Railway rockfall intrusion limit falling point material granularity identification method and system

Abstract

The present application belongs to the field of railway safety monitoring and distributed optical fiber sensing technology, and relates to a railway rockfall intrusion limit drop point material granularity identification method and system. The method comprises high-recall extraction of vibration events based on window-level features of XGBoost; rockfall and non-rockfall binary classification discrimination based on event-level statistical features of XGBoost; drop point material granularity classification based on a deep residual network; and optimization training using an AdamW optimizer. The present application realizes fine monitoring, improves railway rockfall monitoring from single event detection to automatic identification of drop point material granularity; has strong anti-interference ability, significantly improves the engineering usability of the system in complex railway site environments; has high identification accuracy and deep physical feature mining; the introduced ResNet deep network fully utilizes the advantages of residual connection in retaining shallow high-frequency impact details and extracting deep semantic features; the algorithm is lightweight and highly matched, realizes smooth transition from low-dimensional statistical features to high-dimensional physical textures, and has strong overall identification chain stability.

Description

Technical Field

[0001] This invention belongs to the field of railway safety monitoring and distributed optical fiber sensing technology, specifically involving a fine-grained identification method and system for the material of railway rockfall intrusion points based on a phase-sensitive optical time-domain reflectometer and cascaded deep learning. Background Technology

[0002] Foreign object intrusion into railway tracks is a significant threat to train operation safety. Currently, railway monitoring methods based on distributed fiber optic sensing technology using phase-sensitive optical time-domain reflectometry (Φ-OTDR) are mostly limited to coarse-grained event-level classification, only able to distinguish whether vibration signals belong to "falling rocks," "pedestrians," or "trains." However, in actual railway maintenance, the material of the impact point of falling rocks, such as rails, sleepers, ballast, or the ground, is the core basis for determining the accident risk level, and the corresponding handling strategies differ significantly depending on the material.

[0003] Traditional machine learning methods for rockfall identification typically require filtering, noise reduction, and statistical feature compression extraction of the original vibration signal. While removing noise, this process loses a significant amount of fine-grained discriminative features closely related to material properties. Consequently, traditional methods can only identify the occurrence of rockfalls but cannot distinguish the specific material of the impact. Furthermore, conventional shallow neural networks or convolutional neural networks, lacking deep feature extraction capabilities, are prone to gradient degradation during training, making it difficult to effectively capture subtle differences in the time-frequency domain between different materials. Therefore, developing an effective method for automatic, fine-grained identification of the material at the impact point of railway rocks is a core technical challenge that urgently needs to be addressed in the field of railway safety monitoring. Summary of the Invention

[0004] The purpose of this invention is to provide a fine-grained method for identifying the material of railway rockfall intrusion points based on Φ-OTDR and cascaded deep learning. This addresses the problems of existing railway rockfall monitoring methods, which can only identify rockfall events but cannot distinguish the material of the landing point, and the difficulty of traditional machine learning methods in capturing weak material signal features. This invention elevates railway rockfall monitoring from a simple event detection level to a fine-grained classification level for automatic material type identification, providing a quantitative basis for railway risk level assessment.

[0005] This invention is achieved through the following technical solution:

[0006] A method for fine-grained identification of the material at the point of impact of falling rocks on railway tracks includes the following steps:

[0007] S1. High recall extraction of vibration events based on window-level features of XGBoost:

[0008] A Φ-OTDR distributed optical fiber sensing system was used to collect vibration signal data under various scenarios, and a two-dimensional spatial-temporal vibration response matrix along the railway direction was obtained. The spatial channel with the largest energy response and the strongest spatial correlation was selected as the main analysis channel. The first-order difference signal of the channel was divided into a sliding window of length L, and multiple discriminant features of each window were extracted. These features are the features after the first-order difference of the original data. The above features were input into the first XGBoost classifier for binary classification. Windows with a probability greater than a set threshold were merged into candidate vibration event segments.

[0009] S2. Binary classification of rockfall and non-rockfall based on XGBoost event-level statistical features:

[0010] Using the candidate vibration event segment output in step S1 as input, extract five event-level raw data statistical features from the entire candidate event signal, input them into the second XGBoost classifier, complete the binary classification of the candidate vibration event, and determine the candidate event as a rockfall event or a non-rockfall event.

[0011] S3. Fine-grained classification of landing point materials based on deep residual networks:

[0012] The original spatiotemporal two-dimensional response matrix of the rockfall event determined in step S2 is used as the model input. Without feature compression and dimensionality reduction, it is directly input into the pre-trained deep residual network. The deep residual network contains an initial convolutional layer and four residual blocks connected in sequence, which can automatically extract deep discriminative features corresponding to different materials from the two-dimensional matrix. After passing through a fully connected layer, it is connected to the Softmax classifier, and the Softmax function outputs the recognition probability corresponding to each of the four types of landing point materials, thus completing the material classification of the rockfall point.

[0013] S4. Use the AdamW optimizer for training optimization.

[0014] Further, in step S1, the vibration signal data under multiple scenarios covers vibration signal data under five scenarios: rockfall, knocking, walking, train operation and background noise; the discrimination features are the features after the first-order difference of the original data, including the energy proportion, main frequency of the signal and waveform factor of the three frequency bands of 100-200Hz, 200-300Hz and 300-400Hz.

[0015] Further, in step S1, the matrix data type is floating-point, and a single original data file can be represented as a two-dimensional spatial-temporal matrix:

[0016] ;

[0017] Where Q is the total number of spatial sampling points, U is the total number of pulses emitted by the system during the sampling time, xq,u represents the Φ-OTDR amplitude response of the q-th spatial sampling point at the u-th sampling time, and Q=512;

[0018] For a fixed spatial position q, in the q-th row, ... express, This represents the vibration waveform of the spatial channel as it changes over time.

[0019] For a fixed sampling time u, the u-th column is... express, ;

[0020] Spatial response analysis and main channel selection: Channel 138 is selected as the one-dimensional main analysis signal for steps S1 and S2.

[0021] Further, in step S2, the five event-level statistical features include: the standard deviation of the wavelet third-level detail coefficients, the energy of the wavelet second-level detail coefficients, the kurtosis of the wavelet second-level detail coefficients, the total signal energy, and the peak-to-peak value of the signal.

[0022] Further, in step S2, the types of interference include background noise, knocking, pedestrian walking, and train operation that were misjudged as vibration events in step S1.

[0023] Further, in step S3, the size of the input matrix is ​​512 spatial channels multiplied by 120 sampling time points, and it is subjected to amplitude normalization processing.

[0024] Further, in step S3, a deep residual network is used as the classifier for step S3. The network structure is as follows: first, the input passes through a 3×3 convolutional layer to map to 32 channels; then, it passes through four residual blocks in sequence, expanding the number of channels to 64, 128, 256, and 512 respectively; spatial downsampling is achieved between residual blocks through convolutions with a stride of 2; identity mapping within residual blocks is achieved through residual connections; when the number of channels or spatial size of the input feature map and the output feature map are inconsistent, dimension matching is performed through projection transformation; at the end of the network, an adaptive average pooling layer compresses the feature map into a 512×4×4 feature vector.

[0025] Further, in step S3, the four types of landing materials include the ground, ballast, sleepers, or rails.

[0026] Further, in step S4, the initial learning rate for the training phase is set to 0.001, and the weight decay coefficient is set to 1×10. -4The learning rate scheduling adopts a Warmup+Cosine strategy, with linear warming in the first 10 training epochs and cosine decay to zero in the next 90 training epochs. During training, to enhance the model's robustness to noise, time shift, and sample differences, three data augmentation strategies are added: the first is to add Gaussian white noise with an amplitude of 0.02 to the input matrix with a 50% probability; the second is to perform a random time shift of ±5 sampling points on the input matrix with a 30% probability; and the third is to use a Mixup strategy with a mixing parameter α set to 0.2, which reduces the model's tendency to overfit and memorize training samples by performing a convex combination of samples and labels.

[0027] A fine-grained identification system for the material of railway rockfall intrusion points is disclosed. This system employs a Φ-OTDR distributed optical fiber railway rockfall monitoring system, comprising a recall and extraction module, a rockfall feature identification module, a fine-grained material identification module, and a rockfall monitoring module. The recall and extraction module processes collected multi-scenario railway vibration signal data. Based on a spatial-temporal two-dimensional vibration response matrix constructed along the railway alignment, it selects the spatial channel with the largest energy response and strongest spatial correlation as the main analysis channel. It then performs a sliding window segmentation of length L on the first-order difference signal of this channel, extracting multiple discriminative features constructed based on the first-order difference of the original data within each window. These features are then fed into a first XGBoost binary classifier for discrimination. Windows with predicted probabilities higher than a set threshold are merged and output to obtain candidate vibration event segments. The rockfall feature identification module is used to... Five event-level raw data statistical features are extracted from the entire candidate event signal and input into the second XGBoost classifier to complete the binary classification of candidate vibration events, determining whether the candidate events are rockfall events or non-rockfall events. The rockfall monitoring module is used to select the raw spatiotemporal two-dimensional response matrix determined to be a rockfall event and supply it to the fine-grained material recognition module as input. The fine-grained material recognition module is used to receive the raw spatiotemporal two-dimensional response matrix output by the rockfall monitoring module as model input, without feature compression and dimensionality reduction, and directly input it into the pre-trained deep residual network. The deep residual network contains an initial convolutional layer and four sequentially connected residual blocks, which can automatically extract deep discriminative features corresponding to different materials from the two-dimensional matrix. The features are connected to the Softmax classifier through a fully connected layer, and the Softmax function outputs the recognition probability corresponding to each of the four types of landing point materials, completing the material classification of the rockfall point.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] 1. This invention achieves refined monitoring: it elevates railway rockfall monitoring from single event detection to fine-grained automatic identification of the material at the point of impact, which can directly provide the engineering department with a scientific basis for risk level assessment;

[0030] 2. Strong anti-interference capability: The XGBoost binary classification model, based on macroscopic statistical features, achieves absolute zero false alarms against strong periodic interferences such as train operation on site, significantly improving the engineering availability of the system in complex railway site environments;

[0031] 3. High recognition accuracy and in-depth physical feature mining: The introduced ResNet deep network fully leverages the advantages of residual connections in preserving shallow high-frequency impact details and extracting deep semantic features. For material responses with highly similar waveforms, it achieves an event-level material classification accuracy of 98.92%, and the end-to-end cascaded effective recognition rate reaches 93.92%.

[0032] 4. Lightweight algorithm and high matching performance: XGBoost is used to complete the first two stages of coarse screening. By leveraging its efficient processing capability of low-dimensional structured features and combining it with the high-dimensional matrix processing capability of ResNet, a smooth transition from low-dimensional statistical features to high-dimensional physical texture is achieved, resulting in strong overall recognition chain stability. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a structural diagram of a Φ-OTDR distributed optical fiber railway rockfall monitoring system.

[0035] Figure 2 Spatiotemporal distribution of the columnar differential signal for a rockfall event;

[0036] Figure 3 This is a magnified view of the original signal waveform of channel 138 in the rockfall event.

[0037] Figure 4 A 3D waterfall diagram within a 100-180 range of the rockfall event path;

[0038] Figure 5 A three-dimensional energy surface for a rockfall event;

[0039] Figure 6 The differential signal spatiotemporal heatmap is shown within a 0.4s window.

[0040] Figure 7 This is a distribution diagram of the cosine similarity between channel 138 and all 512 spatial channels;

[0041] Figure 8This is a diagram illustrating the overall framework of the three-level cascaded identification method.

[0042] Figure 9 The average time-domain waveform of the original signal within 1 second for five typical vibration scenarios;

[0043] Figure 10 The average amplitude spectrum of 0-500Hz after DC removal for five typical vibration scenarios;

[0044] Figure 11 Box plot for feature selection in step S1, which shows the energy proportion of the differential 100-200Hz frequency band.

[0045] Figure 12 Box plot for feature screening of energy proportion in differential 200-300Hz frequency band step S1;

[0046] Figure 13 Box plot for feature screening of energy proportion in differential 300-400Hz frequency band step S1;

[0047] Figure 14 Box plot for feature selection in differential principal frequency step S1;

[0048] Figure 15 Box plot for feature selection in step S1 of differential waveform factorization;

[0049] Figure 16 The non-train file confusion matrix for step S1;

[0050] Figure 17 The train file obfuscation matrix for step S1;

[0051] Figure 18 Box plot for feature filtering in the S2 step of the third level detail standard deviation of wavelet;

[0052] Figure 19 Box plot for feature filtering in wavelet second-level detail energy step S2;

[0053] Figure 20 Box plots are used to filter features in the S2 step of the wavelet second-level detail kurtosis step.

[0054] Figure 21 Box plot for feature screening in step S2 of total signal energy;

[0055] Figure 22 Box plot for peak-to-peak value step S2 feature selection;

[0056] Figure 23 The confusion matrix diagram for step S2;

[0057] Figure 24 Thermal diagram of the spatiotemporal response to ground rockfall impact;

[0058] Figure 25 Thermal diagram of the spatiotemporal response to the impact of falling rocks on ballast;

[0059] Figure 26 Thermal diagram of the spatiotemporal response to rockfall impact on railway sleepers;

[0060] Figure 27 Thermal diagram of the spatiotemporal response to rockfall impact on rails;

[0061] Figure 28 This is a diagram of the ResNet network architecture in step S3;

[0062] Figure 29 The image shows the ablation experiment results in step S3;

[0063] Figure 30 A diagram showing the sample flow and error propagation in a three-level cascaded link;

[0064] Figure 31 A flowchart illustrating the steps of a method for fine-grained identification of the material at the point of impact of falling rocks on railway tracks. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0066] like Figure 31 As shown, the present invention provides a method for fine-grained identification of the material of the impact point of a falling rock on a railway, comprising the following steps:

[0067] S1. High recall extraction of vibration events based on XGBoost window-level features to complete the monitoring and extraction of vibration event segments.

[0068] A Φ-OTDR distributed fiber optic sensing system was used to collect vibration signal data covering five scenarios: rockfall, impact, walking, train operation, and background noise, to obtain a two-dimensional spatial-temporal vibration response matrix along the railway line. The spatial channel with the largest energy response and the strongest spatial correlation was selected as the main analysis channel. A sliding window was used to segment the first-order difference signal of the main analysis signal, with a window length L set to multiple sampling points. Multiple discriminant features were extracted from each window, which were the features after the first-order difference of the original data. The discriminant features included the energy proportion, main signal frequency, and waveform factor of three frequency bands: 100-200Hz, 200-300Hz, and 300-400Hz. The above features were input into a first XGBoost classifier for binary classification, and windows with output probabilities greater than a set threshold were merged into candidate vibration event segments.

[0069] The matrix data type is floating-point, and a single raw data file can be represented as a two-dimensional space-time matrix:

[0070] ;

[0071] Where Q is the total number of spatial sampling points, U is the total number of pulses emitted by the system during the sampling time, xq,u represents the Φ-OTDR amplitude response of the q-th spatial sampling point at the u-th sampling time, and Q=512;

[0072] For a fixed spatial position q, in the q-th row, ... express, , representing the vibration waveform of the spatial channel as it changes over time.

[0073] For a fixed sampling time u, the u-th column is... express, .

[0074] Spatial response analysis and main channel selection: 138 spatial channels were selected as the main analysis signals. The 138th channel signal was analyzed. Perform first-order difference:

[0075] .

[0076] The Φ-OTDR system possesses distributed spatial sensing capabilities. When an external impact acts near the optical fiber, the local Rayleigh backscattered signal will be significantly disturbed. Since the vibration energy propagates into the adjacent space through the soil, track structure, and optical fiber coupling medium, the event response usually does not appear only in a single spatial channel, but rather produces similar changes in multiple adjacent channels near the event center, and the response intensity gradually decreases from the event center outwards.

[0077] To determine the primary response locations of the rockfall event in the experiment of this invention, firstly, the energy of each spatial channel within the event time range is calculated for the difference matrix D:

[0078] ;

[0079] in, Let represent the differential energy of the i-th spatial channel, where [t1, t2] is the time interval of the event. The spatial channel with the highest energy is defined as:

[0080] ;

[0081] Taking a typical rockfall impact event as an example, within this window, the vibration energy exhibits significant spatial concentration, mainly concentrated near channel 138, and rapidly attenuates to both sides. Figures 4-5 As shown.

[0082] To further verify the representativeness of channel 138, the cosine similarity between channel 138 and other spatial channels can be calculated, and the results are as follows: Figures 6-7 As shown:

[0083] ;

[0084] in, This indicates the degree of similarity between the waveforms of the i-th spatial channel and the 138th channel during the event time period.

[0085] The self-similarity of channel 138 is 1.000, while the similarity of neighboring channels 137 to 141 is all above 0.8, indicating that the spatial channels in this region have highly consistent waveform response characteristics. As the spatial distance increases, the similarity gradually decreases to below 0.3, indicating that the vibration signals received by channels far from the event center have undergone significant distortion. Channel 138 simultaneously satisfies both the maximum energy response and the strongest spatial correlation criteria, which conforms to the basic principle of the spatial correlation enhancement method (ESC)—the location of the maximum response corresponds to the disturbance center, and the signals of neighboring channels are strongly correlated.

[0086] Therefore, this invention selects channel 138 as the one-dimensional main analysis signal for steps S1 and S2:

[0087] ;

[0088] The spatial channels near channel 138 have a high degree of similarity, and the similarity gradually decreases with increasing spatial distance. This indicates that the channel is located in the main response area of ​​the event and can well represent the temporal waveform characteristics of the vibration event.

[0089] This channel corresponds to a fiber optic cable position of approximately 80m, and is about 1m horizontally away from the point of impact of the falling rock. It can effectively characterize the temporal waveform characteristics of vibration events such as falling rocks, knocking, and walking.

[0090] S2. Based on XGBoost event-level statistical features, perform binary classification of falling rocks and non-falling rocks to complete the falling rock identification.

[0091] Using the candidate vibration event segment output in step S1 as input, five event-level statistical features of the original data are extracted from the entire candidate event signal, including: the standard deviation of the wavelet third-level detail coefficients, the energy of the wavelet second-level detail coefficients, the kurtosis of the wavelet second-level detail coefficients, the total signal energy, and the peak-to-peak value of the signal. These five event-level features are then input into the second XGBoost classifier to complete the binary classification of the candidate vibration events, determining whether the candidate events are rockfall events or non-rockfall events. Interference types include background noise that was misclassified as vibration events in step S1, knocking, pedestrian walking, and train operation.

[0092] S3. Fine-grained classification of landing material based on deep residual network to complete the material classification of rock landing points.

[0093] The original spatiotemporal two-dimensional response matrix of the rockfall event determined in step S2, i.e., the spatiotemporal waveform matrix, is used as the model input. Without feature compression or dimensionality reduction, it is directly input into a pre-trained deep residual network (ResNet). The deep residual network contains an initial convolutional layer and four sequentially connected residual blocks. Residual connections overcome gradient degradation, automatically extracting deep discriminative features corresponding to different materials from the two-dimensional matrix. Finally, the features are connected to a Softmax classifier via a fully connected layer, and the Softmax function outputs the recognition probabilities corresponding to the four types of landing materials, including ground, ballast, sleepers, or rails.

[0094] Specifically, the input matrix has a size of 512 spatial channels multiplied by 120 sampling time points, and undergoes amplitude normalization. A deep residual network is used as the classifier in step S3. The network structure is as follows: first, it passes through a 3×3 convolutional layer to map the input to 32 channels; then it passes through four residual blocks, expanding the number of channels to 64, 128, 256, and 512 respectively; spatial downsampling is achieved between residual blocks through convolutions with a stride of 2; identity mapping within residual blocks is achieved through residual connections; when the number of channels or spatial size of the input feature map and the output feature map are inconsistent, dimension matching is performed through projection transformation; the network ends with an adaptive average pooling layer to compress the feature map into a 512×4×4 feature vector.

[0095] S4. The AdamW optimizer is used for training. The initial learning rate during the training phase is set to 0.001, and the weight decay coefficient is set to 1×10⁻⁶. -4 The learning rate scheduling employs a Warmup+Cosine strategy, with linear warming in the first 10 training epochs and cosine decay to zero in the following 90 epochs. During training, to enhance the model's robustness to noise, time shifts, and sample discrepancies, three data augmentation strategies are implemented: the first adds Gaussian white noise with an amplitude of 0.02 to the input matrix with a 50% probability; the second performs a random time shift of ±5 sampling points on the input matrix with a 30% probability; and the third uses a Mixup strategy with a mixing parameter α set to 0.2, reducing the model's tendency to overfit and memorize training samples by performing a convex combination of samples and labels.

[0096] This invention discloses a fine-grained identification system for the material of railway rockfall intrusion points. The system employs the aforementioned fine-grained identification method for railway rockfall intrusion points, utilizing a Φ-OTDR distributed optical fiber railway rockfall monitoring system. This system includes a recall and extraction module, a rockfall feature identification module, a fine-grained material identification module, and a rockfall monitoring module. The recall and extraction module processes collected multi-scenario railway vibration signal data. Based on a spatial-temporal two-dimensional vibration response matrix constructed along the railway alignment, it selects the spatial channel with the largest energy response and strongest spatial correlation as the main analysis channel. It then performs a sliding window segmentation of length L on the first-order difference signal of this channel, extracting multiple discriminative features constructed based on the first-order difference of the original data within each window. These features are then fed into a first XGBoost binary classifier for discrimination. Windows with predicted probabilities higher than a set threshold are merged and output to obtain candidate vibration event segments. The rockfall feature identification module is used for... Five event-level raw data statistical features are extracted from the entire candidate event signal and input into the second XGBoost classifier to complete the binary classification of candidate vibration events, determining whether the candidate events are rockfall events or non-rockfall events. The rockfall monitoring module is used to select the raw spatiotemporal two-dimensional response matrix determined to be a rockfall event and supply it to the fine-grained material recognition module as input. The fine-grained material recognition module is used to receive the raw spatiotemporal two-dimensional response matrix output by the rockfall monitoring module as model input, without feature compression and dimensionality reduction, and directly input it into the pre-trained deep residual network. The deep residual network contains an initial convolutional layer and four sequentially connected residual blocks, which can automatically extract deep discriminative features corresponding to different materials from the two-dimensional matrix. The features are connected to the Softmax classifier through a fully connected layer, and the Softmax function outputs the recognition probability corresponding to each of the four types of landing point materials, completing the material classification of the rockfall point. The monitoring system uses an integrated DVS industrial computer to collect Φ-OTDR distributed optical fiber vibration signals, performs three-stage processing on the signals, and detects the material of the rockfall point. The system sampling frequency is set to 2000Hz, the pulse width is 100ns, and the sensing fiber length is 300m.

[0097] Example 1

[0098] In this embodiment, an integrated DVS industrial control computer is used to complete the Φ-OTDR distributed optical fiber vibration signal acquisition. The system sampling frequency is set to 2000Hz, and the pulse width is 100ns. The sensing optical fiber is 300m long and buried in a laboratory trench at a depth of about 10cm. All vibration events are applied at a position about 80m from the transmitter, and the horizontal distance between the event point and the optical fiber is about 1m.

[0099] This embodiment collected vibration signal data covering five typical scenarios: rockfall, impact, walking, train operation, and background noise. The rockfall samples were further subdivided into four categories based on the impact material: ground, ballast, sleepers, and rails, and included different rock sizes and drop heights. The impact samples covered different impact surfaces; the walking samples included different test personnel; the train operation samples were continuous vibration signals from actual train passages, with vibration intervals manually labeled; and the background samples were used to characterize environmental noise in the absence of events. The vibration signals from these five scenarios generated a total of 1896 valid vibration event segments, constituting the training and testing sample set for this embodiment.

[0100] S1. High recall extraction of window-level vibration events based on XGBoost:

[0101] First, the main analysis signal channels are selected. The sensing signals acquired by the Φ-OTDR distributed fiber optic sensing system are stored in the form of a two-dimensional space-time vibration response matrix, with a total of Q=512 spatial channels. The space-time two-dimensional vibration response matrix is ​​then subjected to time-domain difference to obtain a difference two-dimensional matrix D.

[0102] The sensor signals acquired by the Φ-OTDR system are stored in the form of a two-dimensional matrix, with the matrix data type being floating-point. A single raw data file can be represented as a spatial-temporal two-dimensional matrix:

[0103] ;

[0104] Where Q is the total number of spatial sampling points, U is the total number of pulses emitted by the system during the sampling time, and xq,u represents the Φ-OTDR amplitude response of the q-th spatial sampling point at the u-th sampling time. In the experiment of this invention, Q=512.

[0105] For a fixed spatial position q, in the q-th row, ... express, , representing the vibration waveform of the spatial channel as it changes over time.

[0106] For a fixed sampling time u, the u-th column is... express, .

[0107] Signal preprocessing and automatic event localization: This invention selects q=138 spatial channels as the main analysis signal, and processes the signal of the 138th channel... Perform first-order difference:

[0108] .

[0109] In this embodiment, the energy of each spatial channel within the event time range is calculated for the difference matrix D, and the channel with the highest energy is defined as the main analysis channel. The calculation results show that channel 138 simultaneously satisfies the conditions of maximum energy response and strongest spatial correlation. Therefore, channel 138 is selected as the one-dimensional main analysis signal for steps S1 and S2.

[0110] A sliding window is used to segment the first-order difference signal of the main analysis signal, with a window length L of 120 sampling points. Five window-level features are extracted as input to the first XGBoost classifier:

[0111] Assume differential signal within window The Discrete Fourier Transform (DFT) is Frequency resolution ,in, The sampling rate.

[0112] The first characteristic is the energy proportion in the 100-200 Hz band, the energy proportion in the 200-300 Hz band, and the energy proportion in the 300-400 Hz band:

[0113] ;

[0114] The second characteristic is the dominant frequency, which is the frequency point with the largest amplitude in the differential signal spectrum: ;

[0115] The third characteristic is the waveform factor, which is the ratio of the effective value to the absolute mean of the signal within the window:

[0116] ;

[0117] In this step, the XGBoost classifier parameters are set as follows: a total of 400 decision trees, a learning rate of 0.07, a maximum decision tree depth of 6, a sample subsampling rate of 0.7, a feature column sampling rate of 0.7, an L1 regularization coefficient of 0.1, and an L2 regularization coefficient of 2.0. An event determination threshold of 0.55 is used, and adjacent candidate segments with an interval of less than 180 sampling points are merged. The minimum event length is set to 90 sampling points.

[0118] This step prioritizes high recall, ensuring comprehensive coverage of rockfall events. A small number of false positives will occur, all of which will be passed to the next step. Preferably, in this embodiment, this step achieves a 99.66% coverage and recall rate for rockfall events.

[0119] S2. Binary classification of rockfall and non-rockfall based on XGBoost event-level statistical features:

[0120] This step receives the candidate event segments output from Phase 1, performs signal analysis on the complete event segments, and extracts five event-level statistical features:

[0121] Let the candidate event segment time signal output by the step be s(t), t=1,2,...,T, where T is the event segment length. Perform a db4 wavelet three-level decomposition, and let the j-th level detail coefficient sequence be... , This represents the number of detail coefficients for this layer; the approximate coefficient sequence is... .right Perform a DFT to obtain the spectrum S(f).

[0122] The first feature is the standard deviation of the third level detail in the wavelet. The candidate event segment is decomposed using a db4 wavelet three-level decomposition. The third level detail coefficients correspond to relatively high-frequency components, and their standard deviation reflects the intensity of local high-frequency fluctuations in the impact event.

[0123] ;

[0124] in, This is the mean of the detail coefficients for the third layer.

[0125] The second feature is the wavelet second-level detail energy, which reflects the total energy of the mid-frequency vibrational components within the candidate event segment:

[0126] ;

[0127] in, This represents the number of detail coefficients in the second layer. This is the second-level detail coefficient sequence.

[0128] The third feature is the wavelet second-level detail kurtosis, which reflects the peaks and tails of the data distribution:

[0129] ;

[0130] in, , This is the mean of the detail coefficients for the second layer.

[0131] The fourth characteristic is the total signal energy, which is the sum of the squares of the amplitudes of all sampling points within the event segment:

[0132] ;

[0133] in, The event segment length, This represents the original signal value at the t-th sampling point of the event segment.

[0134] The fifth characteristic is the peak-to-peak value, which is the difference between the maximum and minimum values ​​of the signal within the event segment:

[0135] ;

[0136] in, These represent the maximum and minimum values ​​of the signal within the event segment, respectively.

[0137] The above five features are input into the second XGBoost classifier. In step S2, the parameters of the XGBoost classifier are set as follows: total number of decision trees is 200, learning rate is 0.07, maximum depth is 4, sample subsampling rate is 0.9, feature column sampling rate is 0.7, L1 regularization coefficient is 0.1, and L2 regularization coefficient is 2.0.

[0138] This step completes the binary classification of falling rocks and non-falling rocks (including background noise, knocking, walking, and train operation). Preferably, in this embodiment, the falling rock retention rate in step S2 is 95.24%, and all train operation events are correctly classified as non-falling rocks, with a train false alarm rate of zero.

[0139] S3. Fine-grained classification of landing point materials based on deep residual networks:

[0140] This step receives candidate event segments identified as rockfalls in step S2 and uses the corresponding original spatiotemporal two-dimensional response matrix as model input. The input matrix has the size of 512 spatial channels multiplied by 120 sampling time points and has undergone amplitude normalization processing.

[0141] In this embodiment, a deep residual network is used as the classifier in step S3.

[0142] Residual learning refers to the network no longer directly learning the expected mapping H(x), but instead learning the residual mapping F(x) between the input and output. If the expected mapping is:

[0143] ;

[0144] The residual function can then be expressed as:

[0145] ;

[0146] Where x is the input of the residual block, F(x) is the residual mapping learned by several convolutional layers, and H(x) is the final output of the residual block. The calculation process of a typical residual block can be written as follows:

[0147] ;

[0148] in, Let represent the convolutional layer parameters in the residual branch, and y be the output of the residual block. When the dimensions of the input and output feature maps are the same, an identity mapping can be performed directly. When the number of channels or spatial dimensions of the two are different, linear projection or stride convolution needs to be introduced in the shortcut branch for dimension matching. In this case, the residual block can be represented as:

[0149] ;

[0150] in, These are the projection transformation parameters in the shortcut branch.

[0151] The ResNet classification model first extracts shallow local features through convolutional layers. Then, it progressively expands the number of channels and reduces the feature map size using multiple residual blocks, enabling the network to hierarchically represent textures from low-level impact to high-level material response patterns. After adaptive average pooling, the two-dimensional feature map is compressed into a fixed-length feature vector, which is then input into a fully connected layer to obtain classification scores for each category. Let the output vector of the fully connected layer be... , where C is the total number of categories. The predicted probability of the c-th category can be obtained using the Softmax function:

[0152] ;

[0153] The model is trained using the cross-entropy loss function. For a dataset containing N samples, the loss function is defined as:

[0154] ;

[0155] in, The true label of the i-th sample in class c (if it belongs to that class) =1, otherwise =0), Predict the probability that the i-th sample belongs to the c-th class for the model.

[0156] The network structure is as follows: First, the input is passed through a 3×3 convolutional layer, mapping it to 32 channels. Then, it passes through four residual blocks, expanding the number of channels to 64, 128, 256, and 512 respectively. Spatial downsampling is achieved between residual blocks using convolutions with a stride of 2. Identification within residual blocks is achieved through residual connections, with the basic mapping relationship as follows: When the number of channels or spatial dimensions of the input feature map and the output feature map are inconsistent, dimension matching is performed through projection transformation. At the end of the network, the feature map is compressed into a 512×4×4 feature vector through an adaptive average pooling layer; then, it is input into a fully connected Softmax classifier, and the Softmax function outputs the recognition probability corresponding to the four types of landing material.

[0157] The AdamW optimizer was used during the training phase, with an initial learning rate of 0.001 and a weight decay factor of 1×10⁻⁶. -4 The learning rate is scheduled using a Warmup+Cosine strategy, with linear warming up in the first 10 training epochs and cosine decaying to zero in the following 90 epochs. During training, to enhance the model's robustness to noise, time shifts, and sample discrepancies, three data augmentation strategies are also incorporated:

[0158] The first method involves adding Gaussian white noise with an amplitude of 0.02 to the input matrix with a 50% probability.

[0159] The second method involves randomly shifting the input matrix by ±5 sampling points with a 30% probability.

[0160] The third approach is to use the Mixup enhancement strategy, with the mixing parameter α set to 0.2. This reduces the model's tendency to overfit to training samples by performing a convex combination of samples and labels.

[0161] Preferably, the ResNet in this embodiment achieves a Macro F1 score of 0.9178 on the test set and an event-level material classification accuracy of 98.92%.

[0162] Three-level cascaded links and end-to-end identification:

[0163] By cascading the above three steps, end-to-end automatic identification from the original vibration signal to the impact material is achieved. The effective identification rate R of the end-to-end cascade is defined. cascade The overall probability of a rockfall event being detected and then correctly identified in terms of its final material is calculated using the following formula: in, This represents the coverage and recall rate for the rockfall event in step S1. This represents the retention rate of the covered rocks in step S2. The accuracy of material identification for successfully matched falling rocks in step S3.

[0164] In the actual test of this embodiment, the total number of real rockfall events was 295. Step S1 covered 294 rockfalls; step S2 correctly retained 280 rockfalls from the 294 covered rockfalls; and step S3 correctly identified the material of 276 rockfall samples from the 279 successfully matched end-to-end samples. Based on the above cascaded link statistics, the effective identification rate of the end-to-end cascade was calculated to be 93.92%. Throughout the cascaded process, step S2 achieved strict zero false alarms for strong periodic vibrations of the train, verifying the engineering feasibility and robustness of the method of the present invention under complex railway conditions.

[0165] The structure of the Φ-OTDR distributed fiber optic railway rockfall monitoring system is as follows: Figure 1As shown, the Φ-OTDR system injects narrow-linewidth coherent pulses into the sensing fiber and receives Rayleigh backscattered light generated along the fiber. When external vibrations act near the fiber, they cause local changes in refractive index and optical path, disturbing the phase and interference intensity of the backscattered signal. By continuously acquiring echo curves, a two-dimensional spatial-temporal vibration matrix can be formed, enabling distributed monitoring along the fiber direction.

[0166] The spatiotemporal distribution of the bar chart differential signal in a rockfall event, such as Figure 2 As shown, a partial magnified view of the original signal waveform of channel 138 is shown. Figure 3 As shown.

[0167] The three-dimensional distribution of the spatial response of a rockfall event is as follows: Figure 4 , Figure 5 As shown, taking a typical rockfall impact event as an example, within this window, the vibration energy exhibits significant spatial concentration, mainly concentrated near channel 138, and rapidly attenuates to both sides. Figure 4 It shows how the differential signal changes over time and through the channel.

[0168] The results of spatial channel similarity analysis are as follows: Figure 6 , Figure 7 As shown, the self-similarity of channel 138 is 1.000, while the similarity of neighboring channels 137 to 141 is all above 0.8, indicating that the spatial channels in this region have highly consistent waveform response characteristics. As the spatial distance increases, the similarity gradually decreases to below 0.3, indicating that the vibration signals received by channels far from the event center have undergone significant distortion. Channel 138 simultaneously satisfies both the maximum energy response and the strongest spatial correlation criteria, conforming to the basic principle of the spatial correlation enhancement method (ESC)—the maximum response location corresponds to the disturbance center, and the signals of neighboring channels exhibit strong correlation. Therefore, this invention selects channel 138 as the one-dimensional main analysis signal for both Stage 1 and Stage 2.

[0169] The three-level cascaded identification framework proposed in this invention, such as Figure 8 As shown, the system input is continuously acquired Φ-OTDR spatiotemporal vibration data. In the first stage, candidate vibration time periods are extracted from the continuous signal using an XGBoost binary classification model based on window-level statistical features. In the second stage, the candidate event segments output from the first stage are used as input to further extract event-level statistical features, and an XGBoost binary classification model is used to determine whether the object is a rockfall or not. In the third stage, only candidate samples identified as rockfalls in the second stage are received, and a ResNet model is used to perform fine-grained classification of the four impact point materials: ground, ballast, sleepers, and rails.

[0170] Characteristics of five typical vibration scenarios, for example Figure 9 , Figure 10As shown, a comparison of time-domain waveforms reveals that the amplitudes of the train operation event and the background noise are comparable, exhibiting no obvious impact characteristics. The knocking event is characterized by a sudden, high-amplitude pulse that rapidly decays to the baseline level. The rockfall event also exhibits transient excitation characteristics, but shows a significant oscillating decay process after the initial impact. In contrast, the walking event has the longest duration, exhibiting a non-periodic long-term fluctuation characteristic with an amplitude that first increases and then slowly decreases. From a frequency domain distribution perspective, after mean-reduction processing, the average amplitude of the train operation and background noise signals in the 0-500Hz frequency band is essentially close to zero. The spectral energy of the walking event is highly concentrated in the low-frequency region, particularly in the 0-150Hz range, while the high-frequency components decay rapidly. Both the rockfall and knocking events exhibit broadband impact responses, with their spectral energy covering the entire 0-500Hz frequency band. Furthermore, within the same frequency band, the average amplitude spectrum of the knocking event is generally higher than that of the rockfall event, indicating a shorter pulse duration and a more concentrated high-frequency energy distribution.

[0171] Step S1 Feature Filtering Box Type, such as Figures 11-15 As shown, firstly, the signal of the 138th spatial channel is divided into sliding window segments, and time-domain, frequency-domain, and wavelet-domain features are extracted from the first-order difference signal. After feature selection, the five window-level features with the strongest discriminative power are finally retained.

[0172] Step S1: Confusion matrix as follows Figure 16 , Figure 17 As shown, although most real vibration windows were correctly identified, a considerable number of background windows were misclassified as vibrations. These false positives will all flow into subsequent stages as candidate event segments. Therefore, a high-precision binary classification model was designed in stage two, specifically for identifying falling rocks and filtering out non-falling rock interference from the mixed candidate segments output from stage one, in order to compensate for the insufficient accuracy of stage one.

[0173] Step S2 Feature Filtering Box Type, such as Figures 18-22 As shown, stage two extracts full-segment statistical features from complete event segments. Five optimal features are ultimately selected.

[0174] Step S2: Confusion matrix as follows Figure 23 As shown, the non-rockfall false alarms introduced by the relaxed detection in stage one were effectively eliminated in stage two, indicating that event-level features can effectively compensate for the false detection problem caused by window-level detection, and confirming the significant compensatory effect of the two-level cascaded architecture based on complete event segment features on window-level false detection.

[0175] Thermal diagrams of the spatiotemporal response to rockfall impacts on four typical materials, such as Figures 24-27 As shown, the temporal envelope of falling rocks impacting different media is highly similar to the macroscopic frequency band energy, but there are subtle differences in vibration duration, frequency components, and spatial attenuation modes.

[0176] The ResNet network architecture in step S3 is as follows: Figure 28 As shown, the model input is a normalized 512×120 single-channel waveform matrix. The network first passes through a 3×3 convolutional layer, mapping the input to 32 channels; then it passes through four residual blocks, expanding the number of channels to 64, 128, 256, and 512 respectively, and downsampling is achieved through convolutions with a stride of 2. Residual connections can alleviate the gradient degradation problem in deep network training, enabling the model to learn multi-scale impact features more stably. After adaptive average pooling, the feature map is compressed to 512×4×4, and then input into a fully connected classifier, finally outputting the probabilities of four material types.

[0177] The ablation experiment results in step S3 are as follows: Figure 29 As shown, the Macro F1 of the basic CNN in Stage 3 was only 0.7332; it improved to 0.8378 after deepening the network; it improved to 0.8860 after introducing the residual structure; and its performance continued to improve after adding data augmentation, Warmup+Cosine and Mixup, with the final test set Macro F1 reaching 0.9178.

[0178] The sample flow and error propagation in a three-level cascaded link, such as Figure 30 As shown, to evaluate the error propagation characteristics of the cascaded system under real data flow, this invention connects the three-stage model that has been trained according to the actual processing flow, simulates the whole-link reasoning process from the original signal to material recognition, and statistically analyzes the output data at each stage.

[0179] It will be understood by those skilled in the art that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments, and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for fine-grained identification of the material at the point of impact of falling rocks on railway tracks, characterized in that, Includes the following steps: S1. High recall extraction of vibration events based on window-level features of XGBoost: Vibration signal data under various scenarios were collected using a Φ-OTDR distributed optical fiber sensing system to obtain a two-dimensional spatial-temporal vibration response matrix along the railway line. The spatial channel with the largest energy response and the strongest spatial correlation was selected as the main analysis channel. Its first-order difference signal was divided into a sliding window of length L, and multiple discriminative features of each window were extracted. The above features were input into the first XGBoost classifier for binary classification. Windows with a probability greater than a set threshold were merged into candidate vibration event segments. S2. Binary classification of rockfall and non-rockfall based on XGBoost event-level statistical features: Using the candidate vibration event segment as input, extract five event-level raw data statistical features from the entire candidate event signal, input them into the second XGBoost classifier, complete the binary classification of the candidate vibration event, and determine the candidate event as a rockfall event or a non-rockfall event. S3. Fine-grained classification of landing point materials based on deep residual networks: The original spatiotemporal two-dimensional response matrix that is identified as a rockfall event is used as the model input and fed into a pre-trained deep residual network. The deep residual network contains an initial convolutional layer and four residual blocks connected in sequence, which can automatically extract deep discriminative features corresponding to different materials from the two-dimensional matrix. The rockfall point is connected to the Softmax classifier through a fully connected layer, and the Softmax function outputs the recognition probability of each of the four types of landing materials to complete the material classification of the rockfall point. S4. Use the AdamW optimizer for training optimization.

2. The method for fine-grained identification of the material of the landing point of a railway rockfall as described in claim 1, characterized in that: Step S1: Vibration signal data under multiple scenarios covers five scenarios: rockfall, knocking, walking, train operation, and background noise. The discrimination features are the features after the first-order difference of the original data, including the energy proportion, main signal frequency, and waveform factor of the three frequency bands of 100-200Hz, 200-300Hz, and 300-400Hz.

3. The method for fine-grained identification of the material of the landing point of a railway rockfall as described in claim 1, characterized in that, Step S1: The matrix data type is floating-point. A single raw data file can be represented as a two-dimensional spatial-temporal matrix. ； Where Q is the total number of spatial sampling points, U is the total number of pulses emitted by the system during the sampling time, xq,u represents the Φ-OTDR amplitude response of the q-th spatial sampling point at the u-th sampling time, and Q=512; For a fixed spatial location q, the q-th row: ； This represents the vibration waveform of the spatial channel as it changes over time. For a fixed sampling time u, the u-th column... ； Spatial response analysis and main channel selection: Channel 138 is selected as the one-dimensional main analysis signal for steps S1 and S2.

4. The method for fine-grained identification of the material of the landing point of a railway rockfall as described in claim 1, characterized in that: Step S2, the five event-level statistical features include: the standard deviation of the wavelet third-level detail coefficients, the energy of the wavelet second-level detail coefficients, the kurtosis of the wavelet second-level detail coefficients, the total signal energy, and the peak-to-peak value of the signal.

5. The method for fine-grained identification of the material of the landing point of a railway rockfall as described in claim 1, characterized in that: Step S2, the types of interference include background noise, knocking, pedestrian walking and train operation that were misjudged as vibration events in step S1.

6. The method for fine-grained identification of the material of the landing point of a railway rockfall as described in claim 1, characterized in that: Step S3: The input matrix has a size of 512 spatial channels multiplied by 120 sampling time points, and undergoes amplitude normalization processing.

7. The method for fine-grained identification of the material of the landing point of a railway rockfall as described in claim 1, characterized in that: Step S3: A deep residual network is used as the classifier for step S3. The network structure is as follows: first, the input passes through a 3×3 convolutional layer to map to 32 channels; then, it passes through four residual blocks, expanding the number of channels to 64, 128, 256, and 512 respectively; spatial downsampling is achieved between residual blocks through convolutions with a stride of 2; identity mapping within residual blocks is achieved through residual connections; when the number of channels or spatial size of the input feature map and the output feature map are inconsistent, dimension matching is performed through projection transformation; at the end of the network, an adaptive average pooling layer compresses the feature map into a 512×4×4 feature vector.

8. The method for fine-grained identification of the material of the landing point of a railway rockfall as described in claim 1, characterized in that: Step S3: The four types of landing materials include the ground, ballast, sleepers, or rails.

9. The method for fine-grained identification of the material of the landing point of a railway rockfall as described in claim 1, characterized in that: Step S4: Set the initial learning rate for the training phase to 0.001 and the weight decay coefficient to 1×10. -4 The learning rate scheduling adopts a Warmup+Cosine strategy, with linear warming in the first 10 training rounds and decaying to zero according to a cosine function in the next 90 training rounds. During training, to enhance the model's robustness to noise, time shifts, and sample differences, three data augmentation strategies were implemented: the first was to add Gaussian white noise with an amplitude of 0.02 to the input matrix with a 50% probability; the second was to perform a random time shift of ±5 sampling points on the input matrix with a 30% probability; and the third was to use a Mixup augmentation strategy with a mixing parameter α set to 0.2, which reduces the model's tendency to overfit to training samples by performing a convex combination of samples and labels.

10. A fine-grained identification system for the material of railway rockfall encroachment points, employing the fine-grained identification method for the material of railway rockfall encroachment points as described in claim 1, characterized in that: The identification method employs a Φ-OTDR distributed fiber optic railway rockfall monitoring system, comprising a recall extraction module, a rockfall feature recognition module, a fine-grained material recognition module, and a rockfall monitoring module. The recall extraction module processes collected multi-scenario railway vibration signal data. Based on a spatial-temporal two-dimensional vibration response matrix constructed along the railway alignment, it selects the spatial channel with the largest energy response and strongest spatial correlation as the main analysis channel. It then performs a sliding window segmentation of length L on the first-order difference signal of this channel, extracting multiple discriminative features constructed based on the first-order difference of the original data within each window. These features are then fed into a first XGBoost binary classifier for discrimination. Windows with predicted probabilities higher than a set threshold are merged to obtain candidate vibration event segments. The rockfall feature recognition module extracts five event-level statistical features from the entire candidate event signal segment and inputs them into a second XGBoost classifier to complete the binary classification of candidate vibration events, determining whether the candidate event is a rockfall event or not. The rockfall monitoring module selects the original spatiotemporal two-dimensional response matrix determined to be a rockfall event and provides it as input to the fine-grained material recognition module. The fine-grained material recognition module receives the original spatiotemporal two-dimensional response matrix output by the rockfall monitoring module as model input. Without feature compression or dimensionality reduction, it is directly input into a pre-trained deep residual network. The deep residual network contains an initial convolutional layer and four sequentially connected residual blocks, capable of automatically extracting deep discriminative features corresponding to different materials from the two-dimensional matrix. The rockfall is fed into a Softmax classifier via a fully connected layer, and the Softmax function outputs the recognition probability of each of the four types of landing materials to complete the material classification of the rockfall.

Images