A method and system for monitoring for external breakage sound events

Abstract

This invention provides a method and system for monitoring externally damaged sound source events. The method employs a heterogeneous dual-branch recognition network to process one-dimensional vibration signals and two-dimensional time-frequency maps separately, capturing transient pulses at different physical frequencies and microscopic acoustic signatures, achieving comprehensive coverage of the physical properties and visual features of the sound source signal. Simultaneously, a bidirectional interactive fusion mechanism, using a closed-loop design of "global semantics guiding local details," effectively distinguishes complex sound source events with similar spectra but vastly different operating rhythms. Furthermore, a decision fusion mechanism based on the accuracy weighting of different processing branches determines the initial classification result. Based on this, the category of the sound source event is verified by combining spatiotemporal trajectory data to determine the final classification result. This significantly improves the accuracy of sound source event classification, overcoming the bottleneck of complex sound source identification; greatly reduces the false alarm rate, enhancing the system's anti-interference capability; and balances edge deployment requirements with real-time inference performance.

Description

Technical Field

[0001] This invention relates to the field of fiber optic sensing and security monitoring technology, and in particular to a method and system for monitoring externally damaged sound source events. Background Technology

[0002] Distributed Acoustic Sensing (DAS) technology utilizes the principle of coherent Rayleigh scattering to transform ordinary communication optical cables laid in the field into thousands of "all-time, all-domain" vibration sensors, widely used in long-distance oil and gas pipeline inspection, perimeter security, and rail transit monitoring. However, in complex real-world engineering environments, sound source identification faces the following significant technical challenges: 1) High false alarm rate and poor anti-interference capability: Traditional DAS mainly relies on signal amplitude energy for discrimination, which is easily affected by environmental background noise (such as wind and rain noise, traffic background noise), leading to frequent false alarms. 2) Difficulty in sound source feature extraction: Vibration signals generated by different types of sound sources (such as mechanical excavation, manual excavation, pedestrian movement, and vehicle passage) are intertwined in time and space, making it difficult for simple rule-based algorithms to extract highly recognizable feature vectors. 3) Conflict between real-time performance and computing power: The amount of data generated by DAS systems is huge (usually at the Gbps level). Achieving real-time artificial intelligence (AI) inference with high sampling rate and high spatial resolution at the edge places stringent requirements on the hardware architecture. Summary of the Invention

[0003] This invention provides a method and system for monitoring externally damaged sound source events, which significantly improves the accuracy of sound source event classification, breaks through the bottleneck of complex sound source identification, greatly reduces the false alarm rate, enhances the system's anti-interference capability, and takes into account both edge-side deployment requirements and real-time inference performance.

[0004] Firstly, this embodiment provides a method for monitoring external sound source damage events, the method comprising: The raw vibration data of the test point collected by the distributed acoustic wave sensing device is obtained, and the raw vibration data is preprocessed to obtain a one-dimensional vibration signal and a two-dimensional time-frequency diagram. The one-dimensional vibration signal is processed by the first processing branch in the heterogeneous dual-branch recognition network to extract the global temporal features of the one-dimensional vibration signal, and the two-dimensional time-frequency graph is processed by the second processing branch in the heterogeneous dual-branch recognition network to extract the high-dimensional semantic features of the two-dimensional time-frequency graph. The high-dimensional semantic features are injected into the first processing branch and combined with the global temporal features to perform global temporal logic modeling, thereby obtaining updated global temporal features for global temporal modeling. The updated global temporal features are then injected back into the second processing branch and fused with the high-dimensional semantic features to obtain semantically enhanced updated high-dimensional semantic features. The prediction probability of the first sound source event at the test point is identified based on the updated global temporal features, and the prediction probability of the second sound source event at the test point is identified based on the updated high-dimensional semantic features. Based on the prediction probability of the first sound source event and the prediction probability of the second sound source event, the initial classification result of the sound source event at the test point is determined. When the initial classification result meets the preset alarm conditions, a spatiotemporal waterfall diagram of the test point is constructed, and the final classification result of the sound source event of the test point is determined according to the movement trajectory in the spatiotemporal waterfall diagram, and an alarm decision is executed.

[0005] Secondly, this embodiment provides a system for monitoring external sound source damage events, the system comprising: The preprocessing module is used to acquire the raw vibration data of the test point collected by the distributed acoustic wave sensing device, and to preprocess the raw vibration data to obtain a one-dimensional vibration signal and a two-dimensional time-frequency diagram. The feature extraction module is used to process the one-dimensional vibration signal through the first processing branch in the heterogeneous dual-branch recognition network to extract the global temporal features of the one-dimensional vibration signal, and to process the two-dimensional time-frequency graph through the second processing branch in the heterogeneous dual-branch recognition network to extract the high-dimensional semantic features of the two-dimensional time-frequency graph. The feature fusion module is used to inject the high-dimensional semantic features into the first processing branch and perform global temporal logic modeling with the global temporal features to obtain updated global temporal features for global temporal modeling, and then inject the updated global temporal features back into the second processing branch and fuse them with the high-dimensional semantic features to obtain updated high-dimensional semantic features with enhanced semantics. The initial classification module is used to identify the first sound source event prediction probability of the test point based on the updated global temporal features and the second sound source event prediction probability of the test point based on the updated high-dimensional semantic features, and to determine the initial classification result of the sound source event of the test point based on the first sound source event prediction probability and the second sound source event prediction probability. The final classification module is used to construct a spatiotemporal waterfall diagram of the test point when the initial classification result meets the preset alarm conditions, determine the final classification result of the sound source event of the test point according to the movement trajectory in the spatiotemporal waterfall diagram, and execute an alarm decision.

[0006] This invention provides a method and system for monitoring externally damaged sound sources. The method includes: firstly, acquiring raw vibration data from a test point collected by a distributed acoustic wave sensing device; preprocessing the raw vibration data to obtain a one-dimensional vibration signal and a two-dimensional time-frequency graph; secondly, processing the one-dimensional vibration signal through a first processing branch in a heterogeneous dual-branch recognition network to extract global temporal features of the one-dimensional vibration signal, and processing the two-dimensional time-frequency graph through a second processing branch in the heterogeneous dual-branch recognition network to extract high-dimensional semantic features of the two-dimensional time-frequency graph; and then injecting the high-dimensional semantic features into the first processing branch and performing global temporal logic modeling with the global temporal features to obtain updated global temporal features for global temporal modeling. The updated global temporal features are then injected into the second processing branch and fused with the high-dimensional semantic features to obtain semantically enhanced updated high-dimensional semantic features. Next, the predicted probability of a first sound source event at the test point is identified based on the updated global temporal features, and the predicted probability of a second sound source event at the test point is identified based on the updated high-dimensional semantic features. Based on the predicted probabilities of the first and second sound source events, the initial classification result of the sound source events at the test point is determined. Finally, when the initial classification result meets the preset alarm conditions, a spatiotemporal waterfall diagram of the test point is constructed. The final classification result of the sound source events at the test point is determined based on the movement trajectory in the spatiotemporal waterfall diagram, and an alarm decision is executed. The aforementioned technical solution employs a heterogeneous dual-branch recognition network. One processing branch processes the one-dimensional vibration signal to capture transient pulses at different physical frequencies, while the other processing branch processes the two-dimensional time-frequency map to capture microscopic acoustic patterns, achieving comprehensive coverage of the physical properties and visual features of the sound source signal. Simultaneously, a bidirectional interactive fusion mechanism is used, injecting the high-dimensional semantic features captured by the second processing branch into the first processing branch, and feeding back the global temporal logic understood by the first processing branch to the second processing branch. This closed-loop design, "global semantics guiding local details," effectively distinguishes complex sound source events with similar spectra but vastly different operating rhythms. Furthermore, a decision fusion mechanism based on the accuracy weighting of different processing branches determines the initial classification result. Then, based on this, the category of the sound source event is verified by combining spatiotemporal trajectory verification to determine the final classification result. This distributed acoustic sensing sound source event real-time discrimination method, based on the combination of artificial intelligence neural networks and spatiotemporal trajectory verification, no longer relies on single time-segment acoustic features. Instead, it constructs a three-in-one recognition architecture of "physical features - acoustic texture - spatiotemporal trajectory" to achieve highly reliable discrimination of target events. It significantly improves the accuracy of sound source event classification, breaks through the bottleneck of complex sound source identification, greatly reduces the false alarm rate, and enhances the system's anti-interference capability; and takes into account both edge deployment requirements and real-time inference performance.

[0007] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating a method for monitoring external sound source damage events according to Embodiment 1 of the present invention. Figure 2 This is a structural example diagram of a heterogeneous dual-branch identification network in the execution of a method for monitoring external sound source events, as provided in Embodiment 1 of the present invention. Figure 3 This is a structural example diagram of the data fusion module in the execution of a method for monitoring external sound source events, as provided in Embodiment 1 of the present invention. Figure 4 This is a flowchart illustrating another method for monitoring external sound source incidents provided in Embodiment 2 of the present invention; Figure 5 This is a structural example diagram of a multi-scale convolution module in the execution of a method for monitoring external sound source events, as provided in Embodiment 2 of the present invention. Figure 6 This is a structural example diagram of the first convolution module in the execution of a method for monitoring external sound source events to prevent damage, provided in Embodiment 2 of the present invention. Figure 7 This is a structural example diagram of the data upsampling module in the execution of a method for monitoring external sound source events, as provided in Embodiment 2 of the present invention. Figure 8 This is a structural example diagram of the second convolution module in the execution of a method for monitoring external sound source events, as provided in Embodiment 2 of the present invention. Figure 9 This is a structural example diagram of the data downsampling module in the execution of a method for monitoring external sound source events, as provided in Embodiment 2 of the present invention. Figure 10 This is a structural example diagram of the Transformer module in the execution of a method for monitoring external sound source events, as provided in Embodiment 2 of the present invention. Figure 11 This is a schematic diagram of the structure of an external damage sound source monitoring system provided in Embodiment 3 of the present invention; Figure 12This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of the present invention. Detailed Implementation

[0010] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0011] It should be noted that the terms "first," "second," etc., in the specification, claims, 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 the 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.

[0012] Example 1 Figure 1 This is a flowchart illustrating a method for monitoring external damage sound source events according to Embodiment 1 of the present invention. This method is applicable to situations where external damage sound source events are monitored. This method can be executed by an external damage sound source event monitoring system, which can be implemented in hardware and / or software and is generally integrated into electronic devices.

[0013] like Figure 1 As shown, the method for monitoring external sound source damage events provided in this embodiment can specifically include the following steps: S101. Obtain the original vibration data of the test point collected by the distributed acoustic wave sensing device, preprocess the original vibration data to obtain a one-dimensional vibration signal and a two-dimensional time-frequency diagram.

[0014] In this embodiment, raw phase data of each point to be predicted over a period of time is acquired using a DAS device and recorded as raw vibration data. The raw vibration data is then preprocessed, including but not limited to noise suppression, detrending, bandpass filtering, normalization, data correction, smoothing, and handling of missing values, to obtain a one-dimensional vibration signal. The one-dimensional vibration signal is then time-frequency mapped to a two-dimensional time-frequency graph. The first coordinate of the time-frequency graph is time, and the second coordinate is frequency. The time-frequency graph can be used to reflect the energy distribution of an audio signal over a certain period of time, providing a visual representation of the physical characteristics (frequency and amplitude) of sound. Finally, the normalized one-dimensional vibration signal and the two-dimensional time-frequency graph are obtained and subsequently output to a heterogeneous dual-branch recognition network.

[0015] In this embodiment, the time-frequency diagram is not limited to feature maps generated by Mel spectrograms, Short-Time Fourier Transform (STFT), Wavelet Transform, etc. Its purpose is to visualize the one-dimensional physical properties of sound, and any scheme with such feature transformations falls within the protection scope of this embodiment.

[0016] S102. The one-dimensional vibration signal is processed by the first processing branch in the heterogeneous dual-branch recognition network to extract the global temporal features of the one-dimensional vibration signal, and the two-dimensional time-frequency graph is processed by the second processing branch in the heterogeneous dual-branch recognition network to extract the high-dimensional semantic features of the two-dimensional time-frequency graph.

[0017] In this embodiment, a heterogeneous dual-branch recognition network is pre-constructed and trained, comprising two processing branches, denoted as the first processing branch and the second processing branch. The first processing branch processes the one-dimensional vibration signal to extract its features, denoted as global temporal features. The second processing branch processes the two-dimensional time-frequency graph to extract its features, denoted as high-dimensional semantic features.

[0018] The first processing branch can employ multi-scale convolutional modules, residual neural networks (ResNet), efficient convolutional neural network models (EfficientNet), lightweight neural networks (MobileNet), or traditional convolutional layer stacking as feature vector extraction methods. This serves as an interactive fusion between the "texture extraction branch" and the "semantic sequence branch." No specific restrictions are imposed here, but multi-scale convolutional modules are preferred. The second processing branch can use convolutional modules for feature extraction, equivalent to the semantic sequence branch.

[0019] It should be noted that the training process of the heterogeneous branch recognition network can be described as follows: 1) Data preprocessing: The original vibration data needs to be de-trended, differencing, bandpass filtering, normalization, and data correction to generate a one-dimensional vibration signal, and a Mel-time frequency diagram is generated synchronously for the one-dimensional vibration signal. 2) Dataset definition: According to the actual composition of the dataset, the collected historical alarm signal set is divided into 7 categories, including excavator, manual knocking, electric pick, and environmental noise, through manual screening, and divided into training set and validation set in a 7:3 ratio. In each iteration of model training, the number of samples input into the model for forward propagation and parameter update is set to 64. 3) Model construction: The above model is constructed using a functionally complete framework (PyTorch) for building deep learning models. A truncated normal distribution is used for linear layers, and a Kaiming normal distribution is used for convolutional layers. The Kaiming normal distribution is a neural network weight initialization method designed specifically for ReLU and its variant activation functions, aiming to solve the gradient vanishing or exploding problem in deep networks and maintain the stability of the variance of activation values ​​during forward propagation. 4) Loss Function and Optimizer Selection: Cross-entropy loss is used as the loss function, and AdamW is selected as the optimizer. Cosine annealing combined with Warmup is employed. The learning rate is slowly increased in the early stages of training to stabilize the gradient, and then gradually decayed in the later stages to find the global optimum. AdamW (Adam with Weight Decay) is a deep learning optimization algorithm, and Warmup is a strategy in deep learning that gradually increases the learning rate to accelerate model convergence and improve generalization ability. 5) Training Logic: Time-frequency maps and one-dimensional vibration signals are simultaneously input into the model, and N rounds of feature interaction are performed. The output values ​​of the classification heads of the first and second processing branches are obtained respectively, and the joint loss is calculated. Gradient backpropagation is then performed to update the model parameters. The training environment is deployed on a high-performance computing platform, with 100 iterations and an early stopping mechanism. After each iteration, accuracy and false alarm rate are rigorously monitored on the validation set, serving as core indicators for evaluating the model's classification robustness and environmental generalization ability.

[0020] The trained heterogeneous dual-branch recognition network is then optimized and deployed. For example, L1 regularization penalty (sparse constraint training) is used to prune the model, followed by fine-tuning. The computation graph is scanned using FX Graph Mode and operator fusion is performed, compressing 32-bit floating-point weights and activation values ​​into 8-bit integers to generate a highly compressed Open Neural Network Exchange (ONNX) format model. Depending on the terminal hardware architecture, the quantized model is converted through middleware into high-performance deployment formats such as TFLite (mobile), TensorRT.engine (edge ​​GPU / NPU), or RKNN (embedded chip), ensuring low-latency, low-power real-time inference on edge computing platforms. FX Graph Mode is a static computation graph intermediate representation mode built by PyTorch based on the torch.fx framework.

[0021] S103. The high-dimensional semantic features are injected into the first processing branch and used to perform global temporal logic modeling with the global temporal features to obtain updated global temporal features for global temporal modeling. The updated global temporal features are then injected back into the second processing branch and fused with the high-dimensional semantic features to obtain updated high-dimensional semantic features with enhanced semantics.

[0022] In this embodiment, the first processing branch further includes a deep learning model (Transformer) module based on an attention mechanism. This module is connected to a multi-scale convolution module, and its input is the output of the multi-scale convolution module. In the machine learning model (Transformer) module based on a multi-head attention mechanism, the multi-head attention mechanism (MHSA) can be replaced by a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), or a variant of a speech recognition model (convolution-augmented Transformer). As long as its logic is to capture causal relationships within a long time window through temporal sequence modeling, it is considered an equivalent scheme and falls within the protection scope of this embodiment.

[0023] In this embodiment, the second convolutional module in the second processing branch is referred to as the second convolutional module, and the input of the second convolutional module is the output of the first convolutional module.

[0024] In this embodiment, the process of mutual feature enhancement between the two processing branches can be described as follows: the global temporal features are input into the data upsampling module to generate upsampled output features; the high-dimensional semantic features are input into the second convolution module of the second processing branch to obtain a first refined feature and a second refined feature, and the second refined feature is fused with the upsampled output features to generate the updated high-dimensional semantic features; the first refined feature is input into the data downsampling module to generate downsampled output features; the global temporal features and the downsampled output features are input into the attention-based deep learning module of the first processing branch to output the updated global temporal features; the updated global temporal features and the updated high-dimensional semantic features are cyclically fed back to the second convolution module and the attention-based deep learning module to achieve mutual enhancement of global temporal features and high-dimensional semantic features in a set number of rounds, and finally output the updated global temporal features and the updated high-dimensional semantic features.

[0025] S104. Based on the updated global temporal features, identify the first sound source event prediction probability of the test point and the second sound source event prediction probability of the test point, respectively. Based on the first sound source event prediction probability and the second sound source event prediction probability, determine the initial classification result of the sound source event of the test point.

[0026] In this embodiment, the classification head (i.e., the fully connected layer) of the first processing branch classifies the updated global temporal features to obtain the predicted probability of the first sound source event at the test point. The classification head (i.e., the fully connected layer) of the second processing branch classifies the updated high-dimensional semantic features to obtain the predicted probability of the second sound source event at the test point. Then, the sound source temporal prediction probabilities of the two branches are fused to obtain a comprehensive evaluation score of the sound source event category at the test point. The fusion method can be linear weighted fusion based on class accuracy, or fusion based on DS evidence theory, secondary stacking fusion based on meta-learners, or feature-level fusion based on attention mechanisms; no specific limitation is made here. The linear weighted fusion method based on class accuracy is preferred. Then, the sound source event category corresponding to the maximum value in the comprehensive evaluation score is selected as the initial classification result of the sound source event at the test point.

[0027] For example, Figure 2 This is an example diagram of the structure of a heterogeneous dual-branch identification network in the execution of a method for monitoring external sound source events, as provided in Embodiment 1 of the present invention. Figure 2As shown, the raw vibration data is processed by the data processing module to obtain a one-dimensional vibration signal and a two-dimensional time-frequency plot. The one-dimensional vibration signal is input to the first processing branch for processing. The first processing branch includes a multi-scale convolution module, an attention-based deep learning module, and a classification head, outputting the prediction result of this branch. The two-dimensional time-frequency plot is input to the second processing branch for processing. The second processing branch includes a first convolution module, a second convolution module, and a classification head, outputting the prediction result of this branch. The processing results of the two branches are fused through the result fusion module, and the classification result is output based on the fusion result, corresponding to the initial classification result. Among them, the features output by the attention-based deep learning module and the features output by the second convolution module are further fused through the data fusion module, performing N iterations. After the trained model is pruned and quantized, it is deployed. In subsequent practical applications, the raw vibration data is processed sequentially through the data processing module, the trained model, and the result fusion module to output the initial classification result.

[0028] Figure 3 This is a structural example diagram of the data fusion module in the execution of a method for monitoring external sound source events according to Embodiment 1 of the present invention, as shown below. Figure 3 As shown, the first processing branch (shown as the Transformer branch in the figure) and the second processing branch (shown as the CNN branch in the figure) are fused through a data fusion module, which includes a data downsampling module and a data upsampling module.

[0029] S105. When the initial classification result meets the preset alarm conditions, construct the spatiotemporal waterfall diagram of the test point, determine the final classification result of the sound source event of the test point according to the movement trajectory in the spatiotemporal waterfall diagram, and execute the alarm decision.

[0030] In this embodiment, the system continuously runs the neural network inference of the above steps. When the initial classification result output by the classification head is the target event (entering an excavator, electric pick, manual knocking, etc.), and the probability exceeds a preset threshold (e.g., 0.75), it is determined as a "valid alarm initial judgment," and the wide-area spatiotemporal trajectory recognition mechanism is immediately activated. Simultaneously, the spatial coordinates x0 and timestamp t0 of the current alarm point are recorded. Then, starting from t0, spatiotemporal data within a time period of t0 + a customizable duration (e.g., 300s) is recorded to generate a spatiotemporal waterfall plot. The horizontal axis of the spatiotemporal waterfall plot represents spatial distance (unit: meters), the vertical axis represents time (unit: seconds), and the pixel value represents the normalized vibration amplitude. The spatiotemporal waterfall plot is then preprocessed, sequentially performing median filtering, Gaussian smoothing filtering, signal enhancement, normalization, and size adjustment to obtain the preprocessed spatiotemporal waterfall plot.

[0031] A hard filtering logic based on kinematics is introduced to identify and eliminate traffic flow interference. When the neural network outputs an alarm signal for the first time, the system begins recording spatiotemporal data (spatiotemporal waterfall plot) within a certain range before and after the alarm point to capture the signal evolution process within that area. The system projects subsequent alarm points onto the spatiotemporal waterfall plot's spatial coordinate system in real time. If, within a certain time period, the alarm points are displayed as being on the movement trajectory of the spatiotemporal waterfall plot for a set number of consecutive times (e.g., 3 times), the sequence of signals is determined to be coherent interference caused by an interference source with physical displacement characteristics (such as a car), and no alarm is output. Otherwise, an alarm is output.

[0032] The aforementioned technical solution employs a heterogeneous dual-branch recognition network. One processing branch processes the one-dimensional vibration signal to capture transient pulses at different physical frequencies, while the other processing branch processes the two-dimensional time-frequency map to capture microscopic acoustic patterns, achieving comprehensive coverage of the physical properties and visual features of the sound source signal. Simultaneously, a bidirectional interactive fusion mechanism is used, injecting the high-dimensional semantic features captured by the second processing branch into the first processing branch, and feeding back the global temporal logic understood by the first processing branch to the second processing branch. This closed-loop design, "global semantics guiding local details," effectively distinguishes complex sound source events with similar spectra but vastly different operating rhythms. Furthermore, a decision fusion mechanism based on the accuracy weighting of different processing branches determines the initial classification result. Then, based on this, the category of the sound source event is verified by combining spatiotemporal trajectory verification to determine the final classification result. This distributed acoustic sensing sound source event real-time discrimination method, based on the combination of artificial intelligence neural networks and spatiotemporal trajectory verification, no longer relies on single time-segment acoustic features. Instead, it constructs a three-in-one recognition architecture of "physical features - acoustic texture - spatiotemporal trajectory" to achieve highly reliable discrimination of target events. It significantly improves the accuracy of sound source event classification, breaks through the bottleneck of complex sound source identification, greatly reduces the false alarm rate, and enhances the system's anti-interference capability; and takes into account both edge deployment requirements and real-time inference performance.

[0033] Example 2 Figure 4This is a flowchart illustrating another method for monitoring externally damaged sound sources provided in Embodiment 2 of the present invention. This embodiment is a further optimization of the above embodiment. In this embodiment, the following optimizations are made: "processing the one-dimensional vibration signal through the first processing branch in the heterogeneous dual-branch identification network to extract the global temporal features of the one-dimensional vibration signal, and processing the two-dimensional time-frequency graph through the second processing branch in the heterogeneous dual-branch identification network to extract the high-dimensional semantic features of the two-dimensional time-frequency graph" is further optimized. Additionally, the following optimizations are made: "injecting the high-dimensional semantic features into the first processing branch and performing global temporal logic modeling with the global temporal features to obtain updated global temporal features for global temporal modeling, and then injecting the updated global temporal features inversely into the second processing branch" is also optimized. The optimizations include: merging the processing branch with the high-dimensional semantic features to obtain semantically enhanced updated high-dimensional semantic features; optimizing the following: identifying the prediction probability of the first sound source event of the test point based on the updated global temporal features and the prediction probability of the second sound source event of the test point based on the updated high-dimensional semantic features; determining the initial classification result of the sound source event of the test point based on the prediction probability of the first sound source event and the prediction probability of the second sound source event; and optimizing the following: when the initial classification result meets the preset alarm conditions, constructing a spatiotemporal waterfall diagram of the test point, determining the final classification result of the sound source event of the test point based on the movement trajectory in the spatiotemporal waterfall diagram, and executing an alarm decision.

[0034] like Figure 4 As shown in the figure, this embodiment 2 provides a method for monitoring external sound source damage events, which specifically includes the following steps: S201. Obtain the original vibration data of the test point collected by the distributed acoustic wave sensing device, preprocess the original vibration data to obtain a one-dimensional vibration signal and a two-dimensional time-frequency diagram.

[0035] S202. In the first processing branch, the one-dimensional vibration signal is subjected to global temporal feature extraction by a multi-scale convolution module to generate global temporal features that meet the input requirements of the deep learning module based on the attention mechanism.

[0036] In this embodiment, a multi-scale convolution module is used in the first processing branch to extract features from the one-dimensional vibration signal. The processing logic can be expressed as follows: the one-dimensional vibration signal is input into the multi-scale convolution module, and the output feature matrix is ​​concatenated along the channel dimension using convolution kernels of different sizes. Dynamic normalization, activation, and pooling are then performed. The feature matrix dimensions are adjusted to add classification features (cls_token), and positional encoding is added to meet the input requirements of the subsequent attention-based deep learning model (Transformer) module branch. The final output is a feature matrix that meets the input requirements of the Transformer branch, denoted as the global temporal feature (represented as X).

[0037] It should be noted that the stride is the same as the frame movement of the two-dimensional time-frequency graph, and the dimensions are unified. The stride of the three sets of convolutional kernels adopts the same frame movement design as the two-dimensional time-frequency graph to ensure that the time steps can be aligned when the feature vectors are fused in the future.

[0038] As a specific implementation, the step of extracting global temporal features from the one-dimensional vibration signal using a multi-scale convolutional module to generate global temporal features that meet the input requirements of an attention-based deep learning module can be optimized, including: a1) The high-frequency impact pulse and low-frequency envelope variation features in the one-dimensional vibration signal are captured by at least three sets of one-dimensional convolution kernels configured in parallel in the multi-scale convolution module and spliced ​​into multi-scale features.

[0039] Figure 5 This is a structural example diagram of a multi-scale convolution module in the execution of a method for monitoring external sound source events, as provided in Embodiment 2 of the present invention. Figure 5 As shown, a multi-scale kernel design is implemented for the multi-scale convolution module, with three sets of convolution kernels configured in parallel: a large kernel, a medium kernel, and a small kernel. The short kernel is used to capture high-frequency impulses, while the long kernel is used to capture low-frequency envelope changes. The features extracted by the three kernels are then concatenated into multi-scale features. For example, the sizes of the three sets of convolution kernels are 128, 256, and 512, respectively.

[0040] b1) The multi-scale features are sequentially normalized, activated, pooled, and adjusted in dimension to obtain the processed physical features.

[0041] Continue to refer to Figure 5 The process involves processing multi-scale features through normalization, activation, and pooling layers, followed by dimensionality adjustment to obtain the processed physical features. This step is equivalent to normalizing and downsampling the multi-scale features. Dynamic range normalization is performed using a normalization layer (BatchNorm1d), and max pooling (MaxPool1d) is used to extract salient physical features, compressing the sequence (i.e., the extracted salient physical features) into token form. Then, dimensionality adjustment is performed; for example, the compressed result might be two-dimensional, which is then flattened into a single layer to finally obtain the processed physical features.

[0042] c1) Add position encoding to the physical features to compensate for the lack of time-axis logic perception in the attention mechanism, and generate global temporal features that meet the input requirements of deep learning modules based on the attention mechanism.

[0043] Continue to refer to Figure 5This step is used to inject position encoding into the processed physical features. Sine and cosine position encoding (Position Embedding) is injected into the Token sequence to compensate for the lack of time axis logic awareness of the attention mechanism, thereby generating global temporal features that meet the input requirements of deep learning modules based on the attention mechanism.

[0044] S203. In the second processing branch, the two-dimensional time-frequency map is expanded from the initial number of channels to a high-dimensional semantic space by the first convolution module, and normalization, activation and pooling processes are performed in sequence to generate the high-dimensional semantic features of the two-dimensional time-frequency map.

[0045] In this embodiment, a convolutional module is used in the second processing branch to extract features from the time-frequency map, and this convolutional module is referred to as the first convolutional module. Figure 6 This is a structural example diagram of the first convolutional module in the execution of a method for monitoring external sound source events to prevent damage, as provided in Embodiment 2 of the present invention. Figure 6 As shown, the processing logic of the first convolution module can be expressed as follows: input the two-dimensional time-frequency map into the first convolution module, expand the two-dimensional time-frequency map to a high-dimensional semantic space through multiple convolution kernels (1*1 convolution kernels), perform dynamic normalization, activation and pooling, and output a high-dimensional semantic feature matrix, denoted as the high-dimensional semantic feature of the two-dimensional time-frequency map (represented as Y).

[0046] Understandably, the first convolutional module processes the two-dimensional time-frequency map to achieve dimensional depth, expanding the basic 3-channel time-frequency map (primary sensory information) to high-dimensional semantic features. Activation pooling: By introducing non-linearity through activation layers, the network can learn complex patterns, while aggregating local regions of the feature map (convolutional layer output), significantly reducing the width and height of the feature map.

[0047] S204. Input the global time series features into the data upsampling module to generate upsampled output features.

[0048] It should be noted that the dimensionality of the convolutional module decreases continuously with each N iteration; for example, it decreases from 64 dimensions before fusion to 32, then to 16, and finally to 8. The Transformer module, however, maintains the original dimensionality; for instance, it remains 64 dimensions after fusion. Therefore, downsampling or upsampling is necessary when fusing the features from the two branches. In this embodiment, the global temporal features obtained from the first processing branch and the high-dimensional semantic features obtained from the second processing branch are fused using a data fusion module, which includes a data downsampling module and a data upsampling module. The input to the data upsampling module comes from the global temporal features (denoted as X) of the Transformer module and the second refined feature (denoted as K) output from the second convolutional module; its output is the upsampled output feature (denoted as M).

[0049] As a specific implementation method, the step of inputting the global temporal features into the data upsampling module to generate upsampled output features can be optimized, including: a2) Perform dimensional adjustment on the updated global temporal features to obtain the first reshaped features.

[0050] Figure 7 This is a structural example diagram of the data upsampling module in the execution of a method for monitoring external sound source events, as provided in Embodiment 2 of the present invention. Figure 7 As shown, its processing logic can be expressed as follows: Dimension adjustment is performed on the global temporal feature (represented as X) or the updated global temporal feature (represented as new X), and the dimension-adjusted feature is recorded as the first reshaped feature. Dimension adjustment can be represented as [B,HW+1,C]->[B,HW,C]->[B,C,H,W], where B refers to the batch size, H refers to the feature map height, W refers to the feature map width, and C refers to the feature map dimension size. All letters within [] refer to the above and are unrelated to the letters representing feature map symbols in the text.

[0051] b2) After performing standard convolutional block operations on the first reshaped feature, an upsampling operation is performed to generate global semantic features.

[0052] Continue to refer to Figure 7 The output is then subjected to standard convolutional block operations (1*1 convolutional kernel) and upsampled to obtain new features, which are denoted as global semantic features (G).

[0053] c2) After performing standard convolutional block operations on the second refined feature, add it to the global semantic feature to obtain the added feature.

[0054] Continue to refer to Figure 7The second refined feature (denoted as K) output by the second convolution module is subjected to standard convolution block operation (1*1 convolution kernel) and then added to the global semantic feature (denoted as G) to obtain the added feature.

[0055] d2) The summed features are processed through a standard convolutional block, a convolutional layer, and a normalization layer to generate the upsampled output features.

[0056] Continue to refer to Figure 7 The summed features obtained above are first processed through a standard convolutional block (3*3 kernel), then through a convolutional layer (1*1) and a normalization layer to obtain the final features processed by the data upsampling module, which is denoted as the upsampling output feature (represented as M).

[0057] It's important to note that the data upsampling module utilizes the long-term semantic logic of the 3-second raw signal (extracted by the Transformer branch) to guide the weighted processing of the two-dimensional time-frequency texture features (extracted by the convolution branch). This "global-guided local" mechanism allows the system to semantically enhance subtle spectral details of the target event at the algorithmic level, thus balancing the trade-off between high sensitivity and high false alarm rate in the DAS recognition system. This can be understood as translating the "global semantics" understood by the Transformer back into "spatial language" and feeding it back to the CNN branch.

[0058] S205. Input the high-dimensional semantic features into the second convolution module of the second processing branch to obtain the first thinning feature and the second thinning feature, and fuse the second thinning feature with the upsampled output feature to generate the updated high-dimensional semantic features.

[0059] In this embodiment, another convolutional module in the second processing branch is referred to as the second convolutional module, and the input of the second convolutional module is the output of the first convolutional module. Its input is a high-dimensional semantic feature, and its processing logic is as follows: the high-dimensional semantic feature (represented as Y) is input into the second convolutional module, and after being refined by multiple layers of residual convolutional units, a first refined feature (represented as Y2) and a second refined feature (represented as K) are obtained. The second refined feature is then fused with the upsampled output feature (represented as M) to generate an updated high-dimensional semantic feature. The process can be described as follows: the high-dimensional semantic feature matrix (represented as Y) first passes through two standard convolutional blocks to obtain the first refined feature (represented as Y2). After convolution and normalization, the first refined feature (represented as Y2) is residually connected with the high-dimensional semantic feature matrix (represented as Y) to obtain the second refined feature (represented as K). This K is then added to the upsampled output feature (represented as M) output by the data upsampling module (sent by the Transformer module) to obtain the updated high-dimensional semantic feature (represented as the new Y, the original Y is discarded).

[0060] As a specific implementation, the step of inputting the high-dimensional semantic features into the second convolutional module of the second processing branch to obtain the first refined feature and the second refined feature, and fusing the second refined feature with the upsampled output feature to generate the updated high-dimensional semantic features, can be optimized, including: a3) Input the high-dimensional semantic features into the first standard convolutional block and the second standard convolutional block of the second processing branch in sequence to obtain the first refined features.

[0061] The input for this step is a high-dimensional semantic feature matrix (denoted as Y). Figure 8 This is a structural example diagram of the second convolution module in the execution of a method for monitoring external sound source events to prevent damage, as provided in Embodiment 2 of the present invention. Figure 8 As shown, the processing logic of the second processing branch is as follows: First, the processed feature map is obtained through two standard convolutional blocks, denoted as the first refined feature (represented as Y2). The standard convolutional block includes convolution + normalization + activation, denoted as the first standard convolutional block and the second standard convolutional block, respectively. The kernel size of the first standard convolutional block is 1*1, and the kernel size of the second standard convolutional block is 3*3.

[0062] b3) After performing convolution and normalization on the first refined feature in sequence, a residual connection is made with the high-dimensional semantic feature and activated to generate the second refined feature.

[0063] Continue to refer to Figure 8 The first refined feature (denoted as Y2) is convolved (with a kernel size of 1*1) and normalized, then residually connected with the updated high-dimensional semantic feature (denoted as Y), and activated through an activation layer to obtain a new refined feature, denoted as the second refined feature (denoted as K).

[0064] c3) The second refined feature is fused with the upsampled output feature to generate the updated high-dimensional semantic feature.

[0065] Continue to refer to Figure 8 The second refined feature (denoted as K) is added to the upsampled output feature (denoted as M) output by the data upsampling module (sent by the Transformer module) to obtain the updated high-dimensional semantic feature (denoted as the new Y, the original Y is discarded). By adjusting the stride setting of the standard convolutional block, the high-dimensional semantic feature is updated into a feature map, the size of which can be the same as the original high-dimensional semantic feature matrix (denoted as Y) or the size can be halved.

[0066] It should be noted that the above-mentioned residual convolutional unit based on the bottleneck structure achieves efficient voiceprint encoding of the time-frequency map through the coupling of 1*1 dimensional convolution and 3*3 feature extraction convolution.

[0067] S206. Input the first refined feature into the data downsampling module to generate downsampled output features.

[0068] In this embodiment, the input of the data downsampling module comes from the global temporal features (denoted as X) of the Transformer module, and the output is the downsampled output features (denoted as B).

[0069] As a specific implementation, the step of downsampling the first refined feature input data to generate downsampled output features can be optimized, including: a4) Perform convolutional projection, average pooling, and dimension adjustment on the first refined feature to obtain a second reshaped feature that is dimensionally aligned with the global temporal feature.

[0070] Figure 9 This is a structural example diagram of the data downsampling module in the execution of a method for monitoring external sound source events, as provided in Embodiment 2 of the present invention. Figure 9 As shown, the processing logic of the data downsampling module can be represented as follows: the first refined feature (represented as Y2) is passed through a convolutional layer (1*1 convolutional kernel) and then subjected to average pooling and dimension adjustment to make the dimension consistent with the global temporal feature (represented as X), thus obtaining the dimension-adjusted feature, which is denoted as the second reshaped feature.

[0071] b4) After performing layer normalization and activation on the second reshaped feature, the downsampled output feature is generated.

[0072] Continue to refer to Figure 9 The second reshaped feature after dimension adjustment is layer normalized and activated to obtain the feature output by the data downsampling module, which is denoted as the downsampling output feature (represented as B).

[0073] The dimensional changes in this processing logic can be represented as: [B,C,H,W]->[B,HW,C]->[B,HW+1,C], where B represents the number of images processed each time, C represents the dimension and channels, and HW represents the image size.

[0074] It should be noted that the data downsampling module utilizes 1*1 projective convolution and average pooling operators to transform convolutional feature maps with local topological properties into high-dimensional feature sequences, and then reconstructs these features through layer normalization and nonlinear activation. This design not only preserves the fine texture of the sound source signal in the frequency domain but also transforms it into semantic tokens with temporal correlation. This downsampling module directly "injects" high-dimensional semantic features from the second processing branch (CNN branch) into the first processing branch (Transformer branch), achieving cross-dimensional alignment of two-dimensional spectral texture features to one-dimensional semantic sequence features. In other words, the local texture features extracted by the convolutional branch are mapped to global semantic features and input into the Transformer branch.

[0075] Considering that the Transformer module has a fixed dimension, such as 64, while the convolutional module has a dimension of 128, it needs to be downsampled to 64 dimensions before it can be added to the Transformer module. The downsampling is done by feeding the data from the convolutional module to the Transformer module, and then upsampling is performed to expand the dimension back before adding it to the original features of the convolutional module.

[0076] S207. Input the global temporal features and the downsampled output features into the deep learning module based on the attention mechanism of the first processing branch, and output the updated global temporal features.

[0077] In this embodiment, the first processing branch further includes a deep learning model (Transformer) module based on an attention mechanism. This module is connected to a multi-scale convolution module, and its input is the output of the multi-scale convolution module. The input of the deep learning model based on the attention mechanism is the global temporal feature (denoted as X) that meets the input requirements of the Transformer branch, and the downsampled output feature (denoted as B, whose dimension is the same as X) obtained by the data downsampling module.

[0078] Its processing logic can be expressed as follows: The global temporal features (represented as X) and the downsampled output features (represented as B) are input into the deep learning module based on the attention mechanism in the first processing branch, and residual fusion, layer normalization, multi-head attention mechanism, secondary residual connection, layer normalization, multilayer perceptron nonlinear mapping, and finally residual connection processing are executed in sequence, and the output is updated global temporal features (represented as new X).

[0079] As a specific implementation, the deep learning module based on an attention mechanism that inputs the global temporal features and the downsampled output features into the first processing branch and outputs the updated global temporal features can be optimized, including: a5) Add the global temporal features to the downsampled output features and perform normalization to obtain the first feature matrix.

[0080] Figure 10 This is a structural example diagram of the Transformer module in the execution of a method for monitoring external sound source events to prevent damage, as provided in Embodiment 2 of the present invention. Figure 10 As shown, the processing logic of the deep learning module based on the attention mechanism can be expressed as follows: the global temporal features (represented as X) are added to the downsampled output features (represented as B), and then layer normalization is performed to obtain the feature matrix, which is denoted as the first feature matrix (represented as C).

[0081] b5) Apply a multi-head attention mechanism to the first feature matrix, and perform a residual connection between the output of the attention mechanism and the first feature matrix to obtain the second feature matrix.

[0082] Continue to refer to Figure 10 After applying the multi-head attention mechanism to the first feature matrix (denoted as C), the result of the attention mechanism is concatenated with the residual of the first feature matrix (denoted as C) to obtain a new feature matrix, which is denoted as the second feature matrix (denoted as D).

[0083] c5) Normalize and process the second feature matrix sequentially, and perform residual connection between the output of the multilayer perceptron and the second feature matrix to update the global temporal features.

[0084] Continue to refer to Figure 10 The second feature matrix (denoted as D) is sequentially subjected to layer normalization and Multilayer Perceptron (MLP) processing to obtain the multilayer perceptron output (denoted as E). Then, the second feature matrix (D) and the multilayer perceptron output (E) are residually concatenated to obtain a new feature matrix, denoted as the updated global temporal feature (F). Here, cls_token reuses the cls_token of X, [B,C,H,W]->[B,HW,C]->[B,HW+1,C]. It should be noted that due to the inherent characteristics of Transformer, the updated global temporal feature (denoted as F) and the global temporal feature (denoted as X) have the same matrix dimension, and their names are renamed to the new X (the original X is discarded). This characteristic allows X to be applied to multiple rounds of Transformer modules, i.e. Figure 2 *N in the middle.

[0085] It's important to note that a multi-head attention mechanism is employed, allowing each token in the sequence (representing a specific time segment within 3 seconds) to interact with all other tokens in the sequence. This captures the causal relationships between actions within that 3-second window, rather than simply observing local vibrational textures. Multiple heads are configured to observe the signal from multiple subspaces (each head represents a point of interest). Furthermore, a multi-layer perceptron layer is included, employing a high-magnification feature projection space design. Through coupling a two-layer linear mapping with a Gaussian Error Linear Unit (GELU) nonlinear activation function, the spatiotemporally fused feature vector is nonlinearly reconstructed positionally. Simultaneously, a random deactivation layer effectively suppresses model overfitting caused by complex environmental noise.

[0086] S208. The updated global temporal features and the updated high-dimensional semantic features are cyclically fed back to the second convolutional module and the deep learning module based on the attention mechanism to achieve mutual enhancement of global temporal features and high-dimensional semantic features in a set number of rounds, and output the final updated global temporal features and updated high-dimensional semantic features.

[0087] The value of the set rounds can be set according to the actual situation. In this embodiment, the updated global temporal features and updated high-dimensional semantic features are fed back to the second convolutional module and the deep learning module based on the attention mechanism. Steps S204 to S208 are then re-executed. After performing the set rounds of iteration, the final updated global temporal features and updated high-dimensional semantic features are output.

[0088] S209. Classify the updated global temporal features using the classification head of the first processing branch to obtain the first sound source event prediction probability of the test point.

[0089] In this embodiment, the classification head of the first processing branch performs layer normalization on the updated global temporal features, then extracts the feature vector (cls_token) at the first position, and maps it to the output value of the number of categories through a fully connected layer. The output value is then normalized to obtain the predicted probability distribution of each category of the sound source event, denoted as the predicted probability of the first sound source event.

[0090] S210. The updated high-dimensional semantic features are classified using the classification head of the second processing branch to obtain the second sound source event prediction probability of the test point.

[0091] In this embodiment, the high-dimensional semantic features are updated using global average pooling through the classification head of the second processing branch, then flattened into a one-dimensional vector. Subsequently, a fully connected layer maps the feature dimension to the output value of the number of categories. The output value is then normalized to obtain the predicted probability distribution of each category of the sound source event, denoted as the second sound source event prediction probability.

[0092] S211. Obtain the first accuracy weight of the first processing branch and the second accuracy weight of the second processing branch obtained from the pre-training of the sound source event classification verification set.

[0093] In this embodiment, the classification accuracy of the two processing branches during the pre-training phase on the sound source event classification validation set is obtained, and corresponding weights are constructed. The weight of the classification accuracy corresponding to the first processing branch is denoted as the first accuracy weight, and the weight of the classification accuracy corresponding to the second processing branch is denoted as the second accuracy weight. For example, it is assumed that the accuracy weight corresponding to the first processing branch is 80%, and the accuracy weight corresponding to the second processing branch is 85%.

[0094] S212. Based on the first accuracy weight and the second accuracy weight, the predicted probabilities of the first sound source event and the predicted probabilities of the second sound source event are weighted and summed to obtain the comprehensive evaluation score of the sound source event category of the test point.

[0095] This step employs a weighted decision fusion mechanism based on classification accuracy. By obtaining the accuracy weights of the two processing branches on the validation set, and multiplying them by the predicted probabilities of the sound source events output by each processing branch, the final probability distribution of each predicted category is obtained, denoted as the comprehensive evaluation score of the sound source event category at the test location. The formula for calculating the comprehensive evaluation score can be expressed as: ,in, This indicates the overall evaluation score. and These represent the predicted probabilities of sound source events in the two processing branches, respectively. and These represent the accuracy weights for event classification in the two processing branches, respectively.

[0096] By introducing accuracy-weighted fusion, the sensitivity of convolutional modules to static voiceprints and the ability of attention-based machine learning models to grasp dynamic logic are heterogeneously complementary, effectively making up for the shortcomings of single network branches in the identification of specific complex sound sources, and greatly reducing false alarms caused by transient environmental noise.

[0097] S213. Select the sound source event category corresponding to the maximum value in the comprehensive evaluation score as the initial classification result of the sound source event at the test point.

[0098] In this embodiment, the sound source event category corresponding to the maximum value in the comprehensive evaluation score is selected as the initial classification result of the sound source event at the test point. For example, assuming there are 7 categories of sound source events, the sound source event at the test point should have 7 categories of comprehensive evaluation scores. The category with the highest score among these 7 categories is taken as the initial classification result of the test point.

[0099] S214. When the initial classification result meets the preset alarm conditions, the test point is taken as the first alarm point, a spatiotemporal window is constructed with the spatial location of the first alarm point as the center, and a spatiotemporal waterfall diagram reflecting the change of vibration energy with time and space is generated based on the original vibration data of the time window.

[0100] The preset alarm conditions can be understood as the conditions that trigger the wide-area spatiotemporal trajectory recognition mechanism. These conditions include an initial classification result indicating a target event such as an excavator, electric pick, or manual striking, with a confidence level higher than a preset threshold. Specifically, when the initial classification result meets the preset alarm conditions, the wide-area spatiotemporal actual recognition process is immediately initiated. The point to be tested is taken as the first alarm point, and the initial classification result meeting the preset alarm conditions is taken as the first valid alarm signal output. A spatiotemporal window is formed by extending a set length forward and backward from the spatial position x0 of the first alarm point along the fiber optic path. Simultaneously, recording begins from the current alarm time. For example, if the set length is 1km, then the spatiotemporal window [x0-1000, x0+1000] is formed by extending a 1km range forward and backward from the spatial position x0 of the first alarm point.

[0101] Following the above description, within this spatiotemporal window, the vibration energy intensity of all spatial points over continuous time is extracted from the original vibration data. After detrending and bandpass filtering, a spatiotemporal waterfall plot is drawn, where the horizontal axis represents spatial distance (unit: meters), the vertical axis represents time (unit: seconds), and the pixel value represents the normalized vibration amplitude.

[0102] S215. Preprocess the spatiotemporal waterfall map and use a target detection model to detect the movement trajectory in the preprocessed spatiotemporal waterfall map.

[0103] In this embodiment, the spatiotemporal waterfall image is first preprocessed, including median filtering, Gaussian smoothing, signal enhancement, normalization, and resizing, to obtain the processed spatiotemporal waterfall image. Then, a target detection model is used to detect the movement trajectory in the processed spatiotemporal waterfall image. Specifically, median filtering applies a sliding window (adjusted according to spatial resolution) along the spatial axis of the spatiotemporal waterfall image, replacing the center pixel value with the window's median value; Gaussian smoothing applies a one-dimensional Gaussian kernel along the time axis to each spatial point in the spatiotemporal waterfall image, making the diagonal trajectory more continuous in the time direction and reducing breaks; signal enhancement uses the Sobel operator to convolve the spatiotemporal waterfall image, highlighting the linear trajectory formed by uniformly or quasi-uniformly moving sound sources; and normalization and resizing scale the enhanced spatiotemporal waterfall image to a fixed size required by the target detection model and normalizes the pixel values ​​to the [0,1] interval as input to the target detection model.

[0104] Specifically, the object detection model can be understood as a pre-trained machine learning model that detects movement trajectories on a spatiotemporal waterfall plot. The input to the object detection model is the spatiotemporal waterfall plot, and the output is the movement trajectory. For example, YOLOv8 can be used as the object detection model. The training process of the object detection model is illustrated using YOLOv8 as an example: 1) Label design: In the spatiotemporal waterfall plot, continuous diagonal regions generated by moving sound sources such as vehicles are labeled as axis-aligned bounding boxes (because the diagonal lines have a large span, the axis-aligned boxes are sufficient to cover them). The label includes the upper left corner coordinates (x_min, y_min) and lower right corner coordinates (x_max, y_max) of the bounding box, and the approximate slope k of the movement trajectory. 2) Randomly crop, rotate (±15°), perturb brightness / contrast, and simulate fractures (randomly erasing some line segments) on the real-world spatiotemporal waterfall map to improve the model's robustness to different signal-to-noise ratios and trajectory shapes; 3) Use YOLOv8's default CIOU bounding box loss + binary cross-entropy classification loss, with AdamW as the optimizer. CIOU (Complete Intersection over Union) loss is a high-level loss function used for bounding box regression in object detection. Convert the trained object detection model to a specified format, such as the inference acceleration format for TensorRT or RKNN deep learning models, and deploy it on edge computing devices.

[0105] In this embodiment, a pre-trained object detection model is used, taking the pre-processed spatiotemporal waterfall plot as input, and outputting the bounding boxes of all moving trajectories in the spatiotemporal waterfall plot and the approximate slope K of the trajectories. Post-processing is performed on the bounding boxes; for each retained detection box, detection boxes with abnormal absolute slope values ​​are filtered out using K. For example, a larger |K| indicates a faster object movement speed. If the movement speed exceeds a reasonable range, such as a stationary or ultra-high-speed false trajectory, it is discarded. Simultaneously, short trajectories with a bounding box height (representing the time span) less than a set number (e.g., 10) pixels are discarded to avoid false detections due to transient noise. Based on the bounding boxes and slope K, the moving trajectory is finally determined. For example, the list of verified trajectories can be represented as T = { (x_left, x_right, t_top, t_bottom, k, conf)}, where each element represents a complete moving trajectory region.

[0106] S216. Within a preset time window, project the subsequently generated alarm points onto the coordinate system of the spatiotemporal waterfall diagram.

[0107] In this embodiment, the preset time window can be set according to the actual situation. Specifically, within the preset time window, the subsequently generated alarm points are projected onto the coordinate system of the spatiotemporal waterfall diagram to achieve continuous alarm point monitoring and trajectory projection. During the operation of the time window, the alarm time of each newly arriving alarm target at that location is recorded, and the point is projected onto the spatiotemporal waterfall diagram to determine whether the alarm point is located on the trajectory.

[0108] S217. If multiple alarm points are located on multiple independent movement trajectories, the final classification result of the sound source event at the test point is determined to be a non-external damage event, and the alarm output is suppressed.

[0109] In this embodiment, after the first alarm point is triggered, the alarm fiber optic points within the spatiotemporal window are continuously identified in real time. If the point triggers multiple alarm events within a specified time, and the time of each alarm falls on multiple independent movement trajectory lines in the spatiotemporal waterfall diagram, it proves that the frequent alarms at this location are not caused by in-situ vibration, but rather by misjudgment due to the location's sensitivity to vibration, and are not external damage events. The system will not report this alarm sequence. For example, if three or more consecutive alarms are located on different movement trajectories in the spatiotemporal waterfall diagram, and the slope of the trajectory is 0.5 < v < 40, it is considered to be an interference source with displacement characteristics, rather than a target event of in-situ operation (such as an excavator or electric pick), thereby actively suppressing the alarm.

[0110] S218. Otherwise, determine the final classification result as an external damage event and output an alarm.

[0111] As described above, if the above conditions are not met, then the frequent alarms at this location are determined to be caused by in-situ vibrations, which is an external damage event, and alarm output is required.

[0112] The aforementioned technical solution employs a heterogeneous dual-branch feature parallel extraction approach. The model simultaneously constructs a one-dimensional original phase branch, captures transient pulses of different physical frequencies based on multi-scale convolution, and integrates with a two-dimensional spectral texture branch. The convolution module extracts the microscopic acoustic texture of the time-frequency map, effectively mapping the time-frequency map to a higher-dimensional space, achieving comprehensive coverage of the physical properties and visual features of the sound source signal. Simultaneously, a bidirectional interactive fusion mechanism is used. The downsampling module maps the local spectral texture captured by the convolution module into semantic features (Tokens) which are then injected into an attention-based deep learning model (Transformer). Meanwhile, the upsampling module feeds back the global temporal logic understood by the Transformer to the CNN. This closed-loop design, where "global semantics guides local details," effectively distinguishes complex events with similar spectra but vastly different operational rhythms. Furthermore, a decision fusion mechanism based on accuracy weighting is employed: the classification accuracy of the CNN branch and the Transformer branch on the validation set is obtained, and corresponding weights are constructed. A weighted summation is then used to calculate a comprehensive evaluation score, yielding the initial classification result. Furthermore, a rigid filtering logic based on kinematic laws was introduced to identify and eliminate traffic flow interference. By utilizing artificial intelligence technology and physical motion pattern constraints, multiple types of sound sources were monitored in real time and automatically identified and classified, achieving highly reliable monitoring for external damage incidents.

[0113] To more clearly illustrate the effectiveness of the external sound source monitoring method provided in this embodiment of the invention, existing technical solutions, such as ResNet18 (an 18-layer deep learning model for image recognition), VIT-B / 16 (a pre-trained model based on the visual Transformer), and the Transformer model (a deep learning model), are compared with this technical solution. For example, Table 1 shows the comparison results between the models used in the existing technical solutions and this technical solution: Table 1 The ResNet18, VIT-B / 16, and Transformer models are not the original models; their parameters were technically adjusted to adapt to the dataset. The dataset consists of a collection of historical alarm signals, divided into training and validation sets. Accuracy and false alarm rate were measured on the validation set. Inference time was measured using the Central Processing Unit (CPU) without using the Graphics Processing Unit (GPU).

[0114] As can be seen, this technical solution significantly improves classification accuracy and breaks through the bottleneck of complex sound source recognition. By introducing a multi-scale convolution module into the one-dimensional original signal branch and utilizing parallel short, medium, and long kernel convolution operators, it can simultaneously capture high-frequency transient impacts (such as electric picks) and low-frequency time-varying envelopes (such as excavators), effectively solving the problem of feature extraction difficulties for single-scale algorithms under complex sound sources. According to comparative experiments, the classification accuracy of this technical solution reaches 88.61% on the same test set, which has a significant performance advantage over traditional ResNet18 (84.43%), VIT-B / 16 (83.70%), and the pure Transformer model (81.46%), with an accuracy improvement ranging from 4.18% to 7.15%.

[0115] This technical solution significantly reduces the false alarm rate and enhances the system's anti-interference capability. Utilizing a bidirectional interactive fusion mechanism, a closed-loop design of "global semantics guiding local details" is achieved. The long-term action logic understood by the deep learning module based on the attention mechanism can guide the convolution module to semantically enhance weak spectral details, thereby effectively removing environmental noise. The false alarm rate of this technical solution is only 7.9%, far lower than ResNet18's 11.6% and Transformer's 16.8%. This indicates that the weighted processing of the data upsampling module can balance the contradiction between high sensitivity and low false alarm rate, greatly reducing false alarms caused by transient environmental noise. To address residual false alarms in the neural network (such as coherent interference from passing vehicles), the system introduces a wide-area spatiotemporal trajectory recognition mechanism. When the neural network detects alarm points on the same trajectory N times consecutively in the spatiotemporal waterfall plot, and the trajectory exhibits physical movement characteristics (the slope of the straight line matches the speed), the system determines that the sequence signal is an interference source with displacement characteristics (such as cars or pedestrians), rather than a stationary target event (such as excavators or electric picks), thus actively suppressing alarms.

[0116] This technical solution balances edge deployment requirements with real-time inference performance. It employs a bottleneck-based residual convolutional unit (second convolutional module) and a feature compression strategy (data downsampling module), significantly reducing redundant feature map computation while maintaining high-dimensional semantic extraction capabilities. Pruning and quantization can be performed after training to meet the real-time inference needs of edge devices in long-distance inspection tasks. A single inference time of only 30.3ms ensures that devices can process and issue alarms in real-time during long-distance inspection tasks, avoiding missed events due to computational latency.

[0117] Example 3 Figure 11This is a schematic diagram of a sound source protection monitoring system for external damage incidents provided in Embodiment 3 of the present invention. This device is applicable to situations requiring monitoring of external damage sound source incidents. The sound source protection monitoring system can be implemented in hardware and / or software and is generally integrated into electronic devices. For example... Figure 11 As shown, the device includes: a preprocessing module 31, a feature extraction module 32, a feature fusion module 33, an initial classification module 34, and a final classification module 35, wherein... Preprocessing module 31 is used to acquire the original vibration data of the test point collected by the distributed acoustic wave sensing device, preprocess the original vibration data to obtain a one-dimensional vibration signal and a two-dimensional time-frequency diagram. The feature extraction module 32 is used to process the one-dimensional vibration signal through the first processing branch in the heterogeneous dual-branch recognition network to extract the global temporal features of the one-dimensional vibration signal, and to process the two-dimensional time-frequency graph through the second processing branch in the heterogeneous dual-branch recognition network to extract the high-dimensional semantic features of the two-dimensional time-frequency graph. The feature fusion module 33 is used to inject the high-dimensional semantic features into the first processing branch and perform global temporal logic modeling with the global temporal features to obtain updated global temporal features for global temporal modeling, and to inject the updated global temporal features back into the second processing branch and fuse them with the high-dimensional semantic features to obtain updated high-dimensional semantic features with enhanced semantics. The initial classification module 34 is used to identify the first sound source event prediction probability of the test point based on the updated global temporal features and the second sound source event prediction probability of the test point based on the updated high-dimensional semantic features, and to determine the initial classification result of the sound source event of the test point based on the first sound source event prediction probability and the second sound source event prediction probability. The final classification module 35 is used to construct a spatiotemporal waterfall diagram of the test point when the initial classification result meets the preset alarm conditions, determine the final classification result of the sound source event of the test point according to the movement trajectory in the spatiotemporal waterfall diagram, and execute an alarm decision.

[0118] The aforementioned technical solution employs a heterogeneous dual-branch recognition network. One processing branch processes the one-dimensional vibration signal to capture transient pulses at different physical frequencies, while the other processing branch processes the two-dimensional time-frequency map to capture microscopic acoustic patterns, achieving comprehensive coverage of the physical properties and visual features of the sound source signal. Simultaneously, a bidirectional interactive fusion mechanism is used, injecting the high-dimensional semantic features captured by the second processing branch into the first processing branch, and feeding back the global temporal logic understood by the first processing branch to the second processing branch. This closed-loop design, "global semantics guiding local details," effectively distinguishes complex sound source events with similar spectra but vastly different operating rhythms. Furthermore, a decision fusion mechanism based on the accuracy weighting of different processing branches determines the initial classification result. Then, based on this, the category of the sound source event is verified by combining spatiotemporal trajectory verification to determine the final classification result. This distributed acoustic sensing sound source event real-time discrimination method, based on the combination of artificial intelligence neural networks and spatiotemporal trajectory verification, no longer relies on single time-segment acoustic features. Instead, it constructs a three-in-one recognition architecture of "physical features - acoustic texture - spatiotemporal trajectory" to achieve highly reliable discrimination of target events. It significantly improves the accuracy of sound source event classification, breaks through the bottleneck of complex sound source identification, greatly reduces the false alarm rate, and enhances the system's anti-interference capability; and takes into account both edge deployment requirements and real-time inference performance.

[0119] Optionally, the feature extraction module 32 includes a first extraction unit, which is used for: In the first processing branch, the one-dimensional vibration signal is subjected to global temporal feature extraction through a multi-scale convolution module to generate global temporal features that meet the input requirements of the deep learning module based on the attention mechanism.

[0120] Optionally, the first extraction unit is specifically used for: The high-frequency impact pulse and low-frequency envelope variation features in the one-dimensional vibration signal are captured by at least three sets of one-dimensional convolution kernels configured in parallel in the multi-scale convolution module and spliced ​​into multi-scale features. The multi-scale features are sequentially normalized, activated, pooled, and adjusted in dimension to obtain the processed physical features. Position encoding is added to the physical features to compensate for the lack of time-axis logic awareness in the attention mechanism, thereby generating global temporal features that meet the input requirements of deep learning modules based on the attention mechanism.

[0121] Optionally, the feature extraction module 32 includes a second extraction unit, which is used for: In the second processing branch, the two-dimensional time-frequency map is expanded from the initial number of channels to a high-dimensional semantic space by the first convolution module, and normalization, activation and pooling processes are performed in sequence to generate the high-dimensional semantic features of the two-dimensional time-frequency map.

[0122] Optionally, the feature fusion module 33 includes: The upsampling unit is used to input the global temporal features into the data upsampling module and generate upsampled output features; The first processing unit is used to input the high-dimensional semantic features into the second convolution module of the second processing branch to obtain the first refined features and the second refined features, and to fuse the second refined features with the upsampled output features to generate the updated high-dimensional semantic features; The downsampling unit is used to take the first refined feature input data downsampling module and generate downsampling output features; The second processing unit is used to input the global temporal features and the downsampled output features into the deep learning module based on the attention mechanism of the first processing branch, and output the updated global temporal features. The loop unit is used to cyclically feed back the updated global temporal features and the updated high-dimensional semantic features to the second convolutional module and the deep learning module based on the attention mechanism, so as to realize the mutual enhancement of global temporal features and high-dimensional semantic features in a set number of rounds, and output the final updated global temporal features and updated high-dimensional semantic features.

[0123] Optionally, the first processing unit is specifically used for: The high-dimensional semantic features are sequentially input into the first standard convolutional block and the second standard convolutional block of the second processing branch to obtain the first refined features; After performing convolution and normalization on the first refined feature in sequence, a residual connection is made with the high-dimensional semantic feature and activated to generate the second refined feature; The second refined feature is fused with the upsampled output feature to obtain the updated high-dimensional semantic feature.

[0124] Optionally, the second processing unit is specifically used for: The global temporal features are added to the downsampled output features, and normalization is performed to obtain the first feature matrix; A multi-head attention mechanism is applied to the first feature matrix, and the output of the attention mechanism is residually concatenated with the first feature matrix to obtain the second feature matrix. The second feature matrix is ​​then subjected to normalization and multilayer perceptron processing in sequence. The output of the multilayer perceptron is then residually concatenated with the second feature matrix to update the global temporal features.

[0125] Optionally, the initial classification module 34 is specifically used for: The updated global temporal features are classified using the classification head of the first processing branch to obtain the predicted probability of the first sound source event at the test point. The updated high-dimensional semantic features are classified by the classification head of the second processing branch to obtain the second sound source event prediction probability of the test point; Obtain the first accuracy weight of the first processing branch and the second accuracy weight of the second processing branch obtained from pre-training based on the sound source event classification verification set; Based on the first accuracy weight and the second accuracy weight, the predicted probabilities of the first and second sound source events are weighted and summed to obtain a comprehensive evaluation score for the sound source event category of the test point. The sound source event category corresponding to the maximum value in the comprehensive evaluation score is selected as the initial classification result of the sound source event at the test point.

[0126] Optionally, the final classification module 35 is specifically used for: The test point is taken as the first alarm point. A spatiotemporal window is constructed with the spatial location of the first alarm point as the center. Based on the original vibration data of the time window, a spatiotemporal waterfall diagram reflecting the change of vibration energy with time and space is generated. The spatiotemporal waterfall plot is preprocessed, and the movement trajectory in the preprocessed spatiotemporal waterfall plot is detected using a target detection model. Within a preset time window, the subsequently generated alarm points are projected onto the coordinate system of the spatiotemporal waterfall diagram; If multiple consecutive alarm points are located on multiple independent movement trajectories, the final classification result of the sound source event at the test point is determined to be a non-external damage event, and the alarm output is suppressed. Otherwise, the final classification result is determined to be an external damage event, and an alarm is output.

[0127] The external damage sound source monitoring system provided in this embodiment of the invention can execute the external damage sound source monitoring method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0128] Example 4 Figure 12This is a schematic diagram of an electronic device provided in Embodiment 4 of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0129] like Figure 12 As shown, the electronic device 40 includes at least one processor 41 and a memory, such as a read-only memory (ROM) 42 or a random access memory (RAM) 43, communicatively connected to the at least one processor 41. The memory stores computer programs executable by the at least one processor. The processor 41 can perform various appropriate actions and processes based on the computer program stored in the ROM 42 or loaded from storage unit 48 into the RAM 43. The RAM 43 may also store various programs and data required for the operation of the electronic device 40. The processor 41, ROM 42, and RAM 43 are interconnected via a bus 44. An input / output (I / O) interface 45 is also connected to the bus 44.

[0130] Multiple components in electronic device 40 are connected to I / O interface 45, including: input unit 46, such as keyboard, mouse, etc.; output unit 47, such as various types of monitors, speakers, etc.; storage unit 48, such as disk, optical disk, etc.; and communication unit 49, such as network card, modem, wireless transceiver, etc. Communication unit 49 allows electronic device 40 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0131] Processor 41 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 41 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 41 performs the various methods and processes described above, such as methods for detecting external sound source incidents.

[0132] In some embodiments, the external damage sound source monitoring method can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 48. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 40 via ROM 42 and / or communication unit 49. When the computer program is loaded into RAM 43 and executed by processor 41, one or more steps of the external damage sound source monitoring method described above can be performed. Alternatively, in other embodiments, processor 41 can be configured to perform the external damage sound source monitoring method by any other suitable means (e.g., by means of firmware).

[0133] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0134] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0135] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0136] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0137] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0138] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0139] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the method for monitoring external sound source events as provided in any embodiment of this invention.

[0140] In implementing a computer program product, computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0141] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0142] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for monitoring external noise source incidents, characterized in that, include: The raw vibration data of the test point collected by the distributed acoustic wave sensing device is obtained, and the raw vibration data is preprocessed to obtain a one-dimensional vibration signal and a two-dimensional time-frequency diagram. The one-dimensional vibration signal is processed by the first processing branch in the heterogeneous dual-branch recognition network to extract the global temporal features of the one-dimensional vibration signal, and the two-dimensional time-frequency graph is processed by the second processing branch in the heterogeneous dual-branch recognition network to extract the high-dimensional semantic features of the two-dimensional time-frequency graph. The high-dimensional semantic features are injected into the first processing branch and combined with the global temporal features to perform global temporal logic modeling, thereby obtaining updated global temporal features for global temporal modeling. The updated global temporal features are then injected back into the second processing branch and fused with the high-dimensional semantic features to obtain semantically enhanced updated high-dimensional semantic features. The prediction probability of the first sound source event at the test point is identified based on the updated global temporal features, and the prediction probability of the second sound source event at the test point is identified based on the updated high-dimensional semantic features. Based on the prediction probability of the first sound source event and the prediction probability of the second sound source event, the initial classification result of the sound source event at the test point is determined. When the initial classification result meets the preset alarm conditions, a spatiotemporal waterfall diagram of the test point is constructed, and the final classification result of the sound source event of the test point is determined according to the movement trajectory in the spatiotemporal waterfall diagram, and an alarm decision is executed.

2. The method according to claim 1, characterized in that, The step of processing the one-dimensional vibration signal through the first processing branch of the heterogeneous dual-branch recognition network to extract the global temporal features of the one-dimensional vibration signal includes: In the first processing branch, the one-dimensional vibration signal is subjected to global temporal feature extraction through a multi-scale convolution module to generate global temporal features that meet the input requirements of the deep learning module based on the attention mechanism.

3. The method according to claim 2, characterized in that, The step of extracting global temporal features from the one-dimensional vibration signal using a multi-scale convolution module to generate global temporal features that meet the input requirements of a deep learning module based on an attention mechanism includes: The high-frequency impact pulse and low-frequency envelope variation features in the one-dimensional vibration signal are captured by at least three sets of one-dimensional convolution kernels configured in parallel in the multi-scale convolution module and spliced ​​into multi-scale features. The multi-scale features are sequentially normalized, activated, pooled, and adjusted in dimension to obtain the processed physical features. Position encoding is added to the physical features to compensate for the lack of time-axis logic awareness in the attention mechanism, thereby generating global temporal features that meet the input requirements of deep learning modules based on the attention mechanism.

4. The method according to claim 1, characterized in that, The step of processing the two-dimensional time-frequency graph through the second processing branch of the heterogeneous dual-branch recognition network to extract high-dimensional semantic features of the two-dimensional time-frequency graph includes: In the second processing branch, the two-dimensional time-frequency map is expanded from the initial number of channels to a high-dimensional semantic space by the first convolution module, and normalization, activation and pooling processes are performed in sequence to generate the high-dimensional semantic features of the two-dimensional time-frequency map.

5. The method according to claim 1, characterized in that, The step of injecting the high-dimensional semantic features into the first processing branch and performing global temporal logic modeling with the global temporal features to obtain updated global temporal features for global temporal modeling, and then injecting the updated global temporal features back into the second processing branch and fusing them with the high-dimensional semantic features to obtain semantically enhanced updated high-dimensional semantic features, includes: The global temporal features are input into the data upsampling module to generate upsampled output features; The high-dimensional semantic features are input into the second convolutional module of the second processing branch to obtain the first refined features and the second refined features. The second refined features are then fused with the upsampled output features to generate the updated high-dimensional semantic features. The first refined feature is input into the data downsampling module to generate the downsampled output feature; The global temporal features and the downsampled output features are input into the deep learning module based on the attention mechanism in the first processing branch, and the updated global temporal features are output. The updated global temporal features and the updated high-dimensional semantic features are cyclically fed back to the second convolutional module and the deep learning module based on the attention mechanism, so as to achieve mutual enhancement of global temporal features and high-dimensional semantic features in a set number of rounds, and output the final updated global temporal features and updated high-dimensional semantic features.

6. The method according to claim 5, characterized in that, The step of inputting the high-dimensional semantic features into the second convolutional module of the second processing branch to obtain the first refined feature and the second refined feature, and fusing the second refined feature with the upsampled output feature to generate the updated high-dimensional semantic features includes: The high-dimensional semantic features are sequentially input into the first standard convolutional block and the second standard convolutional block of the second processing branch to obtain the first refined features; After performing convolution and normalization on the first refined feature in sequence, a residual connection is made with the high-dimensional semantic feature and activated to generate the second refined feature; The second refined feature is fused with the upsampled output feature to obtain the updated high-dimensional semantic feature.

7. The method according to claim 5, characterized in that, The step of inputting the global temporal features and the downsampled output features into the attention-based deep learning module of the first processing branch, and outputting the updated global temporal features, includes: The global temporal features are added to the downsampled output features, and normalization is performed to obtain the first feature matrix; A multi-head attention mechanism is applied to the first feature matrix, and the output of the attention mechanism is residually concatenated with the first feature matrix to obtain the second feature matrix. The second feature matrix is ​​then subjected to normalization and multilayer perceptron processing in sequence. The output of the multilayer perceptron is then residually concatenated with the second feature matrix to update the global temporal features.

8. The method according to claim 1, characterized in that, The step of identifying the first sound source event prediction probability of the test point based on the updated global temporal features and the second sound source event prediction probability of the test point based on the updated high-dimensional semantic features, and determining the initial classification result of the sound source event of the test point based on the first and second sound source event prediction probabilities, includes: The updated global temporal features are classified using the classification head of the first processing branch to obtain the predicted probability of the first sound source event at the test point. The updated high-dimensional semantic features are classified by the classification head of the second processing branch to obtain the second sound source event prediction probability of the test point; Obtain the first accuracy weight of the first processing branch and the second accuracy weight of the second processing branch obtained from pre-training based on the sound source event classification verification set; Based on the first accuracy weight and the second accuracy weight, the predicted probabilities of the first and second sound source events are weighted and summed to obtain a comprehensive evaluation score for the sound source event category of the test point. The sound source event category corresponding to the maximum value in the comprehensive evaluation score is selected as the initial classification result of the sound source event at the test point.

9. The method according to claim 1, characterized in that, The process of constructing a spatiotemporal waterfall diagram of the test point, determining the final classification result of the sound source events at the test point based on the spatiotemporal waterfall diagram, and executing an alarm decision includes: The test point is taken as the first alarm point. A spatiotemporal window is constructed with the spatial location of the first alarm point as the center. Based on the original vibration data of the time window, a spatiotemporal waterfall diagram reflecting the change of vibration energy with time and space is generated. The spatiotemporal waterfall plot is preprocessed, and the movement trajectory in the preprocessed spatiotemporal waterfall plot is detected using a target detection model. Within a preset time window, the subsequently generated alarm points are projected onto the coordinate system of the spatiotemporal waterfall diagram; If multiple consecutive alarm points are located on multiple independent movement trajectories, the final classification result of the sound source event at the test point is determined to be a non-external damage event, and the alarm output is suppressed. Otherwise, the final classification result is determined to be an external damage event, and an alarm is output.

10. A system for monitoring external noise sources, characterized in that, include: The preprocessing module is used to acquire the raw vibration data of the test point collected by the distributed acoustic wave sensing device, and to preprocess the raw vibration data to obtain a one-dimensional vibration signal and a two-dimensional time-frequency diagram. The feature extraction module is used to process the one-dimensional vibration signal through the first processing branch in the heterogeneous dual-branch recognition network to extract the global temporal features of the one-dimensional vibration signal, and to process the two-dimensional time-frequency graph through the second processing branch in the heterogeneous dual-branch recognition network to extract the high-dimensional semantic features of the two-dimensional time-frequency graph. The feature fusion module is used to inject the high-dimensional semantic features into the first processing branch and perform global temporal logic modeling with the global temporal features to obtain updated global temporal features for global temporal modeling, and then inject the updated global temporal features back into the second processing branch and fuse them with the high-dimensional semantic features to obtain updated high-dimensional semantic features with enhanced semantics. The initial classification module is used to identify the first sound source event prediction probability of the test point based on the updated global temporal features and the second sound source event prediction probability of the test point based on the updated high-dimensional semantic features, and to determine the initial classification result of the sound source event of the test point based on the first sound source event prediction probability and the second sound source event prediction probability. The final classification module is used to construct a spatiotemporal waterfall diagram of the test point when the initial classification result meets the preset alarm conditions, determine the final classification result of the sound source event of the test point according to the movement trajectory in the spatiotemporal waterfall diagram, and execute an alarm decision.

Images