Pipeline resource state early warning method and system based on intrinsic voiceprint analysis

Abstract

The present application belongs to the technical field of risk early warning, and discloses a pipeline gallery resource state early warning method and system based on intrinsic voiceprint analysis; including obtaining distributed voiceprint sensing signals and processing to obtain original audio signals; inputting the original audio signals into an audio-driven AI early warning model to determine the early warning event type and corresponding risk level of the current pipeline gallery resource state; constructing a voiceprint feature-geographic information-time stamp three-dimensional correlation analysis engine to output the target early warning event and optimized risk level; matching the corresponding early warning response strategy from an early warning response strategy library; the present application expands the spectral richness by combining the frequency multiplication processing technology, and iteratively optimizes the harmonic weight coefficient based on the weighted comprehensive index of the spectral energy ratio and the spectral complexity, finally fuses the enhanced frequency information with the amplitude and phase to obtain a high-fidelity original audio signal, effectively solving the problem of poor restoration quality caused by voiceprint signal attenuation and noise interference in complex urban environments.

Description

Technical Field

[0001] This invention relates to the field of risk warning technology, and more specifically, to a method and system for early warning of the status of utility tunnel resources based on intrinsic voiceprint analysis. Background Technology

[0002] As the core service carrier unit of urban power communication networks, urban power optical cables bear the heavy responsibility of transmitting critical service signals. However, this vital infrastructure is frequently damaged by external construction and cable fires, leading to serious consequences such as power communication network outages and service disruptions, posing a severe challenge to the safe and stable operation of the power grid and reliable communication services. Due to the rapid urbanization leading to a surge in the number of optical cables and the increasing complexity of their laying environments, traditional manual inspections are unable to achieve 100% coverage. In areas with frequent construction, the probability of damage to utility tunnels increases dramatically, resulting in rising labor costs and declining maintenance efficiency. Furthermore, the narrow and enclosed environment of utility tunnels limits the deployment of active sensors, leading to problems such as power supply difficulties, communication interruptions, and susceptibility to moisture damage. Faults can easily trigger a chain of public safety accidents. Moreover, in terms of intelligent operation and maintenance decision-making, existing solutions are mostly based on single-dimensional alarm analysis, lacking multi-dimensional correlation capabilities using voiceprint features, geographic information, and timestamps. This makes it impossible to effectively capture persistent threats, and risk level assessments do not consider the topology and historical evolution of utility tunnels, resulting in crude response strategies and uneven allocation of operation and maintenance resources.

[0003] In view of this, the present invention proposes a method and system for early warning of the status of utility tunnel resources based on intrinsic voiceprint analysis to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: A method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis, the method comprising: The light source signal is acquired and amplified using a two-stage beam splitting and amplification architecture to obtain the amplified light source signal. The amplified light source signal is then split into two optical paths by a beam splitter. Each optical path is independently configured with an acousto-optic modulator and a photodetector. A parallel monitoring channel is constructed using an optical switch to collect distributed acoustic signature sensing signals at the corresponding monitoring positions of each monitoring channel in real time. Phase demodulation and polarization state demodulation are performed on the distributed voiceprint sensing signal to extract frequency information, amplitude information and phase information. The frequency information is processed to obtain enhanced frequency information. The enhanced frequency information, amplitude information and phase information are fused to obtain the original audio signal. An audio-driven AI early warning model is constructed. The original audio signal is input into the audio-driven AI early warning model, and the original multi-dimensional acoustic features are extracted from the original audio signal. The original multi-dimensional acoustic features are matched with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to determine the type of early warning event and the corresponding risk level of the current utility tunnel resource status. Acquire the voiceprint features, alarm point geographic information coordinates and timestamps corresponding to the original audio signal, construct a three-dimensional sound correlation analysis engine, and output target early warning events and optimize risk levels based on the three-dimensional correlation analysis engine; Based on the target warning event and the optimized risk level, the system matches the corresponding warning response strategy from the preset warning response strategy library and generates an alarm work order to push to the operation and maintenance end.

[0005] Furthermore, the method of splitting the amplified light source signal into two optical paths using a beam splitter is as follows: The source signal of a C-band narrow linewidth laser is acquired, and the optical power stability of the source signal is detected. When the optical power fluctuation value of the source signal is detected to be greater than the preset fluctuation threshold, automatic power control compensation is initiated to compensate the optical power of the source signal to obtain a stable source signal. A stable light source signal is input into an erbium-doped fiber amplifier for pre-amplification. The pump current of the erbium-doped fiber amplifier is adjusted to increase the output optical power of the erbium-doped fiber amplifier to a preset first target power, thereby obtaining the pre-amplified signal output by the erbium-doped fiber amplifier. The pre-amplified signal is input into a Raman amplifier for secondary gain. By adjusting the Raman pump power of the Raman amplifier, the optical power of the pre-amplified signal reaches the preset second target power, thus obtaining a secondary gain light source signal. The secondary gain light source signal is input into a 50 / 50 beam splitter for power distribution, and the optical power of the two outputs of the 50 / 50 beam splitter is monitored in real time. The optical power difference between the two outputs is calculated. When the optical power difference is greater than the preset equalization threshold, dynamic fine-tuning compensation of the splitting ratio is initiated until the optical power difference is less than or equal to the preset equalization threshold. The secondary gain light source signal is then split into two independent optical paths with equal optical power, which are called dual optical paths.

[0006] Furthermore, the method for obtaining enhanced frequency information is as follows: Extract fundamental frequency information from frequency information and calculate the spectral concentration of the frequency information; determine the target harmonic order set based on the spectral concentration. Based on the target harmonic order set, the fundamental frequency information is converted into the corresponding harmonic frequencies. The fundamental frequency information and each harmonic frequency are superimposed according to their corresponding amplitude and phase to obtain the frequency-harmonic synthesis signal. Based on the target harmonic order set, the non-target frequency components are determined. The frequency-multiplied synthesized signal is input into a preset stable LCL topology filter, and the output signal of the stable LCL topology filter is used as the filtered synthesized signal. The spectral energy ratio and spectral complexity of the synthesized signal after filtering are calculated. The spectral energy ratio and spectral complexity are weighted to obtain a weighted comprehensive index. The weighted comprehensive index is used as the optimization target. The basic frequency information and the preset basic weight coefficients of each harmonic frequency are iteratively optimized to obtain the optimized weight coefficients. The filtered synthesized signal is decomposed into fundamental frequency components and harmonic components, and the original amplitude and phase of each harmonic component are obtained. The optimized weighting coefficients are multiplied by the original amplitude of the corresponding harmonic components to obtain the adjusted amplitude. The harmonic components are reconstructed based on the adjusted amplitude and the corresponding original phase, and then linearly superimposed to obtain the enhanced frequency information.

[0007] Furthermore, the method for extracting fundamental frequency information from frequency information is as follows: Frequency information is subjected to spectral peak detection, multiple peak frequencies and their corresponding energy amplitudes are extracted, and the peak frequencies are sorted in descending order of energy amplitude to obtain a candidate frequency sequence. Starting from the beginning of the candidate frequency sequence, select one as the candidate fundamental frequency in sorted order, detect whether there are harmonic frequency components in the frequency information that are integer multiples of the candidate fundamental frequency, and count the number of harmonic frequency components. When the number of harmonic frequency components is greater than or equal to the preset harmonic number threshold, the currently selected candidate fundamental frequency is determined as the preliminary fundamental frequency. When the number of harmonic frequency components is less than the preset harmonic number threshold, the next candidate fundamental frequency is selected from the candidate frequency sequence in sorted order, and the detection is repeated until the preliminary fundamental frequency is determined or the candidate frequency sequence is traversed. Obtain the frequency information of the monitoring location corresponding to the preliminary base frequency and its adjacent monitoring locations, and verify the frequency consistency of the preliminary base frequency at adjacent monitoring locations; when the frequency consistency is satisfied, determine the preliminary base frequency as the base frequency information of the original vibration signal.

[0008] Furthermore, the method for obtaining the original audio signal is as follows: Extract the fundamental amplitude value corresponding to the fundamental frequency component and the harmonic amplitude value corresponding to each harmonic component from the amplitude information; extract the fundamental phase value corresponding to the fundamental frequency component and the harmonic phase value corresponding to each harmonic enhancement component from the phase information. The fundamental frequency component, fundamental amplitude value, and fundamental phase value are combined into a sine wave to obtain the fundamental frequency time-domain waveform; each harmonic component, its corresponding harmonic amplitude value, and its corresponding harmonic phase value are combined into a sine wave to obtain the harmonic time-domain waveform. The fundamental frequency time-domain waveform is superimposed with the time-domain waveforms of each harmonic to obtain a preliminary synthesized audio signal. The maximum and minimum amplitudes of the preliminary synthesized audio signal are obtained. Based on the maximum and minimum amplitudes and the preset standard audio amplitude range, the amplitude linear mapping parameter is calculated. Based on the amplitude linear mapping parameter, the amplitude of the preliminary synthesized audio signal is mapped to the standard audio amplitude range to obtain the original audio signal.

[0009] Furthermore, the method for constructing an audio-driven AI early warning model is as follows: A three-layer event recognition architecture is pre-built, which includes a physical layer, an environment layer, and an interaction layer. Historical and real-time acoustic fingerprint sample data of the utility tunnel are collected. Based on the predefined event types of each layer in the three-layer event recognition architecture, the historical and real-time acoustic fingerprint sample data are labeled with event types to obtain the event type labels corresponding to each sample data. Multi-dimensional acoustic features are extracted from the labeled historical and real-time acoustic fingerprint sample data. The multi-dimensional acoustic features are associated with the corresponding event type labels and stored to construct a multi-dimensional event fingerprint database. A deep learning recognition model is built based on a three-layer event recognition architecture. The deep learning recognition model is a CNN-BiGRU hybrid sub-model. The deep learning recognition model is pruned and quantized to obtain a compressed model. The compressed model, the multi-dimensional event fingerprint database and the three-layer event recognition architecture are integrated to obtain an audio-driven AI early warning model.

[0010] Furthermore, the method for determining the type of early warning event and the corresponding risk level of the current utility tunnel resource status is as follows: Original multidimensional acoustic features are extracted from the original audio signal. These original multidimensional acoustic features are then input into the CNN sub-model in the audio-driven AI early warning model for spatiotemporal feature fusion to obtain spatiotemporal fused features. The spatiotemporal fusion features are input into the BiGRU sub-model in the audio-driven AI early warning model to classify events and output the probability distribution of each type of event. The probability distributions of various events are matched and verified with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to obtain the verified event probability distributions. Obtain the current environmental noise and historical false alarm rate, and determine the probability judgment threshold; compare the probability of each event in the verified event probability distribution with the probability judgment threshold, and when the event probability is greater than the probability judgment threshold, determine the corresponding event type as the warning event type; Calculate the probability deviation between the event probability and the probability judgment threshold, and determine the risk level corresponding to the warning event type based on the probability deviation.

[0011] Furthermore, the method for constructing a three-dimensional correlation analysis engine is as follows: The voiceprint features are vectorized to extract event harmonic energy distribution parameters and on-site audio reconstruction parameters, and a voiceprint feature vector is constructed. The geographic coordinates of the alarm points are parsed to extract the GIS coordinates and utility tunnel topology data contained in the geographic coordinates of the alarm points. Spatial index encoding is performed based on the GIS coordinates and utility tunnel topology data to generate spatial grid identifiers and construct a geospatial index. The timestamps are segmented into time windows to generate time segment identifiers and establish a time series index. Alarm event identifiers are generated based on the type and timestamp of the warning event. Voiceprint feature vectors, geospatial indexes, and time series indexes are associated and bound through the alarm event identifiers to obtain a three-dimensional associated dataset. A structured three-dimensional index is constructed based on three dimensions: voiceprint feature vector, geospatial index, and time series index. A bidirectional mapping relationship between dimensions is established based on the structured three-dimensional index to obtain a dimension association mapping table. A cross-dimensional retrieval interface is constructed based on the dimensional association mapping table to obtain the 3D retrieval interface; inter-dimensional association analysis rules are defined based on the dimensional association mapping table to obtain the inter-dimensional association rule set. By integrating the 3D retrieval interface with the set of rules for inter-dimensional association, a 3D association analysis engine is obtained.

[0012] Furthermore, the method for outputting target early warning events and optimizing risk levels based on the 3D correlation analysis engine is as follows: A sliding event window is established based on a 3D correlation analysis engine to perform spatiotemporal consistency verification on the types of early warning events and their corresponding risk levels: the geospatial index sequence within the time window corresponding to the time series index is obtained through spatiotemporal correlation rules, and it is determined whether the geospatial index sequence meets the preset spatial continuity conditions; when the spatial continuity conditions are met, the early warning event type and its corresponding risk level are determined as the event type and risk level after spatiotemporal verification. A two-stage analysis mechanism of initial identification and deep correlation is implemented for event types after spatiotemporal verification: In the initial identification stage, voiceprint-event matching verification is performed based on voiceprint feature vectors and a multi-dimensional event fingerprint database to obtain the initial event type; in the deep correlation stage, geospatial index and time series index are called based on the dimensional correlation mapping table to obtain the associated event records of the initial event type within a preset historical time window; when the number of associated event records reaches a preset correlation threshold, the initial event type is determined as the target warning event; The risk correction coefficient is calculated based on the target early warning event. The risk level after spatiotemporal verification is then weighted and corrected according to the risk correction coefficient to obtain the optimized risk level.

[0013] A utility tunnel resource status early warning system based on intrinsic voiceprint analysis, the system comprising: The sensor signal acquisition module is used to acquire the light source signal. It adopts a two-stage beam splitting and amplification architecture to amplify the light source signal and obtain the amplified light source signal. The amplified light source signal is then split into two optical paths by a beam splitter. Each optical path is independently configured with an acousto-optic modulator and a photodetector. A parallel monitoring channel is constructed using an optical switch to acquire the distributed acoustic signature sensing signal of the corresponding monitoring position of each monitoring channel in real time. The voiceprint restoration and enhancement module is used to perform phase demodulation and polarization state demodulation on the distributed voiceprint sensing signal, extract frequency information, amplitude information and phase information, process the frequency information to obtain enhanced frequency information, and fuse the enhanced frequency information, amplitude information and phase information to obtain the original audio signal; The intelligent identification and early warning module is used to build an audio-driven AI early warning model. The original audio signal is input into the audio-driven AI early warning model, and the original multi-dimensional acoustic features are extracted from the original audio signal. The original multi-dimensional acoustic features are matched with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to determine the type of early warning event and the corresponding risk level of the current utility tunnel resource status. The 3D correlation analysis module is used to obtain the voiceprint features, alarm point geographic information coordinates and timestamps corresponding to the original audio signal, construct a 3D correlation analysis engine of voiceprint features-geographic information-timestamp, and output target early warning events and optimize risk levels based on the 3D correlation analysis engine; The intelligent decision response module is used to match the corresponding early warning response strategy from the preset early warning response strategy library based on the target early warning event and the optimized risk level, and generate an alarm work order to push to the operation and maintenance terminal.

[0014] The technical effects and advantages of the utility tunnel resource status early warning method and system based on intrinsic voiceprint analysis of this invention are as follows: This invention employs a two-stage beam splitting amplification architecture combining EDFA and Raman amplifier, achieving high-gain, low-noise amplification of the light source signal. This effectively overcomes long-distance transmission link losses. Combined with a 50 / 50 beam splitter and dynamic fine-tuning compensation, it ensures strict power balance between the two optical paths. Furthermore, it utilizes two 1×4 MEMS optical switches to construct an 8-channel parallel monitoring network. With a 30-second polling plus 10-second flexible delay strategy, it effectively captures the complete waveform of sudden events while ensuring full coverage of tens of kilometers of cable. This breaks through the bottleneck of single-source multiplexing and provides a high-quality, highly reliable raw sensor data foundation for subsequent acoustic signature analysis.

[0015] This invention extracts the frequency, amplitude, and phase of vibration signals through phase demodulation and polarization state demodulation, expands the spectral richness by combining frequency doubling technology, and iteratively optimizes the harmonic weight coefficients based on a weighted comprehensive index of spectral energy ratio and spectral complexity. Finally, the enhanced frequency information is fused with the amplitude and phase and subjected to dynamic range normalization to obtain a high-fidelity original audio signal. This effectively solves the problem of poor restoration quality caused by attenuation of soundprint signals and noise interference in complex urban environments, realizes passive multiplexing of existing communication optical cables, and effectively solves the problem of difficult deployment of active sensors in confined spaces.

[0016] This invention constructs a three-dimensional correlation analysis engine based on voiceprint features, geographic information, and timestamps, achieving deep correlation mining across three dimensions: voiceprint features, spatial location, and temporal evolution. This overcomes the limitations of traditional single-dimensional alarm analysis and provides multi-dimensional data support for refined operation and maintenance decisions. Simultaneously, it effectively captures persistent threats through sliding event windows and spatiotemporal consistency verification, avoiding misjudgments of single-point instantaneous alarms and improving alarm accuracy. Furthermore, it employs a two-level analysis mechanism of initial judgment and deep correlation, introducing historical correlation verification on the basis of rapid initial judgment. This ensures both the real-time nature of event response and enhances the credibility of event judgment through historical data verification. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the module structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In one embodiment, a method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis is disclosed, such as... Figure 1 As shown, the early warning methods include: The light source signal is acquired and amplified using a two-stage beam splitting amplification architecture to obtain the amplified light source signal. The amplified light source signal is then split into two optical paths by a beam splitter. Each optical path is independently configured with an acousto-optic modulator and a photodetector. A parallel monitoring channel is constructed using an optical switch to collect distributed acoustic signature sensing signals at the corresponding monitoring positions of each monitoring channel in real time.

[0020] The method of splitting the amplified light source signal into two optical paths by using a beam splitter is as follows: acquire the light source signal of C-band narrow linewidth laser, perform optical power stability detection on the light source signal, and when the detected optical power fluctuation value of the light source signal is greater than the preset fluctuation threshold, start automatic power control compensation to compensate the optical power of the light source signal to obtain a stable light source signal; A stable light source signal is input into an erbium-doped fiber amplifier for pre-amplification. The pump current of the erbium-doped fiber amplifier is adjusted to increase the output optical power of the erbium-doped fiber amplifier to a preset first target power, thereby obtaining the pre-amplified signal output by the erbium-doped fiber amplifier. The pre-amplified signal is input into a Raman amplifier for secondary gain. By adjusting the Raman pump power of the Raman amplifier, the optical power of the pre-amplified signal reaches the preset second target power, thus obtaining a secondary gain light source signal. The secondary gain light source signal is input into a 50 / 50 beam splitter for power distribution, and the optical power of the two outputs of the 50 / 50 beam splitter is monitored in real time. The optical power difference between the two outputs is calculated. When the optical power difference is greater than the preset equalization threshold, dynamic fine-tuning compensation of the splitting ratio is initiated until the optical power difference is less than or equal to the preset equalization threshold. The secondary gain light source signal is then split into two independent optical paths with equal optical power, which are called dual optical paths.

[0021] Specifically, the system acquires the light source signal of a C-band narrow-linewidth laser, performs optical power stability detection on the light source signal, and when the detected optical power fluctuation value of the light source signal exceeds a preset fluctuation threshold, it initiates automatic power control compensation to compensate for the optical power of the light source signal, thereby obtaining a stable light source signal. The specific implementation method of optical power stability detection is as follows: the light source signal is continuously sampled, and the optical power variance within the sampling period is calculated. When the optical power variance exceeds the preset fluctuation threshold, it is determined that the optical power stability does not meet the requirements. The preset fluctuation threshold is determined comprehensively based on the factory specifications of the light source device and the system signal-to-noise ratio requirements, for example, it can be set to ±5% of the average optical power. Automatic power control compensation is achieved by feedback adjustment of the light source drive current or an external adjustable optical attenuator, so that the optical power fluctuation value (variance) is restored to below the preset fluctuation threshold, thereby obtaining a stable light source signal.

[0022] A stable light source signal is pre-amplified by an erbium-doped fiber amplifier (EDFA). The pump current of the EDFA is adjusted to increase its output optical power to a preset first target power, resulting in the pre-amplified signal output by the EDFA. The preset first target power is determined as follows: based on the input optical power of the stable light source signal and the initial link loss parameters, the input optical power required for the subsequent Raman amplifier to reach its optimal operating point is calculated, and this required input optical power is defined as the preset first target power. Specifically, the initial link loss parameters include fiber connector loss, fusion splice loss, and intrinsic loss of the transmission fiber, which can be pre-measured using an optical time domain reflectometer (OTDR). The total link loss is obtained by adding the fiber connector loss, fusion splice loss, and intrinsic loss of the transmission fiber. Then, the threshold input power of the Raman amplifier is obtained, which is the threshold input power required to achieve optimal Raman output. The minimum input optical power required for the amplifier to effectively excite stimulated Raman scattering is determined by the Raman amplifier device specifications. An empirically set system margin is added to this threshold input power to obtain the optimal input optical power of the Raman amplifier. This system margin is used to compensate for power fluctuations caused by environmental variations and device aging. Finally, the optimal input optical power of the Raman amplifier is added to the total link loss to obtain the preset first target power. This preset first target power is the optical power level that the erbium-doped fiber amplifier output needs to reach, ensuring that the optical power reaching the Raman amplifier input is exactly at the optimal operating point after link transmission loss. The specific implementation of adjusting the pump current of the erbium-doped fiber amplifier is as follows: monitoring the optical power at the EDFA output, and using a PID closed-loop control algorithm to adjust the driving current of the 980nm or 1480nm pump laser to stabilize the output optical power at the preset first target power.

[0023] The pre-amplified signal is input into a Raman amplifier for secondary gain. By adjusting the Raman pump power of the Raman amplifier, the optical power of the pre-amplified signal reaches the preset second target power, while suppressing amplified spontaneous emission (ASE) noise in the pre-amplified light source signal, thus obtaining the secondary gain light source signal. The Raman amplifier utilizes stimulated Raman scattering to excite optical gain in the transmission fiber itself through a high-power pump laser (usually in the 1450nm band). The preset second target power is determined according to the system's own detection sensitivity requirements and must satisfy that the system signal-to-noise ratio is greater than or equal to a preset signal-to-noise ratio threshold (determined independently based on the environmental noise conditions at the pipe gallery site). The specific method for suppressing ASE noise is to configure a bandpass filter at the output of the Raman amplifier to filter out the ASE noise spectrum near the pump wavelength, thereby obtaining the secondary gain light source signal.

[0024] The secondary gain light source signal is input to a 50 / 50 beam splitter for power distribution, and the optical power at the two outputs of the 50 / 50 beam splitter is monitored in real time. The difference in optical power between the two outputs is calculated. When the difference in optical power exceeds a preset equalization threshold, dynamic fine-tuning compensation of the splitting ratio is initiated, splitting the secondary gain light source signal into two independent optical paths with equal optical power, i.e., dual optical paths. The 50 / 50 beam splitter is either a fused biconical or planar waveguide type beam splitter, with an ideal splitting ratio of 50:50. The preset equalization threshold is determined based on the consistency requirements of the detection sensitivity of the dual optical paths, for example, it can be set to ±1dB. The specific implementation of dynamic fine-tuning compensation of the splitting ratio is as follows: when the difference in optical power between the two outputs is detected to be greater than the preset equalization threshold, the splitting ratio is restored to the range of 50:50 ± the preset equalization threshold by adjusting the temperature control parameters of the beam splitter itself or by using an electrically adjustable beam splitter. The equalization of optical power in the two independent optical paths ensures that the subsequent dual detection channels have the same sensitivity benchmark, avoiding missed detections or false alarms due to differences in optical power.

[0025] An acousto-optic modulator (AOM) and a high-sensitivity photodetector are configured at the output ends of the two independent optical paths, respectively. The acousto-optic modulator is used to perform high-speed pulse modulation on the optical carrier to generate a probe light pulse. The pulse width is determined according to the spatial resolution requirements. For example, a spatial resolution of 1m corresponds to a pulse width of approximately 10ns. The high-sensitivity photodetector adopts an avalanche photodiode (APD) or a photomultiplier tube (PMT). Its responsivity is matched with the optical power of the two independent optical paths to ensure the consistency of response of the dual detection channels. Two 1×4 MEMS optical switches are connected to the two independent optical paths respectively to construct dual parallel monitoring channels. Each channel is independently connected to the fiber optic link to be monitored. The insertion loss of the MEMS optical switches is strictly controlled within 0.8dB, and the switching time is less than 50ms to ensure the high efficiency of channel polling and the maximum utilization of optical power. The two 1×4 MEMS optical switches construct a total of 8 independent monitoring channels. Each channel can cover complex cable lines up to tens of kilometers long, realizing single-source-driven dual-optical-path synchronous detection and 8-channel cable line coverage.

[0026] Furthermore, the channel control software supports independent setting of sampling rate, power gain, and filtering parameters for each channel. The sampling rate is determined based on the Nyquist sampling theorem and the highest frequency of the target acoustic signature signal. For example, if the highest frequency of the target acoustic signature signal is 1kHz, then the sampling rate should be at least 2kHz. The power gain is independently adjusted according to the differences in fiber optic link loss for each channel to ensure consistent output signal strength across all channels. Filtering parameters include the cutoff frequency of the bandpass filter to suppress out-of-band noise. A 30-second polling cycle plus a 10-second flexible delay strategy is implemented, with a basic polling cycle of 30 seconds, allocating 30 seconds of effective monitoring time to each channel. The time window is measured; if the data processing is not completed when the channel is switched, an elastic delay mechanism is triggered to extend the processing time of the current channel by an additional 10 seconds before switching channels. The specific implementation of the elastic delay mechanism is as follows: at the time of channel switching, the data processing status of the current channel is detected. If the data processing task has been completed, a normal channel switch is performed; if the data processing task has not been completed, a delay is automatically triggered. During the delay, the data of the current channel continues to be collected and processed to ensure that the complete waveform and key information of the sudden event are captured without omission. After the delay ends, the channel switch is performed, and the distributed acoustic signature sensing signals of each channel at the corresponding monitoring location are recorded.

[0027] This approach employs a two-stage beam splitting amplification architecture combining EDFA and Raman amplifiers to achieve high-gain, low-noise amplification of the light source signal, effectively overcoming long-distance transmission link losses. Combined with a 50 / 50 beam splitter and dynamic fine-tuning compensation, it ensures strict power balance between the two optical paths. Furthermore, it utilizes two 1×4 MEMS optical switches to construct an 8-channel parallel monitoring network, coupled with a 30-second polling plus 10-second flexible delay strategy. This approach ensures full coverage of tens of kilometers of cable while effectively capturing the complete waveform of sudden events, breaking through the bottleneck of single-source multiplexing and providing a high-quality, highly reliable raw sensor data foundation for subsequent acoustic signature analysis.

[0028] Phase demodulation and polarization demodulation are performed on the distributed voiceprint sensing signal to extract frequency, amplitude and phase information. The frequency information is processed to obtain enhanced frequency information. The enhanced frequency information, amplitude information and phase information are fused to obtain the original audio signal.

[0029] The method for obtaining enhanced frequency information is as follows: extract fundamental frequency information from the frequency information and calculate the spectral concentration of the frequency information; determine the target harmonic order set based on the spectral concentration. Based on the target harmonic order set, the fundamental frequency information is converted into the corresponding harmonic frequencies. The fundamental frequency information and each harmonic frequency are superimposed according to their corresponding amplitude and phase to obtain the frequency-harmonic synthesis signal. Based on the target harmonic order set, the non-target frequency components are determined. The frequency-multiplied synthesized signal is input into a preset stable LCL topology filter, and the output signal of the stable LCL topology filter is used as the filtered synthesized signal. The spectral energy ratio and spectral complexity of the synthesized signal after filtering are calculated. The spectral energy ratio and spectral complexity are weighted to obtain a weighted comprehensive index. The weighted comprehensive index is used as the optimization target. The basic frequency information and the preset basic weight coefficients of each harmonic frequency are iteratively optimized to obtain the optimized weight coefficients. The filtered synthesized signal is decomposed into fundamental frequency components and harmonic components, and the original amplitude and phase of each harmonic component are obtained. The optimized weighting coefficients are multiplied by the original amplitude of the corresponding harmonic components to obtain the adjusted amplitude. The harmonic components are reconstructed based on the adjusted amplitude and the corresponding original phase, and then linearly superimposed to obtain the enhanced frequency information. The method for extracting basic frequency information from frequency information is as follows: perform spectrum peak detection on the frequency information, extract multiple peak frequencies and their corresponding energy amplitudes, sort the peak frequencies in descending order of energy amplitude, and obtain a candidate frequency sequence. Starting from the beginning of the candidate frequency sequence, select one as the candidate fundamental frequency in sorted order, detect whether there are harmonic frequency components in the frequency information that are integer multiples of the candidate fundamental frequency, and count the number of harmonic frequency components. When the number of harmonic frequency components is greater than or equal to the preset harmonic number threshold, the currently selected candidate fundamental frequency is determined as the preliminary fundamental frequency. When the number of harmonic frequency components is less than the preset harmonic number threshold, the next candidate fundamental frequency is selected from the candidate frequency sequence in sorted order, and the detection is repeated until the preliminary fundamental frequency is determined or the candidate frequency sequence is traversed. Obtain the frequency information of the monitoring location corresponding to the preliminary base frequency and its adjacent monitoring locations, and verify the frequency consistency of the preliminary base frequency at adjacent monitoring locations; when the frequency consistency is satisfied, determine the preliminary base frequency as the base frequency information of the original vibration signal; The method for obtaining the original audio signal is as follows: extract the fundamental amplitude value corresponding to the fundamental frequency component and the harmonic amplitude value corresponding to each harmonic component from the amplitude information; extract the fundamental phase value corresponding to the fundamental frequency component and the harmonic phase value corresponding to each harmonic enhancement component from the phase information. The fundamental frequency component, fundamental amplitude value, and fundamental phase value are combined into a sine wave to obtain the fundamental frequency time-domain waveform; each harmonic component, its corresponding harmonic amplitude value, and its corresponding harmonic phase value are combined into a sine wave to obtain the harmonic time-domain waveform. The fundamental frequency time-domain waveform is superimposed with the time-domain waveforms of each harmonic to obtain a preliminary synthesized audio signal. The maximum and minimum amplitudes of the preliminary synthesized audio signal are obtained. Based on the maximum and minimum amplitudes and the preset standard audio amplitude range, the amplitude linear mapping parameter is calculated. Based on the amplitude linear mapping parameter, the amplitude of the preliminary synthesized audio signal is mapped to the standard audio amplitude range to obtain the original audio signal.

[0030] Specifically, phase demodulation is performed on the distributed acoustic signature sensing signal using a Phase Generated Carrier (PGC) algorithm or a 3×3 coupler demodulation algorithm to extract phase change information caused by external vibrations. These external vibrations are mechanical vibrations acting on the optical fiber, and the phase change information is proportional to the displacement of the external vibration, reflecting the instantaneous phase characteristics of the vibration event. Polarization state demodulation is then performed on the distributed acoustic signature sensing signal using a polarization beam splitter combined with differential detection to extract polarization state change information, compensating for signal fading caused by the fiber birefringence effect, thereby improving the stability of the signal demodulation. Based on the phase demodulation and polarization state demodulation results, spectral analysis (such as Fast Fourier Transform, FFT) is performed to extract frequency, amplitude, and phase information. Frequency information includes the distribution of each frequency component in the signal, amplitude information includes the energy intensity of each frequency component, and phase information includes the relative time relationship of each frequency component. The relationship between the sound signal and the vibration signal can be described by the following formula:

[0031] in, For time The sound wave signal on; The amplitude of vibration corresponds to the intensity (loudness) of the sound. The vibration frequency corresponds to the pitch of the sound. Phase refers to the relative time change of sound.

[0032] The process involves extracting fundamental frequency information from frequency information and calculating the spectral concentration of the frequency information. The specific implementation of extracting fundamental frequency information is as follows: Peak frequency detection is performed on the frequency information to extract multiple peak frequencies and their corresponding energy amplitudes. These peak frequencies are then sorted in descending order of energy amplitude to obtain a candidate frequency sequence. Starting from the beginning of the candidate frequency sequence, one frequency is selected as a candidate fundamental frequency according to the sorting order. The presence of harmonic frequency components that are integer multiples of the candidate fundamental frequency is detected, and the number of harmonic frequency components is counted. When the number of harmonic frequency components is greater than or equal to a preset harmonic quantity threshold (usually set to 2 according to the harmonic theory of the fundamental frequency), the currently selected candidate fundamental frequency is determined as the preliminary fundamental frequency. When the number of harmonic frequency components is less than the preset harmonic quantity threshold, the next candidate fundamental frequency is selected from the candidate frequency sequence according to the sorting order, and the detection is repeated until the preliminary fundamental frequency is determined. The system first establishes a base frequency or traverses the candidate frequency sequence. It then obtains the frequency information of the monitoring location corresponding to the initial base frequency and its adjacent monitoring locations. It verifies the frequency consistency of the initial base frequency at adjacent monitoring locations by calculating the frequency deviation value between adjacent monitoring locations. When the frequency deviation value between adjacent locations is less than a preset frequency deviation threshold (e.g., ±0.5Hz, determined based on the frequency resolution of the DAS system and the frequency drift statistics caused by environmental noise), frequency consistency is deemed satisfied. When frequency consistency is satisfied, the initial base frequency is determined as the base frequency information of the original vibration signal. When frequency consistency is not satisfied, the verification range of adjacent monitoring locations is expanded to the next adjacent monitoring location, and frequency consistency verification is performed again. If frequency consistency is still not satisfied after expanding the verification range, the system returns to the candidate frequency sequence to select the next candidate base frequency and performs detection and verification again until the base frequency information is determined or the candidate frequency sequence is traversed.

[0033] The specific implementation method for calculating the spectral concentration of frequency information is as follows: Divide the frequency spectrum into N frequency bands, calculate the energy proportion of each band (i.e., the sum of the squares of the energy amplitudes of all frequencies within each band is taken as the energy value of that band), sum the energy values ​​of the N bands to obtain the total energy value, and divide the energy value of each band by the total energy value to obtain the energy proportion of each band; calculate the information entropy based on the energy proportion of each band. A larger information entropy indicates a more dispersed energy distribution, while a smaller information entropy indicates a more concentrated energy distribution; use the reciprocal of the information entropy as the spectral concentration, which characterizes the degree of concentration of the overall energy distribution of the frequency information. A single index value; the target harmonic order set is determined based on the spectral concentration, which includes at least one of the second, third, and fourth harmonic frequencies; the specific determination method is as follows: when the spectral concentration is greater than or equal to a preset concentration threshold (the spectral concentration value corresponding to when the spectral energy accounts for 70% of the frequency band adjacent to the fundamental frequency is taken as the concentration threshold), the target harmonic order set is determined to include the second, third, and fourth harmonic frequencies; when the spectral concentration is less than the preset concentration threshold, the target harmonic order set is determined to include the second and third harmonic frequencies.

[0034] Based on the target harmonic order set, the fundamental frequency information is converted into the corresponding harmonic frequencies. The fundamental frequency and each harmonic frequency are superimposed according to their corresponding amplitude and phase to obtain the frequency-doubled synthesized signal. The sound signal after frequency doubling is composed of multiple harmonic superpositions. Through frequency doubling processing, the system can extend the characteristics of low-frequency vibrations to the high-frequency part, increasing the richness of the reproduced sound.

[0035] For example: In a DAS system, the effect of frequency multiplication can be explained using Fourier series expansion. Any periodic signal can be represented as a superposition of multiple sine waves, i.e., in Fourier series form:

[0036] In the formula, Indicates the order of harmonics. Indicates the first The frequency of the subharmonic Indicates the first The amplitude of the subharmonic. Indicates the first The phase of the second harmonic; through frequency doubling, the system can artificially increase the amplitude of these harmonics. In particular, low-frequency signals often only have the first or second harmonic. Frequency doubling can artificially introduce higher-order harmonics, thereby enhancing the high-frequency part of the signal and ultimately achieving a clearer and more three-dimensional sound reproduction effect. Taking the fourth harmonic as an example, let's compare two signals: one with harmonic processing and one without; assuming the original vibration signal... It is a simple sine wave signal with a fundamental frequency. and the corresponding amplitude ,but:

[0037] This signal has only one fundamental frequency. In the frequency domain, its spectrum is a single frequency peak centered at [location missing]. ,Right now:

[0038] In the formula, Let be the spectrum function of the signal. The Dirac function is used; without harmonic processing, the signal only contains the fundamental frequency, and its spectrum is very narrow. Therefore, the sound is thin and lacks high-frequency details, and the sound reproduction is not clear enough.

[0039] When we introduce harmonic processing, that is, generate multiple harmonics, the signal contains not only the fundamental frequency, but also higher-order harmonics that are multiples of it; assuming that up to the fourth harmonic is generated, the processed signal... Represented as:

[0040] In the formula, , , , These represent the fundamental frequency and the amplitudes of the second, third, and fourth harmonics, respectively. Generally, the amplitude decreases as the harmonic order increases. , , , As their respective phases, the signal spectrum after harmonic processing contains multiple frequency components, located in , , , Isofrequency points, i.e.:

[0041] The spectrum is significantly wider than the spectrum without harmonics, covering a wider frequency range; after harmonic processing, the signal spectrum includes not only the fundamental frequency. It also extends to , , By multiplying the frequency, the total spectral width of the signal is expanded from the original single frequency to four times the frequency. That is, through frequency multiplication, the spectral width of the signal is increased by three times, and the frequency range covered is significantly expanded.

[0042] The effective harmonic frequency band range is determined based on the target harmonic order set. This range includes the fundamental frequency information and the harmonic frequencies corresponding to each order of the target harmonic order set. All frequency components in the frequency-harmonic synthesized signal outside the effective harmonic frequency band range are defined as non-target frequency components. These include fifth-order and higher harmonic frequencies generated by nonlinear characteristics, as well as non-harmonic spurious frequencies. These non-target frequency components are used as noise components to be filtered out, thus obtaining the noise harmonic frequency band information. The frequency-harmonic synthesized signal is then input into a preset stable LCL topology filter to filter out the non-target frequency components. The stable LCL topology filter is obtained by using... An improved Osprey optimization algorithm is used to optimize the filtering parameters of an LCL topology filter, resulting in an optimized parameter set including resonant frequency, damping coefficient, and quality factor. This optimized parameter set is then configured onto the LCL topology filter, and the closed-loop stability of the LCL topology filter is verified using discrete system control and stability criteria, yielding a stable LCL topology filter configuration. Specifically, a discrete system transfer function model of the LCL topology filter is established. The discrete system transfer function model is a mathematical representation of a continuous-time system converted to a discrete-time system using the Z-transform, a common technique in digital control systems for analyzing system stability and dynamic response characteristics. This will not be described in detail here. The root of the corresponding closed-loop characteristic equation is calculated. When the magnitude of all characteristic roots is less than 1, the closed-loop system is considered stable. The improved osprey optimization algorithm simulates the osprey's predation behavior and, combined with the discrete system transfer function of the LCL filter, iteratively searches for the optimal set of filter parameters that minimizes the total demand distortion rate. Specifically, the discrete system transfer function of the LCL filter is first introduced as a constraint: the resonant frequency, damping coefficient, and quality factor of the filter are used as optimization variables. Based on the discrete system transfer function model, the root of the closed-loop characteristic equation is calculated in each iteration. Only candidate solutions that satisfy the stability criterion (the magnitude of all characteristic roots is less than 1) are retained, and unstable solutions are eliminated to reduce the search space. Then, the total demand is used as the constraint. The distortion rate is used as the fitness function: the total demand distortion rate is used as the optimization objective, replacing the simple distance or position fitness in the standard Osprey optimization algorithm, so that the search direction is directly directed to the filter parameter combination with the optimal fidelity of the voiceprint signal; and the search direction is further guided by the stability criterion: between the exploration and development phases of the Osprey's predation behavior, a feedback mechanism based on discrete system control and stability criteria is introduced. When the candidate solution causes the system to be unstable, the search step size and direction are automatically adjusted to avoid invalid iterations and improve the convergence speed; finally, through the above improvements, the improved Osprey optimization algorithm can quickly search for the optimal filter parameter set that minimizes the total demand distortion rate while ensuring the closed-loop stability of the LCL topology filter.After filtering out non-target frequency components, the total required distortion rate of the output signal of the stably configured LCL topology filter is monitored in real time. When the total required distortion rate is lower than a preset distortion threshold, the output signal is used as the filtered synthesized signal. The preset distortion threshold is determined according to the voiceprint restoration quality requirements, for example, it can be set to 3% to ensure the fidelity of the filtered signal.

[0043] Calculate the spectral energy ratio R and spectral complexity of the synthesized signal after filtering. The spectral energy ratio is defined as the ratio of the high-frequency energy to the total energy, i.e.:

[0044] in, For high-frequency components outside the effective harmonic frequency range, The total energy of the signal is denoted as ; spectral complexity is defined as the number of different frequency components in the synthesized signal after filtering; after harmonic processing, the spectrum expands from a single frequency to multiple frequency components, increasing spectral complexity and thus improving the layering of the sound.

[0045] The weighted comprehensive index is obtained by weighting the spectral energy ratio and spectral complexity. Using this weighted comprehensive index as the optimization target, the fundamental frequency information and the preset fundamental weight coefficients for each harmonic frequency (based on the physical law that the amplitude of each harmonic naturally decays with increasing harmonic order, the initial fundamental weight coefficients are set to 1.0 for the fundamental frequency, 0.5 for the second harmonic, 0.33 for the third harmonic, and 0.25 for the fourth harmonic, i.e., the weight coefficient of each harmonic is the reciprocal of its harmonic order) are iteratively optimized to obtain optimized weight coefficients. The iterative optimization uses an improved Osprey optimization algorithm, which improves the target signal strength by more than 12dB by optimizing the harmonic weights. The filtered synthesized signal is decomposed into fundamental frequency components and harmonic components, and the original amplitude of each harmonic component is obtained. The optimized weight coefficients are multiplied by the original amplitude of the corresponding harmonic components to obtain the adjusted amplitude. The harmonic components are reconstructed based on the adjusted amplitude and the corresponding original phase. The reconstructed harmonic components are then linearly superimposed to obtain the enhanced frequency information.

[0046] Extract the fundamental amplitude value corresponding to the fundamental frequency component and the harmonic amplitude value corresponding to each harmonic component from the amplitude information; extract the fundamental phase value corresponding to the fundamental frequency component and the harmonic phase value corresponding to each harmonic enhancement component from the phase information; synthesize the fundamental frequency component, fundamental amplitude value, and fundamental phase value into a sine wave to obtain the fundamental frequency time-domain waveform; synthesize each harmonic component, its corresponding harmonic amplitude value, and its corresponding harmonic phase value into a sine wave to obtain the harmonic time-domain waveform; superimpose the fundamental frequency time-domain waveform and the harmonic time-domain waveforms to obtain the preliminary synthesized audio signal; obtain the maximum and minimum amplitude values ​​of the preliminary synthesized audio signal; calculate the amplitude linear mapping parameter based on the maximum and minimum amplitude values ​​and the preset standard audio amplitude range; standard audio amplitude... The amplitude range is determined based on the target audio format. For example, the standard amplitude range for 16-bit PCM audio is [-32768, 32767]. The amplitude linear mapping parameter is calculated as follows: obtain the maximum and minimum amplitude of the initial synthesized audio signal, as well as the upper and lower limits of the preset standard audio amplitude range; calculate the amplitude linear mapping parameter based on the maximum amplitude, minimum amplitude, upper limit, and lower limit: divide the difference between the upper and lower limits by the difference between the maximum and minimum amplitudes to obtain the linear scaling factor; subtract the product of the linear scaling factor and the minimum amplitude from the lower limit to obtain the offset; multiply each amplitude in the initial synthesized audio signal by the linear scaling factor and add it to the offset to obtain the corresponding mapped amplitude; combine all mapped amplitudes according to the sampling time sequence to obtain the original audio signal.

[0047] This method extracts the frequency, amplitude, and phase of the vibration signal through phase demodulation and polarization state demodulation. It then expands the spectral richness by combining frequency doubling technology and iteratively optimizes the harmonic weighting coefficients based on a weighted comprehensive index of spectral energy ratio and spectral complexity. Finally, the enhanced frequency information is fused with the amplitude and phase and subjected to dynamic range normalization to obtain a high-fidelity original audio signal. This effectively solves the problem of poor restoration quality caused by soundprint signal attenuation and noise interference in complex urban environments, realizes passive multiplexing of existing communication optical cables, and effectively solves the technical bottleneck of difficult deployment of active sensors in confined spaces.

[0048] An audio-driven AI early warning model is constructed. The original audio signal is input into the audio-driven AI early warning model, and the original multi-dimensional acoustic features are extracted from the original audio signal. The original multi-dimensional acoustic features are matched with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to determine the type of early warning event and the corresponding risk level of the current utility tunnel resource status.

[0049] The method for building an audio-driven AI early warning model is as follows: a three-layer event recognition architecture is pre-built, which includes a physical layer, an environment layer, and an interaction layer. Historical and real-time acoustic fingerprint sample data of the utility tunnel are collected. Based on the predefined event types of each layer in the three-layer event recognition architecture, the historical and real-time acoustic fingerprint sample data are labeled with event types to obtain the event type labels corresponding to each sample data. Multi-dimensional acoustic features are extracted from the labeled historical and real-time acoustic fingerprint sample data. The multi-dimensional acoustic features are associated with the corresponding event type labels and stored to construct a multi-dimensional event fingerprint database. A deep learning recognition model is built based on a three-layer event recognition architecture. The deep learning recognition model is a CNN-BiGRU hybrid sub-model. The deep learning recognition model is pruned and quantized to obtain a compressed model. The compressed model is then integrated with a multi-dimensional event fingerprint database and the three-layer event recognition architecture to obtain an audio-driven AI early warning model. The method for determining the type of early warning event and the corresponding risk level of the current utility tunnel resource status is as follows: extract the original multi-dimensional acoustic features from the original audio signal, input the original multi-dimensional acoustic features into the CNN sub-model in the audio-driven AI early warning model for spatiotemporal feature fusion, and obtain the spatiotemporal fusion features; The spatiotemporal fusion features are input into the BiGRU sub-model in the audio-driven AI early warning model to classify events and output the probability distribution of each type of event. The probability distributions of various events are matched and verified with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to obtain the verified event probability distributions. Obtain the current environmental noise and historical false alarm rate, determine the probability judgment threshold, compare the probability of each event in the verified event probability distribution with the probability judgment threshold, and when the event probability is greater than the probability judgment threshold, determine the corresponding event type as the warning event type; Calculate the probability deviation between the event probability and the probability judgment threshold, and determine the risk level corresponding to the warning event type based on the probability deviation.

[0050] Specifically, a three-layer event recognition architecture is pre-built, consisting of a physical layer, an environment layer, and an interaction layer. The physical layer is used to identify fiber optic cable breakage and bending attenuation events, the environment layer is used to identify construction damage and cable discharge events, and the interaction layer is used to identify maintenance knocking and positioning events and voice communication events. The three-layer event recognition architecture is designed specifically for the power utility tunnel scenario. Through a hierarchical recognition strategy, events with different physical characteristics are categorized to improve recognition accuracy and reduce false alarm rate.

[0051] The system collects historical and real-time acoustic signature data for the utility tunnel. Historical acoustic signature data includes data from over 2000 actively triggered events across more than 10 target categories in a simulated utility tunnel environment, as well as alarm acoustic signature data accumulated during actual tunnel operation. Real-time acoustic signature data consists of data collected in real-time during system operation and manually verified. Based on predefined event types at each layer of the three-layer event recognition architecture, the historical and real-time acoustic signature data are labeled with event types to obtain event type labels for each sample. Semi-automatic labeling tools are used to assist manual labeling. Multi-dimensional acoustic features are extracted from the labeled historical and real-time acoustic signature data, including Mel-frequency cepstral coefficients (MFC). The acoustic fingerprint consists of three parts: Ccepstral Coefficients (CFCs), wavelet packet energy spectrum (time-frequency domain), and spatial domain features. The CFCs are a 128-dimensional feature vector obtained by mapping the spectrum to the Mel scale and performing cepstral analysis, simulating human hearing characteristics. This provides strong differentiation of the timbre and sound quality of the acoustic fingerprint. The wavelet packet energy spectrum (time-frequency domain) features are obtained through multi-resolution wavelet packet decomposition, extracting energy distribution within different frequency bands to reflect the time-frequency local characteristics of the acoustic fingerprint signal. The spatial domain features are obtained by analyzing the differences in acoustic fingerprint propagation characteristics at different monitoring locations along the optical fiber, reflecting the spatial location information of the sound source. These multi-dimensional acoustic features are associated with corresponding event type labels to construct a multi-dimensional event fingerprint database. This database, by fusing time-frequency and spatial domain features, solves the problem of misjudgment caused by complex environmental interference. A deep learning recognition model is constructed based on a three-layer event recognition architecture. The deep learning recognition model is a CNN-BiGRU hybrid sub-model. The CNN sub-model (convolutional neural network sub-model) is used to perform spatiotemporal feature fusion of multi-dimensional acoustic features and extract the long-range temporal dependencies of vibration signals in the corresponding environmental and interaction layers. The CNN sub-model extracts the local time-frequency features of the voiceprint signal through multiple convolutional kernels and uses max pooling layers for dimensionality reduction to obtain a high-level abstract feature representation. The BiGRU sub-model (bidirectional gated recurrent unit sub-model) is used to receive the fused features output by the CNN sub-model and output the probability distribution corresponding to each event type label. The BiGRU sub-model extracts temporal dependencies along both forward and backward time dimensions and fuses the bidirectional outputs through fully connected layers to output the probability values ​​of various events. Among them, the convolutional kernel feature extraction and max pooling dimensionality reduction of the CNN sub-model, as well as the bidirectional temporal modeling and fully connected layer output fusion of the BiGRU sub-model, are all common techniques in the field of deep learning and belong to the existing technologies in this field, so they will not be elaborated here.

[0052] Simultaneously, the deep learning recognition model is pruned and quantized for compression, reducing its size to a preset percentage below the original size, such as 32%. Pruning reduces the number of model parameters by removing neural connections that have little impact on the output. Quantization reduces the model's storage and computational requirements by converting floating-point weights to low-precision integer representations. Floating-point weights refer to the data types of neural network connection parameters in deep learning models, which are 32-bit or 64-bit floating-point numbers (such as FP32 and FP64), occupying a large amount of storage space and having high computational complexity. Quantization converts them to 8-bit integers (INT8) or lower-precision integer representations, significantly reducing storage and computational overhead while maintaining model inference accuracy. This is a common technique in the field of model compression and is an existing technology, so it will not be elaborated further here. The compressed model ensures that the peak inference memory usage on edge devices is strictly controlled below 1.5GB, meeting the resource constraints of edge deployment in the utility tunnel. Finally, the compressed model is integrated with a multi-dimensional event fingerprint database and a three-layer event recognition architecture to obtain an audio-driven AI early warning model.

[0053] Raw multi-dimensional acoustic features are extracted from the original audio signal. The extraction method for raw multi-dimensional acoustic features is consistent with the feature extraction method when constructing the multi-dimensional event fingerprint database, ensuring the consistency of the feature space. The raw multi-dimensional acoustic features are input into the CNN sub-model in the audio-driven AI early warning model for spatiotemporal feature fusion to obtain the corresponding spatiotemporal fusion features. The spatiotemporal fusion features integrate the temporal dynamic characteristics and frequency static characteristics of the voiceprint signal, as well as the spatial correlation characteristics of different monitoring locations. The spatiotemporal fusion features are input into the BiGRU sub-model in the audio-driven AI early warning model for event classification, and the probability distribution of various events is output. Various events include fiber optic breakage events and bending attenuation events at the physical layer, construction damage events and cable discharge events at the environmental layer, and maintenance knocking location events and voice communication events at the interaction layer. The probability distributions of various events are matched and verified with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to obtain the verified event probability distributions. The specific implementation of the matching verification is as follows: the cosine similarity is used to calculate the similarity between the current original multi-dimensional acoustic features and the corresponding multi-dimensional acoustic features of each event type in the multi-dimensional event fingerprint database; the similarity is compared with the preset similarity threshold interval, which is determined based on historical sample statistics. The 5th percentile of the similarity distribution of the correctly matched samples is taken as the lower limit threshold of the interval, and the 95th percentile is taken as the upper limit threshold of the interval. When the similarity is below the lower threshold, the probability of the corresponding event is reduced by multiplying the probability of the corresponding event by a preset penalty coefficient (based on experience, such as 0.8). When the similarity is above the upper threshold, the probability of the corresponding event is increased by multiplying the probability of the corresponding event by a preset reward coefficient (based on experience, such as 1.2). When the similarity is between the lower and upper thresholds, the probability of the corresponding event remains unchanged, thus obtaining the verified event probability distribution. Through fingerprint database matching verification, the model's misjudgment of unknown events is effectively filtered out. Obtain current environmental noise data and historical false alarm rates to determine a probability judgment threshold. The probability judgment threshold is determined as follows: Calculate a noise correction coefficient based on the current environmental noise data, such as by using the noise ratio of the current environmental noise to the historical average noise. When the noise ratio is greater than 1, the noise correction coefficient is the noise ratio; when the noise ratio is less than or equal to 1, the noise correction coefficient is 1. Calculate a false alarm correction coefficient based on the historical false alarm rate, such as by dividing the historical false alarm rate by a preset target false alarm rate (pre-set based on the actual acceptable false alarm frequency of power pipeline maintenance) to obtain the false alarm ratio. When the false alarm ratio is greater than 1, the false alarm correction coefficient is the false alarm ratio; when the false alarm ratio is less than or equal to 1, the false alarm correction coefficient is the product of the false alarm ratio and 0.8. Formulate a benchmark judgment threshold based on the experience data of those in the field. Multiply the benchmark judgment threshold by the noise correction coefficient and the false alarm correction coefficient sequentially to obtain the probability judgment threshold. Then, verify the... The probability of each event in the event probability distribution is compared with a probability judgment threshold. When the event probability is greater than the probability judgment threshold, the corresponding event type is determined as the warning event type. When the event probabilities of multiple event types are all greater than the probability judgment threshold, the event type with the highest event probability is selected as the warning event type. At the same time, the probability deviation value between the event probability and the probability judgment threshold is calculated, and the risk level corresponding to the warning event type is determined based on the probability deviation value. The probability deviation value is the difference between the event probability and the probability judgment threshold, and the difference is normalized and mapped to the 0-1 interval to obtain the probability deviation value. The risk level is determined by setting multiple risk level intervals based on experience, such as low risk interval (0 < probability deviation value ≤ 0.2), medium risk interval (0.2 < probability deviation value ≤ 0.5), and high risk interval (probability deviation value > 0.5). The risk level is determined according to the interval in which the probability deviation value falls.

[0054] This approach, by constructing a three-layer event recognition architecture comprising a physical layer, an environmental layer, and an interaction layer, solves the problem of misjudgment caused by the mixed recognition of events with different physical natures, and achieves refined hierarchical classification of fiber optic body status, external threats, and operation and maintenance interaction events. By employing a CNN-BiGRU hybrid sub-model for spatiotemporal feature fusion and event classification, it addresses the performance imbalance of traditional single models in recognizing continuous and transient events, achieving high-precision recognition of multiple types of events such as construction damage and cable discharge. Furthermore, by introducing a multi-dimensional event fingerprint database for matching and verification, it solves the problem of model misjudgment of unknown events, improving the reliability of event determination. Finally, through an adaptive threshold decision mechanism, it dynamically adjusts the decision threshold based on environmental noise and historical false alarm rates, solving the sensitivity anomaly caused by fixed thresholds in densely populated construction areas, and achieving adaptive recognition under different environments.

[0055] The system acquires the voiceprint features, geographical coordinates and timestamps of the alarm points corresponding to the original audio signals, constructs a three-dimensional correlation analysis engine, and outputs target early warning events and optimizes risk levels based on the three-dimensional correlation analysis engine.

[0056] The method for constructing a three-dimensional correlation analysis engine is as follows: Voiceprint features are processed into feature vectors to extract event harmonic energy distribution parameters and on-site audio reconstruction parameters, thus constructing a voiceprint feature vector; the geographic coordinates of alarm points are parsed to extract GIS coordinates and utility tunnel topology data contained within the alarm point geographic information coordinates; spatial indexing is performed based on the GIS coordinates and utility tunnel topology data to generate spatial grid identifiers and construct a geospatial index; and time stamps are segmented into time windows to generate time segment identifiers and establish a time series index. Alarm event identifiers are generated based on the type and timestamp of the warning event. Voiceprint feature vectors, geospatial indexes, and time series indexes are associated and bound through the alarm event identifiers to obtain a three-dimensional associated dataset. A structured three-dimensional index is constructed based on three dimensions: voiceprint feature vector, geospatial index, and time series index. A bidirectional mapping relationship between dimensions is established based on the structured three-dimensional index to obtain a dimension association mapping table. A cross-dimensional retrieval interface is pre-constructed based on the dimensional association mapping table to obtain a 3D retrieval interface; inter-dimensional association analysis rules are defined based on the dimensional association mapping table to obtain an inter-dimensional association rule set. By integrating the 3D retrieval interface with the set of rules for inter-dimensional association, a 3D association analysis engine is obtained.

[0057] The method for outputting target early warning events and optimizing risk levels based on the 3D correlation analysis engine is as follows: A sliding event window is established based on the 3D correlation analysis engine, and spatiotemporal consistency verification is performed on the early warning event type and the corresponding risk level: The geospatial index sequence within the time window corresponding to the time series index is obtained through spatiotemporal correlation rules, and it is determined whether the geospatial index sequence meets the preset spatial continuity conditions; when the spatial continuity conditions are met, the early warning event type and the corresponding risk level are determined as the event type and risk level after spatiotemporal verification. A two-stage analysis mechanism of initial identification and deep correlation is implemented for event types after spatiotemporal verification: In the initial identification stage, voiceprint-event matching verification is performed based on voiceprint feature vector and multi-dimensional event fingerprint database to obtain the initial event type; In the deep correlation stage, geospatial index and time series index are called based on dimensional correlation mapping table to obtain the associated event records of the initial event type within a preset historical time window. When the number of associated event records reaches the preset correlation threshold, the initial event type is determined as the target warning event. Based on the dimensional correlation mapping table, spatiotemporal clustering analysis is performed on historical alarm events. Combined with the pipe gallery topology data, the topological impact range and historical frequency of the target early warning event are calculated to obtain the risk correction coefficient. Based on the risk correction coefficient, the risk level after spatiotemporal verification is weighted and corrected to obtain the optimized risk level.

[0058] Specifically, voiceprint features are extracted from the original audio signal. These features include event harmonic energy distribution characteristics and on-site audio reconstruction characteristics. Event harmonic energy distribution characteristics are obtained by analyzing the energy proportion of each harmonic in the original audio signal, reflecting the spectral structure characteristics of the voiceprint. On-site audio reconstruction characteristics are the effective audio segments after noise reduction of the original audio signal, generated by removing invalid noise during construction intervals to ensure the purity of audio information during the alarm phase. The geographical coordinates of the alarm point are obtained by fusing GIS coordinates with the pipe gallery topology data to generate alarm point location and distribution information in real time. GIS coordinates are geographical coordinates obtained through the Global Positioning System or BeiDou Navigation Satellite System, and pipe gallery topology data are spatial relationship data of optical cable routing, joint locations, and branch nodes within the pipe gallery. By mapping the GIS coordinates to the pipe gallery topology data, the three-dimensional coordinates of the precise location of the alarm point within the pipe gallery are obtained, which are the geographical coordinates of the alarm point. A timestamp is obtained, which is a precise time marker for the time of distributed voiceprint sensor signal acquisition. The system clock synchronization protocol (such as PTP or NTP) ensures that the time synchronization accuracy of each monitoring channel reaches the millisecond level.

[0059] The voiceprint features are processed into feature vectors to extract the event harmonic energy distribution parameters and the on-site audio reconstruction parameters, and a voiceprint feature vector is constructed. The event harmonic energy distribution parameters include the energy amplitude, energy ratio and phase relationship between each harmonic. The on-site audio reconstruction parameters include the duration, signal-to-noise ratio and spectral centroid of the audio segment. The voiceprint feature vector is a normalized numerical representation of the above parameters, which is used for subsequent similarity calculation and matching verification. The geographic coordinates of alarm points are parsed to extract GIS coordinates and utility tunnel topology data contained within them. Spatial indexing is then performed based on the GIS coordinates and utility tunnel topology data to generate spatial grid identifiers and construct a geospatial index. The specific implementation of spatial indexing is as follows: the three-dimensional space of the utility tunnel is divided into equally spaced spatial grids and numbered. The spatial grid number to which the GIS coordinates of the alarm point belong is obtained, and the spatial grid number is associated with the node identifier in the utility tunnel topology data to generate a spatial grid identifier. The geospatial index supports fast retrieval based on spatial range, such as retrieving all historical alarm records within a certain utility tunnel section. The time stamp is segmented into time windows to generate time segment identifiers and establish a time series index. The specific implementation of time window segmentation is as follows: the set continuous time axis is divided into time windows of equal length and numbered (for example, each time window is 10 minutes), the time window number to which the timestamp belongs is obtained, and the time segment identifier is generated. The time series index supports fast retrieval based on time range, such as retrieving all alarm events within a certain time period.

[0060] An alarm event identifier is generated based on the warning event type and timestamp. Specifically, the warning event type code (e.g., construction damage code EP001, cable discharge code EF002) is concatenated with the timestamp string representation, and then a fixed-length unique identifier is generated using a hash function. The alarm event identifier is used to uniquely identify the current warning event, ensuring consistent association of the same event across different dimensions of data. The voiceprint feature vector, geospatial index, and time series index are associated and bound through the alarm event identifier to obtain a three-dimensional associated dataset. Specifically, in the database or data structure, the alarm event identifier is used as the primary key to establish associated records for the voiceprint feature vector, geospatial index, and time series index, so that the voiceprint feature vector, geospatial index, and time series index under the same alarm event identifier form a one-to-one correspondence.

[0061] A three-dimensional data index cube is constructed based on three dimensions: voiceprint feature vector, geospatial index, and time series index, resulting in a structured three-dimensional index. Specifically, a three-dimensional coordinate system is constructed with the voiceprint feature vector dimension as the X-axis, the geospatial index dimension as the Y-axis, and the time series index dimension as the Z-axis. Each record in the three-dimensional associated dataset is mapped to its corresponding position in the three-dimensional coordinate system, forming a structured three-dimensional index. The structured three-dimensional index supports fast positioning based on any dimension and cross-dimensional navigation. A bidirectional mapping relationship between dimensions is established based on the structured three-dimensional index, resulting in a dimension association mapping table. The bidirectional mapping relationship between dimensions includes the spatiotemporal mapping relationship between the time series index and the geospatial index. The system establishes the geospatial index and the voiceprint feature vector mapping relationship between geospatial indexes and voiceprint feature vectors. The spatiotemporal mapping relationship is established as follows: for a given time segment identifier, all corresponding geospatial indexes are searched through the three-dimensional data index cube to establish a time-to-space mapping; conversely, for a given geospatial index, all corresponding time segment identifiers are searched to establish a space-to-time mapping. The voice-to-ground mapping relationship is established as follows: for a given geospatial index, the corresponding voiceprint feature vector is searched through the three-dimensional data index cube to establish a space-to-voiceprint mapping; conversely, for a given voiceprint feature vector, the corresponding geospatial index is searched to establish a voiceprint-to-space mapping.

[0062] Based on the dimensional association mapping table, a cross-dimensional retrieval interface is pre-constructed to obtain a three-dimensional retrieval interface. The cross-dimensional retrieval interface supports querying related data in other dimensions using any dimension as the retrieval entry point. For example, using a time range as the retrieval entry point, the spatial distribution and voiceprint characteristics of all alarm events within that time range can be queried; using a spatial range as the retrieval entry point, the temporal evolution and voiceprint characteristics of all historical alarms within that utility tunnel section can be queried; using voiceprint characteristics as the retrieval entry point, the spatial location and occurrence time of similar voiceprint events can be queried. Based on the dimensional association mapping table, inter-dimensional association analysis rules are defined to obtain an inter-dimensional association rule set. These rules include spatiotemporal association rules that call the corresponding geospatial index based on the time series index, and acoustic-geographic association rules that call the corresponding voiceprint feature vector based on the geospatial index. The spatiotemporal association rules define the matching logic between time windows and spatial ranges. For example, if the same spatial grid identifier experiences alarm events within multiple consecutive time windows, it is determined to be a persistent threat. The acoustic-geographic association rules define the matching logic between spatial location and voiceprint features. For example, if the same geospatial index shows similar voiceprint features multiple times in historical records, it is determined to be a high-frequency risk point. By integrating the 3D retrieval interface with the inter-dimensional association rule set, a 3D association analysis engine is obtained. The 3D association analysis engine obtains cross-dimensional data by calling the 3D retrieval interface, and performs association analysis and reasoning through the inter-dimensional association rule set to achieve deep association mining of three dimensions: voiceprint features, geographic information, and timestamps.

[0063] After establishing the 3D correlation analysis engine, a sliding event window is created based on the engine to perform spatiotemporal consistency verification on the types of early warning events and their corresponding risk levels. The sliding event window is created by setting the window length (e.g., 5 minutes) and the sliding step size (e.g., 1 minute), and sliding along the time axis to extract continuous time segment identifiers. The specific implementation of the spatiotemporal consistency verification is as follows: the geospatial index sequence within the time window corresponding to the time series index is obtained through spatiotemporal correlation rules, and it is determined whether the geospatial index sequence meets the preset spatial continuity condition. The determination method for the spatial continuity condition is as follows: the spatial distance between the geospatial indices of each alarm event within the corresponding time window is calculated. When the spatial distance is less than the preset distance threshold (preset based on experience), the spatial continuity condition is determined to be met, indicating that the alarm events originate from continuous activities in the same spatial location. When the spatial continuity condition is met, the early warning event type and its corresponding risk level are determined as the event type and risk level after spatiotemporal verification.

[0064] A two-stage analysis mechanism of initial identification and deep association is implemented for event types after spatiotemporal verification. In the initial identification stage, voiceprint feature vectors are matched and verified with a multi-dimensional event fingerprint database to obtain the initial event type. The specific implementation of voiceprint-event matching verification is as follows: the cosine similarity between the current voiceprint feature vector and each event type in the multi-dimensional event fingerprint database is calculated, and the event type with the highest cosine similarity is selected as the initial event type. In the deep association stage, the geospatial index and time series index are called based on the dimensional association mapping table to obtain the associated event records of the initial event type within a preset historical time window. The associated event records include the number of similar events occurring in the same geospatial index within the historical time period, the time distribution pattern, and the risk evolution trend. When the number of associated event records reaches the preset association threshold (preset based on experience), it indicates that the event type has historically high incidence characteristics in this location, and the initial event type is determined as the target warning event.

[0065] Based on a dimensional correlation mapping table, spatiotemporal clustering analysis is performed on historical alarm events. Events of the same type that occur multiple times (e.g., more than 3 times) within a preset historical time window using the same geospatial index are aggregated into spatiotemporal clusters. The spatiotemporal distribution characteristics and frequency of occurrence of each cluster are extracted, constructing historical event spatiotemporal clustering data and storing it in the 3D correlation analysis engine. Based on the target warning event, the historical event spatiotemporal clustering data is retrieved, and combined with the utility tunnel topology data, the topological influence range and historical frequency of the target warning event are calculated to obtain the risk correction coefficient. Specifically, the topological influence range is calculated as follows: using the geospatial index of the target warning event as the central node, a breadth-first search is performed along the connectivity relationships in the utility tunnel topology data. The search depth is determined according to the event type (e.g., cable discharge events). The search depth for event detection is 3 levels (and the search depth for construction-related external damage events is 2 levels). All nodes (early warning events) within the search range are defined as the topology influence range. The historical occurrence frequency is the sum of the number of times the same type of event occurs at each node within the topology influence range in the historical time period. Correspondingly, the risk correction coefficient is calculated as follows: Risk Correction Coefficient = Number of Nodes in Topology Influence Range × First Weight + Historical Occurrence Frequency × Second Weight, where the first and second weights are preset according to the event type and the importance level of the utility tunnel, and their sum is equal to one. Finally, the risk level after spatiotemporal verification is weighted and corrected according to the risk correction coefficient to obtain the optimized risk level. The specific implementation of the weighted correction is as follows: Optimized Risk Level = Risk Level After Spatiotemporal Verification × (1 + Risk Correction Coefficient).

[0066] This approach, by constructing a three-dimensional correlation analysis engine encompassing voiceprint features, geographic information, and timestamps, achieves deep correlation mining across three dimensions: voiceprint features, spatial location, and temporal evolution. This overcomes the limitations of traditional single-dimensional alarm analysis, providing multi-dimensional data support for refined operation and maintenance decisions. Simultaneously, the sliding event window and spatiotemporal consistency verification effectively capture persistent threats, avoiding misjudgments of single-point instantaneous alarms and improving alarm accuracy. The adoption of a two-level analysis mechanism—initial judgment and deep correlation—introduces historical correlation verification on top of rapid initial judgment, ensuring both real-time event response and improved accuracy through historical data verification. The system enhances the credibility of event determination. Furthermore, by calculating the impact range of the topology and statistically analyzing historical occurrences, it enables dynamic adjustments to risk levels. This allows risk assessment to consider not only the current event intensity but also the tunnel topology and historical evolution patterns, making risk levels more aligned with actual operational needs. Finally, the system supports flexible retrieval and correlation analysis by operational personnel using any dimension as an entry point through a 3D search interface and inter-dimensional association rule sets. This provides multi-dimensional data support and data-driven decision-making basis for inspection optimization and external collaboration, significantly improving the accuracy of threat identification, the credibility of risk determination, and the refinement of operational decisions.

[0067] Based on the target warning event and the optimized risk level, the system matches the corresponding warning response strategy from the preset warning response strategy library and generates an alarm work order which is then pushed to the operation and maintenance end.

[0068] Specifically, the early warning response strategy library includes a set of response strategies for different types of early warning events and different optimization risk levels, including alarm notification strategies, inspection optimization strategies, and external collaboration strategies. Among them, the alarm notification strategy defines the generation rules, push methods, and recipients of alarm work orders. For example, when the optimization risk level is high, an alarm work order containing the event type and location coordinates is automatically pushed to the operations manager and on-duty personnel via the SMS module; when the optimization risk level is medium, it is pushed to the regional operations personnel; when the optimization risk level is low, it is recorded in the system log for periodic summarization. The inspection optimization strategy defines rules for adjusting inspection routes based on risk weights. For example, it increases the inspection frequency to once a day for sections with high-risk events, maintains the regular inspection frequency for sections with medium-risk events, and reduces the inspection frequency for sections with low-risk events. The external collaboration strategy defines rules for coordinated handling with construction units. For example, when the target warning event is construction damage, it automatically generates a "Construction Notification" and an "Electronic Fence Warning" and pushes them to the construction unit, simultaneously updating inspection records. It also retrieves response strategies from the warning response strategy library that match the event type and optimized risk level of the target warning event and generates a notification. Alarm work orders are pushed to the operations and maintenance (O&M) end via methods including but not limited to SMS, APP push, email, and voice call. The content of the alarm work order includes: a description of the event type of the target warning event, an optimization risk level identifier, geographic information coordinates of the alarm point (including GIS coordinates and precise location inside the utility tunnel), timestamp, voiceprint feature summary (including key parameters of event harmonic energy distribution), suggested response measures, and statistics of historical related events. The format of the alarm work order is adaptively adjusted according to the type of O&M end device. For example, a simplified text format is used when pushing to mobile devices, and a detailed graphic format is used when pushing to computers.

[0069] In one embodiment, a utility tunnel resource status early warning system based on intrinsic voiceprint analysis is also disclosed, such as... Figure 2 As shown, the system includes: The sensor signal acquisition module is used to acquire the light source signal. It adopts a two-stage beam splitting and amplification architecture to amplify the light source signal and obtain the amplified light source signal. The amplified light source signal is then split into two optical paths by a beam splitter. Each optical path is independently configured with an acousto-optic modulator and a photodetector. A parallel monitoring channel is constructed using an optical switch to acquire the distributed acoustic signature sensing signal of the corresponding monitoring position of each monitoring channel in real time. The voiceprint restoration and enhancement module is used to perform phase demodulation and polarization state demodulation on the distributed voiceprint sensing signal, extract frequency information, amplitude information and phase information, process the frequency information to obtain enhanced frequency information, and fuse the enhanced frequency information, amplitude information and phase information to obtain the original audio signal; The intelligent identification and early warning module is used to build an audio-driven AI early warning model. The original audio signal is input into the audio-driven AI early warning model, and the original multi-dimensional acoustic features are extracted from the original audio signal. The original multi-dimensional acoustic features are matched with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to determine the type of early warning event and the corresponding risk level of the current utility tunnel resource status. The 3D correlation analysis module is used to obtain the voiceprint features, alarm point geographic information coordinates and timestamps corresponding to the original audio signal, build a 3D correlation analysis engine, and output target early warning events and optimize risk levels based on the 3D correlation analysis engine; The intelligent decision response module is used to match the corresponding early warning response strategy from the preset early warning response strategy library based on the target early warning event and the optimized risk level, and generate an alarm work order to push to the operation and maintenance terminal.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0071] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0072] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis, characterized in that, The early warning method includes: The light source signal is acquired and amplified using a two-stage beam splitting and amplification architecture to obtain the amplified light source signal. The amplified light source signal is then split into two optical paths by a beam splitter. Each optical path is independently configured with an acousto-optic modulator and a photodetector. A parallel monitoring channel is constructed using an optical switch to collect distributed acoustic signature sensing signals at the corresponding monitoring positions of each monitoring channel in real time. Phase demodulation and polarization state demodulation are performed on the distributed voiceprint sensing signal to extract frequency information, amplitude information and phase information. The frequency information is processed to obtain enhanced frequency information. The enhanced frequency information, amplitude information and phase information are fused to obtain the original audio signal. An audio-driven AI early warning model is constructed. The original audio signal is input into the audio-driven AI early warning model, and the original multi-dimensional acoustic features are extracted from the original audio signal. The original multi-dimensional acoustic features are matched with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to determine the type of early warning event and the corresponding risk level of the current utility tunnel resource status. Acquire the voiceprint features, alarm point geographic information coordinates and timestamps corresponding to the original audio signal, construct a three-dimensional sound correlation analysis engine, and output target early warning events and optimize risk levels based on the three-dimensional correlation analysis engine; Based on the target warning event and the optimized risk level, the system matches the corresponding warning response strategy from the preset warning response strategy library and generates an alarm work order which is then pushed to the operation and maintenance end.

2. The method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis according to claim 1, characterized in that, The method of splitting the amplified light source signal into two optical paths using a beam splitter is as follows: The source signal of a C-band narrow linewidth laser is acquired, and the optical power stability of the source signal is detected. When the optical power fluctuation value of the source signal is detected to be greater than the preset fluctuation threshold, automatic power control compensation is initiated to compensate the optical power of the source signal to obtain a stable source signal. A stable light source signal is input into an erbium-doped fiber amplifier for pre-amplification. The pump current of the erbium-doped fiber amplifier is adjusted to increase the output optical power of the erbium-doped fiber amplifier to a preset first target power, thereby obtaining the pre-amplified signal output by the erbium-doped fiber amplifier. The pre-amplified signal is input into a Raman amplifier for secondary gain. By adjusting the Raman pump power of the Raman amplifier, the optical power of the pre-amplified signal reaches the preset second target power, thus obtaining a secondary gain light source signal. The secondary gain light source signal is input into a 50 / 50 beam splitter for power distribution, and the optical power of the two outputs of the 50 / 50 beam splitter is monitored in real time. The optical power difference between the two outputs is calculated. When the optical power difference is greater than the preset equalization threshold, dynamic fine-tuning compensation of the splitting ratio is initiated until the optical power difference is less than or equal to the preset equalization threshold. The secondary gain light source signal is then split into two independent optical paths with equal optical power, which are called dual optical paths.

3. The method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis according to claim 1, characterized in that, The method for obtaining enhanced frequency information is as follows: Extract fundamental frequency information from frequency information and calculate the spectral concentration of the frequency information; determine the target harmonic order set based on the spectral concentration. Based on the target harmonic order set, the fundamental frequency information is converted into the corresponding harmonic frequencies. The fundamental frequency information and each harmonic frequency are superimposed according to their corresponding amplitude and phase to obtain the frequency-harmonic synthesis signal. Based on the target harmonic order set, the non-target frequency components are determined. The frequency-multiplied synthesized signal is input into a preset stable LCL topology filter, and the output signal of the stable LCL topology filter is used as the filtered synthesized signal. The spectral energy ratio and spectral complexity of the synthesized signal after filtering are calculated. The spectral energy ratio and spectral complexity are weighted to obtain a weighted comprehensive index. The weighted comprehensive index is used as the optimization target. The basic frequency information and the preset basic weight coefficients of each harmonic frequency are iteratively optimized to obtain the optimized weight coefficients. The filtered synthesized signal is decomposed into fundamental frequency components and harmonic components, and the original amplitude and phase of each harmonic component are obtained. The optimized weighting coefficients are multiplied by the original amplitude of the corresponding harmonic components to obtain the adjusted amplitude. Based on the adjusted amplitude and the corresponding original phase, the harmonic components are reconstructed and linearly superimposed to obtain the enhanced frequency information.

4. The method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis according to claim 3, characterized in that, The method for extracting fundamental frequency information from frequency information is as follows: Frequency information is subjected to spectral peak detection, multiple peak frequencies and their corresponding energy amplitudes are extracted, and the peak frequencies are sorted in descending order of energy amplitude to obtain a candidate frequency sequence. Starting from the beginning of the candidate frequency sequence, select one as the candidate fundamental frequency in sorted order, detect whether there are harmonic frequency components in the frequency information that are integer multiples of the candidate fundamental frequency, and count the number of harmonic frequency components. When the number of harmonic frequency components is greater than or equal to the preset harmonic number threshold, the currently selected candidate fundamental frequency is determined as the preliminary fundamental frequency. When the number of harmonic frequency components is less than the preset harmonic number threshold, the next candidate fundamental frequency is selected from the candidate frequency sequence in sorted order, and the detection is repeated until the preliminary fundamental frequency is determined or the candidate frequency sequence is traversed. Obtain the frequency information of the monitoring location corresponding to the preliminary base frequency and its adjacent monitoring locations, and verify the frequency consistency of the preliminary base frequency at adjacent monitoring locations; when the frequency consistency is satisfied, determine the preliminary base frequency as the base frequency information of the original vibration signal.

5. The method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis according to claim 4, characterized in that, The method for obtaining the original audio signal is as follows: Extract the fundamental amplitude value corresponding to the fundamental frequency component and the harmonic amplitude value corresponding to each harmonic component from the amplitude information; extract the fundamental phase value corresponding to the fundamental frequency component and the harmonic phase value corresponding to each harmonic enhancement component from the phase information. The fundamental frequency component, fundamental amplitude value, and fundamental phase value are combined into a sine wave to obtain the fundamental frequency time-domain waveform; each harmonic component, its corresponding harmonic amplitude value, and its corresponding harmonic phase value are combined into a sine wave to obtain the harmonic time-domain waveform. The fundamental frequency time-domain waveform is superimposed with the time-domain waveforms of each harmonic to obtain a preliminary synthesized audio signal; the maximum and minimum amplitudes of the preliminary synthesized audio signal are obtained, and the amplitude linear mapping parameters are calculated based on the maximum and minimum amplitudes and the preset standard audio amplitude range; Based on the amplitude linear mapping parameter, the amplitude of the initially synthesized audio signal is mapped to the standard audio amplitude range to obtain the original audio signal.

6. The method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis according to claim 1, characterized in that, The method for constructing an audio-driven AI early warning model is as follows: A three-layer event recognition architecture is pre-built, which includes a physical layer, an environment layer, and an interaction layer. Collect historical and real-time voiceprint sample data of the utility tunnel. Based on the predefined event types of each layer in the three-layer event recognition architecture, label the historical and real-time voiceprint sample data with event types to obtain the event type label corresponding to each sample data. Multi-dimensional acoustic features are extracted from labeled historical and real-time voiceprint sample data. These multi-dimensional acoustic features are then associated with and stored with corresponding event type labels to construct a multi-dimensional event fingerprint database. A deep learning recognition model is built based on a three-layer event recognition architecture. The deep learning recognition model is a CNN-BiGRU hybrid sub-model. The deep learning recognition model is pruned and quantized to obtain a compressed model. The compressed model, the multi-dimensional event fingerprint database and the three-layer event recognition architecture are integrated to obtain an audio-driven AI early warning model.

7. The method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis according to claim 6, characterized in that, The method for determining the type of early warning event and the corresponding risk level of the current utility tunnel resource status is as follows: Original multidimensional acoustic features are extracted from the original audio signal. These original multidimensional acoustic features are then input into the CNN sub-model in the audio-driven AI early warning model for spatiotemporal feature fusion to obtain spatiotemporal fused features. The spatiotemporal fusion features are input into the BiGRU sub-model in the audio-driven AI early warning model to classify events and output the probability distribution of each type of event. The probability distributions of various events are matched and verified with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to obtain the verified event probability distributions. Obtain the current environmental noise and historical false alarm rate, and determine the probability judgment threshold; compare the probability of each event in the verified event probability distribution with the probability judgment threshold, and when the event probability is greater than the probability judgment threshold, determine the corresponding event type as the warning event type; Calculate the probability deviation between the event probability and the probability judgment threshold, and determine the risk level corresponding to the warning event type based on the probability deviation.

8. The method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis according to claim 1, characterized in that, The method for constructing a 3D correlation analysis engine is as follows: The voiceprint features are processed into feature vectors to extract the event harmonic energy distribution parameters and on-site audio reconstruction parameters, and a voiceprint feature vector is constructed. The geographic information coordinates of the alarm point are parsed to extract the GIS coordinates and utility tunnel topology data contained in the geographic information coordinates of the alarm point. Based on the GIS coordinates and utility tunnel topology data, spatial index coding is performed to generate spatial grid identifiers and construct a geospatial index. Perform time window segmentation on the timestamps, generate time segment identifiers, and establish a time series index; Alarm event identifiers are generated based on the type and timestamp of the warning event. Voiceprint feature vectors, geospatial indexes, and time series indexes are associated and bound through the alarm event identifiers to obtain a three-dimensional associated dataset. A structured three-dimensional index is constructed based on three dimensions: voiceprint feature vector, geospatial index, and time series index. A bidirectional mapping relationship between dimensions is established based on the structured three-dimensional index to obtain a dimension association mapping table. A cross-dimensional retrieval interface is constructed based on a dimension association mapping table to obtain a three-dimensional retrieval interface; Based on the dimension association mapping table, define the inter-dimensional association analysis rules to obtain the inter-dimensional association rule set; By integrating the 3D retrieval interface with the set of rules for inter-dimensional association, a 3D association analysis engine is obtained.

9. The method for early warning of utility tunnel resource status based on intrinsic voiceprint analysis according to claim 8, characterized in that, The method for outputting target early warning events and optimizing risk levels based on a 3D correlation analysis engine is as follows: A sliding event window is established based on a 3D correlation analysis engine to perform spatiotemporal consistency verification on the types of early warning events and their corresponding risk levels: the geospatial index sequence within the time window corresponding to the time series index is obtained through spatiotemporal correlation rules, and it is determined whether the geospatial index sequence meets the preset spatial continuity conditions; when the spatial continuity conditions are met, the early warning event type and its corresponding risk level are determined as the event type and risk level after spatiotemporal verification. A two-stage analysis mechanism of initial identification and deep correlation is implemented for event types after spatiotemporal verification: In the initial identification stage, voiceprint-event matching verification is performed based on voiceprint feature vectors and a multi-dimensional event fingerprint database to obtain the initial event type; in the deep correlation stage, geospatial index and time series index are called based on the dimensional correlation mapping table to obtain the associated event records of the initial event type within a preset historical time window; when the number of associated event records reaches a preset correlation threshold, the initial event type is determined as the target warning event; The risk correction coefficient is calculated based on the target early warning event. The risk level after spatiotemporal verification is then weighted and corrected according to the risk correction coefficient to obtain the optimized risk level.

10. A utility tunnel resource status early warning system based on intrinsic voiceprint analysis, implementing the utility tunnel resource status early warning method based on intrinsic voiceprint analysis as described in any one of claims 1-9, characterized in that, The system includes: The sensor signal acquisition module is used to acquire the light source signal. It adopts a two-stage beam splitting and amplification architecture to amplify the light source signal and obtain the amplified light source signal. The amplified light source signal is then split into two optical paths by a beam splitter. Each optical path is independently configured with an acousto-optic modulator and a photodetector. A parallel monitoring channel is constructed using an optical switch to acquire the distributed acoustic signature sensing signal of the corresponding monitoring position of each monitoring channel in real time. The voiceprint restoration and enhancement module is used to perform phase demodulation and polarization state demodulation on the distributed voiceprint sensing signal, extract frequency information, amplitude information and phase information, process the frequency information to obtain enhanced frequency information, and fuse the enhanced frequency information, amplitude information and phase information to obtain the original audio signal; The intelligent identification and early warning module is used to build an audio-driven AI early warning model. The original audio signal is input into the audio-driven AI early warning model, and the original multi-dimensional acoustic features are extracted from the original audio signal. The original multi-dimensional acoustic features are matched with the multi-dimensional event fingerprint database in the audio-driven AI early warning model to determine the type of early warning event and the corresponding risk level of the current utility tunnel resource status. The 3D correlation analysis module is used to obtain the voiceprint features, alarm point geographic information coordinates and timestamps corresponding to the original audio signal, construct a 3D correlation analysis engine of voiceprint features-geographic information-timestamp, and output target early warning events and optimize risk levels based on the 3D correlation analysis engine; The intelligent decision response module is used to match the corresponding early warning response strategy from the preset early warning response strategy library based on the target early warning event and the optimized risk level, and generate an alarm work order to push to the operation and maintenance terminal.

Images