal time interval sequence; The resampled DAS, DTS and FBG data are reconstructed and sorted in chronological order, and cross-sensor channel alignment mapping is performed on the time axis to generate multi-source fiber optic sensing raw data with a unified timestamp. In the process of time alignment, abnormal timestamp data is removed or interpolated and corrected, and the corrected timestamp sequence is arranged continuously.

4. The method for human behavior recognition based on fiber optic sensor data as described in claim 1, characterized in that, Specifically, S2 is: Based on the actual deployment path of the DTS / DAS composite optical cable in the courtyard environment and the installation position of the FBG sensor at the key nodes of the structure, the collected DAS vibration data, DTS temperature data and FBG strain and displacement data are processed for spatial coordinate calibration. The continuous sampling position along the optical fiber is mapped to the spatial geometric position of the courtyard building structure, so that each segment of optical fiber sensing data corresponds to a specific courtyard space area. After completing the spatial calibration, the courtyard space is divided according to its functional attributes, including the perimeter protection area, indoor living area, personnel passage area and structural safety monitoring area. Based on the role of different areas in the monitoring task, corresponding initial weight coefficients are set for the three types of sensor data: DAS, DTS and FBG. The weight coefficients are preset according to the matching relationship between the sensor type and the spatial area. The weighted raw data of multi-source fiber optic sensing is processed by time slicing according to a unified time reference. DAS, DTS and FBG data are synchronously segmented using a fixed time window or a sliding time window. Data with different sampling frequencies are aligned within the same time segment. During time alignment, DAS vibration data is resampled according to the spatial position of the optical fiber to arrange the vibration signal in a spatially continuous manner. DTS temperature data is interpolated and reconstructed according to the distribution position of the optical fiber to form a continuous spatial sequence. FBG strain and displacement data are aligned in sequence according to the structural node number to make the discrete structural response data correspond one-to-one with the spatial position. Various types of sensor data are fused and stitched together according to both spatial location and time dimensions, so that DAS, DTS and FBG data in the same spatial area within the same time window are jointly organized and a unified data representation is generated. The final output is multi-source fiber optic sensor fusion data consistent with the courtyard spatial structure. This fusion data includes spatial location index, time window index, and corresponding DAS vibration data, DTS temperature data, and FBG structural response data sequences.

5. The method for human behavior recognition based on fiber optic sensor data as described in claim 4, characterized in that, The specific steps of fusing and stitching together various types of sensor data according to both spatial location and time dimensions are as follows: After spatial coordinate calibration and time alignment, the DAS vibration data, DTS temperature data, and FBG structural response data are classified and organized according to the spatial location index to form a multi-type sensor data set corresponding to each spatial location. Within the multi-type sensor data set, different types of sensor data are sorted by time dimension according to a unified time window, so that DAS vibration data, DTS temperature data and FBG structural response data are arranged in corresponding time segments. A two-dimensional index structure is constructed based on spatial location index and time window index, and multiple types of sensor data are combined and mapped based on the two-dimensional index structure, so that each spatial location corresponds to each time window to form a data combination unit. The data combination units are arranged sequentially according to the spatial location dimension and serialized and spliced ​​in the time dimension to form multi-source fiber optic sensing fusion data. The output includes multi-source fiber optic sensor fusion data containing spatial location index, time window index, and corresponding DAS vibration data, DTS temperature data, and FBG structural response data.

6. The method for human behavior recognition based on fiber optic sensor data as described in claim 1, characterized in that, Specifically, S3 is: The DAS vibration signal in the multi-source fiber optic sensor fusion data after spatiotemporal alignment is segmented according to a preset time window, and the vibration data in each time window is processed by short-time Fourier transform (STFT) to calculate the time-frequency distribution matrix of the vibration signal in different time periods and extract the spectral amplitude features in the corresponding time window. Based on the STFT processing, the DAS vibration signal is further processed by continuous wavelet transform (WT). The vibration signal is decomposed into multiple scales according to different scale factors, the wavelet coefficients at each scale are calculated, and the transient change components and local abrupt change components of the vibration signal are extracted. The DTS temperature data is serialized according to spatial sampling points, and the temperature difference change sequence between adjacent sampling points is calculated. At the same time, the first-order difference is calculated on the temperature sequence to form a temperature change gradient feature sequence. The FBG strain data and displacement data are processed synchronously to calculate the strain change and displacement change of each structural monitoring node in adjacent time windows. The structural response data are then arranged in a time series to form a structural response change characteristic sequence. The DAS vibration time-frequency features obtained by STFT processing, the DAS wavelet coefficient features obtained by WT transform, the DTS temperature change features, and the FBG strain and displacement change features are spliced ​​together by feature dimensions, and the spliced ​​high-dimensional features are uniformly normalized to construct a multi-parameter fusion feature vector.

7. The method for human behavior recognition based on fiber optic sensor data as described in claim 1, characterized in that, Specifically, S4 is: The multi-parameter fused feature vector is input into a preset feature representation model. The input feature vector is serialized and organized according to the time dimension, arranged into a time-series feature sequence according to a unified time window order, and an initial behavioral feature representation is established based on the sequence structure. In the feature representation process, based on the propagation attenuation characteristics of DAS vibration signals in optical fiber medium, the vibration characteristics corresponding to different spatial locations are constrained and modeled. Among them, the spatial attenuation constraint term is constructed according to the correspondence between optical fiber distance and vibration signal intensity, and the feature representation update process is introduced. Meanwhile, based on the physical consistency relationship between FBG sensors in structural strain and displacement response, the structural response characteristics are constrained and modeled. The synchronous change relationship between strain change and displacement change is embedded as a constraint condition in the feature optimization process, and the consistency correction of the structural response characteristics is performed. Under the constraints, the multi-parameter fusion feature vector is iteratively optimized. The input features are nonlinearly transformed by the feature mapping function, so that different types of sensing features are reconstructed and fused in a unified feature space to form an intermediate feature representation with structural constraints. Temporal consistency verification is performed on the intermediate feature representations after constraint optimization, the feature change trend is smoothed according to continuous time windows, and feature jumps between time windows are eliminated. The final output is a semantic representation of human behavior, which is a unified representation that integrates DAS vibration features, DTS temperature features, and FBG structural response features.

8. The method for human behavior recognition based on fiber optic sensor data as described in claim 1, characterized in that, Specifically, S5 is: The semantic representation of the personnel behavior is arranged in chronological order to construct continuous time series input data, and the time series input data is subjected to sliding segmentation based on a preset time window; The segmented semantic feature sequence is input into the improved Res-GRU temporal behavior recognition model for temporal modeling processing to generate a global temporal behavior feature vector. Based on the global temporal behavior feature vector, classification mapping is performed. The temporal features are mapped to the personnel behavior category space through a preset behavior category mapping relationship, and the corresponding personnel behavior category and confidence parameters are output.

9. A method for human behavior recognition based on fiber optic sensor data as described in claim 8, characterized in that, The step of inputting the segmented semantic feature sequence into the improved Res-GRU temporal behavior recognition model for temporal modeling processing to generate a global temporal behavior feature vector is as follows: The segmented semantic feature sequence is input into the improved Res-GRU temporal behavior recognition model for temporal modeling processing to generate a global temporal behavior feature vector. The improved Res-GRU temporal behavior recognition model includes an input processing module, a time-dependent recursive modeling module, a state update control module, a temporal enhancement propagation module, and a global feature aggregation module. The input processing module receives the segmented semantic feature sequence and arranges the semantic features corresponding to each time window in chronological order to form a standard time-series input vector sequence, while maintaining the semantic continuity index relationship between time windows; The time-dependent recursive modeling module recursively models the dependencies of the input vector sequence in the time dimension. At each time step, it jointly calculates the input vector of the current time step and the hidden state of the previous time step to generate the candidate hidden state of the current time step. The state update control module is based on the reset gate and update gate mechanism, introduces residual information path, fuses and updates candidate hidden states and historical hidden states, and constructs an adaptive residual adjustment factor based on the semantic feature sequence. The residual fusion weight is dynamically adjusted according to the semantic state change intensity of different time windows to generate the hidden state of the current time step. The step of constructing an adaptive residual adjustment factor based on the semantic feature sequence includes: extracting the change intensity of semantic features in the current time window; calculating the Euclidean distance between the semantic vector of the current time window and the semantic vectors of adjacent time windows; quantifying the intensity of temporal fluctuations in the current window; and inputting the quantized change intensity into a single-layer fully connected mapping layer and outputting it after passing through a Sigmoid activation function. , which serves as the adaptive residual adjustment factor for the current time step; The hidden state at the current time step is represented as: ,in, Indicates an update to the door. This indicates the hidden state at the previous time step. Indicates the candidate hidden state at the current time step; The temporal enhancement propagation module performs cross-temporal enhancement fusion propagation of the hidden state at the current time step and the historical hidden state based on the residual connection mechanism. The temporal enhancement propagation module receives the adaptive residual adjustment factor corresponding to the current time step and uses it as the dynamic weight benchmark for cross-temporal residual fusion. The cross-temporal enhancement fusion strength of the historical hidden state and the current hidden state is synchronously linked with the dynamic weight benchmark. The global feature aggregation module performs global aggregation processing on the hidden state sequence output by all time steps based on a combination of terminal hidden state extraction and temporal dimension attention weighted aggregation to generate a global temporal behavior feature vector.

10. A method for human behavior recognition based on fiber optic sensor data as described in claim 1, characterized in that, Specifically, S6 is: The personnel behavior categories and confidence parameters are synchronously input with DTS temperature data and FBG structural response data, and the three types of data are time-aligned based on a unified timestamp. The current behavior state is initially determined based on the personnel behavior categories, and personnel behavior category identification information within the corresponding time window is extracted. The personnel behavior category identification information includes normal walking behavior, stationary behavior, falling behavior, gathering behavior, and abnormal intrusion behavior. The DTS temperature data is divided into time windows, and the temperature change, temperature gradient change and temperature change characteristics within each time window are calculated. Based on the preset temperature threshold, DTS temperature anomaly information is generated. The FBG structural response data is divided into time synchronization segments to calculate the strain and displacement changes of each structural monitoring node, and FBG structural anomaly information is generated based on the structural response threshold. Spatial location mapping and time series alignment are performed on personnel behavior category identification information, DTS temperature anomaly information and FBG structure anomaly information to construct a multi-source fusion judgment data structure. Based on the multi-source fusion judgment data structure, hierarchical rule matching processing is performed to verify the rules for different categories of personnel behavior. Among them, fall behavior and abnormal intrusion behavior are given priority for high-level risk judgment, and cross-validation is performed by combining DTS temperature anomaly information and FBG structure anomaly information. The risk level output is generated based on the combination relationship between personnel behavior category, DTS temperature anomaly information and FBG structural anomaly information, and the first-level warning information, second-level warning information or third-level warning information is output according to the preset priority order. The warning information is sent to the courtyard safety monitoring terminal or remote management platform, and the time windows corresponding to fall behavior and abnormal intrusion behavior are marked and stored.

Images