Method and system for preventing external force damage event prediction based on optical fiber acoustic wave sensing

Abstract

The application relates to the field of optical cable operation and maintenance, in particular to a method and system for preventing external damage event prediction based on optical fiber acoustic wave sensing, which comprises the following steps: collecting optical fiber audio data and converting the same into a to-be-processed frequency spectrum feature map, dividing the to-be-processed frequency spectrum feature map into a plurality of frequency band feature maps according to a preset frequency range, inputting the frequency band feature maps into a corresponding external damage event identification model to obtain an external damage event probability distribution of each frequency range; vectorizing the to-be-processed frequency spectrum feature map to obtain a frequency spectrum feature vector, and inputting the frequency spectrum feature vector into a gating network to obtain a feature weight of each frequency range; weighting and summing the external damage event probability distribution and the corresponding feature weight to obtain a global external damage event probability distribution, and obtaining a target external damage event type according to the global external damage event probability distribution. Compared with the prior art, the application can dynamically adjust the contribution degree of the external damage event probability distribution of each frequency band through fine identification and adaptive weighted fusion of acoustic wave features of different frequency components, and the accuracy of external damage event prediction based on optical fiber acoustic waves is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of optical cable maintenance, and more specifically, to a method and system for predicting external force damage events based on optical fiber acoustic wave sensing. Background Technology

[0002] With the popularization of fiber optic technology, fiber optic cables are increasingly used in people's daily lives, and the number of fiber optic cables laid is constantly increasing. However, this increase in the number of fiber optic cables also makes them more susceptible to damage from external factors such as construction or urban redevelopment, thus affecting the user experience. Therefore, it is necessary to predict external damage events to fiber optic cables in order to mitigate their impact.

[0003] Current methods for predicting external fiber optic cable damage events involve collecting acoustic signals from the cable and analyzing these signals as a whole. However, the acoustic signals generated by external damage events often contain a rich and wide-ranging set of frequency components. Different frequency bands of acoustic signals exhibit different vibrational characteristics and physical meanings. Analyzing only the overall frequency band of the acoustic signal makes it difficult to simultaneously address the identification needs of different frequency signals, thus limiting the accuracy of the identification. Summary of the Invention

[0004] This invention provides a method and system for predicting external force damage events based on fiber optic acoustic wave sensing, which is used to improve the accuracy of external damage event prediction based on fiber optic acoustic waves.

[0005] According to a first aspect of this application, a method for predicting external force damage events based on fiber optic acoustic wave sensing is provided, the method comprising: Acquire audio data from optical fibers and convert the audio data into a spectral feature map to be processed; The spectrum feature map to be processed is divided into blocks according to a preset frequency range to obtain several frequency band feature maps; The frequency band feature map is input into the external damage event identification model of the corresponding frequency range for processing to obtain the probability distribution of external damage events in the frequency range. The spectral feature map to be processed is vectorized to obtain the spectral feature vector. The spectral feature vector is input into a pre-trained gating network to obtain the feature weights for each frequency range. The global external failure event probability distribution is obtained by weighting and summing the feature weights of the external failure event probability distribution with respect to the corresponding frequency range. Based on the global external damage event probability distribution, the target external damage event type is obtained.

[0006] By converting fiber optic audio data into a spectral feature map to be processed and dividing it into blocks according to frequency range, and combining it with a pre-trained gating network to obtain the feature weights of each frequency band, the fine-grained identification and adaptive weighted fusion of acoustic wave features of different frequency components are realized. In this way, the contribution of each frequency band to the probability distribution of external damage events can be dynamically adjusted according to the frequency components with a large span in the acoustic signal and their physical meaning. This allows for more accurate extraction of effective features in complex noise environments, significantly improving the accuracy of external damage event prediction based on fiber optic acoustic waves.

[0007] Optionally, the step of performing feature vectorization on the spectral feature map to be processed to obtain a spectral feature vector includes: The spectral feature map to be processed is divided according to a preset grid to obtain several sub-spectral feature maps; The sub-spectral feature map is subjected to self-attention processing to obtain a sub-spectral self-attention feature map; The sub-spectral self-attention feature map is fused with its neighboring sub-spectral self-attention feature maps to obtain a fused feature map. The fused feature map is compressed using global average pooling to obtain a sub-spectral feature vector. The sub-spectral feature vectors are concatenated to obtain the concatenated spectral feature vector. The spectral feature vector is obtained by processing the spliced ​​spectral feature vector through nonlinear mapping.

[0008] By dividing the spectral feature map to be processed into a grid and combining it with a self-attention mechanism and feature fusion processing, the local features within the sub-spectral feature map and their correlation information with adjacent sub-spectrals can be effectively captured, enhancing the ability to represent the subtle features of complex acoustic signals. Furthermore, by performing global average pooling compression on the fused feature map and subsequent feature splicing and nonlinear mapping processing, the high-dimensional spectral spatial information is transformed into a compact spectral feature vector. This not only significantly reduces data redundancy but also retains key quality features that are sensitive to environmental noise, interference, and fiber quality. This provides a high-quality input vector for the subsequent gated network to accurately calculate the feature weights of each frequency range, thereby improving the accuracy of external damage event prediction.

[0009] Optionally, the step of fusing the sub-spectral self-attention feature map with its neighboring sub-spectral self-attention feature maps to obtain a fused feature map includes: For each of the sub-spectral self-attention feature maps, obtain the sparse connection topology between the sub-spectral self-attention feature map and its neighboring sub-spectral self-attention feature maps; Based on the multi-head attention mechanism and the sparse connection topology, the sub-spectral self-attention feature map is fused with its neighboring sub-spectral self-attention feature maps to obtain a fused feature map.

[0010] By introducing sparse connection topology and multi-head attention mechanism for feature fusion, it is possible to further model long-distance dependencies and nonlinear correlation patterns between adjacent sub-spectrums while retaining the detailed features captured by local window self-attention processing. This enhances the model's ability to perceive the coordinated changes in cross-band acoustic signals, making the fused feature map not only contain local fine structures but also incorporate cross-regional contextual information. This improves the completeness and discriminativeness of spectral feature representation, laying a high-quality feature foundation for the accurate calculation of the probability distribution of subsequent external damage events.

[0011] Optionally, converting the audio data into a spectral feature map to be processed includes: The audio data is decoded into time-domain data; The time-domain data is subjected to a preset bandpass filter to obtain the processed time-domain data; The processed time-domain data is resampled to a preset sampling rate to obtain the time-domain data at the preset sampling rate; The time-domain data at a preset sampling rate is converted into a Mel spectrum, and features are extracted from the Mel spectrum to obtain the spectrum feature map to be processed.

[0012] By performing a series of preprocessing operations on the audio data, such as decoding, bandpass filtering, resampling, and Mel spectrum conversion, the interference of environmental noise and invalid frequency bands is effectively removed, ensuring the purity and consistency of the time domain data. Furthermore, by extracting features from the Mel spectrum to obtain the spectrum feature map to be processed, the spectral distribution characteristics of the sound wave signal can be more closely matched with the characteristics of human auditory perception, enhancing the distinguishability of acoustic fingerprints of different types of external damage events, and providing a high-quality and highly representative input data foundation for subsequent frequency band segmentation and accurate identification.

[0013] Optionally, the external damage event recognition model includes a depthwise separable convolutional module, a frequency domain attention module, a residual connection module, and a classification output module connected in sequence; The depthwise separable convolution module is used to perform convolutional feature extraction processing on the input frequency band feature map to obtain an initial feature map; The frequency domain attention module is used to perform channel weight calibration processing on the initial feature map using a global attention mechanism to obtain an attention-weighted feature map; The residual connection module is used to perform an identity mapping on the attention-weighted feature map, and to perform feature superposition processing on the mapped attention-weighted feature map and the frequency band feature map to obtain an enhanced feature map; The classification output module is used to perform category probability mapping processing on the enhanced feature map to obtain the probability distribution of the external damage event.

[0014] The described external damage event recognition model, through its depthwise separable convolutional module, can efficiently extract convolutional features from the input frequency band feature map, significantly reducing computational complexity while preserving key local features. Simultaneously, the global attention mechanism of the frequency domain attention module calibrates the channel weights of the initial feature map, adaptively strengthening frequency domain feature channels that contribute highly to event recognition and suppressing redundant information. Furthermore, by combining the residual connection module with identity mapping and feature superposition of the attention-weighted feature map, the model effectively alleviates the gradient vanishing problem in deep networks, enhancing feature propagation and reuse.

[0015] Optionally, the external damage event recognition model is obtained through pre-training; The pre-training of the external failure event identification model includes: Collect historical audio data from optical fibers and convert the historical audio data into historical spectral feature maps; Obtain the external damage event type and external damage scene environment information corresponding to the historical audio data; The historical spectrum feature map is divided into blocks according to a preset frequency range to obtain several frequency band historical feature maps; The historical feature map of the frequency band in the frequency range is input into the external damage event recognition model to be trained in the corresponding frequency range for processing, so as to obtain the historical external damage event probability distribution and predicted external damage site environment information in the frequency range. For each frequency range, based on the external damage event type and the historical external damage event probability distribution, the first training loss of the external damage event recognition model to be trained is calculated; based on the external damage site environment information and the predicted external damage site environment information, the second training loss of the external damage event recognition model to be trained is calculated; and based on the first training loss and the second training loss, the parameters of the external damage event recognition model to be trained are updated to obtain the external damage event recognition model trained for the frequency range.

[0016] By introducing external damage scene environment information as an auxiliary supervision signal during the training process, the external damage event recognition model not only learns the mapping relationship between frequency band historical feature maps and event types, but also simultaneously perceives the influence of different scene environments on sound wave propagation characteristics. This enhances the robustness of the external damage event recognition model to complex environmental interference, enabling it to adaptively adjust the recognition strategy according to environmental characteristics. As a result, it can more accurately distinguish the acoustic characteristics of external damage events under different environments in practical applications, improving the model's generalization ability and recognition accuracy under diverse scene conditions.

[0017] Optionally, obtaining the target external failure event type based on the global external failure event probability distribution includes: The external failure event type with the highest probability value is obtained from the global external failure event probability distribution. If the highest probability value exceeds a preset threshold, the external failure event type with the highest probability value is taken as the target external failure event type.

[0018] By selecting the external failure event with the highest probability value from the global external failure event probability distribution and combining it with a preset threshold for validity determination, the most likely type of external failure event can be accurately identified. At the same time, false alarms with low confidence can be effectively filtered out, thereby improving the accuracy and reliability of the determination of the target external failure event.

[0019] According to a second aspect of this application, a system for predicting external force damage events based on fiber optic acoustic wave sensing is provided, the system comprising: The data acquisition module is used to acquire audio data from the optical fiber and convert the audio data into a spectral feature map to be processed. The data segmentation module is used to segment the spectrum feature map to be processed according to a preset frequency range to obtain several frequency band feature maps; The model processing module is used to input the frequency band feature map into the external damage event identification model of the corresponding frequency range for processing, and obtain the probability distribution of external damage events in the frequency range. The vectorization module is used to perform feature vectorization processing on the spectrum feature map to be processed to obtain a spectrum feature vector. The weight calculation module is used to input the spectral feature vector into the pre-trained gating network to obtain the feature weights of each frequency range. The event probability calculation module is used to perform a weighted summation of the external failure event probability distribution and the feature weights of its corresponding frequency range to obtain the global external failure event probability distribution; The event determination module is used to obtain the target external damage event type based on the global external damage event probability distribution.

[0020] According to a third aspect of this application, an electronic device is provided, comprising: Memory, used to store one or more computer programs; The processor, when the one or more computer programs are executed by the processor, implements the method for predicting external force damage events based on fiber optic acoustic wave sensing as described in the first aspect above.

[0021] According to a fourth aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions, the computer instructions being configured to cause a processor to execute and implement the method for predicting external force damage events based on fiber optic acoustic wave sensing as described in the first aspect.

[0022] Based on any of the above aspects, the fiber optic acoustic wave sensing-based method, system, electronic device, and computer storage medium for predicting external force damage events provided in this application convert fiber optic audio data into a spectrum feature map to be processed and process it in blocks according to frequency range. By combining a pre-trained gating network to obtain the feature weights of each frequency band, it achieves refined identification and adaptive weighted fusion of acoustic wave features of different frequency components. Furthermore, it can dynamically adjust the contribution of each frequency band to the probability distribution of external damage events based on the frequency components with a large span in the acoustic wave signal and their physical meaning. This allows for more accurate extraction of effective features in complex noise environments and significantly improves the accuracy of external damage event prediction based on fiber optic acoustic waves. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the steps of the prediction method provided in this embodiment.

[0025] Figure 2 This is a schematic diagram of the steps for obtaining the spectral feature map to be processed provided in this embodiment.

[0026] Figure 3 This is a schematic diagram of the model structure of the external failure event identification model provided in this embodiment.

[0027] Figure 4 This is a schematic diagram illustrating the steps of pre-training the external damage event recognition model provided in this embodiment.

[0028] Figure 5 This is a schematic diagram illustrating the steps for obtaining the spectral feature vector provided in this embodiment.

[0029] Figure 6This is a schematic diagram of the functional modules of the prediction system provided in this embodiment.

[0030] Figure 7 This is a schematic diagram of the device structure of the electronic device provided in this embodiment. Detailed Implementation

[0031] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this application. To better illustrate the following embodiments, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product; it is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0032] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application 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 the embodiments of this application 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 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.

[0034] With the popularization of fiber optic technology, fiber optic cables are increasingly used in people's daily lives, and the number of fiber optic cables laid is constantly increasing. However, this increase in the number of fiber optic cables also makes them more susceptible to damage from external factors such as construction or urban redevelopment, thus affecting the user experience. Therefore, it is necessary to predict external damage events to fiber optic cables in order to mitigate their impact.

[0035] Current methods for predicting external fiber optic cable damage events involve collecting acoustic signals from the cable and analyzing these signals as a whole. However, the acoustic signals generated by external damage events often contain a rich and wide-ranging set of frequency components. Different frequency bands of acoustic signals exhibit different vibrational characteristics and physical meanings. Analyzing only the overall frequency band of the acoustic signal makes it difficult to simultaneously address the identification needs of different frequency signals, thus limiting the accuracy of the identification.

[0036] This embodiment provides a technical solution that can solve the above problems. The specific implementation of this application will be described in detail below with reference to the accompanying drawings.

[0037] like Figure 1 As shown in the figure, this embodiment provides a method for predicting external force damage events based on fiber optic acoustic wave sensing, which may include the following steps: S1: Acquire audio data from the optical fiber and convert the audio data into a spectral feature map to be processed; In this embodiment, audio data from optical fibers can be acquired using DAS (Distributed Acoustic Sensing) technology. DAS technology can serve as a highly sensitive optical fiber sensing and detection tool, effectively analyzing acoustic signals in optical fibers and reconstructing them into audio data.

[0038] In this embodiment, the audio data may include, but is not limited to, audio files in formats such as MP3 (MPEG-1 Audio Layer III), AAC (Advanced Audio Coding), and WAV (Waveform Audio File Format). The format of the audio files in the audio data can be set according to requirements.

[0039] In one implementation, such as Figure 2 As shown, converting the audio data into a spectral feature map to be processed may include the following sub-steps: S11: Decode the audio data into time-domain data; In this embodiment, the time-domain data can be understood as the changes in sound wave intensity around the optical fiber at different points in time. Therefore, decoding the audio data into the time-domain data allows the sound to be restored to its most original physical form, facilitating subsequent processing steps.

[0040] S12: Perform a preset bandpass filter on the time-domain data to obtain the processed time-domain data; In this embodiment, the time-domain data can be bandpass filtered using a bandpass filter with a preset bandpass frequency to eliminate environmental noise. Preferably, the preset bandpass frequency can be set to 10Hz-20kHz.

[0041] S13: Resample the processed time-domain data to a preset sampling rate to obtain the time-domain data at the preset sampling rate; In this embodiment, by resampling the time-domain data to a uniform preset sampling rate, data from different frequencies can be unified, ensuring the consistency of the time-domain data and facilitating subsequent identical processing. Preferably, the preset sampling rate can be set to 14kHz~18kHz.

[0042] S14: Convert the time-domain data at the preset sampling rate into a Mel spectrum, and extract features from the Mel spectrum to obtain the spectrum feature map to be processed.

[0043] Understandably, the Mel spectrum can effectively simulate the auditory perception characteristics of the human ear. Therefore, by converting the time-domain data into a Mel spectrum and then extracting features from the Mel spectrum to obtain a spectral feature map to be processed, the obtained spectral feature map can more accurately represent the spectral distribution characteristics of the sound wave signal in accordance with the auditory perception characteristics of the human ear.

[0044] S2: Divide the spectrum feature map to be processed into blocks according to a preset frequency range to obtain several frequency band feature maps; In this embodiment, the spectrum feature map to be processed can be divided into several frequency band feature maps according to a preset frequency range, wherein each frequency band feature map corresponds to a frequency range.

[0045] In one implementation, the preset frequency range may include 0-5kHz, 5-15kHz, and 15-20kHz. Here, 0-5kHz represents the low-frequency range, 5-15kHz represents the mid-frequency range, and 15-20kHz represents the high-frequency range. Therefore, the spectral feature map to be processed is divided into frequency band feature maps of low-frequency, mid-frequency, and high-frequency ranges. These band feature maps each contain acoustic information with different physical properties, enabling differentiated processing based on the energy distribution characteristics of different frequency bands. Specifically, the low-frequency range typically concentrates the main impact energy and structural vibration characteristics of external damage events; the mid-frequency range carries the harmonic structure and transition characteristics of the event; and the high-frequency range reflects more of the environmental background noise, transient details generated by mechanical friction, and edge information of the signal. Through this frequency band decoupling method, effective information can be extracted from different frequency ranges to identify external damage events, thereby effectively improving the accuracy of identification.

[0046] S3: Input the frequency band feature map into the external damage event identification model of the corresponding frequency range for processing to obtain the probability distribution of external damage events in the frequency range; In this embodiment, an external damage event recognition model is pre-trained for each frequency range. The external damage event recognition model can identify the frequency band feature image of its corresponding frequency range and predict the probability of various external damage event types occurring in the current optical fiber.

[0047] In this embodiment, the external failure event identification model can be constructed based on existing expert models, such as... Figure 3 As shown, the external damage event recognition model may include a depthwise separable convolutional module 1, a frequency domain attention module 2, a residual connection module 3, and a classification output module 4 connected in sequence. The depthwise separable convolution module 1 is used to perform convolutional feature extraction processing on the input frequency band feature map to obtain an initial feature map. Specifically, the depthwise separable convolution module 1 is used to perform two-dimensional convolution on each frequency channel of the frequency band feature map within its frequency range, so as to extract the local texture changes of the frequency band feature map using each frequency channel, obtain the convolutional features of each frequency channel, and then perform pointwise convolution between the convolutional features of each frequency channel to obtain the initial feature map.

[0048] The frequency domain attention module 2 is used to perform channel weight calibration processing on the initial feature map using a global attention mechanism to obtain an attention-weighted feature map. Specifically, the frequency domain attention module 2 is used to acquire the feature channels of the initial feature map, and perform global channel attention processing based on the feature channels to obtain the channel attention weights of each feature channel of the initial feature map. Then, the initial feature map is divided into several sub-band feature maps according to a preset sub-band frequency range, and the sub-band energy corresponding to each sub-band feature map is calculated. The frequency weight of each sub-band feature map is calculated based on the sub-band energy of each sub-band feature map. The attention calibration weight of the initial feature map is calculated based on the frequency weight and the channel attention weight. The initial feature map is then weighted and calibrated based on the attention calibration weight to obtain the attention-weighted feature map.

[0049] The residual connection module 3 is used to perform an identity mapping on the attention-weighted feature map, and then perform feature superposition processing on the mapped attention-weighted feature map and the frequency band feature map to obtain an enhanced feature map. Specifically, the residual connection module 3 is used to perform an identity mapping on the attention-weighted feature map, and then add the mapped attention-weighted feature map and the frequency band feature map element by element to obtain the enhanced feature map.

[0050] The classification output module 4 is used to perform category probability mapping processing on the enhanced feature map to obtain the probability distribution of the external damage event. Specifically, the classification output module 4 is used to calculate the correlation between the enhanced feature map and a preset external damage event type to obtain the probability distribution of the external damage event.

[0051] Understandably, in this embodiment, by configuring a separate external failure event identification model for each frequency range, the external failure event identification model can perform refined modeling and focused identification of the acoustic vibration signals within its corresponding frequency range. Specifically, since the acoustic signals generated by external failure events often contain rich and wide-ranging frequency components, and acoustic signals in different frequency bands contain different vibration characteristics and physical meanings, an independent external failure event identification model is configured for each preset frequency range (such as the low-frequency, mid-frequency, and high-frequency ranges mentioned above). Each model only needs to focus on signal feature extraction and pattern discrimination within its assigned frequency band, avoiding cross-frequency band feature aliasing and strong energy masking, thereby significantly improving the detection sensitivity of weak external failure features within each frequency band. Ultimately, this effectively improves the accuracy of external failure event prediction and reduces the false alarm and missed alarm rates.

[0052] In one implementation, such as Figure 4 As shown, the pre-training of the external damage event recognition model may include the following steps: A1: Collect historical audio data from optical fibers and convert the historical audio data into a historical spectral feature map; It is understood that the acquisition of historical audio data and the conversion of historical spectral feature maps can be implemented with reference to the acquisition of audio data and the conversion of the spectral feature maps to be processed described in step S1 above, and will not be elaborated further here.

[0053] A2: Obtain the external damage event type and external damage scene environment information corresponding to the historical audio data; In this embodiment, in addition to acquiring the external damage event type from the historical audio data, external damage site environmental information is also acquired. This external damage site environmental information refers to the environment in which the historical external damage event occurred; understandably, it may include sandy terrain, hard soil, and rainy weather. By acquiring this external damage site environmental information, it can be used as an environmental label during the training of the external damage event recognition model, supervising the model's training. This allows the model to learn the differentiated acoustic feature distribution of the same external damage event type under different environmental conditions. Consequently, the model can perceive and adapt to the sound wave propagation characteristics and background noise patterns in different environments, thereby enhancing its generalization ability and robustness in complex and changing field environments.

[0054] A3: Divide the historical spectrum feature map into blocks according to a preset frequency range to obtain several frequency band historical feature maps; A4: Input the historical feature map of the frequency band in the frequency range into the external damage event recognition model to be trained in the corresponding frequency range for processing, and obtain the historical external damage event probability distribution and predicted external damage site environment information of the frequency range. In one embodiment, the classification output module 4 of the external damage event identification model can be configured with at least two output task heads. Each output task head consists of one or more fully connected layers. Each output task head enables the classification output module 4 to output a type of data. Therefore, the historical probability distribution of external damage events in the frequency range can be output through one output task head of the classification output module 4, and the predicted external damage site environment information can be output through the other output task head.

[0055] A5: For each frequency range, calculate the first training loss of the external damage event recognition model to be trained based on the external damage event type and the historical external damage event probability distribution; calculate the second training loss of the external damage event recognition model to be trained based on the external damage site environment information and the predicted external damage site environment information; and update the parameters of the external damage event recognition model to be trained based on the first training loss and the second training loss to obtain the external damage event recognition model trained for the frequency range.

[0056] In one implementation, the first training loss can be calculated using the Focal Loss function, the formula of which can be expressed as: In the formula, This represents the Focal Loss function. This represents the probability distribution of the historical external damage events. This indicates the type of external disruption event corresponding to the historical audio data. This represents the probability value of the external failure event type in the probability distribution of historical external failure events. This represents a learnable category balance factor corresponding to the aforementioned external failure event type. This represents the learnable focusing parameters.

[0057] It's understandable, when When the value approaches 1, As the value approaches zero, the loss is significantly suppressed; while When the value approaches 0, As the value approaches 1, the loss is preserved or even amplified, forcing the external damage event recognition model to focus on the historical feature map of the input frequency band at this time. Ultimately, this makes the external damage event recognition model pay more attention to the few or difficult-to-classify input samples, thereby significantly improving the classification performance of imbalanced data.

[0058] Understandably, after the external failure event identification model is completed, the external failure event identification model does not explicitly output the predicted external failure site environment information during the actual inference process. This predicted external failure site environment information will be integrated into the prediction of the external failure event probability distribution. This allows the external failure event identification model to effectively incorporate the perception of the site environment during the prediction of the external failure event probability distribution, thereby improving the prediction accuracy in different scenarios.

[0059] S4: Perform feature vectorization on the spectrum feature map to be processed to obtain the spectrum feature vector; In this embodiment, as Figure 5 As shown, step S4 may include the following sub-steps: S41: Divide the spectrum feature map to be processed according to a preset grid to obtain several sub-spectrum feature maps; S42: Perform self-attention processing on the sub-spectral feature map to obtain a sub-spectral self-attention feature map; S43: The sub-spectral self-attention feature map is fused with its neighboring sub-spectral self-attention feature maps to obtain a fused feature map; S44: Perform global average pooling compression on the fused feature map to obtain the sub-spectral feature vector; S45: Perform feature concatenation on the sub-spectral feature vectors to obtain a spectrum concatenation feature vector; S46: The spectrum splicing feature vector is processed by nonlinear mapping to obtain the spectrum feature vector.

[0060] In this embodiment, the spectral feature map to be processed is processed through a preset grid, which divides the spectral feature map to be processed into M×N or N×N sub-spectral feature maps. The sub-spectral feature maps obtained by the division are equivalent to containing local features of the spectral feature map to be processed. Therefore, performing self-attention processing on the sub-spectral feature maps is equivalent to using the sub-spectral feature maps as local windows to perform local window self-attention processing on the spectral feature map to be processed. By performing self-attention processing on the sub-spectral feature maps, the local features in the spectral feature map to be processed can be captured very well.

[0061] However, relying solely on local features in the spectral feature map to be processed is insufficient to capture the long-range dependencies and overall distribution patterns of external damage events in the frequency domain. It is prone to misjudgment due to the similarity between local textures and complex environmental noise, and fails to perceive the macroscopic spatiotemporal context of the event occurrence. This results in a lack of global semantic support for the identification of the external damage event type, thereby affecting the accuracy and robustness of the historical external damage event probability distribution output by the model. Therefore, in this embodiment, the sub-spectral self-attention feature map is fused with its adjacent sub-spectral self-attention feature maps to capture the dependencies between local features.

[0062] Furthermore, in this embodiment, step S43 may include the following steps: For each of the sub-spectral self-attention feature maps, the sparse connection topology between the sub-spectral self-attention feature map and its neighboring sub-spectral self-attention feature maps is obtained; based on the multi-head attention mechanism and the sparse connection topology, the sub-spectral self-attention feature map and its neighboring sub-spectral self-attention feature map are fused to obtain a fused feature map.

[0063] In this embodiment, the sparse connection topology refers to the feature connection relationship between specific sub-spectral self-attention feature maps established based on a preset connection relationship. Specifically, this topology defines selective connections between sub-spectral self-attention feature maps: that is, a specific subset of feature points in one sub-spectral self-attention feature map is only allowed to establish computational connections with a specific associated subset of feature points in another sub-spectral self-attention feature map, while ignoring other non-associated feature points. It can be understood that the sparse connection topology has a dual constraint: it macroscopically expresses the connectivity structure between sub-spectral self-attention feature maps and microscopically constrains the interaction range between feature points. In this embodiment, each sub-spectral self-attention feature map is connected to its adjacent sub-spectral self-attention feature maps, and then the interaction range between the feature points of each sub-spectral self-attention feature map and its adjacent sub-spectral self-attention feature maps is confirmed, thereby obtaining the sparse connection topology.

[0064] The multi-head attention mechanism is configured to perform cross-regional feature interactions between associated sub-spectral self-attention feature maps based on specific connection paths defined by the sparse connection topology. Specifically, the multi-head attention mechanism utilizes a multi-head attention algorithm, using feature points in the current sub-spectral self-attention feature map as query vectors, and only calculates attention weights and aggregates features in the associated sub-spectral self-attention feature maps specified in the topology as key vectors or value vectors, thereby achieving precise injection and fusion of specific global context information into the current feature map.

[0065] Understandably, in this embodiment, by utilizing the multi-head attention mechanism and the sparse connection topology, combined with the local window self-attention mechanism, the complexity of traditional fusion based on global relationships is reduced to the complexity based on the sparse connection topology, thereby greatly reducing the computational load required for the external damage event recognition model inference, while also effectively capturing the long-range dependencies between sub-spectral self-attention feature maps.

[0066] S5: Input the spectral feature vector into the pre-trained gating network to obtain the feature weights of each frequency range; In this embodiment, the gated network is essentially a frequency importance evaluator. This network includes at least one lightweight feature processing sub-network, such as a Multilayer Perceptron (MLP). The gated network can identify target frequency ranges containing key information about external damage events based on the input spectral feature vector, and distinguish non-target frequency ranges containing significant background noise; thereby quantifying the importance of each frequency range to generate feature weights corresponding to each frequency range.

[0067] In one implementation, the loss function of the gated network can be constructed based on KL divergence, and is expressed as: In the formula, This represents the KL divergence formula. Let be the output of the gating network, representing the probability distribution of each gating option z in the gating network given the spectral feature vector x of the input. Let z represent the prior distribution of the gating option. The weighting coefficients representing the KL divergence. This represents the loss between the gated network and the real labels. This represents the total loss of the gated network.

[0068] Understandably, the loss function of the gated network constructed using KL divergence can sparsify the feature weights of each frequency range output by the gated network, thus avoiding the feature weights of each frequency range being too average.

[0069] In one implementation, the gated network includes two softmax layers. One softmax layer outputs the feature weights for each frequency range, and the other softmax layer outputs the environmental features predicted based on the spectral feature vectors. During training of the gated network, in addition to calculating the loss of the feature weights for each frequency range, the loss of the environmental features is also calculated. It is understood that by outputting the environmental features and calculating the loss of the environmental features during training of the gated network, environmental information is incorporated into the calculation of the feature weights for each frequency range. This allows the probability distribution of external damage events output by each external damage event identification model to be adjusted based on environmental information, further improving the accuracy of external damage event type prediction.

[0070] S6: The probability distribution of external failure events is weighted and summed with the feature weights of its corresponding frequency range to obtain the global probability distribution of external failure events; In this embodiment, the probability distribution of external damage events includes the predicted probability of each type of external damage event identified based on the frequency band feature map of the frequency range. The feature weight represents the confidence or importance of the frequency range. Therefore, the probability distribution of external damage events can be multiplied by its corresponding frequency range. Specifically, the predicted probability of each type of external damage event in the probability distribution of external damage events is multiplied by the feature weight to obtain the local probability distribution of external damage events for each frequency range. Then, the predicted probabilities of the corresponding external damage events in each local probability distribution of external damage events are added together to obtain the global probability distribution of external damage events.

[0071] S7: Obtain the target external damage event type based on the global external damage event probability distribution.

[0072] In this example, step S7 may include: The external failure event type with the highest probability value is obtained from the global external failure event probability distribution. If the highest probability value exceeds a preset threshold, the external failure event type with the highest probability value is taken as the target external failure event type.

[0073] Understandably, in order to ensure the accuracy of the identified external damage event type, in this embodiment, after obtaining the external damage event type with the highest probability value in the global external damage event probability distribution, the probability value of the obtained external damage event type is compared with a preset threshold. If the probability value exceeds the preset threshold, it indicates that the identification result of the external damage event type has a high degree of confidence, and the external damage event type is taken as the target external damage event type. Understandably, if the maximum probability value does not exceed the preset threshold, it means that the identification result of the external damage event type has a low confidence level. In this case, it can be understood that the audio data does not contain the preset external damage event type and needs to be re-predicted.

[0074] like Figure 6 As shown in the illustration, this application also provides a system for predicting external force damage events based on fiber optic acoustic wave sensing. Optionally, the prediction system may include: The data acquisition module 11 is used to acquire audio data from the optical fiber and convert the audio data into a spectrum feature map to be processed. In this embodiment, the data acquisition module 11 can be used to perform... Figure 1 For a detailed description of the data acquisition module 11 shown in step S1, please refer to the description of step S1.

[0075] Data segmentation module 12 is used to segment the spectrum feature map to be processed according to a preset frequency range to obtain several frequency band feature maps; In this embodiment, the data segmentation module 12 can be used to perform... Figure 1 For a detailed description of the data segmentation module 12 shown in step S2, please refer to the description of step S2.

[0076] Model processing module 13 is used to input the frequency band feature map into the external damage event identification model of the corresponding frequency range for processing, and obtain the probability distribution of external damage events in the frequency range. In this embodiment, the model processing module 13 can be used to execute... Figure 1 For a detailed description of the model processing module 13 shown in step S3, please refer to the description of step S3.

[0077] Vectorization module 14 is used to perform feature vectorization processing on the spectrum feature map to be processed to obtain spectrum feature vector; In this embodiment, the vectorization module 14 can be used to execute Figure 1 For a detailed description of step S4, please refer to the description of step S4.

[0078] The weight calculation module 15 is used to input the spectral feature vector into the pre-trained gating network to obtain the feature weights of each frequency range. In this embodiment, the weight calculation module 15 can be used to perform... Figure 1 For a detailed description of the weight calculation module 15 shown in step S5, please refer to the description of step S5.

[0079] Event probability calculation module 16 is used to perform a weighted summation of the external failure event probability distribution and the feature weights of its corresponding frequency range to obtain a global external failure event probability distribution; In this embodiment, the event probability calculation module 16 can be used to perform... Figure 1 For a detailed description of the event probability calculation module 16 shown in step S6, please refer to the description of step S6.

[0080] Event determination module 17 is used to obtain the target external damage event type based on the global external damage event probability distribution; In this embodiment, the event determination module 17 can be used to perform... Figure 1 For a detailed description of the event determination module 17 shown in step S7, please refer to the description of step S7.

[0081] This application provides an electronic device with the following structure: Figure 7 As shown.

[0082] The electronic device includes a memory 21, a processor 22, a communication module 23, and an input / output interface 24, etc. Optionally, the memory 21, the processor 22, the communication module 23, and the input / output interface 24 can be connected and communicate with each other through a bus 25.

[0083] The memory 21 is used to store one or more computer programs and transmit the code of the computer programs to the processor 22; when the one or more computer programs are executed by the processor 22, the method for predicting external force damage events based on fiber optic acoustic wave sensing in this application embodiment is implemented.

[0084] Optionally, the electronic device can be connected to a network via communication module 23 to communicate with other devices, such as terminals or servers, to achieve data interaction. The electronic device can be various forms of digital computers, exemplarily such as desktop computers, servers, workbenches, mainframes, or other types of computers. The electronic device can also be various forms of mobile terminals, exemplarily such as smartphones, tablets, wearable devices (such as helmets, glasses, watches, etc.), and other similar mobile terminals.

[0085] Optionally, the electronic device can connect to required input / output devices, such as a keyboard or display device, via the input / output interface 24. The electronic device itself may have a display device, and other display devices can also be connected externally via the input / output interface 24. Optionally, a storage device, such as a hard disk, can also be connected via the input / output interface 24 to store data from the electronic device, read data from the storage device, or store data from the storage device in the memory 21. It is understood that the input / output interface 24 can be a wired interface or a wireless interface. Depending on the actual application scenario, the device connected to the input / output interface 24 can be a component of the electronic device or an external device connected to the electronic device when needed.

[0086] Optionally, the memory 21 may be a volatile memory and / or a non-volatile memory. The volatile memory may be a random access memory, etc., and the non-volatile memory may be a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, or a flash memory, etc.

[0087] Optionally, the computer program stored in the processor 22 can be divided into one or more modules, which are stored in the memory 21 and executed by the processor 22 to perform the method provided in this embodiment. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the electronic device.

[0088] Optionally, the processor 22 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the processor 22 include, but are not limited to, a central processing unit, a graphics processing unit, a digital signal processor, various special-purpose artificial intelligence computing chips, various processors running machine learning model algorithms, and can also be any suitable controller, microcontroller, processor, etc. The processor 22 executes the various methods and processes of this embodiment, exemplarily, such as a method for predicting external force damage events based on fiber optic acoustic wave sensing according to an embodiment of this application.

[0089] Optionally, the bus 25 may include a path for transmitting information. Depending on its function, the bus 25 may be divided into an address bus, a data bus, a control bus, etc.

[0090] In an optional implementation, this application embodiment also provides a computer storage medium storing a computer program thereon, which, when executed by a computer, enables the computer to perform the methods described in the above-described method embodiments. Part or all of the computer program can be loaded and / or installed on the memory 21 of an electronic device. When the computer program is executed by the processor 22, one or more steps of a fiber optic acoustic wave sensing-based method for predicting external force damage events according to this application embodiment can be performed.

[0091] Optionally, the computer-readable storage medium may be a random access memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, etc.

[0092] Obviously, the above embodiments of this application are merely examples for clearly illustrating the technical solution of this application, and are not intended to limit the specific implementation of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of this application should be included within the protection scope of the claims of this application.

Claims

1. A method for predicting external force damage events based on fiber optic acoustic wave sensing, characterized in that, The method includes: Acquire audio data from optical fibers and convert the audio data into a spectral feature map to be processed; The spectrum feature map to be processed is divided into blocks according to a preset frequency range to obtain several frequency band feature maps; The frequency band feature map is input into the external damage event identification model of the corresponding frequency range for processing to obtain the probability distribution of external damage events in the frequency range. The spectral feature map to be processed is vectorized to obtain the spectral feature vector. The spectral feature vector is input into a pre-trained gating network to obtain the feature weights for each frequency range. The global external failure event probability distribution is obtained by weighting and summing the feature weights of the external failure event probability distribution with respect to the corresponding frequency range. Based on the global external damage event probability distribution, the target external damage event type is obtained.

2. The method for predicting external force damage events based on fiber optic acoustic wave sensing according to claim 1, characterized in that, The step of performing feature vectorization on the spectral feature map to obtain a spectral feature vector includes: The spectral feature map to be processed is divided according to a preset grid to obtain several sub-spectral feature maps; The sub-spectral feature map is subjected to self-attention processing to obtain a sub-spectral self-attention feature map; The sub-spectral self-attention feature map is fused with its neighboring sub-spectral self-attention feature maps to obtain a fused feature map. The fused feature map is compressed using global average pooling to obtain a sub-spectral feature vector. The sub-spectral feature vectors are concatenated to obtain the concatenated spectral feature vector. The spectral feature vector is obtained by processing the spliced ​​spectral feature vector through nonlinear mapping.

3. The method for predicting external force damage events based on fiber optic acoustic wave sensing according to claim 2, characterized in that, The step of fusing the sub-spectral self-attention feature map with its neighboring sub-spectral self-attention feature maps to obtain a fused feature map includes: For each of the sub-spectral self-attention feature maps, obtain the sparse connection topology between the sub-spectral self-attention feature map and its neighboring sub-spectral self-attention feature maps; Based on the multi-head attention mechanism and the sparse connection topology, the sub-spectral self-attention feature map is fused with its neighboring sub-spectral self-attention feature maps to obtain a fused feature map.

4. The method for predicting external force damage events based on fiber optic acoustic wave sensing according to any one of claims 1-3, characterized in that, The step of converting the audio data into a spectral feature map to be processed includes: The audio data is decoded into time-domain data; The time-domain data is subjected to a preset bandpass filter to obtain the processed time-domain data; The processed time-domain data is resampled to a preset sampling rate to obtain the time-domain data at the preset sampling rate; The time-domain data at a preset sampling rate is converted into a Mel spectrum, and features are extracted from the Mel spectrum to obtain the spectrum feature map to be processed.

5. The method for predicting external force damage events based on fiber optic acoustic wave sensing according to any one of claims 1-3, characterized in that, The external damage event recognition model includes a depthwise separable convolutional module, a frequency domain attention module, a residual connection module, and a classification output module connected in sequence. The depthwise separable convolution module is used to perform convolutional feature extraction processing on the input frequency band feature map to obtain an initial feature map; The frequency domain attention module is used to perform channel weight calibration processing on the initial feature map using a global attention mechanism to obtain an attention-weighted feature map; The residual connection module is used to perform an identity mapping on the attention-weighted feature map, and to perform feature superposition processing on the mapped attention-weighted feature map and the frequency band feature map to obtain an enhanced feature map; The classification output module is used to perform category probability mapping processing on the enhanced feature map to obtain the probability distribution of the external damage event.

6. The method for predicting external force damage events based on fiber optic acoustic wave sensing according to any one of claims 1-3, characterized in that, The external damage event identification model is obtained through pre-training; The pre-training of the external failure event identification model includes: Collect historical audio data from optical fibers and convert the historical audio data into historical spectral feature maps; Obtain the external damage event type and external damage scene environment information corresponding to the historical audio data; The historical spectrum feature map is divided into blocks according to a preset frequency range to obtain several frequency band historical feature maps; The historical feature map of the frequency band in the frequency range is input into the external damage event recognition model to be trained in the corresponding frequency range for processing, so as to obtain the historical external damage event probability distribution and predicted external damage site environment information in the frequency range. For each frequency range, based on the external damage event type and the historical external damage event probability distribution, the first training loss of the external damage event recognition model to be trained is calculated; based on the external damage site environment information and the predicted external damage site environment information, the second training loss of the external damage event recognition model to be trained is calculated; and based on the first training loss and the second training loss, the parameters of the external damage event recognition model to be trained are updated to obtain the external damage event recognition model trained for the frequency range.

7. The method for predicting external force damage events based on fiber optic acoustic wave sensing according to any one of claims 1-3, characterized in that, The step of obtaining the target external damage event type based on the global external damage event probability distribution includes: The external failure event type with the highest probability value is obtained from the global external failure event probability distribution. If the highest probability value exceeds a preset threshold, the external failure event type with the highest probability value is taken as the target external failure event type.

8. A system for predicting external force damage events based on fiber optic acoustic wave sensing, characterized in that, The system includes: The data acquisition module is used to acquire audio data from the optical fiber and convert the audio data into a spectral feature map to be processed. The data segmentation module is used to segment the spectrum feature map to be processed according to a preset frequency range to obtain several frequency band feature maps; The model processing module is used to input the frequency band feature map into the external damage event identification model of the corresponding frequency range for processing, and obtain the probability distribution of external damage events in the frequency range. The vectorization module is used to perform feature vectorization processing on the spectrum feature map to be processed to obtain a spectrum feature vector. The weight calculation module is used to input the spectral feature vector into the pre-trained gating network to obtain the feature weights of each frequency range. The event probability calculation module is used to perform a weighted summation of the external failure event probability distribution and the feature weights of its corresponding frequency range to obtain the global external failure event probability distribution; The event determination module is used to obtain the target external damage event type based on the global external damage event probability distribution.

9. An electronic device, characterized in that, include: Memory, used to store one or more computer programs; A processor, when the one or more computer programs are executed by the processor, implements the method for predicting external force damage events based on fiber optic acoustic wave sensing as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the method for predicting external force damage events based on fiber optic acoustic wave sensing as described in any one of claims 1-7.

Images