Optical fiber sensing signal diagnosis method based on TimesNet two-dimensional time-frequency transformation

Abstract

The invention relates to the field of optical fiber sensing, and discloses an optical fiber sensing signal diagnosis method based on TimesNet two-dimensional time-frequency transformation, which comprises the following steps: S1, inputting an original signal, and carrying out adaptive time-frequency denoising through learnable wavelet packet transformation; s2, extracting a dominant frequency component based on the enhanced signal, and constructing a two-dimensional time-frequency tensor through zero padding and remodeling; s3, using depth separable convolution to extract features, using grouping convolution to capture local space-time coupling, using point-by-point convolution to fuse abnormal features, and combining a SimAM attention mechanism to enhance the weight of a key area; s4, modeling time sequence evolution through minLSTM, and realizing parallel long-range dependence modeling by adopting a logarithm domain correlation scanning algorithm; and S5, performing joint training through a gradient conflict coordination mechanism based on the shared features by using a multi-task output head, and outputting fault detection, type classification and positioning results at the same time. According to the method, two-dimensional orthogonal expression is constructed through physical prior dimension raising and noise reduction, and the signal-to-noise ratio, the feature interpretability and the diagnosis precision are remarkably improved.

Description

Technical Field

[0001] This application belongs to the field of fiber optic sensing, and specifically relates to a fiber optic sensing signal diagnosis method based on TimesNet two-dimensional time-frequency transformation. Background Technology

[0002] In the field of fault diagnosis for distributed fiber optic sensing systems, existing technologies generally employ one-dimensional time series modeling methods, such as LSTM, TCN, or Transformer. While these methods have achieved some success, their fundamental flaw stems from the inherent limitations of one-dimensional signal structures. Because fiber optic sensing signals exhibit strong periodicity and multi-scale coupling characteristics, and one-dimensional convolutional or recursive structures can only capture local dependencies between adjacent time points, long-period trends and short-period abrupt changes cannot be modeled simultaneously. This structural defect further triggers the following chain reaction: First, the model must be stacked with multiple layers or the hidden dimension increased to expand the receptive field, resulting in a surge in the number of parameters; second, high-frequency noise and low-frequency fault characteristics are superimposed in a single dimension, causing early, weak faults to be submerged; ultimately, this manifests as a high false diagnosis rate, large localization error, and poor real-time performance.

[0003] More importantly, traditional methods rely on manually designed features (such as wavelet coefficients and spectral energy), and this "human-driven" approach is prone to subjective bias. When faced with new types of faults (such as progressive fiber aging caused by composite stress), the existing feature library becomes ineffective, and the system immediately falls into an "unknown fault blind zone." The ultimate contradiction of this technical approach lies in the mismatch between one-dimensional modeling capabilities and the nature of two-dimensional faults—the essence of fiber optic signals is a two-dimensional coupled process of intra-periodic variations (such as scattering intensity jitter caused by micro-bending) and inter-periodic variations (such as trend drift caused by temperature gradients). Forcibly modeling with one-dimensional methods is equivalent to measuring a curved surface with a ruler, inevitably leading to information distortion.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To address or at least alleviate one or more of the above problems, a diagnostic method for fiber optic sensing signals based on TimesNet two-dimensional time-frequency transform is provided.

[0006] To achieve the above objectives, according to the first aspect of this application, a method for diagnosing fiber optic sensor signals based on TimesNet two-dimensional time-frequency transform is provided, comprising the following steps: S1, Signal Preprocessing: The original one-dimensional fiber optic sensing signal is input, and adaptive time-frequency denoising is performed through learnable wavelet packet transform (LWPT) to output an enhanced time-domain signal. S2, TimesNet periodic transformation: Based on the enhanced time-domain signal, the top K dominant frequency components are extracted by fast Fourier transform; for each dominant frequency component, a two-dimensional time-frequency tensor is constructed by zero-padding and reshaping operations, wherein the column dimension of the two-dimensional time-frequency tensor represents the transient micro-variation characteristics within a single period, and the row dimension represents the slow macro-variation characteristics across periods. S3, Feature Extraction: Feature extraction is performed on the two-dimensional time-frequency tensor using depthwise separable convolution, wherein grouped convolutional layers are used to extract local spatiotemporal coupling patterns in the row dimension, and pointwise convolutional layers are used to fuse cross-channel anomalous features; a SimAM attention mechanism is introduced to adaptively enhance the weights of key feature regions related to fiber faults. S4. Temporal Evolution Modeling: The extracted features are modeled temporally using a minLSTM network. The minLSTM network uses the Heinsen logarithmic domain association scanning algorithm to decompose the hidden state into a forward logarithmic flow and a backward logarithmic flow, thereby achieving parallelized long-range dependency modeling. S5. Multi-task output header output results: The classifiers, which share a common underlying feature representation, are jointly trained through a gradient conflict coordination mechanism to simultaneously complete multiple target tasks. The classifiers include a detection head, a classification head, and a localization head. The detection head outputs whether a fault exists, the classification head outputs the fault type, and the localization head outputs the fault location.

[0007] By adopting the above technical solution, this application has the following beneficial effects compared with the prior art:

[0008] In this application, the input raw one-dimensional vibration signal is first processed by LWPT. LWPT adopts a multi-layer decomposition structure (two layers in this application), using parallel low-pass and high-pass convolution kernels to simulate the traditional wavelet packet decomposition process, and introducing learnable soft threshold functions in each frequency band channel to achieve nonlinear compression. This effectively suppresses background noise and non-discriminatory components in the signal, while preserving the main frequency domain structure related to the event, thereby significantly improving the signal-to-noise ratio and discriminability of the signal before entering the subsequent convolutional feature extraction stage. The fiber optic sensing signal is upscaled to a two-dimensional time-frequency space through the TimesNet architecture, enabling intra-period and inter-period variations to be orthogonally expressed for the first time. This upscaling not only solves the feature aliasing problem, but also achieves synchronous capture of microsecond-level fault pulses and hourly-level temperature drift through the local receptive field of two-dimensional convolution, ultimately constructing a final fiber optic sensing signal diagnostic method that "requires no manual features, no large number of parameters, and no sacrifice of real-time performance." TimesNet, with its physical prior filtering, provides high-quality, interpretable 2D input; asymmetric depthwise separable convolutions enable efficient, targeted feature extraction from the 2D input; and a unique serialization interface ensures lossless and structured transformation of 2D features into 1D temporal context. This end-to-end collaborative design for a specific problem domain produces a synergistic effect that goes beyond the simple sum of its components, solving the problems of low signal-to-noise ratio and multi-scale coupling that cannot be overcome by a single technique.

[0009] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. Attached Figure Description

[0010] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application. The illustrative embodiments and descriptions of the application are used to explain the application, but do not constitute an undue limitation of the application. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0011] In the attached diagram: Figure 1 This is a flowchart illustrating the fiber optic sensor signal diagnosis method based on TimesNet two-dimensional time-frequency transformation in this specific embodiment. Figure 2 This is a flowchart illustrating the multi-periodicity and two-dimensional temporal variation of the time series in this specific embodiment; Figure 3 This is a schematic diagram illustrating the process of converting the original one-dimensional time series into a structured two-dimensional tensor in this specific embodiment. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0013] Please see Figure 1 This application provides a method for diagnosing fiber optic sensor signals based on TimesNet two-dimensional time-frequency transform, comprising the following steps: S1, Signal Preprocessing: The original one-dimensional fiber optic sensing signal is input, and adaptive time-frequency denoising is performed through learnable wavelet packet transform (LWPT) to output an enhanced time-domain signal. S2, TimesNet periodic transformation: Based on the enhanced time-domain signal, the top K dominant frequency components are extracted by fast Fourier transform; for each dominant frequency component, a two-dimensional time-frequency tensor is constructed by zero-padding and reshaping operations, wherein the column dimension of the two-dimensional time-frequency tensor represents the transient micro-variation characteristics within a single period, and the row dimension represents the slow macro-variation characteristics across periods. S3, Feature Extraction: Feature extraction is performed on the two-dimensional time-frequency tensor using depthwise separable convolution, wherein grouped convolutional layers are used to extract local spatiotemporal coupling patterns in the row dimension, and pointwise convolutional layers are used to fuse cross-channel anomalous features; a SimAM attention mechanism is introduced to adaptively enhance the weights of key feature regions related to fiber faults. S4. Temporal Evolution Modeling: The extracted features are modeled temporally using a minLSTM network. The minLSTM network uses the Heinsen logarithmic domain association scanning algorithm to decompose the hidden state into a forward logarithmic flow and a backward logarithmic flow, thereby achieving parallelized long-range dependency modeling. S5. Multi-task output header output results: The classifiers, which share a common underlying feature representation, are jointly trained through a gradient conflict coordination mechanism to simultaneously complete multiple target tasks. The classifiers include a detection head, a classification head, and a localization head. The detection head outputs whether a fault exists, the classification head outputs the fault type, and the localization head outputs the fault location.

[0014] In practical applications, the following scenario is defined: φ-OTDR pipeline monitoring. A combination of minor bends and vehicle traffic vibrations is observed at kilometer 23.6.

[0015] Acquire the input, i.e., the original one-dimensional fiber optic sensing signal: ; Wherein, the sampling rate is defined. Window length , This represents a complete one-dimensional fiber optic sensing time series. This represents the sensor signal vector at the t-th sampling time in the time series, with dimension . , Indicates the index of the sampling time. This indicates the number of physical dimensions of the fiber optic sensing signal.

[0016] In some embodiments, an adaptive time-frequency denoising module using a Learnable Wavelet Packet Transform (LWPT) is first employed to output an enhanced time-domain signal. The LWPT is then used to perform learnable time-frequency denoising and structural enhancement on the distributed acoustic sensing (DAS) signal. Unlike traditional fixed filtering or static wavelet transform, the LWPT module learns the most suitable filter parameters and threshold coefficients for the task scenario through end-to-end training, enabling it to adapt to different noise environments and signal types.

[0017] The adaptive time-frequency denoising system consists of three steps: wavelet packet decomposition (encoding), adaptive threshold denoising, and wavelet packet reconstruction (decoding). Since effective components and noise are often mixed together in time-domain vibration signals and difficult to distinguish directly, we first use a wavelet packet transform-like approach to decompose the signal layer by layer into multiple frequency bands. Unlike traditional wavelet packets, the filters here are not fixed but replaced by learnable convolutional kernels, allowing the model to automatically learn a more suitable frequency band division method during training. Next, a learnable threshold "shrinkage function" is added after each frequency band to suppress low-amplitude (more likely noise) coefficients and retain high-amplitude (more likely effective signal) coefficients, achieving adaptive denoising by frequency band. Finally, the processed coefficients are synthesized layer by layer through the reconstruction structure to obtain the final denoised signal.

[0018] The input raw one-dimensional fiber optic sensing signal, such as a raw one-dimensional vibration signal, first passes through the LWPT denoising module. The LWPT denoising module employs a multi-layer decomposition structure (two layers in this embodiment), utilizing parallel low-pass and high-pass convolutional kernels to simulate the traditional wavelet packet decomposition process, and introducing learnable soft thresholding functions in each frequency band to achieve nonlinear compression. This processing effectively suppresses background noise and non-discriminatory components in the signal while preserving the main event-related frequency domain structure, thus significantly improving the signal-to-noise ratio and discriminability of the signal before entering the subsequent convolutional feature extraction stage.

[0019] In terms of network structure, the "denoised time-domain sequence" output by the LWPT denoising module is directly fed into the multi-layer TimesNet periodic feature analysis module to form a continuous information flow. The LWPT denoising module plays the role of "information purifier" in the whole framework, providing more stable and higher-quality input features for subsequent deep representation learning.

[0020] From an experimental perspective, the introduction of LWPT significantly improved the event recognition accuracy of the model under low signal-to-noise ratio conditions, indicating that the LWPT denoising module can effectively suppress environmental noise, transmission noise and system noise in distributed acoustic signals, providing a more robust foundation for feature learning in deep networks.

[0021] In some embodiments, the TimesNet periodic transformation is performed as follows: based on the enhanced time-domain signal, the top K dominant frequency components are extracted by fast Fourier transform; for each dominant frequency component, a two-dimensional time-frequency tensor is constructed by zero-padding and reshaping operations, wherein the column dimension of the two-dimensional time-frequency tensor represents the transient micro-variation characteristics within a single period, and the row dimension represents the slowly varying macro-variation characteristics across periods.

[0022] like Figure 2 As shown, there are multiple periodicities in time series. By identifying the dominant period, a one-dimensional time series can be unfolded into a two-dimensional structure in the time dimension. The column direction (within the period) captures transient micro-changes, while the row direction (during the period) captures gradual trends.

[0023] Specifically, the enhanced time-domain signal is first subjected to a Fast Fourier Transform (FFT) through an FFT period extraction module, automatically identifying the top k dominant frequencies in the frequency domain amplitude spectrum. These dominant frequencies, like the "exons" of a gene, determine the basic rhythm of the signal. Through intelligent zero-padding and reshaping operations, the one-dimensional time series is transformed into k two-dimensional time-frequency tensors (BxCxPixFi), where column vectors carry small variations within a period (such as scattering phase jitter on the order of 0.1 ns), and row vectors carry macro variations between periods (such as temperature gradients on the order of 10 minutes). This structure, for the first time, enables the mathematical orthogonal separation of the transient impact of fiber micro-bending and the gradual trend of soil thermal diffusion.

[0024] In some embodiments, based on the enhanced time-domain signal, the top K dominant frequency components are extracted using Fast Fourier Transform (FFT), including: performing FFT on each channel of the input enhanced time-domain signal, wherein the time-domain signal is a one-dimensional time series, and calculating the frequency domain amplitude spectrum of the one-dimensional time series; identifying and selecting the top K dominant frequency components with the largest amplitude from the frequency domain amplitude spectrum. Each peak value corresponds to a dominant frequency; the signal period corresponding to each dominant frequency is calculated based on the sampling rate.

[0025] In practical applications, for each dominant frequency component, a two-dimensional time-frequency tensor is constructed through zero-padding and reshaping operations, including: for each dominant frequency and its period, calculating the maximum number of complete periods of the one-dimensional time series under the given period, and determining the target length after zero-padding; and setting the starting point of the original sequence... Each data point is retained, and zeros are padded to the end to the length. To align an integer number of periods; reshape the zero-padded one-dimensional time series into a two-dimensional time-frequency tensor. Among them, column dimensions Represents the number of sampling points within a period, used to represent the subtle variations within the period; row dimension. This represents the number of complete cycles, used to carry the macro-variation characteristics between cycles, and C is the number of input channels.

[0026] To further explain, Figure 3 This paper demonstrates an example process for converting a univariate time series into a two-dimensional structure, including period identification, zero-padding alignment, and two-dimensional reshaping.

[0027] Specifically, the master cycle automatically extracts the FFT: The frequency domain amplitude spectrum is calculated for each channel and expressed as: ; in, express The frequency domain representation (complex form). Indicates to Perform a Fourier transform. Represents the discrete-time sequence sampling points in the original time domain. Represents the frequency domain amplitude spectrum; from Select the first peak value dominant frequency Corresponding period , This indicates the sampling rate.

[0028] Intelligent zero-filling and cycle reshaping: To align integer periods, zero-padding is performed: Let ,in, This indicates the length of the target sequence after padding with zeros. Indicates dominant frequency The corresponding time-domain period length is padded with zeros at the end. Then, the one-dimensional time series is reshaped into a two-dimensional time-frequency tensor, represented as: ; in, Represents a two-dimensional time-frequency tensor. express , express .

[0029] In some embodiments, the feature extraction of the two-dimensional time-frequency tensor through depthwise separable convolution includes: grouped convolutional layers for extracting local spatiotemporal coupling patterns in the row dimension, and pointwise convolutional layers for fusing cross-channel anomalous features. This includes: performing channel-wise two-dimensional convolution on the input two-dimensional time-frequency tensor, independently extracting local coupling patterns in the spatiotemporal dimension of each channel; and processing the feature tensor after depthwise convolution. A 1×1 convolution operation is applied to achieve cross-channel feature fusion, resulting in a fused feature tensor. .

[0030] Specifically, lightweight 2D convolutional feature extraction: For each Extracting local patterns using depthwise separable convolution and fusing them across channels: Channel-wise depthwise convolution, where each channel is convolved independently, captures the local coupling between periodicity and time frequency, and is represented as: ; in, Represents the convolution window, This represents the feature tensor after depthwise convolution. Feature fusion is performed using 1×1 point convolution to fuse information between channels, resulting in... , represented as: ; in, This represents the feature tensor after fusion.

[0031] In some embodiments, the introduction of the SimAM attention mechanism to adaptively enhance the weights of key feature regions associated with fiber faults includes: applying the fused feature tensor to... Input a SimAM attention mechanism to generate a fused feature tensor. Attention weight maps of the same dimension are used, and the fused feature tensor is multiplied element-wise. Weighting is applied to enhance key feature regions relevant to the sensing event and suppress noise, resulting in a refined feature output. , for k periods Multi-scale fusion is performed to obtain a two-dimensional feature map. .

[0032] Attention weighting (such as channel attention or pixel attention) suppresses global noise and enhances key regions related to "micro-bending / vibration," represented as: ; in, Indicates the characteristics of refined products. This represents the attention weight generation function. These represent the height, width, and number of channel dimensions, respectively. Multi-scale fusion is represented as: ; in, Indicates the number of multi-scale features. Indexes representing multi-scale features Represents a two-dimensional feature map.

[0033] It should be noted that the lightweight 2D convolution and attention, used to "amplify" useful features and suppress noise in the 2D time-frequency plane, are the key engineering implementation points of this invention.

[0034] In some embodiments, the temporal modeling of the extracted features using a minLSTM network includes: processing the two-dimensional feature map... Expand the sequence into a one-dimensional sequence in row or column order; process the one-dimensional sequence using a minimal long short-term memory network minLSTM, wherein the hidden state update of the minLSTM is implemented through a parallelized association scan algorithm, and the long-range dependencies in the one-dimensional sequence are modeled with linear computational complexity.

[0035] Specifically, two-dimensional feature maps (In row / column or raster order) fed into minLSTM for efficient modeling, capturing long-range temporal dependencies across kilometers with a small number of parameters: The minLSTM form is represented as: ; ; ; ; ; ; in, For the current time step The input vector, i.e., from Sampling or pooling sequence features , , , This represents the weight matrix of the input gate, the weight matrix of the forget gate, the weight matrix of the output gate, and the weight matrix of the candidate memory. , , , Let represent the bias vectors of the input gate, the forget gate, the output gate, and the candidate memory. , , This represents the input gate, forget gate, and output gate. The memory unit representing the current time step. This represents the memory unit from the previous time step. Indicates candidate memory units. Indicates a hidden state. This represents the sigmoid activation function. This represents the hyperbolic tangent activation function; the engineering implementation employs parallelized correlation scans to approximate or equivalently capture long-range dependencies and reduce complexity.

[0036] In some embodiments, S5, the multi-task output head outputs the result: the result is jointly output by multiple classifiers that share the same underlying feature representation. The multiple classifiers are jointly trained through a gradient conflict coordination mechanism to complete multiple target tasks simultaneously. The classifier includes a detection head, a classification head, and a localization head. The detection head outputs whether a fault exists, the classification head outputs the fault type, and the localization head outputs the fault location.

[0037] The classification header outputs the fault type, which is represented as follows: ; in, This represents the fault detection result (binary classification probability output), with a value range of (0,1). The closer it is to 1, the higher the probability of a fault exists; the closer it is to 0, the higher the probability of no fault exists. By setting a threshold (such as 0.5), it can be determined whether there is a "fault (1)" or "no fault (0)". This represents the weight matrix of the detection head. This represents the final temporal aggregated features after minLSTM processing. This represents the bias vector of the detection head, used to adjust the offset of the linear mapping and improve the model's fitting ability.

[0038] Whether the output of the detection head is faulty is indicated as follows: ; in, This represents the fault type classification result (multi-class probability distribution). This represents a multi-class activation function used to convert the result of a linear mapping into a probability distribution. This represents the weight matrix of the classification heads. This represents the bias vector of the classification head. Represents a K-dimensional simplex space; The fault location output by the positioning head is indicated as follows: ; in, Represents the position score vector. Represents the position mapping function, The output indicates the final fault location. This represents the maximum value index function, used to select the index of the sampling point with the highest confidence from the position score vector s.

[0039] Joint losses These are represented as loss weight coefficients, all of which are positive numbers, used to balance the loss contributions of the three tasks: fault detection, type classification, and location localization. ; in, The cross-entropy loss indicates that it is used to calculate the loss of a fault detection task. This represents the cross-entropy loss, used to calculate the loss for fault classification tasks. This represents the smoothed L1 loss, used to calculate the loss in fault location tasks. , , These represent the actual labels for fault detection, type classification, and location positioning, respectively.

[0040] The gradient conflict coordination mechanism includes calculating the loss of each task relative to the shared backbone network parameters. The gradients of the detection, classification, and localization tasks are obtained, and a dot product is performed on the gradients of each task pair. For example, if the product of the detection and classification gradients is less than zero, then the gradients of the detection and classification tasks are considered to be within the range of zero. There is a conflict on a certain dimension.

[0041] Prioritize ensuring the gradient direction of the detection task, and remove the components that are negatively correlated with the detection task gradient and the localization task gradient from the classification task gradient and the localization task gradient. Specifically, project the detection task gradient onto a plane orthogonal to the classification task gradient to obtain the harmonic classification task gradient and the localization task gradient.

[0042] The reciprocal of the norm of the harmonic gradient for each task is calculated and used as a weighting factor. The original loss is then reweighted using dynamic weights to obtain the loss used for final backpropagation. Based on the harmonic gradient and the total loss calculated using the weighting factor, standard backpropagation is performed. The gradient is first applied to the shared backbone network. The parameters are aggregated according to the aforementioned coordination mechanism and then propagated to the private parameters of the three task heads. The optimizer AdamW is used to update all network parameters; the hyperparameter settings include an initial learning rate of 1e−4 to... The batch size is 16 or 32, the number of training rounds is 100-200, and the norm threshold is set to 1.0 to prevent gradient explosion.

[0043] The end-to-end input-output link of this method is clear and complete: the initial input is the sampled one-dimensional fiber optic sensing signal. (in These represent the physical dimensions of a signal, such as intensity, phase, and polarization. For window length, (Sampling rate); The process involves several key steps, first using Learnable Wavelet Packet Transform (LWPT) for adaptive time-frequency denoising to enhance signal quality, and then using TimesNet periodic transform to extract the previous signal. Each dominant frequency component is zero-padded and reshaped to construct a two-dimensional time-frequency tensor (the column dimension carries transient micro-variation features within a period, and the row dimension carries slowly varying macro-variation features across periods). Local spatiotemporal coupling patterns are then extracted using depthwise separable convolution and fused with cross-channel anomaly features. The SimAM attention mechanism is then used to enhance the weights of fault-related key regions. The resulting two-dimensional feature map is then unfolded into a one-dimensional sequence in row / column or grid order, input into a minLSTM network, and parallelized long-range dependency modeling is achieved using the Heinsen logarithm domain correlation scanning algorithm. Finally, temporal aggregated features are obtained. The final output is a triple. These correspond to the detection results of whether a fault exists, the classification results of the fault type, and the location results of the physical location of the fault, respectively, realizing the fully automated processing of fiber optic sensor signal fault diagnosis.

[0044] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although this application has disclosed preferred embodiments as above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-mentioned technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of this application.

Claims

1. A method for diagnosing optical fiber sensing signals based on TimesNet two-dimensional time-frequency transform, characterized in that, The method comprises the following steps: S1, signal preprocessing: An original one-dimensional optical fiber sensing signal is input, adaptive time-frequency denoising is performed on the original one-dimensional optical fiber sensing signal through a learnable wavelet packet transform (LWPT), and an enhanced time-domain signal is output; S2, TimesNet period transformation: Based on the enhanced time-domain signal, the first K dominant frequency components are extracted through a fast Fourier transform; for each of the dominant frequency components, a two-dimensional time-frequency tensor is constructed through zero padding and reshaping, wherein the column dimension of the two-dimensional time-frequency tensor represents the transient micro-variation characteristics in a single period, and the row dimension represents the slowly-varying macro-variation characteristics across periods; S3, feature extraction: The two-dimensional time-frequency tensor is subjected to feature extraction through a deep separable convolution, wherein a deep convolution layer is used to extract a local space-time coupling mode in the row dimension, and a point-by-point convolution layer is used to fuse abnormal features across channels; a SimAM attention mechanism is introduced to adaptively enhance the weight of a key feature area related to an optical fiber fault; S4, time series evolution modeling: The extracted features are subjected to time series modeling through a minLSTM network, wherein the minLSTM network adopts a Heinsen logarithmic domain correlation scanning algorithm to decompose a hidden state into a forward logarithmic flow and a reverse logarithmic flow, so as to realize parallel long-range dependence modeling; S5, output results through a multi-task output head: Joint output is performed through multiple classifiers sharing a bottom layer feature representation, the multiple classifiers are jointly trained through a gradient conflict coordination mechanism, and multiple target tasks are simultaneously completed; wherein the classifiers include a detection head, a classification head, and a positioning head, the detection head outputs whether a fault exists, the classification head outputs a fault type, and the positioning head outputs a fault position.

2. The method of claim 1, wherein, Based on the enhanced time-domain signal, the first K dominant frequency components are extracted through a fast Fourier transform, which comprises: performing a fast Fourier transform on each channel of the input enhanced time-domain signal, wherein the time-domain signal is a one-dimensional time series, and calculating a frequency-domain amplitude spectrum of the one-dimensional time series; identifying and selecting a number of peaks with the largest amplitudes from the frequency-domain amplitude spectrum, wherein the peaks correspond to dominant frequencies; and calculating a signal period corresponding to each dominant frequency according to a sampling rate. ​ 3. The method of claim 2, wherein, For each of the dominant frequency components, a two-dimensional time-frequency tensor is constructed through zero padding and reshaping, which comprises: For each of the dominant frequency components and the period thereof, the maximum complete period number of the one-dimensional time series in the period is calculated, and a target length after zero padding is determined; Start of the original sequence Each data point is retained, and zeros are padded to the end to the length. To align an integer number of cycles; The one-dimensional time sequence after zero padding is reshaped into a two-dimensional time-frequency tensor ; wherein the column dimension represents the number of sampling points in a cycle, used to carry the micro-variation characteristics in the cycle; the row dimension represents the number of complete cycles, used to carry the macro-variation characteristics between cycles, and C is the number of input channels.

4. The method of claim 1, wherein, The two-dimensional time-frequency tensor is subjected to feature extraction through a deep separable convolution, wherein a deep convolution layer is used to extract a local space-time coupling mode in the row dimension, and a point-by-point convolution layer is used to fuse abnormal features across channels, which comprises: The two-dimensional convolution operation is performed on the input two-dimensional time-frequency tensor channel by channel, and the local coupling mode is extracted on the space-time dimension of each channel independently A 1x1 convolution operation is applied to realize cross-channel feature fusion, and a fused feature tensor is obtained .

5. The method of claim 4, wherein, The SimAM attention mechanism is introduced to adaptively enhance the weight of a key feature area related to an optical fiber fault, which comprises: The fused feature tensor Input a SimAM attention mechanism to generate a fused feature tensor. Attention weight maps of the same dimension are used, and the fused feature tensor is multiplied element-wise. Weighting is applied to enhance key feature regions relevant to the sensing event and suppress noise, resulting in a refined feature output. The k dominant frequency components correspond to Multi-scale fusion is performed to obtain a two-dimensional feature map. .

6. The method of claim 5, wherein, The extracted features are subjected to time series modeling through a minLSTM network, which comprises: The two-dimensional feature map Expand the sequence into a one-dimensional sequence in row or column order; process the one-dimensional sequence using a minimal long short-term memory network minLSTM, wherein the hidden state update of the minLSTM is implemented through a parallelized association scan algorithm to model the long-range dependencies in the one-dimensional sequence with linear computational complexity.

Images