Method and system for ai sound online analysis and recognition based on fiber sensing

Abstract

The application discloses an AI sound online analysis and identification method and system based on optical fiber sensing, which comprises the following steps: capturing sound signals of a target area by a distributed optical fiber sensing unit and converting the sound signals into optical fiber transmission optical signals, and completing photoelectric signal conversion to form an acoustic electric signal stream through a sound signal conversion transmission unit. Hierarchical voiceprint feature extraction is sequentially performed through an optical fiber voiceprint feature self-encoding model, feature refinement and redundancy elimination are realized by using a voiceprint feature distillation optimization algorithm, time sequence feature correlation is constructed through a sound signal time sequence prediction identification model, and finally online analysis and category determination are completed by a sound signal AI real-time identification engine. The application fully adapts to the characteristics of optical fiber sensing signals, strengthens the synergy of feature processing and time sequence modeling, improves the identification response speed and accuracy, effectively solves the problems of insufficient feature synergy and time sequence adaptation in the prior art, and can be widely applied to online analysis of sound signals in complex scenes such as industrial monitoring and security warning.

Description

Technical Field

[0001] This invention relates to the field of sound analysis and recognition technology, and in particular to an AI-based online sound analysis and recognition method and system based on fiber optic sensing. Background Technology

[0002] With the increasing demands for real-time and accurate sound signal analysis in fields such as industrial monitoring, security early warning, and environmental sensing, traditional sound sensing technologies are gradually showing limitations in terms of adaptability to complex environments and stability over long distances. Fiber optic sensing technology, with its advantages of resistance to electromagnetic interference, low transmission loss, and distributed sensing, has become the preferred solution for sound signal capture. Furthermore, the development of artificial intelligence technology provides core support for the accurate extraction and rapid identification of voiceprint features. AI-based online sound analysis and recognition technology, based on fiber optic sensing, combines the physical sensing capabilities of fiber optic sensors with the intelligent analysis capabilities of AI models. This enables real-time capture, feature extraction, refinement, optimization, and category determination of sound signals in complex scenarios, meeting the urgent needs of multiple fields for online sound signal monitoring. Its technical system includes key components such as sensing units, signal conversion, feature processing, model building, and real-time recognition, forming a complete sound signal analysis chain.

[0003] Existing technologies have two significant drawbacks: First, the synergy between voiceprint feature extraction and optimization is insufficient. Traditional technologies lack deep integration between feature extraction models and optimization algorithms. The feature extraction process fails to fully incorporate the inherent characteristics of fiber optic sensing signals, such as amplitude, frequency, and phase, resulting in a large amount of redundant information in high-dimensional feature vectors. Key voiceprint features are masked, affecting subsequent recognition accuracy. Second, the adaptability of temporal feature modeling to real-time recognition is lacking. Existing methods do not specifically design temporal prediction models for the continuous transmission characteristics of fiber optic sensing acoustic signals, and the interaction link with the real-time recognition engine lacks an efficient collaborative mechanism. This makes it difficult for the modeling results of temporal features to quickly adapt to recognition needs, making it difficult to meet the real-time requirements of online acoustic signal analysis in complex scenarios and limiting the application scope of the technology in highly dynamic scenarios. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of existing technologies, the purpose of this invention is to provide an AI-based online sound analysis and recognition method and system based on fiber optic sensing.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The AI-based online sound analysis and recognition method based on fiber optic sensing includes the following steps: S1, captures acoustic signals in the target area through distributed optical fiber sensing units, and dynamically captures scattered light signals during the sound-to-light conversion process using an optical fiber sensing array, thereby obtaining optical fiber transmission light signals including acoustic signature features. S2, the optical signal transmitted through the optical fiber is transmitted to the signal receiving module of the AI ​​real-time recognition engine for acoustic signals. The conversion from optical signal to electrical signal is completed through the conversion link formed by the optical fiber circulator and the balanced detector, forming a continuous acoustic-electric signal flow. S3, call the fiber optic acoustic signature feature autoencoder model to extract features from the acoustic signal stream, and perform hierarchical mapping of the amplitude, frequency and phase features related to acoustic signature in the signal stream through the encoder network layer to generate a high-dimensional acoustic signature feature vector. S4, the high-dimensional voiceprint feature vector is compressed and its features are enhanced by the voiceprint feature distillation optimization algorithm. The feature information is filtered and refined by dual-path parallel computing of the teacher network and the student network. S5, the access sound signal time-series prediction and recognition model performs time-series correlation analysis on the refined feature vector, and completes the time-series feature modeling of the sound signal through the sliding time window and feature sequence matching mechanism; S6 uses an AI real-time acoustic signal recognition engine to perform online analysis and recognition of the time-series modeling results. It completes the classification and information feedback of the acoustic signal through feature matching threshold determination and recognition result output link.

[0007] Preferably, the expression for the fiber acoustic signature feature autoencoder model in S3 is: in, The high-dimensional voiceprint feature vector output by the type. For the input acoustic signal stream, These are the feature weight coefficients. It is the Sigmoid activation function. These are the weight parameters of the LSTM encoder network layer. For Long Short-Term Memory (LSTM) encoder functions, For time step parameters, For bias terms, The number of feature mapping layers. For the first Layer feature adjustment coefficient, This is the fiber acoustic signature feature mapping function. For the first The fiber optic sensing characteristic adaptation parameters of the layer.

[0008] Preferably, the expression for the voiceprint feature distillation optimization algorithm in S4 is: ,in, This is the refined feature vector after distillation. This is the distillation equilibrium coefficient. For temperature parameters, For normalization function, Output feature vectors for the teacher network. Output feature vectors for the student network. The number of feature dimensions. For the first Distillation weights of dimensional features It is the L2 norm. This is the distillation regularization parameter.

[0009] Preferably, the expression for the acoustic signal temporal prediction and recognition model in S5 is: ,in, for The temporal prediction feature results at each time point. These are the convolution weight coefficients. This is a one-dimensional convolution operation function. for The feature vector at time step, For the kernel size parameter, For attention mechanism weights, The function for calculating temporal attention is... This is the attention decay coefficient. For time-series changes in weight, for Time and The characteristic changes over time.

[0010] Preferably, the sound signal AI real-time recognition engine in S6 adopts a multi-channel parallel processing architecture, and the recognition expression is: ,in To identify the output results, A set of sound signal categories, For specific category labels, and To identify the weighting coefficients, This is the cosine similarity calculation function. For category Standard feature vector library, It is an L1 norm.

[0011] Preferably, the optical signal to electrical signal conversion process in S2 incorporates a dynamic gain adjustment mechanism, the adjustment expression of which is: ,in The adjusted electrical signal output gain. The initial gain coefficient, To convert the optical signal into an initial electrical signal, This refers to the fiber optic intensity adaptation coefficient. For the input optical signal power, To adjust the sensitivity for gain, This is the amplitude of the reference electrical signal.

[0012] Preferably, step S3 includes the following sub-steps: S31, receiving the acoustic-electric signal stream output from S2 through the input layer of the fiber optic acoustic signature feature autoencoder model, and orderly cutting the continuous signal stream according to a preset signal segment length to form a fixed-length acoustic-electric signal segment sequence; S32, inputting the acoustic-electric signal segment sequence into the first convolutional layer of the encoder network, and preliminarily extracting local acoustic features in the signal segments through the set convolution kernel parameters to form a low-dimensional local feature matrix; S33, inputting the local feature matrix into the deep LSTM layer of the encoder network, and capturing the temporal correlation of the feature matrix through the time step parameter control to generate an intermediate feature vector including temporal information; S34, inputting the intermediate feature vector into the output layer of the encoder network, and completing the final generation and output of the high-dimensional acoustic signature feature vector through the combined action of the activation function and weight parameters.

[0013] Preferably, step S4 includes the following sub-steps: S41, starting the teacher network of the voiceprint feature distillation optimization algorithm, taking the high-dimensional voiceprint feature vector output from S3 as input, and fully mining and strengthening the calibration information in the feature vector through a deep network structure to generate a teacher feature vector rich in details; S42, starting the student network of the algorithm, using a simplified network structure to quickly process the same high-dimensional voiceprint feature vector to generate a simplified student feature vector; S43, through temperature parameter adjustment and loss function calculation, quantitatively analyzing the difference between the teacher feature vector and the student feature vector to determine the key areas and intensity of feature distillation; S44, based on the quantitative analysis results, optimizing and adjusting the student feature vector through a weight allocation mechanism, integrating the calibration information in the teacher feature vector, and generating a refined voiceprint feature vector.

[0014] Preferably, step S5 includes the following sub-steps: S51, arranging the refined feature vectors output from S4 in chronological order to construct a continuous feature time sequence, and determining the size and step size of the sliding time window; S52, inputting the feature time sequence into the input layer of the acoustic signal time sequence prediction and recognition model, and segmenting the sequence segment by segment through the sliding time window to form multiple overlapping time sequence feature segments; S53, performing correlation calculation on each time sequence feature segment, distinguishing the importance of features at each time point within the segment through the attention mechanism in the model, and assigning different weight coefficients; S54, inputting the weighted time sequence feature segments into the prediction layer of the model, and completing the time sequence pattern mining of the feature segments through the time sequence modeling algorithm to generate the time sequence prediction result corresponding to each segment.

[0015] This system is an AI-based online sound analysis and recognition system based on fiber optic sensing. It includes: a distributed fiber optic acoustic signal capture unit connected to an acoustic signal conversion and transmission unit, which captures acoustic signals from a target area via a fiber optic sensor array and converts them into optical signals for fiber optic transmission; an acoustic signal conversion and transmission unit connected to both the distributed fiber optic acoustic signal capture unit and the acoustic signature feature extraction unit, which converts the optical signals into an acoustic-electric signal stream via a fiber optic circulator and a balanced detector; and an acoustic signature feature extraction unit connected to both the acoustic signal conversion and transmission unit and the acoustic signature feature refinement unit. The system employs a fiber optic acoustic signature feature autoencoder model to extract hierarchical features from the acoustic signal stream. An acoustic signature feature refinement unit, connected to both the acoustic signature feature extraction unit and the temporal feature modeling unit, filters and refines high-dimensional acoustic signature feature vectors using an acoustic signature feature distillation optimization algorithm. A temporal feature modeling unit, connected to both the acoustic signature feature refinement unit and the AI ​​real-time recognition unit, performs temporal correlation analysis on the refined feature vectors using an acoustic signal temporal prediction and recognition model. Finally, an AI real-time recognition unit, connected to the temporal feature modeling unit, performs online analysis and category determination of the temporal modeling results using an acoustic signal AI real-time recognition engine, and outputs the recognition results.

[0016] Beneficial effects This invention proposes an AI-based online sound analysis and recognition method and system based on fiber optic sensing. Through the collaborative operation of distributed fiber optic sensing units and sound signal conversion and transmission units, it fully leverages the anti-interference and low-loss characteristics of fiber optic sensing to ensure the stability and integrity of sound signal capture in complex environments. By deeply integrating a fiber optic acoustic signature feature autoencoding model with an acoustic signature feature distillation optimization algorithm, it fully adapts to the inherent characteristics of fiber optic sensing signals during feature extraction. Through hierarchical feature mapping and key information refinement, it significantly reduces redundant components in feature vectors, highlighting core acoustic signature features and completely solving the problem of insufficient synergy between feature extraction and optimization in traditional technologies. Finally, it achieves recognition through sound signal temporal prediction. The model is specifically designed for the continuous transmission characteristics of acoustic signals from fiber optic sensors. Combined with an efficient interactive link with an AI real-time acoustic signal recognition engine, it achieves accurate adaptation between time-series feature modeling results and real-time recognition requirements, significantly improving the response speed of online acoustic signal analysis and effectively compensating for the shortcomings of existing technologies in the adaptability of time-series modeling and real-time recognition. At the same time, the entire technical system forms a closed-loop operation from signal capture, conversion, feature processing to recognition output through the organic connection of various units, models, and algorithms. This not only improves the accuracy of acoustic signal recognition but also enhances the adaptability to complex and highly dynamic scenarios, fully meeting the core needs of online acoustic signal analysis in multiple fields such as industrial monitoring and security early warning. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method steps of the present invention.

[0018] Figure 2 This is a system unit composition diagram of the present invention; Detailed Implementation It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, the AI-based online sound analysis and recognition method based on fiber optic sensing includes the following steps: S1, capturing sound signals in the target area using distributed fiber optic sensing units, dynamically capturing scattered light signals during the sound-to-light conversion process using a fiber optic sensing array, and obtaining fiber optic transmitted optical signals including acoustic signature features; S2, transmitting the fiber optic transmitted optical signal to the signal receiving module of the sound signal AI real-time recognition engine, completing the conversion from optical signal to electrical signal through a conversion link composed of a fiber optic circulator and a balanced detector, forming a continuous acoustic-electric signal stream; S3, calling the fiber optic acoustic signature feature autoencoder model to extract features from the acoustic-electric signal stream, and using an encoder network layer to analyze the amplitude of acoustic signature-related features in the signal stream. S4. Values, frequencies, and phase features are hierarchically mapped to generate high-dimensional voiceprint feature vectors; S5. The high-dimensional voiceprint feature vectors are compressed and calibrated using a voiceprint feature distillation optimization algorithm, and feature information is filtered and refined through dual-path parallel computing of teacher and student networks; S6. The refined feature vectors are analyzed for temporal correlation by accessing the audio signal temporal prediction and recognition model, and the temporal feature modeling of the audio signal is completed through a sliding time window and feature sequence matching mechanism; S7. The temporal modeling results are analyzed and recognized online by the audio signal AI real-time recognition engine, and the category determination and information feedback of the audio signal are completed through feature matching threshold determination and recognition result output link.

[0020] Step S1 is implemented by using a distributed optical fiber sensing unit to capture acoustic signals in the target area and acquire optical signals transmitted through optical fibers. In the specific implementation process, the distributed optical fiber sensing unit uses an optical fiber sensing array as the core sensing component. The sensing spacing of the array is set to 1.2 meters, and the array covers a range of 50 meters to 2000 meters of the target monitoring area. It can realize distributed continuous capture of acoustic signals in the target area, and the frequency range of the captured acoustic signals covers 20 Hz to 20 kHz, which fully matches the frequency range of various sound signals in the daily environment. During the acoustic-to-optical conversion process, the sensing array receives the acoustic pressure signal propagating in the target area. The acoustic pressure signal acts on the center wavelength of the optical fiber reflection, causing a wavelength shift in the reflected optical signal. The magnitude of the shift is linearly related to the intensity of the acoustic pressure signal. The sensing unit captures the dynamic changes of this wavelength shift to achieve the conversion of the acoustic signal into an optical signal, and finally obtains the optical fiber transmission optical signal containing complete acoustic signature features. The optical power of this optical signal is maintained in the range of 10 milliwatts to 15 milliwatts to ensure the stability and recognizability of the signal during transmission, providing a raw and complete input data source for subsequent signal conversion and feature processing.

[0021] Step S2 performs the conversion and transmission of optical signals from fiber optic transmission to acoustic-electric signals. This is carried out using a conversion link consisting of the signal receiving module of the acoustic signal AI real-time recognition engine, an optical fiber circulator, and a balanced detector. Specifically, port 1 of the optical fiber circulator is connected to the output of the distributed optical fiber sensing unit to receive the optical signals transmitted through the fiber optic transmission and transmit them to port 2. Port 2 is connected to the optical signal input of the balanced detector. The response time of the balanced detector is set to 1 nanosecond to 5 nanoseconds, which enables rapid response and conversion of optical signals. The balanced detector performs a preliminary conversion from optical to electrical signals through the photoelectric effect, generating an initial electrical signal with an amplitude ranging from 0.1 volts to 5 volts. Subsequently, a dynamic gain adjustment mechanism is introduced into the conversion link to perform gain adaptation processing on the initial electrical signal. During the adjustment process, the amplitude of the reference electrical signal is maintained at 2.5 volts. The gain adjustment sensitivity parameter is adapted to the changes in fiber optic intensity adaptation coefficient and input optical signal power, ultimately forming a continuous acoustic-electric signal stream. The transmission rate of the signal stream is set to 100 megabits per second. During transmission, the signal is transmitted through an optical fiber medium to achieve distortion-free transmission, ensuring that the acoustic-electric signal stream completely retains the characteristic information of the original acoustic signal, providing a high-quality signal input for subsequent voiceprint feature extraction.

[0022] Step S3 completes the hierarchical acoustic signature feature extraction of the acoustic signal stream through the fiber optic acoustic signature feature autoencoder model. In the specific implementation process, the input layer of the model receives the acoustic signal stream output from step S2. The signal segment length is preset to 20 milliseconds. The continuous signal stream is cut into segments of fixed length according to this length to form a sequence of acoustic signal segments. Each segment sequence includes 441 sampling points and the sampling frequency is 22.05 kHz. The first layer of the encoder network uses 16 convolutional kernels with a size of 3×3 and a stride of 1 to initially extract local voiceprint features from the signal segment, forming a low-dimensional local feature matrix with dimensions of 16×147×1. This local feature matrix is ​​then fed into the deep LSTM layer of the encoder network. The LSTM layer contains 128 hidden units with a time step parameter of 10, which controls the capture of the temporal correlation of the feature matrix, generating an intermediate feature vector containing temporal information. This intermediate feature vector has a dimension of 128×1. Finally, the intermediate feature vector is input into the output layer of the encoder network. The output layer uses a sigmoid activation function, combined with weight parameters, to complete the final generation and output of a high-dimensional voiceprint feature vector with a dimension of 256, ensuring that all core voiceprint feature information, including the amplitude, frequency, and phase of the sound signal, is included.

[0023] Step S4 uses a voiceprint feature distillation optimization algorithm to filter and refine the high-dimensional voiceprint feature vector output in step S3. In the specific implementation process, the algorithm first starts the teacher network and the student network. The teacher network adopts a deep network structure including 5 fully connected layers, with the number of neurons in each fully connected layer being 256, 128, 64, 32, and 16 respectively. The high-dimensional voiceprint feature vector is used as input, and the key information in the feature vector is fully mined and enhanced to generate a teacher feature vector rich in details. The student network adopts a simplified structure including 3 fully connected layers, with the number of neurons in each layer being 128, 64, and 32 respectively. The same high-dimensional voiceprint feature vector is processed quickly to generate a simplified student feature vector. The algorithm sets the temperature parameter to 10 and uses this parameter to quantify the differences between the teacher's and student's feature vectors. The L2 norm is used to calculate these differences, and a distillation weight is assigned to each dimension corresponding to the feature dimension. The weight value is dynamically allocated based on the importance of the feature in that dimension. After the quantification analysis is completed, the student's feature vector is optimized and adjusted through the weight allocation mechanism, while key information from the teacher's feature vector is integrated. Finally, a refined voiceprint feature vector is generated, with the vector dimension compressed to 64, significantly reducing redundant information and retaining core voiceprint features. This provides efficient and accurate feature input for subsequent time-series feature modeling.

[0024] Step S5 uses the acoustic signal temporal prediction and recognition model to perform temporal correlation analysis on the refined feature vectors output in step S4. In the specific implementation process, the refined feature vectors are first arranged in chronological order to construct a continuous feature temporal sequence with a sequence length of 500. The sliding time window size is determined to be 50 and the step size is 10. Then, the feature temporal sequence is input into the input layer of the model. The sequence is segmented by segment through the sliding time window to form multiple overlapping temporal feature segments. Each segment includes 50 continuous refined feature vectors. The attention mechanism layer in the model includes 32 attention heads, which distinguish the importance of features at different time points within a segment and assign different weight coefficients to each feature. The weight values ​​are determined based on the contribution of the feature to the time sequence. The weighted time-series feature segments are then input into the prediction layer of the model. The prediction layer employs a time-series modeling algorithm, which combines one-dimensional convolution with attention calculation. The convolution kernel size is set to 3. The time-series modeling algorithm completes the time-series pattern mining of feature segments and generates time-series prediction results for each segment. The dimension of the prediction results is consistent with the dimension of the refined feature vector, ensuring complete preservation of time-series feature information and providing accurate time-series feature data support for subsequent real-time recognition.

[0025] In step S6, the real-time AI recognition engine for sound signals performs online analysis and category determination on the time-series modeling results output in step S5. In the specific implementation process, the real-time AI recognition engine for sound signals adopts a multi-channel parallel processing architecture, including 8 parallel processing channels. Each channel corresponds to a category of sound signals. The set of sound signal categories includes 32 categories such as industrial equipment operation sounds, environmental natural sounds, and human activity sounds. The engine first loads a preset standard feature vector library, which includes standard feature vectors corresponding to each category of sound signals. The vector dimension is consistent with the dimension of the time-series prediction results. Subsequently, a multi-channel parallel processing mechanism is used to match the time-series prediction results with each standard feature vector in the standard feature vector library. During the calculation, cosine similarity and L1 norm are combined to assign different weights to different matching indicators. The weight of cosine similarity is 0.6 and the weight of L1 norm is 0.4. The matching score of each category is obtained by weighted calculation. The category with the lowest matching score is selected as the recognition result of the sound signal. The response time of the entire recognition process is controlled within 10 milliseconds, realizing online real-time analysis and category determination of sound signals. Finally, the output link of the recognition result is completed, and the determination result is fed back to the front-end monitoring system or data storage unit to realize the closed-loop processing of sound signal analysis.

[0026] The expression for the fiber optic acoustic signature feature autoencoder model in S3 is as follows: in, The high-dimensional voiceprint feature vector output by the type. For the input acoustic signal stream, These are the feature weight coefficients. It is the Sigmoid activation function. These are the weight parameters of the LSTM encoder network layer. For Long Short-Term Memory (LSTM) encoder functions, For time step parameters, For bias terms, The number of feature mapping layers. For the first Layer feature adjustment coefficient, This is the fiber acoustic signature feature mapping function. For the first The fiber optic sensing characteristic adaptation parameters of the layer.

[0027] Specifically, the fiber optic acoustic signature feature autoencoder model in step S3 achieves the generation of high-dimensional acoustic signature feature vectors of acoustic and electrical signal streams through multi-layer network structure and parameter coordination, and the core parameter settings are tailored to the characteristics of fiber optic sensing acoustic signals. The model input is the acoustic-electric signal stream output from step S2, and the feature weight coefficient is set to 0.85 to ensure the dominance of core information during feature extraction. The Sigmoid activation function is used to perform nonlinear mapping on the encoding result to enhance the discriminative power of the features. The LSTM encoder network layer weight parameter is set to 0.7, and the time step parameter is set to 12 to adapt to the continuous temporal characteristics of the acoustic-electric signal stream and achieve effective capture of acoustic features at different time points. The bias term is set to 0.15 to fine-tune the encoding output and reduce system errors. The number of feature mapping layers is set to 8, and the deep features of the acoustic signal are gradually mined through multi-level mapping. The feature adjustment coefficient of the k-th layer increases from 0.1 to 0.9 in a hierarchical manner to ensure the basic role of shallow features and the enhancement effect of deep features. The fiber optic acoustic feature mapping function is specifically adapted to the amplitude, frequency, and phase characteristics of the fiber optic sensing signal. The fiber optic sensing feature adaptation parameter of the k-th layer is dynamically adjusted according to the feature type of the corresponding layer, with a value range between 0.2 and 0.8. This model achieves accurate extraction and high-dimensional mapping of voiceprint features in acoustic signal streams through the synergistic effect of parameters. The generated high-dimensional voiceprint feature vector completely retains the core information of the original acoustic signal, providing high-quality input for subsequent feature refinement. Its implementation method is to embed the voiceprint feature extraction unit through software programming and complete real-time calculation using hardware computing resources.

[0028] The expression for the voiceprint feature distillation optimization algorithm in S4 is as follows: ,in, This is the refined feature vector after distillation. This is the distillation equilibrium coefficient. For temperature parameters, For normalization function, Output feature vectors for the teacher network. Output feature vectors for the student network. The number of feature dimensions. For the first Distillation weights of dimensional features It is the L2 norm. This is the distillation regularization parameter.

[0029] Specifically, in step S4, the voiceprint feature distillation optimization algorithm refines high-dimensional feature vectors through multi-parameter collaboration and a dual-network architecture. Parameter settings balance feature preservation and computational efficiency. The distillation balance coefficient is set to 0.6 to reasonably allocate the weight ratio of the output features of the teacher network and the student network; the temperature parameter is set to 15 to adjust the output distribution of the Softmax function, allowing the student network to learn the feature information of the teacher network more fully; the output feature vector of the teacher network originates from the comprehensive mining of high-dimensional voiceprint features by the deep network, while the output feature vector of the student network is quickly generated by simplifying the network structure, and both dimensions are consistent with the input high-dimensional feature vector; the feature dimension is set to 256, matching the dimension of the high-dimensional voiceprint feature vector output in step S3; the distillation weight of the i-th feature is allocated according to the importance of that dimension feature, with a value range between 0.001 and 0.01, and higher weights corresponding to core features; the L2 norm is used to calculate the difference between the output feature vectors of the teacher network and the student network, quantifying the feature deviation between the two; the distillation regularization parameter is set to 0.005 to prevent the algorithm from overfitting and improve the stability of feature refinement. The algorithm is implemented through a dual-network parallel computing architecture. The teacher network and the student network process the same high-dimensional voiceprint feature vector simultaneously. Based on the parameters, feature difference calculation, weight allocation and feature fusion are completed, and finally a refined feature vector is output. This effectively removes redundant information and strengthens the core voiceprint features. The implementation process utilizes the hardware resources of the voiceprint feature refinement unit to ensure a balance between computational efficiency and feature refinement effect.

[0030] The expression for the acoustic signal time-series prediction and recognition model in S5 is as follows: ,in, for The temporal prediction feature results at each time point. These are the convolution weight coefficients. This is a one-dimensional convolution operation function. for The feature vector at time step, For the kernel size parameter, For attention mechanism weights, The function for calculating temporal attention is... This is the attention decay coefficient. For time-series changes in weight, for Time and The characteristic changes over time.

[0031] Specifically, the acoustic signal temporal prediction and recognition model in step S5 achieves accurate prediction of temporal features through parameter optimization and algorithm design, with parameter settings adapted to the continuous transmission characteristics of fiber optic sensing acoustic signals. The convolution weight coefficient is set to 0.75 to highlight the role of one-dimensional convolution operation in extracting temporal features; the one-dimensional convolution operation function uses a 3×3 convolution kernel size and a stride of 1 to extract local features from the feature vector of the previous time step and capture the correlation between features at adjacent time steps; the feature vector at time t-1 is used as the temporal input of the model, maintaining the same dimensionality as the feature vector at time t, both being 64-dimensional; the attention mechanism weight is set to 0.6, strengthening the weight ratio of key temporal features through the attention calculation function to improve prediction accuracy; the attention decay coefficient is set to 0.05 to adjust the decay rate of the attention weight, avoiding over-focusing on recent features while ignoring historical features; the temporal change weight is set to 0.5 to balance the impact of temporal change on the prediction results; the feature change at time t and time t-1 is calculated by the difference between the two, reflecting the dynamic change trend of features at adjacent time steps. In the implementation of this model, the feature vector at time t-1 is first loaded, local temporal features are extracted through one-dimensional convolution operation, and different weights are assigned to features at different times through an attention mechanism. Combined with the temporal change, the temporal prediction feature result at time t is generated. The implementation architecture embeds a temporal feature modeling unit, and the real-time modeling and prediction of temporal features are realized through hardware computing power support, providing accurate temporal feature data for subsequent real-time identification.

[0032] The S6's real-time AI recognition engine for sound signals employs a multi-channel parallel processing architecture, and its recognition expression is as follows: ,in To identify the output results, A set of sound signal categories, For specific category labels, and To identify the weighting coefficients, This is the cosine similarity calculation function. For category Standard feature vector library, It is an L1 norm.

[0033] Specifically, in step S6, the recognition mechanism and parameter configuration of the real-time AI recognition engine for sound signals achieve rapid recognition of sound signals through multi-parameter collaboration and category matching algorithms. The parameter settings balance recognition accuracy and response speed. The sound signal category set includes 32 common sound signal categories, including typical sound types in scenarios such as industrial monitoring and security early warning. Specific category labels assign unique identifiers to each sound signal category, facilitating the output and feedback of recognition results. In the recognition weight coefficients, the cosine similarity weight is set to 0.6, and the L1 norm weight is set to 0.4. By weighted fusion of the calculation results of the two matching indicators, the accuracy of category determination is improved. The cosine similarity calculation function is used to measure the similarity between the time-series predicted feature results and the standard feature vectors of each category in the standard feature vector library. The value ranges from -1 to 1, with a higher matching degree indicating a closer value to 1. The standard feature vector library includes standard feature vectors corresponding to 32 types of sound signals. Each vector has a 64-dimensional dimension, consistent with the dimension of the time-series predicted feature results. The L1 norm is used to calculate the absolute error between the time-series predicted feature results and the standard feature vectors, quantifying the feature differences between the two. The engine is implemented through a multi-channel parallel processing architecture. It simultaneously matches the time-series predicted feature results with all vectors in the standard feature vector library, calculates weighted scores based on parameters, and selects the category corresponding to the highest score as the recognition output. The implementation process utilizes the hardware resources of the AI ​​real-time recognition unit to ensure that the recognition response time is controlled within 10 milliseconds, thereby achieving online real-time recognition of sound signals.

[0034] In S2, the optical signal to electrical signal conversion process introduces a dynamic gain adjustment mechanism, the adjustment expression of which is: ,in The adjusted electrical signal output gain. The initial gain coefficient, To convert the optical signal into an initial electrical signal, This refers to the fiber optic intensity adaptation coefficient. For the input optical signal power, To adjust the sensitivity for gain, This is the amplitude of the reference electrical signal.

[0035] Specifically, in step S2, the optical signal to electrical signal conversion process incorporates a dynamic gain adjustment mechanism. Parameter optimization ensures the stability and adaptability of the electrical signal output, with parameter settings tailored to the dynamic characteristics of optical signal transmission via fiber optics. The initial gain coefficient is set to 10, providing a basic gain for the conversion of the optical signal to the initial electrical signal. The initial electrical signal originates from the photoelectric conversion output of the balanced detector, with an amplitude range between 0.1 volts and 5 volts. The fiber optic intensity adaptation coefficient is set to 0.8 to adapt to the intensity variations of the optical signal transmitted via fiber optics, ensuring the effectiveness of gain adjustment under different light intensities. The input optical signal power is the actual power of the optical signal transmitted via fiber optics, ranging from 10 milliwatts to 15 milliwatts, and is acquired in real-time by an optical power detector. The gain adjustment sensitivity is set to 0.01 to control the fineness of gain adjustment, ensuring the smoothness of the electrical signal output. The reference electrical signal amplitude is fixed at 2.5 volts, serving as the benchmark value for gain adjustment, stabilizing the output electrical signal amplitude near this benchmark value. During the implementation of this mechanism, the input optical signal power and the amplitude of the initial electrical signal are collected in real time. The dynamic gain value is calculated based on the parameters, and the gain of the initial electrical signal is adjusted in real time through the gain adjustment module. Finally, a stable acoustic-electric signal stream is output. The implementation process is integrated into the acoustic signal conversion and transmission unit to ensure the efficiency of the optical-electric conversion process and the stability of the signal output, providing a high-quality electrical signal input for subsequent feature extraction.

[0036] S3 includes the following sub-steps: S31, receiving the acoustic-electric signal stream output from S2 through the input layer of the fiber optic acoustic signature feature autoencoder model, and orderly cutting the continuous signal stream according to the preset signal segment length to form a fixed-length acoustic-electric signal segment sequence; S32, inputting the acoustic-electric signal segment sequence into the first convolutional layer of the encoder network, and preliminarily extracting local acoustic features in the signal segments through the set convolution kernel parameters to form a low-dimensional local feature matrix; S33, inputting the local feature matrix into the deep LSTM layer of the encoder network, and capturing the temporal correlation of the feature matrix through the time step parameter control to generate an intermediate feature vector including temporal information; S34, inputting the intermediate feature vector into the output layer of the encoder network, and completing the final generation and output of the high-dimensional acoustic signature feature vector through the combined action of the activation function and weight parameters.

[0037] Specifically, the feature extraction process of the fiber optic acoustic signature autoencoder model in step S3 achieves hierarchical feature extraction through parameter setting and process integration. In S31, the model input layer accurately receives the acoustic signal stream output from step S2. The preset signal segment length is 20 milliseconds. The continuous and uninterrupted acoustic signal stream is cut into orderly and non-overlapping segments according to this fixed length to form a standardized acoustic signal segment sequence. The number of sampling points in each segment sequence is strictly controlled to 441, and the sampling frequency is set to 22.05 kHz to ensure that each segment completely retains the acoustic signature information within the corresponding time period. In S32, the generated acoustic signal segment sequence is input into the first convolutional layer of the encoder network. This convolutional layer is configured with 16 convolutional kernels, with a kernel size of 3×3 and a stride of 1. The local acoustic signature features in the signal segments are initially screened and extracted through convolution operations, and transformed into a low-dimensional local feature matrix with a fixed matrix dimension of 16×147×1, realizing the initial dimensionality reduction and structuring of the acoustic signature features. In S33, the low-dimensional local features are processed... The matrix is ​​fed into the deep LSTM layer of the encoder network. This LSTM layer includes 128 hidden units, and the time step parameter is set to 10. This parameter precisely controls the depth and range of capturing temporal correlation information in the feature matrix, explores the intrinsic relationship between local features at different time nodes, and generates an intermediate feature vector containing complete temporal information with a vector dimension of 128×1. S34 inputs the intermediate feature vector into the output layer of the encoder network. The output layer embeds a Sigmoid activation function and performs nonlinear transformation and dimension enhancement on the intermediate feature vector with preset fixed weight parameters. Finally, a high-dimensional voiceprint feature vector with a dimension of 256 is generated, which fully includes the core features of the sound signal such as amplitude, frequency, and phase. Each step ensures the completeness and accuracy of feature extraction through data flow and parameter coordination. The implementation process utilizes the hardware computing power of the voiceprint feature extraction unit to complete efficient computation.

[0038] S4 includes the following sub-steps: S41, activating the teacher network of the voiceprint feature distillation optimization algorithm, taking the high-dimensional voiceprint feature vector output from S3 as input, and fully mining and strengthening the calibration information in the feature vector through a deep network structure to generate a teacher feature vector rich in detail; S42, activating the student network of the algorithm, using a simplified network structure to quickly process the same high-dimensional voiceprint feature vector to generate a simplified student feature vector; S43, through temperature parameter adjustment and loss function calculation, quantitatively analyzing the difference between the teacher feature vector and the student feature vector to determine the key areas and intensity of feature distillation; S44, based on the quantitative analysis results, optimizing and adjusting the student feature vector through a weight allocation mechanism, integrating the calibration information in the teacher feature vector, and generating a refined voiceprint feature vector.

[0039] Specifically, the voiceprint feature distillation optimization algorithm in step S4 achieves feature refinement through dual-network collaboration and parameter control. S41 activates the teacher network of the voiceprint feature distillation optimization algorithm. This teacher network employs a deep network structure with 5 fully connected layers, where the number of neurons in each fully connected layer is set to 256, 128, 64, 32, and 16 respectively. The high-dimensional voiceprint feature vector with a dimension of 256 output from step S3 is used as the sole input. Through multi-level operations of the deep network, key information in the feature vector is deeply mined, strengthened, and integrated to generate a teacher feature vector rich in detailed information, with the vector dimension consistent with the input high-dimensional feature vector. S42 simultaneously activates the student network of the algorithm. This student network employs a simplified network structure with 3 fully connected layers, where the number of neurons in each layer is set to 128, 64, and 32 respectively. It performs fast parallel processing on the same high-dimensional voiceprint feature vector, simplifying the computation process while ensuring no loss of core features, generating a simplified student feature vector with a dimension of 32. S43 adjusts So through a preset temperature parameter (valued at 15). The output distribution of the ftmax function, combined with the L2 norm, calculates the difference between the teacher's feature vector and the student's feature vector, and performs quantitative analysis on the feature deviation between the two. This accurately locates the missing key information and redundant components in the student's feature vector, and determines the key areas and reinforcement intensity of feature distillation. Based on the feature difference results obtained from the quantitative analysis, S44 assigns differentiated weights to different feature dimensions through a dynamic weight allocation mechanism. The weight of the core feature dimension ranges from 0.005 to 0.01, and the weight of the non-core feature dimension ranges from 0.001 to 0.003. Based on the weights, the student's feature vector is optimized and adjusted in a targeted manner, and key details in the teacher's feature vector are integrated, while redundant components are removed. Finally, a refined voiceprint feature vector with a dimension of 64 is generated. Each step achieves a balance between feature refinement efficiency and effect through parallel computation of dual networks and precise parameter control. The implementation process utilizes the hardware resources of the voiceprint feature refinement unit to complete real-time processing.

[0040] S5 includes the following sub-steps: S51, arranging the refined feature vectors output from S4 in chronological order to construct a continuous feature time sequence, and determining the size and step size of the sliding time window; S52, inputting the feature time sequence into the input layer of the acoustic signal time sequence prediction and recognition model, and segmenting the sequence segment by segment through the sliding time window to form multiple overlapping time sequence feature segments; S53, performing correlation calculation on each time sequence feature segment, and distinguishing the importance of features at each time point within the segment through the attention mechanism in the model, assigning different weight coefficients; S54, inputting the weighted time sequence feature segments into the prediction layer of the model, and completing the time sequence pattern mining of the feature segments through the time sequence modeling algorithm to generate the time sequence prediction result corresponding to each segment.

[0041] Specifically, the acoustic signal temporal prediction and recognition model modeling process in step S5 achieves accurate modeling through temporal processing and an attention mechanism. S51 first receives the refined feature vectors of dimension 64 output from step S4, arranges them in order according to the time sequence of feature generation, constructing a continuous feature temporal sequence of length 500. Simultaneously, the sliding time window size is preset to 50, and the step size is set to 10, providing standard parameters for subsequent temporal segment truncation. S52 inputs the constructed feature temporal sequence into the input layer of the acoustic signal temporal prediction and recognition model. The continuous feature temporal sequence is truncated segment by segment through the preset sliding time window, forming multiple overlapping temporal feature segments. Each segment includes 50 continuous refined feature vectors, ensuring the continuity and integrity of the temporal information. S53 performs temporal correlation calculation on each temporal feature segment. The attention mechanism layer in the model is configured with 32 attention heads. An attention calculation algorithm quantifies the importance of features at each moment within the segment in the overall temporal pattern, assigning a value to each feature. Differentiated weight coefficients are assigned to features at each time step, with weight values ​​ranging from 0.01 to 0.05. The weight coefficients for core temporal features are higher, strengthening the influence of key temporal information. S54 inputs the weighted temporal feature segments into the prediction layer of the model. This prediction layer integrates a composite algorithm of one-dimensional convolution and attention calculation. The convolution kernel size is set to 3. Local temporal features within the segments are further extracted through one-dimensional convolution operations. The weight allocation results of the attention mechanism are combined to mine the changing patterns and development trends of temporal features, complete the temporal pattern modeling of feature segments, and generate the temporal prediction results corresponding to each segment. The dimension of the prediction results is consistent with the refined feature vector, which is 64. Each step ensures the accuracy and relevance of temporal modeling through the orderly connection of temporal sequence construction, segment truncation, weight allocation, and pattern mining. The implementation process embeds the temporal feature modeling unit and uses hardware computing power to achieve real-time temporal analysis.

[0042] like Figure 2As shown, an AI-based online sound analysis and recognition system based on fiber optic sensing is applied to an AI-based online sound analysis and recognition method based on fiber optic sensing. The system includes: a distributed fiber optic acoustic signal capture unit connected to an acoustic signal conversion and transmission unit, which captures acoustic signals from a target area via a fiber optic sensor array and converts them into optical signals for fiber optic transmission; an acoustic signal conversion and transmission unit connected to both the distributed fiber optic acoustic signal capture unit and the acoustic signature feature extraction unit, which converts the optical signals for fiber optic transmission into an acoustic-electric signal stream via a fiber optic circulator and a balanced detector; and an acoustic signature feature extraction unit connected to both the acoustic signal conversion and transmission unit and the acoustic signature feature refinement unit. A hierarchical feature extraction method is used to extract acoustic and electrical signal streams through an optical fiber acoustic signature feature autoencoder model. An acoustic signature feature refinement unit, connected to both the acoustic signature feature extraction unit and the temporal feature modeling unit, filters and refines high-dimensional acoustic signature feature vectors using an acoustic signature feature distillation optimization algorithm. A temporal feature modeling unit, connected to both the acoustic signature feature refinement unit and the AI ​​real-time recognition unit, performs temporal correlation analysis on the refined feature vectors using an acoustic signal temporal prediction and recognition model. Finally, an AI real-time recognition unit, connected to the temporal feature modeling unit, performs online analysis and category determination of the temporal modeling results using an acoustic signal AI real-time recognition engine, and outputs the recognition results.

[0043] This paper presents an AI-based online sound analysis and recognition method and system based on fiber optic sensing. It employs a distributed fiber optic sensing unit and a sound signal conversion and transmission unit to construct the front-end sensing link. Utilizing a fiber optic sensing array and a dedicated conversion link, it achieves high-fidelity capture and stable transmission of sound signals, fully leveraging the inherent advantages of fiber optic sensing technology, such as resistance to electromagnetic interference and long transmission distance. Through a progressive collaboration of a fiber optic acoustic signature feature self-encoding model, an acoustic signature feature distillation optimization algorithm, and a sound signal temporal prediction and recognition model, it completes a refined process from high-dimensional feature extraction and redundant information filtering to temporal pattern modeling, ensuring the integrity and relevance of acoustic signature features. The multi-channel parallel processing architecture and efficient interaction mechanism of the sound signal AI real-time recognition engine enable rapid connection between feature processing and category determination, balancing analysis accuracy and response efficiency. The entire system features clearly defined and tightly connected units, forming a closed-loop technical link that can adapt to the online sound signal analysis needs of various complex scenarios.

[0044] This method and system address the issue of insufficient synergy between feature extraction and optimization by deeply integrating the fiber optic acoustic signature feature autoencoding model with the acoustic signature feature distillation optimization algorithm. During the feature extraction stage, it fully adapts to the amplitude, frequency, and phase characteristics of the fiber optic sensing signal. Through hierarchical mapping and dual-path network refinement, it accurately selects core acoustic signature features, significantly reducing redundant information interference. Addressing the lack of adaptability between temporal modeling and real-time recognition, the acoustic signal temporal prediction and recognition model is specifically adapted to the continuous transmission characteristics of fiber optic sensing acoustic signals. It strengthens temporal correlation through a sliding time window and attention mechanism. Simultaneously, through the dedicated output link of the acoustic signal AI real-time recognition engine, it achieves rapid matching between modeling results and recognition requirements, significantly improving the real-time performance of online analysis and breaking the application limitations of traditional technologies in highly dynamic scenarios.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based online sound analysis and recognition method based on fiber optic sensing, characterized in that, Includes the following steps: S1, captures acoustic signals in the target area through distributed optical fiber sensing units, and dynamically captures scattered light signals during the sound-to-light conversion process using an optical fiber sensing array, thereby obtaining optical fiber transmission light signals including acoustic signature features. S2, the optical signal transmitted through the optical fiber is transmitted to the signal receiving module of the AI ​​real-time recognition engine for acoustic signals. The conversion from optical signal to electrical signal is completed through the conversion link formed by the optical fiber circulator and the balanced detector, forming a continuous acoustic-electric signal flow. S3, call the fiber optic acoustic signature feature autoencoder model to extract features from the acoustic signal stream, and perform hierarchical mapping of the amplitude, frequency and phase features related to acoustic signature in the signal stream through the encoder network layer to generate a high-dimensional acoustic signature feature vector. S4, the high-dimensional voiceprint feature vector is compressed and its features are enhanced by the voiceprint feature distillation optimization algorithm. The feature information is filtered and refined by dual-path parallel computing of the teacher network and the student network. S5, the access sound signal time-series prediction and recognition model performs time-series correlation analysis on the refined feature vector, and completes the time-series feature modeling of the sound signal through the sliding time window and feature sequence matching mechanism; S6 uses an AI real-time acoustic signal recognition engine to perform online analysis and recognition of the time-series modeling results. It completes the classification and information feedback of the acoustic signal through feature matching threshold determination and recognition result output link.

2. The AI ​​sound online analysis and recognition method based on fiber optic sensing according to claim 1, characterized in that, The expression for the fiber optic acoustic signature autoencoder model in S3 is as follows: in, The high-dimensional voiceprint feature vector output by the type. For the input acoustic-electric signal stream, These are the feature weight coefficients. It is the Sigmoid activation function. These are the weight parameters of the LSTM encoder network layer. For Long Short-Term Memory (LSTM) encoder functions, For time step parameters, For bias terms, The number of feature mapping layers. For the first Layer feature adjustment coefficient, This is the fiber acoustic signature feature mapping function. For the first The fiber optic sensing characteristic adaptation parameters of the layer.

3. The AI ​​sound online analysis and recognition method based on fiber optic sensing according to claim 1, characterized in that, The expression for the voiceprint feature distillation optimization algorithm in S4 is: ,in, This is the refined feature vector after distillation. This is the distillation equilibrium coefficient. For temperature parameters, For normalization function, Output feature vectors for the teacher network. Output feature vectors for the student network. The number of feature dimensions. For the first Distillation weights of dimensional features It is the L2 norm. This is the distillation regularization parameter.

4. The AI ​​sound online analysis and recognition method based on fiber optic sensing according to claim 1, characterized in that, The expression for the acoustic signal temporal prediction and recognition model in S5 is: ,in, for Temporal prediction feature results at each time point. These are the convolution weight coefficients. This is a one-dimensional convolution operation function. for The feature vector at time step, For the kernel size parameter, For attention mechanism weights, For calculating the temporal attention function, This is the attention decay coefficient. For time-varying weights, for Time and The characteristic changes over time.

5. The AI ​​sound online analysis and recognition method based on fiber optic sensing according to claim 1, characterized in that, The real-time AI recognition engine for acoustic signals in S6 adopts a multi-channel parallel processing architecture, and the recognition expression is: ,in To identify the output results, A set of sound signal categories, For specific category labels, and To identify the weighting coefficients, This is the cosine similarity calculation function. For category Standard feature vector library, It is an L1 norm.

6. The AI ​​sound online analysis and recognition method based on fiber optic sensing according to claim 1, characterized in that, The optical signal to electrical signal conversion process in S2 introduces a dynamic gain adjustment mechanism, the adjustment expression of which is: ,in The adjusted electrical signal output gain. The initial gain coefficient, To convert the optical signal into an initial electrical signal, This refers to the fiber optic intensity adaptation coefficient. For the input optical signal power, To adjust the sensitivity for gain adjustment, This is the amplitude of the reference electrical signal.

7. The AI ​​sound online analysis and recognition method based on fiber optic sensing according to claim 1, characterized in that, S3 includes the following steps: S31 receives the acoustic-electric signal stream output by S2 through the input layer of the fiber optic acoustic signature feature autoencoder model, and cuts the continuous signal stream into segments of fixed length according to the preset signal segment length to form a sequence of acoustic-electric signal segments. S32, the sequence of acoustic and electrical signal segments is input into the first convolutional layer of the encoder network, and the local acoustic features in the signal segments are initially extracted by setting the convolution kernel parameters to form a low-dimensional local feature matrix; S33, the local feature matrix is ​​fed into the deep LSTM layer of the encoder network, and the temporal correlation of the feature matrix is ​​captured by controlling the time step parameter to generate an intermediate feature vector including temporal information. S34 inputs the intermediate feature vector into the output layer of the encoder network. Through the combined action of the activation function and weight parameters, the high-dimensional voiceprint feature vector is finally generated and output.

8. The AI ​​sound online analysis and recognition method based on fiber optic sensing according to claim 1, characterized in that, S4 includes the following steps: S41, the teacher network of the voiceprint feature distillation optimization algorithm is started. The high-dimensional voiceprint feature vector output by S3 is used as input. The calibration information in the feature vector is fully mined and enhanced through the deep network structure to generate a teacher feature vector rich in details. S42, Start the student network of the algorithm, use a simplified network structure to quickly process the same high-dimensional voiceprint feature vector, and generate a simplified student feature vector. S43, through temperature parameter adjustment and loss function calculation, quantitatively analyzes the difference between teacher feature vectors and student feature vectors to determine the key areas and intensity of feature distillation; S44, based on the quantitative analysis results, optimizes and adjusts the student feature vector through a weight allocation mechanism, integrates the calibration information in the teacher feature vector, and generates a refined voiceprint feature vector.

9. The AI ​​sound online analysis and recognition method based on fiber optic sensing according to claim 1, characterized in that, S5 includes the following steps: S51, arrange the refined feature vectors output by S4 in chronological order to construct a continuous feature time sequence, and determine the size of the sliding time window and the moving step size; S52, the feature time sequence is input into the input layer of the acoustic signal time sequence prediction and recognition model, and the sequence is segmented by segment through a sliding time window to form multiple overlapping time sequence feature segments; S53, perform correlation calculations on each temporal feature segment, and use the attention mechanism in the model to distinguish the importance of features at each time point within the segment and assign different weight coefficients; S54: The weighted temporal feature segments are input into the prediction layer of the model. The temporal pattern of the feature segments is mined through the temporal modeling algorithm, and the temporal prediction result corresponding to each segment is generated.

10. An AI-based online sound analysis and recognition system based on fiber optic sensing, characterized in that, This system is applied to the AI ​​sound online analysis and recognition method based on fiber optic sensing as described in claim 1, comprising: A distributed fiber optic acoustic signal capture unit, connected to an acoustic signal conversion and transmission unit, captures acoustic signals in the target area through a fiber optic sensor array and converts them into optical signals for fiber optic transmission. The acoustic signal conversion and transmission unit is connected to the distributed optical fiber acoustic signal capture unit and the acoustic fingerprint feature extraction unit, respectively. It converts the optical fiber transmitted optical signal into an acoustic-electric signal stream and transmits it through an optical fiber circulator and a balanced detector. The voiceprint feature extraction unit is connected to the acoustic signal conversion and transmission unit and the voiceprint feature refinement unit, respectively, and performs hierarchical feature extraction on the acoustic-electric signal stream through the fiber optic voiceprint feature autoencoder model. The voiceprint feature refinement unit is connected to the voiceprint feature extraction unit and the temporal feature modeling unit, respectively. It uses the voiceprint feature distillation optimization algorithm to filter and refine the high-dimensional voiceprint feature vector. The temporal feature modeling unit is connected to the voiceprint feature refinement unit and the AI ​​real-time recognition unit, respectively, and performs temporal correlation analysis on the refined feature vector through the audio signal temporal prediction and recognition model; The AI ​​real-time recognition unit is connected to the temporal feature modeling unit. It performs online analysis and category determination of the temporal modeling results through the sound signal AI real-time recognition engine and outputs the recognition results.

Images