Feature synthesis-based phi-otdr zeroth event recognition method, device, system and medium

Abstract

The application relates to the technical field of optical fiber sensing, in particular to a Phi-OTDR zero event recognition method, device and system based on feature synthesis and a medium, the method comprises the following steps: feature extraction is performed on a known vibration sample of a known vibration type to obtain a training feature data set; a semantic vector of each vibration type is generated according to language definition, mixed vibration samples are obtained by selecting similar vibration types, and a semantic space conversion relationship is calculated; a common feature extractor is trained by using the training feature data set, and a unique target feature extractor is trained by using the mixed vibration samples and a guiding training method; the outputs of the two feature extractors are added and combined according to the semantic space conversion relationship weight, a synthetic feature vector is obtained, a classifier is trained, and unknown vibration samples are recognized; the application solves the problem of lacking vibration samples in the neural network training process, and enables the optical fiber early warning system to still have vibration recognition capability without prior training.

Description

Technical Field

[0001] This invention relates to the field of fiber optic sensing technology, and in particular to a method, apparatus, system, and medium for zero-event identification of Φ-OTDR based on feature synthesis. Background Technology

[0002] Distributed fiber optic sensing systems are widely used in long-distance oil and gas pipeline monitoring, perimeter security, and building structural health monitoring due to their high sensitivity, high positioning accuracy, and electromagnetic immunity. Phase-sensitive optical time-domain reflectometer (Φ-OTDR) technology uses a long-coherent light source to detect the coherence of the returned light pulse. Its interferometric method effectively achieves dynamic response, simultaneously achieving high positioning accuracy and high sensitivity detection. It is particularly suitable for detecting weak disturbance signals, making it ideal for pipeline early warning systems.

[0003] With the development of artificial intelligence, Φ-OTDR has introduced an event recognition method based on machine learning and deep learning, demonstrating good vibration event recognition capabilities and robustness in field tests. All machine learning or deep learning classifiers are data-driven and require a training dataset containing all types of events to be tested. However, in practical applications, it is difficult to obtain a perfect training dataset with a large number of known vibration samples and all event types. We only know the abstract concepts of some event types, but cannot obtain known vibration samples of them in advance. If the training dataset does not contain known vibration samples of a specific event type, these data-driven recognition methods cannot identify the event in tests.

[0004] On the other hand, zero-order learning is an effective method for solving the problem of having few or no sample instances. Instance synthesis is a type of zero-order learning method used for tasks involving the identification of known vibration samples without training or for identifying unknown classes. However, existing zero-order learning methods cannot be directly applied to fiber optic sensing applications because there is no semantic set to describe vibration or sensing signals in these applications, and the effectiveness of ordinary instance synthesis methods is limited. Currently, there is no effective method for solving the problem of classification without samples in the field of fiber optic sensing. Therefore, a method is needed that can complete vibration classification tasks without pre-training on specific vibrations. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a Φ-OTDR zero-event identification method, device, system, and medium based on feature synthesis, aiming to achieve high-quality feature synthesis and vibration type identification of unknown vibration samples, effectively improving the vibration identification capability of fiber optic early warning systems in the absence of specific vibration training samples.

[0006] On one hand, embodiments of the present invention provide a Φ-OTDR zero-event recognition method based on feature synthesis, the method comprising the following steps: Acquire known vibration samples of known vibration types and determine unknown vibration samples of vibration types to be identified. Extract features from the known vibration samples of known vibration types to obtain a training feature dataset. Obtain the linguistic definitions of various vibrations, generate a semantic vector for each type of vibration based on the linguistic definitions, select other vibration type samples similar to the unknown vibration sample based on similarity calculation, mix the training samples of the other vibration type samples to obtain a mixed vibration sample, and calculate the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space. A general feature extractor is trained using the training feature dataset, and a unique target feature extractor is trained using the mixed vibration samples and the guided training method. The samples in the training feature dataset are input into a unique target feature extractor and a regular feature extractor. The mixed vibration samples are input into the corresponding unique target feature extractor. The outputs of the unique target feature extractor are weighted and summed according to the transformation relationship in the semantic space. Then, the outputs of the regular feature extractor are combined to obtain a synthetic feature vector. The synthetic feature vector is used as a training sample to train a classifier. The trained classifier is used to identify the vibration type of unknown vibration samples.

[0007] Optionally, the step of extracting features from known vibration samples of the known vibration type to obtain a training feature dataset includes: The phase signal of Rayleigh scattered light at the time of vibration is obtained by using a fiber optic early warning system, and multiple Rayleigh scattered light phase signals are rearranged into a matrix form to obtain a spatiotemporal matrix signal. The vibration in the spatiotemporal matrix signal is located using a positioning method. Regional data of the corresponding position in the spatiotemporal matrix signal is extracted, and the column direction of the regional data is high-pass filtered to remove the DC component, thus obtaining the filtered regional data. Perform Mel-Cepstral Transform on the filtered regional data to obtain the Mel-Cepstral plot; The encoder and decoder are trained using a self-supervised learning method, with the Mel cepstral as input, the reconstructed Mel cepstral as output, and the difference between the reconstructed Mel cepstral and the input Mel cepstral as the loss function. The trained encoder is used to encode the Mel cepstral, and the encoding result is recorded as a feature data sample. At the same time, the vibration type of the feature data sample is recorded as a label to obtain the training feature dataset.

[0008] Optionally, performing Mel-Cepstral Transform on the filtered regional data to obtain a Mel-Cepstral Map includes: Each frame of the filtered regional data is multiplied by a Hamming window to obtain the windowed signal. For each short-term analysis window, the spectrum of different time window distributions on the time axis is obtained by discrete Fourier transform; The spectrum is passed through a Mel filter bank to obtain a Mel cepstral plot, wherein the Mel filter bank consists of multiple triangular filters; Record the energy value, and then transform the signal using a discrete cosine transform to obtain the Mel-Cepstral.

[0009] Optionally, the step of generating a semantic vector for each type of vibration according to the language definition, selecting other vibration type samples similar to the unknown vibration sample based on similarity calculation, mixing the training samples of the other vibration type samples to obtain a mixed vibration sample, and calculating the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space includes: The linguistic descriptions of all known vibration samples and unknown vibration samples are input into the semantic model. The semantic model is guided by prompt words to output a semantic vector for each type of vibration. The dimension of the semantic vector is preset, and each value in the semantic vector is a discrete value. Calculate the Euclidean distance between the semantic vector of each type of vibration and the semantic vector of the unknown vibration sample, select the vibration with the closer distance as the hybrid sub-vibration, mix the training samples of the hybrid sub-vibration with each other to obtain the hybrid vibration sample, and combine the hybrid vibration sample with other single vibration samples to form the training set of the feature extractor. The semantic vectors of the mixed sub-vibrations are summed and the average value is taken as the semantic vector of the mixed vibration sample. The transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space is calculated.

[0010] Optionally, calculating the transformation relationship between the mixed vibration samples and the unknown vibration samples in the semantic space includes: Calculate the semantic vector of the hybrid vibration sample, where the semantic vector of the hybrid vibration sample is the average of the sum of the semantic vectors of the hybrid sub-vibrations; Calculate the semantic spatial distance between the mixed vibration sample and the unknown vibration sample; Calculate the average semantic spatial distance between all the mixed vibration samples and the unknown vibration samples; Based on the semantic space distance and the average value, the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space is calculated.

[0011] Optionally, training the unique target feature extractor using the hybrid vibration samples and guided training method includes: The hybrid vibration samples are used as the target class, and the samples are input into a deep Q-learning network. When the true label of the sample is the label of the target class, if the predicted label generated by the deep Q-learning network is equal to the true label, a positive reward is fed back to the deep Q-learning network; otherwise, a negative reward is fed back to the deep Q-learning network. When the true label of the sample is not the target class label, if the predicted label generated by the deep Q-learning network is not equal to the target class label, a positive reward is fed back to the deep Q-learning network; otherwise, a negative reward is fed back to the deep Q-learning network. After training the deep Q-learning network, the first to second-to-last layers of the deep Q-learning network are used as the sole target feature extractor.

[0012] Optionally, the step of inputting samples from the training feature dataset into a unique target feature extractor and a regular feature extractor, inputting the mixed vibration samples into the corresponding unique target feature extractor, weighting and summing the outputs of the unique target feature extractor according to the transformation relationship in the semantic space, and then combining them with the output of the regular feature extractor to obtain a synthetic feature vector includes: The samples in the training feature dataset are respectively input into the unique target feature extractor and the ordinary feature extractor, and the mixed vibration samples are respectively input into the unique target feature extractor with corresponding recognition capabilities; The transformation relationship in the semantic space is used as the output weight of the unique target feature extractor. The output of the unique target feature extractor for the training feature dataset samples is added according to the weight and concatenated with the output of the ordinary feature extractor to synthesize a synthetic feature vector of known vibration type. Based on the transformation relationship in the semantic space, the outputs of the unique target feature extractor for the mixed vibration sample are added according to the weights, and the weighted sum is copied and spliced ​​to synthesize a synthetic feature vector of unknown vibration type. The synthetic feature vectors of the known vibration type and the unknown vibration type are used as the synthetic feature vectors.

[0013] On the other hand, embodiments of the present invention provide a Φ-OTDR zero-event recognition device based on feature synthesis, comprising: The first module is used to acquire known vibration samples of known vibration types and determine unknown vibration samples of vibration types to be identified, and to extract features from the known vibration samples of known vibration types to obtain a training feature dataset. The second module is used to obtain the linguistic definitions of various vibrations, generate a semantic vector for each type of vibration based on the linguistic definitions, select other vibration type samples similar to the unknown vibration sample based on similarity calculation, mix the training samples of the other vibration type samples to obtain a mixed vibration sample, and calculate the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space. The third module is used to train a general feature extractor using the training feature dataset and to train a unique target feature extractor using the mixed vibration samples and the guided training method. The fourth module is used to input samples from the training feature dataset into a unique target feature extractor and a regular feature extractor, input mixed vibration samples into the corresponding unique target feature extractor, add the outputs of the unique target feature extractor according to the transformation relationship in the semantic space, and then combine them with the output of the regular feature extractor to obtain a synthetic feature vector. The synthetic feature vector is used as a training sample to train a classifier, and the trained classifier is used to identify the vibration type of unknown vibration samples.

[0014] On the other hand, embodiments of the present invention provide a Φ-OTDR zero-event recognition system based on feature synthesis, comprising: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0015] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the above-described method.

[0016] The embodiments of the present invention have the following beneficial effects: This invention utilizes a Φ-OTDR zero-order event recognition method based on feature synthesis, employing a hybrid class approach and a reinforcement learning-based guided training method to achieve high-quality synthesis of unknown vibration features. Specifically, mixing multiple vibration samples highlights the common features of the mixed vibration samples while weakening other marginal features. The common features of the mixed sub-vibrations are precisely the important features of the unknown vibration, expanding the feature space while emphasizing common features. Furthermore, the reinforcement learning-based guided training method ensures the network can only recognize the target class, blurring other classes. This makes the network more focused on the target class, resulting in a purer output feature vector that better preserves the feature information of the target class. Simultaneously, using a large model for semantic vector construction avoids human cognitive errors. Moreover, the transformation relationships and similarities between vibrations in the semantic space also apply to the feature space. Therefore, by mixing vibration samples and the transformation relationships of the unknown vibration in the semantic space, the feature vector of the unknown vibration can be synthesized in the feature space.

[0017] This invention solves the problem of insufficient vibration samples during neural network training, enabling fiber optic early warning systems to identify specific vibrations even without prior training. The system uses multiple acousto-optic modulators to adjust the heterodyne beat frequency, allowing the heterodyne signal to be directly acquired by a data acquisition card, reducing the complexity and cost of the acquisition module. This is currently a relatively effective solution for addressing the problem of classification without samples in the field of fiber optic sensing. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0019] Figure 1 This is a schematic diagram of the optical path of the fiber optic early warning system provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the steps of a Φ-OTDR zero-event recognition method based on feature synthesis provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the deep Q-learning network provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a known vibration sample obtained by the method of an embodiment of the present invention.

[0020] Figure 5 This is a logical schematic diagram of a Φ-OTDR zero-event recognition device based on feature synthesis provided in an embodiment of the present invention.

[0021] Explanation of reference numerals in the attached figures: 1: Narrow linewidth light source; 2: First 50:50 coupler; 3: Frequency-shifting acousto-optic modulator; 4: 90:10 coupler; 5: Electro-optic modulator; 6: First circulator; 7: Filter grating; 8: First optical amplifier; 9: Second circulator; 10: Sensing fiber; 11: Pulse-modulated acousto-optic modulator; 12: Second optical amplifier; 13: Third circulator; 14: Second 50:50 coupler; 15: Photodetector; 16: Data acquisition card and host computer. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] It should be noted that although the device diagram shows a modular division and the flowchart illustrates a logical order, in some cases, the steps shown or described may be performed in a different order than the modular division in the device or the order shown in the flowchart. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0024] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the invention.

[0025] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0026] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0027] In this embodiment of the invention, the fiber optic early warning system uses Brillouin distributed amplification to improve the sensing distance and utilizes multiple acousto-optic modulators to adjust the optical frequency shift of the probe light and the local oscillator light, thereby generating a low-frequency heterodyne detection beat frequency. Detection can be completed directly using a conventional acquisition card. The system is simple, reliable, and inexpensive.

[0028] like Figure 1 As shown, the fiber optic early warning system includes: a narrow linewidth light source 1, a first 50:50 coupler 2, a frequency-shifting acousto-optic modulator 3, a 90:10 coupler 4, an electro-optic modulator 5, a first circulator 6, a filter grating 7, a first optical amplifier 8, a second circulator 9, a sensing fiber 10, a pulse-modulated acousto-optic modulator 11, a second optical amplifier 12, a third circulator 13, a second 50:50 coupler 14, a photodetector 15, and a data acquisition card and a host computer 16.

[0029] Narrow linewidth light source 1 is connected to a first 50:50 coupler 2. The first 50:50 coupler 2 is connected to a frequency-shifting acousto-optic modulator 3 and a pulse-modulated acousto-optic modulator 11. The pulse-modulated acousto-optic modulator 11 is connected to a second optical amplifier 12, and then to a third circulator 13. The output of the third circulator 13 is connected to the beginning of the sensing fiber 10, and the return end of the third circulator 13 is connected to a second 50:50 coupler 14. The frequency-shifting acousto-optic modulator 3 is connected to a 90:10 coupler 4. One 10% output of the 90:10 coupler 4 is connected to the second 50:50 coupler 14, and the other 90% output is connected to an electro-optic modulator 5. Then, it is connected to a filter grating 7 through a first circulator 6, and then returns to the first optical amplifier 8 through the first circulator 6. Finally, it is connected to the end of the sensing fiber 10 through a second circulator 9.

[0030] A narrow-linewidth light source 1 emits continuous laser light (optionally, the wavelength of the continuous laser light is 1550 nm), which is split into pump light and probe light by a first 50:50 coupler 2. The probe light is modulated into probe pulse light by a pulse-modulated acousto-optic modulator 11, generating a frequency shift f1 in the MHz range (optionally, this frequency shift is 200 MHz). After passing through a second optical amplifier 12, it is injected into the sensing fiber 10 by a third circulator 13. The probe pulse light generates backscattered Rayleigh light in the sensing fiber 10, which reaches the second 50:50 coupler 14 via the third circulator 13. The pump light passes through a frequency-shifted acousto-optic modulator 3, generating another frequency shift f2 in the MHz range (optionally, this frequency shift is 180 MHz), and then reaches the 90:10 coupler 4. After passing through the 90:10 coupler 4, 10% of the pump light becomes the local oscillator light, and the remaining 90% reaches the electro-optic modulator 5, generating a double-sideband frequency shift of approximately 10.8 GHz. The specific value of this frequency shift is determined by the Brillouin frequency shift value of the sensing fiber 10. The double-sideband-shifted pump light passes through the first circulator 6 and the filter grating 7 to filter out the upper sideband. It then undergoes power amplification through the first optical amplifier 8 and is injected into the sensing fiber 10 from the reverse end via the second circulator 9. There, it undergoes energy transfer with the pulse probe light within the sensing fiber 10, completing the distributed optical amplification of the probe pulse light and increasing the effective sensing distance. The local oscillator light interferes with the Rayleigh scattering of the probe light at the second 50:50 coupler 14, forming a modulated signal with a carrier beat frequency of |f1-f2|. The energy of the frequency-shifted local oscillator light and the backscattered Rayleigh light each account for 50%. After the modulated signal is converted into an electrical signal by the photodetector 15, it is orthogonally demodulated by the acquisition card and the host computer 16 using sine and cosine waves with frequencies of |f1-f2| (optionally, the acquisition card and the host computer 16 acquire the electrical signal at a preset sampling rate of 100MHz and use sine and cosine waves with frequencies of 20MHz to orthogonally demodulate the electrical signal). After demodulation, the Rayleigh scattering phase signal TR is obtained for subsequent use.

[0031] like Figure 2 As shown in the figure, the Φ-OTDR zero-event recognition method based on feature synthesis provided by the present invention includes the following steps: S100: Obtain known vibration samples of known vibration types and determine unknown vibration samples of vibration types to be identified; extract features from the known vibration samples of known vibration types to obtain a training feature dataset. S200, obtain the language definitions of various vibrations, generate the semantic vector of each type of vibration according to the language definitions, select other vibration type samples similar to the unknown vibration sample according to the similarity calculation, mix the training samples of the other vibration type samples to obtain mixed vibration samples, and calculate the transformation relationship between the mixed vibration samples and the unknown vibration samples in the semantic space. S300, a general feature extractor is trained using the training feature dataset, and a unique target feature extractor is trained using the mixed vibration samples and the guided training method; S400, input the samples in the training feature dataset into the unique target feature extractor and the ordinary feature extractor, input the mixed vibration samples into the corresponding unique target feature extractor, add the outputs of the unique target feature extractor according to the transformation relationship in the semantic space, and then combine them with the output of the ordinary feature extractor to obtain a synthetic feature vector. Use the synthetic feature vector as a training sample to train the classifier, and use the trained classifier to identify the vibration type of the unknown vibration sample.

[0032] This invention proposes a Φ-OTDR zero-order event recognition method based on feature synthesis. It establishes the prerequisite of zero-order learning by acquiring known vibration samples of known vibration types and identifying unknown vibration samples of the type to be identified. Features are extracted from the known vibration samples of known vibration types to obtain a training feature dataset, providing foundational data support for subsequent feature synthesis and classifier training. Similarity calculation is used to select known vibration types similar to the unknown vibration samples for mixing. By calculating transformation relationships in the semantic space, semantic similarity is mapped to the feature space, achieving an effective transformation from semantic space to feature space. Mixing multiple vibration samples highlights common features and weakens marginal features, expanding the feature space range. A unique target feature extractor is trained using a guided training method, focusing the network on the target class, resulting in a purer output feature vector that effectively suppresses interference from non-target class features. A general feature extractor is trained using the training feature dataset, providing basic feature representation and complementing the unique target feature extractor. By weighting and combining the outputs of the two feature extractors, the synthesized feature vector possesses both discriminative power for the target class and retains sufficient basic information. Feature synthesis is guided by transformation relationships in the semantic space, ensuring that the distribution of the synthesized features in the feature space remains consistent with the true feature distribution of the unknown vibration samples, thereby improving the classifier's accuracy in recognizing unknown vibration samples. The synthesized feature vector is used as training samples for the classifier. After training, the classifier is used to identify the vibration type of unknown vibration samples, obtaining the recognition results. This solves the problem of insufficient vibration samples during neural network training, enabling the fiber optic early warning system to still identify the corresponding vibration type even without pre-training on specific vibrations.

[0033] In some embodiments, the specific steps in S100 for extracting features from known vibration samples of known vibration types to obtain a training feature dataset are as follows: S110, using an optical fiber early warning system to obtain the Rayleigh scattering phase signal when the vibration occurs, rearranging multiple Rayleigh scattering phase signals into a matrix form to obtain a spatiotemporal matrix signal.

[0034] S120, the vibration in the spatiotemporal matrix signal is located using a positioning method, the corresponding regional data in the spatiotemporal matrix signal is extracted, and the column direction of the regional data is high-pass filtered to remove the DC component, so as to obtain the filtered regional data.

[0035] In this embodiment, multiple Rayleigh scattering phase signals obtained within a certain time period are rearranged into a matrix form to obtain a spatiotemporal matrix signal. Assuming the repetition frequency of the detection light pulse in the fiber optic early warning system is 1 kHz, a total of 1000 Rayleigh scattering phase signals are obtained within 1 second, and the resulting spatiotemporal matrix signal is denoted as... The phase signal of the i-th Rayleigh scattering light is denoted as The superscript T indicates transpose. In this case, the row dimension of the spatiotemporal matrix signal represents the sensing space dimension, and the column dimension represents the temporal order of the Rayleigh scattered light phase signals. Traditional positioning methods (including but not limited to the moving differential averaging method and wavelet entropy positioning method) are used to locate vibrations and extract the spatiotemporal matrix signal. The data in the corresponding location area is denoted as For spatiotemporal matrix signals High-pass filtering is performed in the column direction to remove the DC component.

[0036] S130 performs Mel-Cepstral Transform on the filtered regional data to obtain the Mel-Cepstral Map.

[0037] S140, the encoder and decoder are trained using the Mel cepstral as input, the reconstructed Mel cepstral as output, and the difference between the reconstructed Mel cepstral and the input Mel cepstral as the loss function, through a self-supervised learning method.

[0038] In this embodiment, the size of the Mel-Cepstral plot is adjusted to 256×256×3 for known vibration samples, and the type of vibration is recorded as the label for the known vibration sample. A training feature dataset O is constructed, and the resulting known vibration samples are as follows: Figure 3 As shown, an encoder and decoder are constructed, taking the Mel-Cepstral as input and the reconstructed Mel-Cepstral as output. The difference between the reconstructed Mel-Cepstral and the input Mel-Cepstral is used as the loss function. The encoder and decoder are trained through self-supervised learning.

[0039] S150, the trained encoder is used to encode the Mel cepstral, the encoding result is recorded as a feature data sample, and the vibration type of the feature data sample is recorded as a label to obtain the training feature dataset.

[0040] In this embodiment, a trained encoder is used to encode all the obtained Mel cepstral data, and the encoding results are recorded together with the original vibration types to construct a training feature dataset O.

[0041] Specifically, a fiber optic early warning system was used to obtain known vibration samples, totaling 11,703 samples across seven vibration categories: vibration samples from sunny days, vibration samples from rainy days, vibration samples from people walking, vibration samples from people jumping, vibration samples from water jets, vibration samples from shovels striking the ground, and vibration samples from bicycles passing by. The known vibration samples were preprocessed and converted into Mel-Cepstral Maps (MCMs). These MCMs were used as input to train an encoder and decoder network with the goal of reconstructing the MCMs. Both the encoder and decoder were 5-layer neural networks. All MCMs were sequentially input into the trained encoder to obtain their encoded features, and the corresponding vibration sample labels were recorded, denoted as the original training set O. The steps S110 to S150 described above train the encoder and decoder through self-supervised learning, enabling them to learn effective feature representations of vibration data without relying on external labels. Using the encoder as a general feature extractor allows for efficient feature extraction from known vibration samples of known vibration types, providing fundamental features for subsequent feature synthesis. Simultaneously, the Mel-Cepstral Transform (MCT) converts the time-domain vibration signal into a time-frequency domain representation, better capturing the spectral characteristics of the vibration signal and facilitating subsequent feature extraction and classification.

[0042] In some embodiments, S130 involves performing a Mel-Cepstral Transform on the filtered regional data to obtain a Mel-Cepstral Map, including: S131, multiply each frame of the spatiotemporal matrix signal by a Hamming window to obtain the windowed signal.

[0043] The formula is: , ;in, It is the initial signal. The signal is windowed, where U is the window size and n represents the number of signal frames. This formula enables the initial signal to be windowed using a Hamming window.

[0044] S132, for each short-term analysis window, the spectrum of different time window distributions on the time axis is obtained by discrete Fourier transform.

[0045] The formula for the Discrete Fourier Transform (DFT) of a signal is: ;in, This represents the signal after the Discrete Fourier Transform. NF represents the signal before the Discrete Fourier Transform (DFT), and NF represents the number of points in the Fourier Transform. The Discrete Fourier Transform (DFT) converts the time-domain signal into a frequency-domain signal.

[0046] S133 passes the spectrum through a Mel filter bank to obtain a Mel cepstral plot, which consists of multiple triangular filters.

[0047] S134 records the energy value. After transforming the signal through discrete cosine transform, the Mel-Cepstral plot is obtained.

[0048] In this embodiment, after the signal undergoes Discrete Cosine Transform (DCT), the energy is concentrated in the low-frequency range. The formula is: ;in, Let M represent the logarithmic energy of the output of the m-th triangular filter, M be the number of triangular filters, L be the order of the coefficients of the Mel filter bank, and NC represent the number of sampling points of the transform. Using this formula, the logarithmic energy of the output of the m-th triangular filter is transformed by discrete cosine transform to obtain the Mel cepstral coefficients.

[0049] In some embodiments, the specific steps for obtaining the language definitions of various vibrations in S200 are as follows: This paper describes the physical process of vibration events using written language, and re-encodes the written description of each type of vibration event into a semantic vector with discrete feature values ​​using a semantic model. The length of the semantic vector is limited by manual input. Let the semantic vector of the k-th type of vibration be denoted as... Each vector includes the properties of the vibration signal, and each property value is discrete.

[0050] Specifically, known vibration types are denoted as sunny days, rainy days, people walking, people jumping, water washing, shovels hitting the ground, and bicycles passing by. Unknown vibration samples are denoted as shovel digging. The physical processes of each vibration type are described in words, and these descriptions are input into the semantic model Deepseek-V3. The semantic model is required to generate discrete semantic vectors for each vibration type based on the descriptions, considering characteristics such as whether the material is flexible, whether the pulse duration is short, whether it is a strong vibration, whether it is a pulse-like waveform, whether there are aftershocks, whether the impact time is short, and whether it decays rapidly. The semantic vector for the k-th vibration type is denoted as... .

[0051] Based on the occurrence process of vibration categories, the known vibration types and the unknown vibration samples are defined using language. Semantic vectors are generated through a semantic model, avoiding accidental and biased errors that may occur when semantic features are manually set. Using language to describe the physical process of vibration events can transform human cognitive experience of vibration events into machine-understandable semantic representations, providing a reliable semantic foundation for subsequent similarity calculations and feature synthesis.

[0052] In some embodiments, the specific steps in S300 of generating a semantic vector for each type of vibration according to the language definition, selecting other vibration type samples similar to the unknown vibration sample based on similarity calculation, mixing the training samples of the other vibration type samples to obtain a mixed vibration sample, and calculating the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space are as follows: S310, input the linguistic descriptions of all known vibration samples and unknown vibration samples into the semantic model, and guide the semantic model to output the semantic vector of each type of vibration through prompt words. The dimension of the semantic vector is preset, and each value in the semantic vector is a discrete value.

[0053] Specifically, the semantic vector of the k-th type of vibration is manually defined. In this example, each vector includes seven attributes of the vibration signal. Each attribute value is discrete, specifically whether it is a flexible material, whether it is a short pulse duration, whether it is a strong vibration, whether it is a pulse waveform, whether there is aftershock, whether it is a short impact time, and whether it decays rapidly.

[0054] S320, calculate the Euclidean distance between the semantic vector of each type of vibration and the semantic vector of the unknown vibration sample, select the vibration with the closer distance as the hybrid sub-vibration, mix the training samples of the hybrid sub-vibration with each other to obtain the hybrid vibration sample, and combine the hybrid vibration sample with other single vibration samples to form the training set of the feature extractor.

[0055] Specifically, the Euclidean distance between the semantic vectors of each known vibration type and the unknown vibration sample is calculated. The three vibration types with the closest distance are selected as mixed sub-vibrations, namely bicycle, walking, and digging. The sub-vibrations are then combined pairwise to form mixed vibration samples. These mixed vibration samples are combined with other single vibration samples to form the training set for the unique target feature extractor. ={Bicycles and pedestrians, and the other five types of single vibrations} ={Bicycles and shovels for digging, and five other categories of single vibration} ={shovel digging and human walking, the other five types of single vibration}.

[0056] The formula for calculating similarity is as follows: , ; in, The distance between the semantic vectors of known vibration types and unknown vibration samples is represented by , where i represents the i-th attribute of the semantic vector, K represents the total number of vibration types, and k represents the index of the vibration type.

[0057] S330, sum the semantic vectors of the mixed sub-vibrations and take the average value as the semantic vector of the mixed vibration sample, and calculate the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space.

[0058] Specifically, the semantic vectors of the mixed vibration samples are calculated, along with the transformation relationship between the mixed vibration samples and the unknown vibration samples. The semantic vector of the mixed vibration sample is the average of the sums of the semantic vectors of the mixed sub-vibrations. The average semantic spatial distance between all mixed vibration samples and the unknown vibration samples is used to calculate the transformation relationship.

[0059] The calculation formulas are as follows: ; ; in, This represents the semantic vector of the j-th type of mixed vibration sample. , The semantic vector representing the vibration of the hybrid subsystem. denoted as the average semantic spatial distance between all mixed vibration samples and unknown vibrations, and v represents the number of selected mixed sub-vibrations.

[0060] The steps S310 to S330 above generate semantic vectors through a semantic model, avoiding accidental and biased errors that may occur when semantic features are manually set. By using similarity calculations to select known vibration types similar to the unknown vibration sample for mixing, it is ensured that the mixed vibration sample maintains a high similarity to the unknown vibration sample in the semantic space, thus making the subsequently synthesized feature vector closer to the feature distribution of the real unknown vibration. Simultaneously, by calculating transformation relationships in the semantic space, semantic-level similarity can be mapped to the feature level, achieving an effective transformation from the semantic space to the feature space.

[0061] refer to Figure 4 In some embodiments, the specific steps of training a general feature extractor using the training feature dataset in step S400, and training a unique target feature extractor using the mixed vibration samples and guided training method are as follows: S410, the mixed vibration sample is used as the target class, and the sample is input into the deep Q-learning network.

[0062] Specifically, AlexNet, pre-trained on the ImageNet dataset, is used as the base network for the deep Q-learning network in the guided training method.

[0063] S420, when the true label of the sample is the label of the target class, if the predicted label generated by the deep Q-learning network is equal to the true label, a positive reward is given to the deep Q-learning network; otherwise, a negative reward is given to the deep Q-learning network. When the true label of the sample is not the label of the target class, if the predicted label generated by the deep Q-learning network is not equal to the target class label, a positive reward is given to the deep Q-learning network; otherwise, a negative reward is given to the deep Q-learning network.

[0064] Specifically, the mixed vibration samples are used as the target class. These samples are input into a deep Q-learning network using a guided training method. When the sample's true label matches the target class label, a positive reward is given to the network if the predicted label generated by the deep Q-learning network equals the true label; otherwise, a negative reward is given. When the sample's true label does not match the target class label, a positive reward is given to the network if the predicted label generated by the deep Q-learning network does not equal the target class label; otherwise, a negative reward is given. The network is updated each time it receives feedback, using three mixed vibration samples. , , Three deep Q-learning networks were trained separately.

[0065] S430, after completing the training of the deep Q-learning network, the first to second-to-last layers of the deep Q-learning network are used as the sole target feature extractor.

[0066] Specifically, the first to second-to-last layers of the deep Q-learning network are taken as feature extractors, and the output of the feature extractors is a 1×10 vector, resulting in unique target feature extractor 1, unique target feature extractor 2 and unique target feature extractor 3 that can only identify a unique target class.

[0067] S410 to S430 above train a unique target feature extractor using a guided training method, enabling the network to identify only the target class while blurring other classes, thus focusing the network more intently on the target class. This results in a purer output feature vector that better preserves the feature information of the target class. Compared to ordinary training methods, guided training effectively suppresses interference from non-target class features, improves the discriminative power and purity of the feature vector, and provides a high-quality target class feature representation for subsequent feature synthesis.

[0068] In some embodiments, training a general feature extractor using the training feature dataset in step S400 specifically involves: Based on the pre-trained AlexNet, classification training is performed using the training feature dataset O and the normal training method. The first to the penultimate layers of AlexNet are taken as ordinary feature extractors, and the output of the feature extractors is a 1×10 vector.

[0069] In some embodiments, the specific steps in step S500 of inputting samples from the training feature dataset into a unique target feature extractor and a regular feature extractor, inputting mixed vibration samples into the corresponding unique target feature extractor, weighting and summing the outputs of the unique target feature extractor according to the transformation relationship in the semantic space, and then combining them with the output of the regular feature extractor to obtain the synthesized feature vector are as follows: S510, the samples in the training feature dataset are respectively input into the unique target feature extractor and the ordinary feature extractor, and the mixed vibration samples are respectively input into the unique target feature extractor with corresponding recognition capabilities; Specifically, the samples in the training feature dataset are input into each of the feature extractors to obtain four feature vectors of size 1×10. The mixed vibration samples are input into a unique target feature extractor with corresponding recognition capabilities to obtain three feature vectors of size 1×10.

[0070] S520, the transformation relationship in the semantic space is used as the output weight of the unique target feature extractor. The output of the unique target feature extractor for the training feature dataset samples is added according to the weight and concatenated with the output of the ordinary feature extractor to synthesize the synthetic feature vector of the known vibration type. Specifically, the transformation relationship in the semantic space is used as the output weight of the three unique target feature extractors. The outputs of the three unique target feature extractors are added together according to their weights, and then concatenated with the output of the ordinary feature extractor to obtain a 1×20 feature vector, which is used as a training sample for known vibration types.

[0071] S530, based on the transformation relationship in the semantic space, the output of the unique target feature extractor for the mixed vibration sample is added according to the weight, and the weight addition result is copied and spliced ​​to synthesize a synthetic feature vector of unknown vibration type; Specifically, based on the transformation relation w calculated in the semantic space, the feature vectors of the mixed vibration samples are added together according to their weights. To maintain consistent vector lengths, the output of any unique target feature extractor is copied once and concatenated to obtain a 1×20 feature vector. This result serves as the synthetic feature vector for the unknown vibration sample. The calculation formula is as follows: ;in, The mixed vibration samples are input into the output of the unique target feature extractor. The output weights of the unique target feature extractor The synthesized feature vector is the training sample for the final classifier, and N is the number of mixed vibration samples.

[0072] S540, the synthetic feature vector of the known vibration type and the synthetic feature vector of the unknown vibration type are used as the synthetic feature vector.

[0073] Specifically, the synthesized feature vectors are used as training samples for a support vector classifier to train the classifier, thus completing the training of the final classifier. The trained classifier is then used to identify the vibration type of the unknown vibration sample to obtain the identification result.

[0074] The steps S510 to S540 above, by weighting and combining the outputs of the unique target feature extractor and the ordinary feature extractor, can fully utilize the advantages of both types of feature extractors. The pure target class features provided by the unique target feature extractor complement the basic features provided by the ordinary feature extractor, making the synthesized feature vector both discriminative of the target class and retaining sufficient basic information. By guiding feature synthesis through transformation relations in the semantic space, it can be ensured that the distribution of synthesized features in the feature space is consistent with the feature distribution of real unknown vibration samples, thereby improving the classifier's accuracy in recognizing unknown vibration samples.

[0075] In this example, the fiber optic early warning system adjusts the heterodyne beat frequency using two acousto-optic modulators, allowing the heterodyne signal to be directly acquired by the acquisition card, reducing the complexity and cost of the acquisition module. A semantic model is used to generate semantic feature vectors, avoiding human-initiated random and biased errors. The Φ-OTDR zero-order intelligent recognition method based on feature synthesis utilizes a hybrid class approach and a guided training method to achieve high-quality feature synthesis of unknown vibration samples. This solves the problem of insufficient vibration samples during neural network training, enabling the fiber optic early warning system to still identify specific vibrations even without prior training.

[0076] See Figure 5 This invention provides a Φ-OTDR zero-event recognition device based on feature synthesis, comprising: The first module is used to acquire known vibration samples of known vibration types and determine unknown vibration samples of vibration types to be identified, and to extract features from the known vibration samples of known vibration types to obtain a training feature dataset. The second module is used to obtain the linguistic definitions of various vibrations, generate a semantic vector for each type of vibration based on the linguistic definitions, select other vibration type samples similar to the unknown vibration sample based on similarity calculation, mix the training samples of the other vibration type samples to obtain a mixed vibration sample, and calculate the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space. The third module is used to train a general feature extractor using the training feature dataset and to train a unique target feature extractor using the mixed vibration samples and the guided training method. The fourth module is used to input samples from the training feature dataset into a unique target feature extractor and a regular feature extractor, input mixed vibration samples into the corresponding unique target feature extractor, add the outputs of the unique target feature extractor according to the transformation relationship in the semantic space, and then combine them with the output of the regular feature extractor to obtain a synthetic feature vector. The synthetic feature vector is used as a training sample to train a classifier, and the trained classifier is used to identify the vibration type of unknown vibration samples.

[0077] It is evident that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented in this device embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0078] This invention provides a Φ-OTDR zero-event recognition system based on feature synthesis, comprising: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0079] It is evident that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0080] Furthermore, embodiments of the present invention also disclose a computer program product or computer program stored in a computer-readable storage medium. A processor of a computer device can read the computer program from the computer-readable storage medium, and the processor executes the computer program, causing the computer device to perform the described method. Similarly, the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0081] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0083] The terms "first," "second," "third," "fourth," etc. (if present) in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0084] It should be understood that in this invention, "at least one (item)" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0085] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0086] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0087] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the claims of the present invention.

Claims

1. A Φ-OTDR zero-event recognition method based on feature synthesis, characterized in that, The method includes the following steps: Acquire known vibration samples of known vibration types and determine unknown vibration samples of vibration types to be identified. Extract features from the known vibration samples of known vibration types to obtain a training feature dataset. Obtain the linguistic definitions of various vibrations, generate a semantic vector for each type of vibration based on the linguistic definitions, select other vibration type samples similar to the unknown vibration sample based on similarity calculation, mix the training samples of the other vibration type samples to obtain a mixed vibration sample, and calculate the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space. A general feature extractor is trained using the training feature dataset, and a unique target feature extractor is trained using the mixed vibration samples and the guided training method. The samples in the training feature dataset are input into a unique target feature extractor and a regular feature extractor. The mixed vibration samples are input into the corresponding unique target feature extractor. The outputs of the unique target feature extractor are weighted and summed according to the transformation relationship in the semantic space. Then, the outputs of the regular feature extractor are combined to obtain a synthetic feature vector. The synthetic feature vector is used as a training sample to train a classifier. The trained classifier is used to identify the vibration type of unknown vibration samples.

2. The method according to claim 1, characterized in that, The step of extracting features from known vibration samples of the known vibration type to obtain a training feature dataset includes: The phase signal of Rayleigh scattered light at the time of vibration is obtained by using a fiber optic early warning system, and multiple Rayleigh scattered light phase signals are rearranged into a matrix form to obtain a spatiotemporal matrix signal. The vibration in the spatiotemporal matrix signal is located using a positioning method. Regional data of the corresponding position in the spatiotemporal matrix signal is extracted, and the column direction of the regional data is high-pass filtered to remove the DC component, thus obtaining the filtered regional data. Perform Mel-Cepstral Transform on the filtered regional data to obtain the Mel-Cepstral plot; The encoder and decoder are trained using a self-supervised learning method, with the Mel cepstral as input, the reconstructed Mel cepstral as output, and the difference between the reconstructed Mel cepstral and the input Mel cepstral as the loss function. The trained encoder is used to encode the Mel cepstral, and the encoding result is recorded as a feature data sample. At the same time, the vibration type of the feature data sample is recorded as a label to obtain the training feature dataset.

3. The method according to claim 2, characterized in that, The step of performing Mel-Cepstral Transform on the filtered regional data to obtain a Mel-Cepstral Map includes: Each frame of the filtered regional data is multiplied by a Hamming window to obtain the windowed signal. For each short-term analysis window, the spectrum of different time window distributions on the time axis is obtained by discrete Fourier transform; The spectrum is passed through a Mel filter bank to obtain a Mel cepstral plot, wherein the Mel filter bank consists of multiple triangular filters; Record the energy value, and then transform the signal using a discrete cosine transform to obtain the Mel-Cepstral.

4. The method according to claim 1, characterized in that, The process of generating a semantic vector for each type of vibration based on the language definition, selecting other vibration type samples similar to the unknown vibration sample based on similarity calculation, mixing the training samples of the other vibration type samples to obtain a mixed vibration sample, and calculating the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space includes: The linguistic descriptions of all known vibration samples and unknown vibration samples are input into the semantic model. The semantic model is guided by prompt words to output a semantic vector for each type of vibration. The dimension of the semantic vector is preset, and each value in the semantic vector is a discrete value. Calculate the Euclidean distance between the semantic vector of each type of vibration and the semantic vector of the unknown vibration sample, select the vibration with the closer distance as the hybrid sub-vibration, mix the training samples of the hybrid sub-vibration with each other to obtain the hybrid vibration sample, and combine the hybrid vibration sample with other single vibration samples to form the training set of the feature extractor. The semantic vectors of the mixed sub-vibrations are summed and the average value is taken as the semantic vector of the mixed vibration sample. The transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space is calculated.

5. The method according to claim 4, characterized in that, The calculation of the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space includes: Calculate the semantic vector of the hybrid vibration sample, where the semantic vector of the hybrid vibration sample is the average of the sum of the semantic vectors of the hybrid sub-vibrations; Calculate the semantic spatial distance between the mixed vibration sample and the unknown vibration sample; Calculate the average semantic spatial distance between all the mixed vibration samples and the unknown vibration samples; Based on the semantic space distance and the average value, the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space is calculated.

6. The method according to claim 1, characterized in that, The process of training a unique target feature extractor using the hybrid vibration samples and guided training method includes: The hybrid vibration samples are used as the target class, and the samples are input into a deep Q-learning network. When the true label of the sample is the label of the target class, if the predicted label generated by the deep Q-learning network is equal to the true label, a positive reward is fed back to the deep Q-learning network; otherwise, a negative reward is fed back to the deep Q-learning network. When the true label of the sample is not the target class label, if the predicted label generated by the deep Q-learning network is not equal to the target class label, a positive reward is fed back to the deep Q-learning network; otherwise, a negative reward is fed back to the deep Q-learning network. After training the deep Q-learning network, the first to second-to-last layers of the deep Q-learning network are used as the sole target feature extractor.

7. The method according to claim 1, characterized in that, The process involves inputting samples from the training feature dataset into a unique target feature extractor and a regular feature extractor, inputting the mixed vibration samples into the corresponding unique target feature extractor, weighting the outputs of the unique target feature extractor according to the transformation relationship in the semantic space, and then combining them with the output of the regular feature extractor to obtain a synthetic feature vector, including: The samples in the training feature dataset are respectively input into the unique target feature extractor and the ordinary feature extractor, and the mixed vibration samples are respectively input into the unique target feature extractor with corresponding recognition capabilities; The transformation relationship in the semantic space is used as the output weight of the unique target feature extractor. The output of the unique target feature extractor for the training feature dataset samples is added according to the weight and concatenated with the output of the ordinary feature extractor to synthesize a synthetic feature vector of known vibration type. Based on the transformation relationship in the semantic space, the outputs of the unique target feature extractor for the mixed vibration sample are added according to the weights, and the weighted sum is copied and spliced ​​to synthesize a synthetic feature vector of unknown vibration type. The synthetic feature vectors of the known vibration type and the unknown vibration type are used as the synthetic feature vectors.

8. A Φ-OTDR zero-event recognition device based on feature synthesis, characterized in that, include: The first module is used to acquire known vibration samples of known vibration types and determine unknown vibration samples of vibration types to be identified, and to extract features from the known vibration samples of known vibration types to obtain a training feature dataset. The second module is used to obtain the linguistic definitions of various vibrations, generate a semantic vector for each type of vibration based on the linguistic definitions, select other vibration type samples similar to the unknown vibration sample based on similarity calculation, mix the training samples of the other vibration type samples to obtain a mixed vibration sample, and calculate the transformation relationship between the mixed vibration sample and the unknown vibration sample in the semantic space. The third module is used to train a general feature extractor using the training feature dataset and to train a unique target feature extractor using the mixed vibration samples and the guided training method. The fourth module is used to input samples from the training feature dataset into a unique target feature extractor and a regular feature extractor, input mixed vibration samples into the corresponding unique target feature extractor, add the outputs of the unique target feature extractor according to the transformation relationship in the semantic space, and then combine them with the output of the regular feature extractor to obtain a synthetic feature vector. The synthetic feature vector is used as a training sample to train a classifier, and the trained classifier is used to identify the vibration type of unknown vibration samples.

9. A Φ-OTDR zero-event recognition system based on feature synthesis, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the method as described in any one of claims 1 to 7.

Images