Optical fiber vibration signal time identification and classification method, system and device based on Mamba-YOLOv10 neural network and medium

Abstract

The invention discloses an optical fiber vibration signal time identification and classification method, system and device based on a Mam-YOLOv10 neural network and a medium, and belongs to the technical field of optical fiber communication, and the method comprises the steps: obtaining OTDR signals of a plurality of targets based on an optical fiber vibration signal collection system, and carrying out the denoising of the OTDR signals; performing short-time Fourier transform on the de-noised OTDR signal to obtain a two-dimensional time-frequency diagram of the vibration signal; based on the two-dimensional time-frequency diagram, multi-scale features of the optical fiber vibration signals are extracted; and inputting the multi-scale features into a Mam-YOLOv10 neural network containing a Mama attention block to complete identification and classification of the optical fiber vibration signals. According to the method, the influence of interference fading on the pattern recognition accuracy is effectively reduced, meanwhile, the hardware complexity is reduced, the attention weight is automatically adjusted according to different fault scenes and data characteristics, and the accuracy of optical fiber fault recognition and classification is improved.

Description

Technical Field

[0001] This invention relates to the field of power system energy storage equipment technology, specifically to a method, system, device, and medium for time identification and classification of optical fiber vibration signals based on Mamba-YOLOv10 neural network. Background Technology

[0002] Fiber optic networks are one of the fundamental infrastructures of today's information society. Failures can lead to serious problems such as communication interruptions and data loss, making timely and accurate fault diagnosis crucial. Currently, in engineering practice, optical time-domain reflectometers (OTDRs) are commonly used to acquire one-dimensional waveform data reflecting the fiber's condition. Then, manual analysis is performed to determine the type and location of the fault based on characteristics such as reflection peaks and signal attenuation.

[0003] However, in practical applications, it has been found that this method of extracting features based on human experience is prone to problems such as difficulty in identification and high false alarm rates when dealing with complex scenarios, especially when multiple faults are mixed or when there is strong background noise. As a result, the diagnostic efficiency is low and the accuracy cannot be guaranteed. Although some studies have attempted to introduce other sensor data such as temperature, vibration, and sound to improve the comprehensiveness and accuracy of diagnosis, most of them simply splice or weightedly fuse various types of data, without truly exploring the intrinsic temporal and spatial correlations between different modal data. The overall diagnostic effect still has considerable room for improvement. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: how to solve the problems that existing technologies are prone to identification difficulties, high false alarm rates, low diagnostic efficiency, and inability to guarantee accuracy when dealing with complex scenarios, especially when multiple faults are mixed or the background noise is strong.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a time-based identification and classification method for optical fiber vibration signals based on a Mamba-YOLOv10 neural network, comprising: acquiring OTDR signals of multiple targets based on an optical fiber vibration signal acquisition system; denoising the OTDR signals; performing a short-time Fourier transform on the denoised OTDR signals to obtain a two-dimensional time-frequency diagram of the vibration signals; extracting multi-scale features of the optical fiber vibration signals based on the two-dimensional time-frequency diagram; and inputting the multi-scale features into a Mamba-YOLOv10 neural network that combines a structured state-space model with a YOLOv10 NMS-free consistent dual-assignment detection head to complete the identification and classification of optical fiber vibration signals.

[0007] As a preferred embodiment of the optical fiber vibration signal time identification and classification method based on the Mamba-YOLOv10 neural network described in this invention, the optical fiber vibration signal acquisition system acquires OTDR signals of multiple targets and performs noise reduction processing on the OTDR signals, including: acquiring electrical signals using the optical fiber vibration signal acquisition system; sampling the electrical signals using a data acquisition device; and decomposing and reconstructing the acquired signals.

[0008] As a preferred embodiment of the optical fiber vibration signal time identification and classification method based on the Mamba-YOLOv10 neural network described in this invention, the denoising process includes decomposing and reconstructing the acquired signal, removing interference components, and retaining key information in the original data.

[0009] As a preferred embodiment of the optical fiber vibration signal time identification and classification method based on the Mamba-YOLOv10 neural network described in this invention, the step of performing short-time Fourier transform on the denoised OTDR signal includes dividing the signal into frames and performing frequency domain transformation within each frame to form a data representation in the form of an image.

[0010] As a preferred embodiment of the optical fiber vibration signal time recognition and classification method based on the Mamba-YOLOv10 neural network described in this invention, the extraction of multi-scale features of the optical fiber vibration signal includes: sequentially inputting a two-dimensional time-frequency map into two convolutional layers to extract low-level features, obtaining a first feature map and a second feature map; inputting the first and second feature maps into a cross-stage partial connection and fusion module to generate a third feature map; further processing the third feature map through a convolution and fusion module to obtain a fourth feature map; inputting the fourth feature map into a spatial channel joint downsampling module to obtain a fifth feature map; inputting the fifth feature map into a cross-stage fusion module containing an inverted bottleneck structure to obtain a sixth feature map; inputting the fourth to sixth feature maps into a spatial pyramid feature fusion module to perform multi-scale pooling and channel splicing respectively; and after channel compression of the spliced ​​feature map through a convolutional layer, inputting it into a pyramid squeezing attention module to complete channel mapping processing.

[0011] The advantages of this preferred solution are as follows: by processing the two-dimensional time-frequency map at different scales through the multi-level feature extraction module, the feature separation and combination of the fiber vibration signal at multiple spatial resolution levels can be realized. Combined with the channel compression and attention module, the feature map has the ability to express dimensions across scales and adapt to the input requirements of scenarios with complex signal structures.

[0012] As a preferred embodiment of the optical fiber vibration signal time recognition and classification method based on the Mamba-YOLOv10 neural network described in this invention, the step of inputting multi-scale features into the Mamba-YOLOv10 neural network containing Mamba attention blocks includes: upsampling the extracted feature map and concatenating it with the corresponding layer feature map output by the backbone network to form a fused feature map; inputting the fused feature map into a cross-stage fusion module containing an inverted bottleneck structure for convolutional interaction processing; inputting the interactively processed feature map into an attention module containing structured state space modeling, reducing the tensor dimension using average pooling and generating feature tensors in two directions; multiplying the feature tensors in the two directions with the original feature map respectively and concatenating them, and outputting an enhanced feature map by adjusting the number of channels through convolution.

[0013] The beneficial effects of this preferred scheme are as follows: by fusing the backbone network output and the upsampled feature map, and introducing an attention module with structured state space modeling to perform dual-channel product fusion of the tensor dimension direction, the feature map input to the detection head has both spatial expression consistency and channel information expression integrity, which meets the structural requirements of the detection network front end for coupling features of different dimensions.

[0014] As a preferred embodiment of the optical fiber vibration signal time recognition and classification method based on the Mamba-YOLOv10 neural network described in this invention, the process of completing the recognition and classification of optical fiber vibration signals includes: inputting the enhanced feature map into a detection module based on a consistent dual allocation strategy, performing multi-target detection without non-maximum suppression; processing feature maps of different scales using the multi-resolution structure of the detection module; performing classification labeling on the target region based on the event category information defined during training; and outputting a data structure containing event category and location.

[0015] The advantages of this preferred solution are as follows: by using a detection module based on a consistent dual allocation strategy, combining a multi-resolution structure to divide the input feature map into ranges and match events, and combining the output structure to provide event target type and location information, the recognition process can achieve a structured integration of the three sub-steps of target partitioning, matching, and output while maintaining structural integrity, thus adapting to the processing needs of multi-category input features.

[0016] This invention provides a time-based identification and classification system for fiber optic vibration signals based on the Mamba-YOLOv10 neural network.

[0017] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a fiber optic vibration signal time recognition and classification system based on Mamba-YOLOv10 neural network, comprising:

[0018] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the aforementioned optical fiber vibration signal time recognition and classification method based on Mamba-YOLOv10 neural network.

[0019] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the aforementioned optical fiber vibration signal time recognition and classification method based on a Mamba-YOLOv10 neural network.

[0020] The beneficial effects of this invention are as follows: This invention can achieve feature extraction, event recognition, and classification of optical fiber vibration signals, and output classification results. It utilizes the Mamba model to achieve multi-scale OTDR data fusion processing, effectively reducing the impact of interference fading on pattern recognition accuracy while reducing hardware complexity. The Mamba attention mechanism enhances the feature fusion effect, and automatically adjusts the attention weights according to different fault scenarios and data characteristics, thereby improving the accuracy of optical fiber fault identification and classification. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 The following is a flowchart of an overall method for time identification and classification of optical fiber vibration signals based on a Mamba-YOLOv10 neural network, provided as an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the structure of an OTDR signal acquisition method based on the Mamba-YOLOv10 neural network for time identification and classification of optical fiber vibration signals, provided as an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the OTDR signal denoising process of an optical fiber vibration signal time recognition and classification method based on Mamba-YOLOv10 neural network, provided as an embodiment of the present invention.

[0025] Figure 4 This is a flowchart illustrating the short-time Fourier transform of a time-based classification method for optical fiber vibration signals based on a Mamba-YOLOv10 neural network, as provided in one embodiment of the present invention.

[0026] Figure 5 This is a schematic diagram of the basic structure of the Mamba model for a time-based identification and classification method of optical fiber vibration signals based on the Mamba-YOLOv10 neural network, provided as an embodiment of the present invention.

[0027] Figure 6 This is a schematic diagram of the structure of a Mamba attention block in a time-based classification method for optical fiber vibration signals based on a Mamba-YOLOv10 neural network, provided as an embodiment of the present invention.

[0028] Figure 7 This is a schematic diagram of a time recognition and classification system for optical fiber vibration signals based on a Mamba-YOLOv10 neural network, provided as an embodiment of the present invention. Detailed Implementation

[0029] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0030] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a method for time-based identification and classification of optical fiber vibration signals based on a Mamba-YOLOv10 neural network, including:

[0031] S1. Based on the fiber optic vibration signal acquisition system, acquire OTDR signals from multiple targets and perform noise reduction processing on the OTDR signals.

[0032] S2. Perform a short-time Fourier transform on the denoised OTDR signal to obtain a two-dimensional time-frequency diagram of the vibration signal.

[0033] S3. Extract multi-scale features of fiber optic vibration signals based on two-dimensional time-frequency diagrams.

[0034] S4. Input the multi-scale features into the Mamba module that combines the structured state-space model and the Mamba-YOLOv10 neural network with the NMS-free consistent dual-assignment detection head of YOLOv10 to complete the identification and classification of fiber optic vibration signals.

[0035] The Mamba-YOLOv10 neural network combines the Mamba module of the structured state-space model with the NMS-free consistent dual-assignment detection head of YOLOv10. It enhances the expression of key features through the Mamba attention mechanism and achieves high-precision recognition and classification in multi-event scenarios based on the time-domain, frequency-domain, and spatial-domain features of vibration signals.

[0036] OTDR signals of multiple targets are acquired using an optical fiber vibration signal acquisition system.

[0037] It should be noted that while OTDR technology can obtain fiber optic vibration signals from multiple targets, some noise data is present. During data analysis, useful data may be obscured by noise, leading to calculation errors. Therefore, wavelet transform can be used to denoise the acquired OTDR data.

[0038] Based on the denoised signal, short-time Fourier transform analysis is used to process the one-dimensional time-series data, resulting in a two-dimensional time-frequency graph containing the vibration signal's characteristics in both the time and frequency domains. The YOLOv10 backbone network extracts features from the input two-dimensional image information to obtain feature data of fiber optic vibration signals from multiple targets. Based on this feature data, it is input into a detection head containing a Mamba attention module for feature optimization, using an attention mechanism to highlight important features and suppress irrelevant ones. The optimized feature data is then input into the YOLOv10 detection module for target recognition and classification, and finally, output is generated.

[0039] This invention combines fiber optic vibration signal processing with deep learning algorithms. Based on the acquired OTDR signal, wavelet transform and short-time Fourier transform are used to denoise and convert the signal to time-frequency, resulting in two-dimensional image data. This effectively removes noise and retains useful information. Furthermore, features are extracted through the YOLOv10 backbone network, combined with Mamba attention block optimization to highlight important features and suppress irrelevant features. The YOLO detection head classifies intrusion events such as wind power, lightning strikes, and icing based on the trained model, achieving high classification accuracy and strong generalization ability. Ultimately, this invention realizes the monitoring and classification of fiber optic vibration signals, which can be widely applied in fields such as fiber optic communication and power transmission, improving the intelligence level of fiber optic sensing systems.

[0040] Example 2, refer to Figures 2-6 As one embodiment of the present invention, based on the previous embodiment, a method for time identification and classification of optical fiber vibration signals based on a Mamba-YOLOv10 neural network is provided, including:

[0041] In this embodiment of the application, step S1 involves acquiring OTDR signals from multiple targets based on an optical fiber vibration signal acquisition system, and then performing noise reduction processing on the OTDR signals, including the following steps A1-A3:

[0042] A1. An optical fiber vibration signal acquisition system is used to acquire electrical signals.

[0043] The fiber optic vibration signal acquisition system is configured to acquire information on the fiber optic vibration status of multiple targets.

[0044] Specifically, this invention employs a domestically produced fiber optic vibration signal acquisition system. This system consists of a narrow linewidth laser (NLL), an acousto-optic modulator (AOM), an erbium-doped fiber amplifier (EDFA), a circulator, a photodetector (PD), a bandwidth acquisition quantizer (BAQ) with a 50MHz sampling rate, and a PC. Light is injected into the fiber through a 3kHz linewidth NLL. It is then modulated into a pulse by the AOM with a 200MHz frequency shift and compensated insertion loss by the EDFA. The forward propagation of the optical pulse simultaneously generates Rayleigh backscattered light for backward propagation, which is then guided through the same optical circulator into the PD. The PD converts the optical signal into an electrical signal and sends it to the BAQ.

[0045] like Figure 2 As shown, the components of the fiber optic vibration signal acquisition system include an NLL, AOM, EDFA, Rayleigh backscattering, optical circulator, PD, and BAQ. Among them, NLL is an ultra-narrow linewidth laser, AOM is an acousto-optic modulator, EDFA is an erbium-doped fiber amplifier, PD is a photodetector, and BAQ is a bandpass filter and analog-to-digital converter.

[0046] Light is injected into the optical fiber through an NLL with a linewidth of 3 kHz. The narrow linewidth light source provides a highly coherent optical signal, and its narrow linewidth characteristic helps to improve the spatial resolution and measurement accuracy of the system, ensuring accurate location and analysis of minute disturbances or changes in long-distance fiber optic sensing.

[0047] The optical signal is then injected into the AOM (Acousto-Optical Array), which modulates the light into pulses with a 200MHz frequency shift. The AOM alters the light frequency through an acousto-optic effect, generating the desired frequency shift and providing a reference frequency for the system to distinguish and identify backscattered light signals in subsequent processing. Simultaneously, the AOM can also pulse-modulate the optical signal, controlling the pulse width and repetition frequency to suit the system's sampling and detection requirements.

[0048] The modulated optical pulse enters the EDFA to compensate for the insertion loss introduced in the AOM. The EDFA amplifies the optical signal through stimulated emission, ensuring that the energy of the optical pulse is strong enough during transmission to effectively excite Rayleigh backscattering. Simultaneously, the gain characteristics of the EDFA need to be precisely controlled to avoid nonlinear effects caused by excessively strong signals, thus ensuring system stability and measurement accuracy.

[0049] When an optical pulse propagates forward in an optical fiber, it interacts with the medium within the fiber, generating Rayleigh backscattered light. Rayleigh scattering is caused by minute refractive index inhomogeneities in the fiber, and the intensity of the scattered light depends on the characteristics of the optical pulse and the physical properties of the fiber. Rayleigh backscattered light contains information about various disturbances along the fiber, such as temperature, strain, and vibration, and is a crucial signal for sensing in Φ-OTDR systems.

[0050] The generated Rayleigh backscattered light is guided through the same optical circulator into the PD. The circulator separates and guides the forward-propagating light and the backscattered light at different ports, ensuring that the backscattered light can be efficiently coupled into the PD. The PD converts the received optical signal into an electrical signal, providing the basis for subsequent signal processing and analysis.

[0051] The converted electrical signal is sent to a bandpass filter and an analog-to-digital converter (BAQ) for further processing and analysis. The BAQ system first performs bandpass filtering to remove unwanted noise and interference, extracting frequency bands containing useful information. Then, the analog signal is converted to a digital signal via analog-to-digital conversion, enabling more precise analysis using digital signal processing techniques. These processing steps help improve the system's signal-to-noise ratio and enhance its ability to detect weak signals, thereby achieving high-sensitivity and high-precision monitoring of various disturbances along the optical fiber.

[0052] In an alternative implementation, the fiber optic vibration signal acquisition system may also employ an integrated architecture with a frequency-modulated laser injection mechanism and a gain compensation module to meet the echo acquisition requirements of long-distance fiber optic vibration data.

[0053] In another alternative implementation, the fiber optic vibration signal acquisition system can also be configured with multi-channel signal detection paths and wide dynamic range sampling units to support parallel acquisition and real-time processing of multi-point vibration targets.

[0054] A2. Based on electrical signals, sampling is performed through a data acquisition device.

[0055] Based on the received electrical signal, it is sampled by a high-precision data acquisition device to provide a solid data foundation for subsequent signal processing.

[0056] The acquired data encompasses information from multiple modes, including optical time-domain reflectometry (OTDR) signals, fiber temperature variations, fiber scattering signal intensity, wavelength shifts, and pressure changes. These are all time-series functions, collectively forming the complex characteristic space of the fiber optic vibration signal.

[0057] A3. Decompose and reconstruct the acquired signals.

[0058] Denoising processing involves decomposing and reconstructing the acquired signal, removing interference components, and retaining key information in the original data.

[0059] Based on the collected vibration waveform data of different types, wavelet transform and short-time Fourier transform are used to enhance the time-frequency characteristics of the vibration signals, converting different types of vibration signals into time-frequency graphs. These graphs are then input into a neural network for feature extraction and classification. The spectrum obtained from time-domain analysis displays combined information from both the time and frequency domains.

[0060] The decomposition and reconstruction of the acquired signal includes wavelet decomposition and wavelet reconstruction.

[0061] Specifically, the decomposition and reconstruction of the acquired signal includes the following steps B1-B3:

[0062] B1. Perform wavelet decomposition on the initial OTDR data to obtain wavelet decomposition coefficients at multiple levels.

[0063] B2. Threshold quantization is performed on the wavelet decomposition coefficients at multiple levels to obtain multiple wavelet decomposition coefficients after threshold quantization.

[0064] B3. Wavelet reconstruction is performed based on multiple wavelet decomposition coefficients after threshold quantization to obtain the target OTDR data.

[0065] like Figure 3 As shown, Figure 3 This is a schematic diagram of the OTDR signal denoising process provided by the present invention.

[0066] Specifically, the fiber optic vibration signal acquisition system acquires initial OTDR signal data.

[0067] Furthermore, wavelet decomposition is performed on the initial OTDR signal data to obtain wavelet decomposition coefficients at multiple levels.

[0068] We set the wavelet decomposition level P and perform wavelet decomposition on the data in each level. Therefore, the formula for calculating the wavelet decomposition coefficients is as follows:

[0069]

[0070] Where k represents ground noise; t represents time; y i y M w represents the wavelet coefficients of the i-th and M-th layers at position P, respectively; p The wavelet decomposition coefficients at position P are represented.

[0071] Furthermore, the wavelet decomposition coefficients at multiple levels are threshold-quantized to obtain multiple wavelet decomposition coefficients after threshold quantization.

[0072] The formula for calculating the threshold quantization of wavelet decomposition coefficients is as follows:

[0073]

[0074] Among them, w P Characterized by the wavelet decomposition coefficients at position P; y M (t) represents the wavelet coefficients of the Mth layer at time t; u(t) represents the wavelet decomposition coefficients at time t; Characterizes the wavelet decomposition coefficients after threshold quantization.

[0075] Furthermore, the fiber core splice box information input device performs wavelet reconstruction based on multiple wavelet decomposition coefficients after threshold quantization processing.

[0076] Determine the low-frequency coefficient d at position P, and combine it with the wavelet decomposition coefficients after threshold quantization. Wavelet reconstruction is performed; therefore, the calculation method for wavelet reconstruction is as follows:

[0077]

[0078] Among them, P d Characterized by the wavelet decomposition hierarchy at low-frequency coefficients d; P t The wavelet decomposition level is represented at the current time; y(t) represents the target OTDR data.

[0079] The embodiments of the present invention effectively process and optimize the original OTDR signal through signal decomposition and reconstruction, namely wavelet decomposition and wavelet reconstruction, which can significantly reduce noise and improve signal quality, so that the subsequent input into the M-YoLo model can improve the accuracy of vibration time identification and classification.

[0080] In one alternative implementation, the decomposition and reconstruction of the acquired signal can also be performed by empirical mode decomposition to adaptively decompose the vibration signal in the time domain, extract the main vibration mode functions, remove high-frequency interference components, and then reconstruct the signal. This method can be used to replace the wavelet method for denoising under complex background noise conditions.

[0081] In another alternative implementation, the decomposition and reconstruction of the acquired signal can also be combined with the sliding window segmented reconstruction method. The OTDR signal is reconstructed in segments according to a fixed time window, feature-preserving denoising is performed in each window, and then the segments are spliced ​​together to form a complete sequence for use in step S3.

[0082] This invention ensures that detailed information of the OTDR signal is preserved during spatial transmission.

[0083] In this embodiment of the application, step S2 involves performing a short-time Fourier transform on the denoised OTDR signal to obtain a two-dimensional time-frequency diagram of the vibration signal, including the following steps C1-C2:

[0084] Based on the collected fiber optic vibration signal data, wavelet transform and short-time Fourier transform are used for denoising and time-frequency conversion to obtain two-dimensional image data.

[0085] C1. The acquired vibration signal is preprocessed for denoising using wavelet transform. Through wavelet decomposition and reconstruction, the original OTDR signal is effectively processed and optimized, effectively suppressing broadband noise and improving signal quality. When input into the Mamba-YOLOv10 model, it can improve the accuracy of vibration time identification and classification, reduce the adverse effects of noise on spectrum classification, and provide a clearer signal foundation for signal processing.

[0086] C2. Short-Time Fourier Transform (STFT) is used for time-frequency analysis of the vibration signal. STFT converts the one-dimensional vibration signal into a time-frequency graph, visually displaying the signal's variation characteristics in the time and frequency domains. Through STFT processing, the complex features of the vibration signal are transformed into a form that can be input into a neural network for subsequent classification tasks.

[0087] Specifically, STFT captures both the time and frequency domain characteristics of a signal by dividing it into local segments within short time windows and performing a Fourier transform on each segment. Its core idea is to analyze the frequency components of the signal within a fixed time window, using a sliding window to cover the entire signal time axis.

[0088] The mathematical definition of STFT is as follows:

[0089]

[0090] Here, x(t) is the time-domain signal, and w(t) is the window function, typically a Hamming window or a Gaussian window centered at zero. τ is the center of the window function. When the STFT window function is selected, the resolution is fixed as long as the window length is constant. Since time resolution and frequency resolution are inversely proportional, they cannot both be optimal simultaneously. As the window length increases, the time resolution decreases; as the window length decreases, the frequency resolution increases. Therefore, choosing an appropriate window function type and an appropriate window length is crucial.

[0091] After transforming the signal in each time domain to the frequency domain using a Fast Fourier Transform (FFT), the amplitude and frequency are rotated by 90 degrees, mapping the spectral amplitude to a gray value (0-255). This time-frequency representation of the signal is called a spectrogram.

[0092] Combination Figure 4 , Figure 4 This is a schematic diagram of the STFT process provided by the present invention.

[0093] Specifically, the vibration signal is converted into a spectrum using STFT, and then input into a neural network for classification, including the following steps D1-D3:

[0094] D1. Framing the denoised signal. Due to the randomness and complexity of the spectral characteristics of vibration signals, a Hamming window function is used to divide the signal into smaller frames. The Hamming window function can reduce discontinuities at frame edges, eliminate high-frequency interference and energy leakage. The mathematical definition of the Hamming window is:

[0095]

[0096] Where 0≤n≤N-1, and N is the length of the window.

[0097] D2. Use STFT to generate a time-frequency diagram.

[0098] Based on the amplitude value of the spectrum of each frame, it is rotated 90 degrees and mapped to grayscale values, where the larger the amplitude, the darker the corresponding pixel. Based on the time order, the spectra of all frames are arranged to form a time-frequency image, resulting in a two-dimensional time-frequency diagram.

[0099] D3. Convert the STFT result into a grayscale image and use it as input to the neural network.

[0100] The signal is processed by dividing it into frames, and frequency domain transformation is performed within each frame to form a data representation in the form of an image.

[0101] In one alternative implementation, the intra-frame frequency domain transformation to form an image-like data representation can also employ Gabor transformation, using a Gaussian window function as the modulation kernel to apply frequency-shifting convolution to the local signal of each frame, thereby achieving joint analysis of the frequency domain and time, and providing richer feature information for different types of vibration events.

[0102] In another alternative implementation, the multi-window overlapping Fourier transform method can be used to perform frequency domain transformation within a frame to form a data representation in the form of an image. By setting different window function lengths and sliding step sizes, multiple combinations of time-frequency resolutions can be constructed, and the results are then merged to generate a two-dimensional feature map in a unified format, which is then input into the neural network.

[0103] The present invention provides a method for converting one-dimensional time series signals into two-dimensional time-frequency graphs using STFT, which provides structured input for subsequent feature extraction and classification of neural networks.

[0104] In the embodiments of this application, in step S3, multi-scale features of the optical fiber vibration signal are extracted based on the two-dimensional time-frequency diagram.

[0105] The image is input into the YOLOv10 backbone network for feature extraction.

[0106] The two-dimensional time-frequency map is sequentially input into two convolutional layers to extract low-level features, resulting in a first feature map P1 and a second feature map P2. The first feature map P1 and the second feature map P2 are then input into a cross-stage partial connection fusion module to generate a third feature map P3. The third feature map P3 is further processed by a convolution and fusion module to obtain a fourth feature map P4. The fourth feature map P4 is input into a spatial channel joint downsampling module to obtain a fifth feature map P5. The fifth feature map P5 is input into a cross-stage fusion module containing an inverted bottleneck structure to obtain a sixth feature map P6. P4, P5, and P6 are input into a spatial pyramid feature fusion module, where multi-scale pooling and channel concatenation are performed respectively. The concatenated feature maps are then processed by channel compression through a convolutional layer and input into a pyramid squeezing attention module to complete channel mapping processing.

[0107] Based on the two-dimensional time-frequency diagram, multi-scale features of the optical fiber vibration signal are extracted, specifically including the following steps E1-E6:

[0108] E1. Based on the input image, two consecutive convolutional layers are used to initially extract low-level features, resulting in preliminary feature maps P1 and P2. These are then input into the Cross Stage Partial Fusion (C2f) module, an improved version of the CSP structure in YOLOv8. By fusing shallow and deep features through cross-stage partial connections, computational redundancy is reduced and gradient flow is enhanced, achieving a balance between accuracy and speed, resulting in feature map P3. The features are then further refined and enhanced through a convolutional layer and the C2f module, resulting in feature map P4.

[0109] E2, based on the Spatial Channel Downsample (SCDown) module, performs dimensionality compression and channel dimension adjustment through spatial downsampling, 1×1 convolution, efficiently reducing feature map resolution, and adjusting the number of channels at the same time.

[0110] E3. Obtain feature map P5, perform secondary downsampling to further compress the resolution, and prepare for deep abstract feature extraction.

[0111] E4. The C2f (C2f with Cross-stage Inverted Bottleneck, C2fCIB) module, based on a cross-stage inverted bottleneck structure, extracts high-level semantic features through dimensionality increase and decrease operations. It balances computational efficiency with feature representation capability based on cross-stage feature reuse and the inverted bottleneck design.

[0112] E5. The extracted feature information is input into the Spatial Pyramid Feature Fusion (SPFF) module. Max pooling at three scales of 5×5, 9×9 and 13×13 is used to perform multi-scale feature fusion. The working principle is to first pool, then concatenate the channels through Concat, and finally perform dimensionality reduction through convolutional layers.

[0113] E6. Based on the Pyramid Sequeeze Attention (PSA) squeezing attention mechanism, multi-scale feature maps are obtained through pooling operations at different scales. Then, channel attention weights are calculated to dynamically enhance important feature channels and strengthen multi-scale feature interactions.

[0114] In an alternative implementation, extracting multi-scale features of fiber vibration signals may further include introducing a spatial channel joint downsampling module and an inverted bottleneck structure between multi-resolution feature maps to construct a feature representation process from local to global.

[0115] In another alternative implementation, extracting multi-scale features of the fiber vibration signal may also include performing max pooling operations at different scales using a spatial pyramid structure, and forming a feature tensor in a uniform format through channel splicing and compression.

[0116] This invention constructs a feature extraction path with distinct levels and rich scales, enabling the separation and expression of different frequencies and time-domain variation patterns in fiber optic vibration signals in the feature space, thus providing structurally complete data support for subsequent neural network structure input.

[0117] In this embodiment of the application, in step S4, multi-scale features are input into the Mamba-YOLOv10 neural network containing Mamba attention blocks to complete the identification and classification of fiber optic vibration signals.

[0118] Based on the extracted image features, the data is input into a detection head containing Mamba attention blocks for feature optimization. The optimized features are then input into the detection module, which uses a trained model (containing three types of intrusion events: wind power, lightning strikes, and icing) for identification and classification.

[0119] The Mamba-YOLOv10 neural network combines the Mamba module of the structured state-space model with the consistent dual-assignment detection head of YOLOv10 with non-maximum suppression (NMS). It enhances the expression of key features through the Mamba attention mechanism and achieves high-precision recognition and classification in multi-event scenarios based on the time-domain, frequency-domain, and spatial-domain features of vibration signals.

[0120] The process involves inputting multi-scale features into a Mamba-YOLOv10 neural network containing a Mamba attention block. This includes upsampling the extracted feature maps and concatenating them with the corresponding layer feature maps output from the backbone network to form a fused feature map. The fused feature map is then input into a cross-stage fusion module containing an inverted bottleneck structure for convolutional interaction. The processed feature map is then input into an attention module containing structured state space modeling, where average pooling is used to reduce the tensor dimension and generate feature tensors in two directions. The feature tensors in both directions are multiplied by the original feature map and concatenated. Finally, the number of channels is adjusted through convolution to output an enhanced feature map.

[0121] The identification and classification of fiber optic vibration signals involves inputting the enhanced feature map into a detection module based on a consistent dual-assignment strategy, performing multi-target detection without non-maximum suppression; processing feature maps of different scales using the multi-resolution structure of the detection module; performing classification and labeling of target regions based on event category information defined during training; and outputting a data structure containing event category and location.

[0122] Based on the feature map input from the backbone network, the P3-P5 feature levels are fused step by step through repeated upsampling and concatenation. The C2fCIB module reduces redundant computation through cross-stage interaction and introduces the Mamba attention mechanism to achieve effective fusion of feature information under different pooling strategies, further optimizing the feature extraction process.

[0123] Based on the low-level feature maps from the backbone network, upsampling is performed to obtain upsampled feature maps. These upsampled feature maps are then concatenated with the original low-level feature maps from the backbone network before being input into the C2fCIB module. This module includes two convolutions, one interactive connection, and batch normalization for feature enhancement and channel adjustment. Mamba Attention, a state-space model-based attention mechanism, is used to capture long-range dependencies and enhance the feature weights of key regions.

[0124] The Selective State-Space Model (Mamba) incorporates a selection mechanism into the State-Space Model (SSM). The SSM model abstracts the feature extraction process as a feature update process, continuously updating the current state by scanning the input sequence, and influencing the output through this state. The SSM model mainly consists of state equations and output equations, as shown in the mathematical formulas below:

[0125] h′(t)=Ah(t)+Bx(t)

[0126] y(t)=Ch′(t)+Dx(t)

[0127] In this model, matrix A represents the matrix required to update the state equation, matrix B represents the matrix that affects the input, matrix C represents the matrix that transforms the state into the output, and matrix D represents the matrix that affects the output. x(t) represents the input, and h(t) represents the state matrix. The Mamba improves the SSM by parameterizing matrices B and C, initializing them by mapping the input; their values ​​can be updated during network training. The Mamba is able to selectively identify key information from the input and enhance its ability to realize that key information.

[0128] Mamba's operating procedure is as follows Figure 5 As shown, the steps F1-F4 are included:

[0129] F1. Input the sequence into the Mamba, map it to matrices B, C, and Δ, through a linear layer, as shown below:

[0130] B,C,Δ=Linear B (x),Linear C (x),Linear Δ (x)

[0131] F2. Discretize the continuous parameter matrices A and B based on the discretization parameter Δ. The formula is shown below:

[0132]

[0133] F3. Update the state matrix by scanning the input sequence.

[0134] F4. Based on the output equation, perform operations on the state matrix and the input sequence to obtain the final output.

[0135] The Mamba algorithm possesses strong global feature extraction capabilities and low computational requirements. Therefore, this invention combines the Mamba algorithm with an attention mechanism, proposing a novel MBAM module composed of MSAM and MCAM, such as... Figure 6 As shown. MSAM and MCAM represent spatial dimension attention and channel dimension attention, respectively.

[0136] Based on the Spatial Attention Module (SAM), spatial features are extracted by pooling the feature map using 3×3 or 7×7 convolutions. MSAM replaces traditional convolutions with the SS2D module to extract long-range semantic information. Average pooling and max pooling are performed on the feature map in the spatial dimension, and then the pooled feature maps are concatenated. The concatenated feature map is input into the SS2D module in four scanning directions. The SS2D module scans the feature map in four directions, allowing each pixel to more fully integrate long-range information. Then, the spatial information is converted into spatial weights using the Sigmoid function, followed by a 3×3 convolution to enhance local feature extraction while compressing the number of channels to 1. Finally, it is concatenated with the input feature map. The calculation formula for MSAM is shown below:

[0137] f s ′=f×σ(Conv 3×3 (SS2D(Concat(f avg ,f max ))))

[0138] Where, f,f s ′∈R [B,W,H,C] These represent the input and output feature maps, respectively. avg ,f max ∈R [B,W,H,1] This represents the feature map after performing average pooling and max pooling operations along the channel dimension. `Concat(·)` represents the concatenation of feature maps along the channel dimension, `SS2D(·)` represents cross-selective state scanning, `σ(·)` represents the Sigmoid function, and `Conv3×3(·)` represents a convolution operation with a kernel size of 3×3. The channel information of an image contains a large number of abstract features. By focusing on the channel information, we can enhance our focus on key abstract features.

[0139] The ChannelAttentionModule (CAM) transforms the feature map into a one-dimensional vector along the channel dimension using globally adaptive max pooling and globally adaptive average pooling operations. MCAM converts the feature map into a one-dimensional vector through global average pooling and global max pooling layers. Then, the SS1D module, with bidirectional scanning capability, scans the one-dimensional vector, fully integrating information between channels. The sigmoid function of the scanned one-dimensional vector is then converted into channel weights and multiplied with the input feature map. The calculation formula for the MCAM module is as follows:

[0140] f C ′=f×σ(SS1D(Concat(AVG(f),MAX(f)))

[0141] Where, f,f C ∈R [B,W,H,C] These represent the input and output feature maps, respectively. AVG(·) represents the global average pooling operation, SS1D(·) represents the bidirectional selection state scan, and σ(·) represents the Sigmoid function.

[0142] Furthermore, the optimized features are obtained and input into the detection head.

[0143] It should be noted that the YOLOv10 detection head adopts a multi-scale feature fusion strategy, which includes multiple different resolution layers.

[0144] The Mamba attention mechanism provided by this invention is applied to features at multiple different scales. By processing the Mamba attention block, it enhances the model's ability to filter features.

[0145] Furthermore, the optimized features are input into the detection module for accurate identification and classification of fiber optic vibration signal events.

[0146] This invention provides a Mamba module for subsequently constructing the Mamba attention mechanism. The Mamba module is an innovative sequence modeling architecture method, and its core uses a structured state-space model.

[0147] The Mamba module comprises four parts: an input layer module, a state space model layer module, a feature fusion layer module, and an output layer module. The input layer module receives the preprocessed signal, providing input for subsequent processing. The state space model layer module extracts and models data features, capturing dynamic features in the data through a structured state space model. The feature fusion layer module fuses feature information from different levels to fully utilize the information at each level. The output layer module adjusts the dimensions of the output feature map to adapt to the needs of subsequent processing. The Mamba module provided by this invention includes:

[0148] The input layer module, used as input to the Mamba module in the fiber optic vibration signal event recognition and classification task, uses preprocessed signals, such as denoised signals, as input to the Mamba module. These time-frequency plots contain characteristic information of the signal at different frequencies and times.

[0149] The state-space model layer module is used for data feature extraction and modeling. The input time-frequency plot is fed into the Mamba module, where the state-space model is used to model the sequence data. The Mamba module can extract and model data features from different dimensions, obtaining multiple feature representations. In the M-YOLO model, the Mamba module outputs two tensors, Xx and Xy, which capture feature representations of the data from different dimensions.

[0150] The feature fusion layer module is used for feature fusion. The features extracted by the Mamba module are fused with the original features. Xx and Xy are multiplied by the original input tensor to obtain new feature representations. Then, further feature fusion and dimensionality adjustment are performed through concatenation and convolution operations, and the final output is a feature map used for subsequent fiber optic vibration event recognition and classification.

[0151] The output layer module is used for dimensionality adjustment. Through operations such as convolution, the fused features are adjusted to obtain an output feature map with the same dimensions as the input, which is then used for subsequent fiber optic vibration event recognition and classification.

[0152] The Mamba attention mechanism reduces tensor dimensionality through average pooling and utilizes Mamba modules to obtain feature representations. It fuses feature information from different dimensions and adjusts the dimensionality through convolution operations, ensuring consistency between the input and output of the attention mechanism. This enhances the model's ability to identify and extract key features and significantly optimizes the overall model performance. By introducing an attention mechanism, the Mamba attention mechanism enables the model to focus more on key features, thereby improving the model's recognition accuracy and helping it sift through a large number of features to extract crucial information, thus improving classification accuracy.

[0153] The Mamba-YOLOv10 neural network combines the backbone network of YOLOv10 with the Mamba attention mechanism to achieve the identification and classification of optical fiber vibration signals.

[0154] YOLOv10 is an advanced object detection network with a backbone network capable of feature extraction. In the task of fiber optic vibration signal recognition and classification, the YOLOv10 backbone network can effectively extract image features, providing support for the classification task. Furthermore, YOLOv10 proposes a consistent dual-assignment strategy without NMS, featuring dual-label assignment and a consistent matching metric, which solves the problem of redundant prediction in post-processing and improves the model's performance.

[0155] YOLOv10 introduces a lightweight classification head, spatial-channel decoupling downsampling, and a rank-guided block design, reducing computational redundancy and achieving a more efficient architecture, thus improving efficiency. The lightweight classification head consists of two depthwise separable convolutions with a kernel size of 3×3, followed by a 1×1 convolution. This decouples spatial reduction and channel augmentation operations, enabling more efficient downsampling. Specifically, point-directed convolutions are first used to modulate the channel dimension, and then depthwise convolutions are used to perform spatial downsampling. The rank-guided block allocation strategy, given a model, sorts all its stages in ascending order based on its intrinsic rank. It further examines the performance change of replacing the basic blocks of the leading stage with CIBs. If there is no performance degradation compared to the given model, the next stage is replaced; otherwise, the process stops. This allows for adaptive compact block design across stages and model scales, achieving higher efficiency without compromising performance.

[0156] To fully leverage the advantages of YOLOv10 and the Mamba attention mechanism, this invention inputs the extracted image features into a detection head containing Mamba attention blocks. The Mamba attention blocks enhance the model's ability to recognize key features by filtering and strengthening them.

[0157] Finally, the optimized image features are input into the detection module to complete the identification and classification of fiber optic vibration events. The detection module performs comprehensive analysis of the input features to accurately determine the type and location of fiber optic vibration events, thus realizing real-time monitoring and early warning of fiber optic vibration signals.

[0158] Example 3, referring to Figure 7 This is one embodiment of the present invention, which provides a fiber optic vibration signal time recognition and classification system based on the Mamba-YOLOv10 neural network, including a data acquisition and processing module, a time-frequency conversion module, a feature extraction module and a classification and recognition module.

[0159] The data acquisition and processing module is used to acquire OTDR signals from multiple targets based on the fiber optic vibration signal acquisition system and to perform noise reduction processing on the OTDR signals.

[0160] The time-frequency conversion module is used to perform short-time Fourier transform on the denoised OTDR signal to obtain a two-dimensional time-frequency diagram of the vibration signal.

[0161] The feature extraction module is used to extract multi-scale features of optical fiber vibration signals based on two-dimensional time-frequency diagrams.

[0162] The classification and recognition module is used to input multi-scale features into the Mamba-YOLOv10 neural network containing Mamba attention blocks to complete the recognition and classification of fiber optic vibration signals. The Mamba-YOLOv10 neural network combines the Mamba module of the structured state-space model with the YOLOv10 non-NMS consistent dual-assignment detection head. It enhances the expression of key features through the Mamba attention mechanism and achieves high-precision recognition and classification in multi-event scenarios based on the time-domain, frequency-domain, and spatial-domain features of the vibration signals.

[0163] This embodiment also provides an electronic device applicable to a time-based identification and classification method for fiber optic vibration signals based on a Mamba-YOLOv10 neural network, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the time-based identification and classification method for fiber optic vibration signals based on a Mamba-YOLOv10 neural network as proposed in the above embodiment.

[0164] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a time-based identification and classification method for fiber optic vibration signals based on a Mamba-YOLOv10 neural network as proposed in the above embodiment.

[0165] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for time identification and classification of optical fiber vibration signals based on Mamba-YOLOv10 neural network proposed in the above embodiment. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0166] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0167] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A time-based identification and classification method for fiber optic vibration signals based on a Mamba-YOLOv10 neural network, characterized in that: include, Based on the fiber optic vibration signal acquisition system, OTDR signals of multiple targets are acquired, and the OTDR signals are denoised. A short-time Fourier transform is performed on the denoised OTDR signal to obtain a two-dimensional time-frequency diagram of the vibration signal. Multi-scale features of optical fiber vibration signals are extracted based on two-dimensional time-frequency diagrams. By inputting multi-scale features into the Mamba module, which combines a structured state-space model with the Mamba-YOLOv10 neural network that uses a dual-assignment detection head without NMS, the identification and classification of fiber optic vibration signals can be completed.

2. The fiber optic vibration signal time identification and classification method based on Mamba-YOLOv10 neural network as described in claim 1, characterized in that: The fiber optic vibration signal acquisition system acquires OTDR signals from multiple targets and performs noise reduction processing on the OTDR signals, including: An optical fiber vibration signal acquisition system is used to acquire electrical signals; Based on electrical signals, sampling is performed using a data acquisition device; The acquired signals are decomposed and reconstructed.

3. The fiber optic vibration signal time recognition and classification method based on Mamba-YOLOv10 neural network as described in claim 2, characterized in that: The noise reduction process includes decomposing and reconstructing the acquired signal, removing interference components, and retaining key information in the original data.

4. The fiber optic vibration signal time identification and classification method based on Mamba-YOLOv10 neural network as described in claim 3, characterized in that: The short-time Fourier transform of the denoised OTDR signal includes dividing the signal into frames and performing frequency domain transformation within each frame to form a data representation in the form of an image.

5. The fiber optic vibration signal time identification and classification method based on Mamba-YOLOv10 neural network as described in claim 4, characterized in that: The extraction of multi-scale features from fiber optic vibration signals includes, The two-dimensional time-frequency map is sequentially input into two convolutional layers to extract low-level features, resulting in the first feature map and the second feature map. The first and second feature maps are input into the cross-stage partial connection and fusion module to generate the third feature map; The third feature map is further processed by a convolution and fusion module to obtain the fourth feature map; The fourth feature map is input into the spatial channel joint downsampling module to obtain the fifth feature map; The fifth feature map is input into the cross-stage fusion module containing the inverted bottleneck structure to obtain the sixth feature map; The fourth to sixth feature maps are input into the spatial pyramid feature fusion module, where multi-scale pooling and channel stitching are performed respectively. The concatenated feature maps are compressed through convolutional layers and then input into the pyramid squeezing attention module to complete the channel mapping process.

6. The fiber optic vibration signal time identification and classification method based on Mamba-YOLOv10 neural network as described in claim 5, characterized in that: The input of multi-scale features into the Mamba-YOLOv10 neural network containing Mamba attention blocks includes, The extracted feature maps are upsampled and then concatenated with the corresponding layer feature maps output by the backbone network to form a fused feature map. The fused feature map is input into a cross-stage fusion module containing an inverted bottleneck structure for convolutional interactive processing. The interactively processed feature map is input into the attention module, which includes structured state space modeling. The tensor dimension is reduced by average pooling and feature tensors in two directions are generated. The feature tensors in both directions are multiplied by the original feature map and then concatenated. The number of channels is adjusted by convolution to output the enhanced feature map.

7. The fiber optic vibration signal time identification and classification method based on Mamba-YOLOv10 neural network as described in claim 6, characterized in that: The process of identifying and classifying fiber optic vibration signals includes, The enhanced feature map is input into the detection module based on the consistent dual allocation strategy, and multi-target detection is performed without nonmaximum suppression. The multi-resolution structure of the detection module is used to process feature maps at different scales; Classification and labeling of the target region are performed based on the event category information defined during training; The output contains a data structure that includes the event type and location.

8. A fiber optic vibration signal time identification and classification system based on a Mamba-YOLOv10 neural network, employing the fiber optic vibration signal time identification and classification method based on a Mamba-YOLOv10 neural network as described in any one of claims 1 to 7, characterized in that, include: Data acquisition and processing module, time-frequency conversion module, feature extraction module, and classification and recognition module; The data acquisition and processing module is used to acquire OTDR signals from multiple targets based on the fiber optic vibration signal acquisition system and to perform noise reduction processing on the OTDR signals. The time-frequency conversion module is used to perform a short-time Fourier transform on the denoised OTDR signal to obtain a two-dimensional time-frequency diagram of the vibration signal. The feature extraction module is used to extract multi-scale features of optical fiber vibration signals based on two-dimensional time-frequency diagrams; The classification and recognition module is used to input multi-scale features into the Mamba-YOLOv10 neural network containing Mamba attention blocks to complete the recognition and classification of fiber optic vibration signals. The Mamba-YOLOv10 neural network combines the Mamba module of the structured state-space model with the YOLOv10 non-NMS consistent dual-assignment detection head. It enhances the expression of key features through the Mamba attention mechanism and achieves high-precision recognition and classification in multi-event scenarios based on the time-domain, frequency-domain, and spatial-domain features of the vibration signal.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the fiber optic vibration signal time recognition and classification method based on the Mamba-YOLOv10 neural network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the fiber optic vibration signal time recognition and classification method based on the Mamba-YOLOv10 neural network as described in any one of claims 1 to 7.

Images