A method for identifying inter-fusion layer consistent open set events

Abstract

The application discloses a fusion interlayer consistency open set event identification method and relates to the technical field of distributed optical fiber acoustic sensing, which comprises the following steps: step one, collecting vibration response signals distributed along an optical fiber; step two, constructing a two-dimensional phase event image based on the vibration response signals; step three, constructing a phi-OTDR open set event identification model; the phi-OTDR open set event identification model comprises an MSRS post-processing module, and the MSRS post-processing module comprises a multilayer Mahalanobis distance branch, an interlayer consistency branch and an MSP branch; step four, inputting the two-dimensional phase event image into the phi-OTDR open set event identification model, obtaining a final reliability score, and performing open set discrimination. The application has the characteristics of significantly improving the unknown class rejection capability and the overall open set identification accuracy.

Description

Technical Field

[0001] This invention relates to the field of distributed fiber optic acoustic sensing technology, and more specifically, to a method for identifying open set events with fused interlayer consistency. Background Technology

[0002] Phase-sensitive optical time-domain reflectometer (φ-OTDR) is a distributed vibration sensing technology based on fiber Rayleigh backscattering and coherent detection principles. This technology enables long-distance, continuous, and distributed monitoring of vibration or acoustic events along a single sensing fiber, and has been widely applied in fields such as pipeline safety early warning, perimeter intrusion detection, rail transit monitoring, water safety monitoring, geological disaster monitoring, and health monitoring of critical infrastructure.

[0003] With the development of deep learning, convolutional neural networks, transfer learning, lightweight networks, and data augmentation methods have been gradually applied to φ-OTDR event recognition tasks. Existing methods typically train classification models under the closed-set assumption, that is, they assume that the samples in the training and testing phases belong to the same set of known event categories and obtain the category probabilities through softmax probability normalization. However, the softmax probability normalization process assumes that the input samples belong to the set of known categories. Therefore, relying solely on the maximum softmax probability for open environment judgment is unreliable.

[0004] For the problems of open set identification and out-of-distribution detection, existing methods include maximum softmax probability, input perturbation methods, Mahalanobis distance, energy functions, OpenMax, ViM, and deep nearest neighbors. These methods detect unknown samples from the perspectives of output probability, feature space distance, energy score, or virtual logit matching, respectively. However, in the engineering application of φ-OTDR, the following limitations still exist:

[0005] 1. High confidence in the output layer does not equate to true reliability in an open environment:

[0006] Closed-set classifiers typically use the maximum softmax probability as the prediction confidence. However, softmax is a probability allocation that is normalized within a known set of categories. Even if an unknown sample does not belong to any known category, it may still be forced to be mapped to a certain known category and output a high softmax confidence, which can lead to false alarms, missed alarms, or high-confidence false alarms. For security monitoring applications, such misjudgments may reduce the reliability of system alarms and affect the judgment of on-site maintenance personnel on real events. Therefore, deep neural networks trained under the closed-set assumption cannot directly deal with unknown category events, out-of-distribution inputs, or abnormal signals.

[0007] 2. A single output layer score is insufficient to reflect the multi-layer characteristic shift of the φ-OTDR signal:

[0008] φ-OTDR phase event images have multi-scale spatiotemporal structures. Unknown events may not necessarily be represented as low confidence in the final output layer, but may show different degrees of shift in shallow phase texture, mid-level local strip structure, or high-level event semantic representation. Using only output layer scores such as Energy, Entropy, or maximum softmax probability is difficult to capture the distribution shift of shallow texture, mid-level local spatiotemporal structure, and high-level event representation in φ-OTDR phase images, and cannot fully utilize the multi-layer representation information within the network.

[0009] 3. Modeling based solely on the features of the last layer may result in the loss of discriminative information from intermediate layers:

[0010] The intermediate layers of convolutional networks contain richer spatial texture, local structure, and phase perturbation pattern information. For φ-OTDR phase maps, these intermediate layer features are of great value for identifying unknown perturbations. However, OpenMax, ViM, or other post-processing methods based on the features of the last layer usually focus on the features before the classification head or the activation vector of the last layer, which may ignore the abnormal changes of unknown events in the process of the network's intermediate layer representation evolution. This may result in the inability to capture the abnormal performance of unknown samples at different network stages if only the features of the last layer are relied upon.

[0011] 4. Existing open-set methods may require additional training or complex parameter settings:

[0012] Some open set identification methods require external unknown samples, negative sample generation, model retraining, or complex hyperparameter search. Moreover, they usually require the adjustment of many parameters, and the threshold or scoring function is complex. For already deployed φ-OTDR / DAS monitoring systems, this will increase the difficulty of system modification and deployment costs, and is not conducive to stable use in long-term online systems in the field.

[0013] 5. Existing methods lack characterization of the stability of inter-layer representation evolution:

[0014] For known φ-OTDR phase event images, the network will form a relatively stable feature extraction path during training. That is, there is usually a relatively smooth and stable evolutionary relationship between shallow phase texture, mid-level local structure and high-level event representation. Although unknown samples may be close to known classes in a certain layer, their cross-layer representation changes may be abnormal. Existing technologies usually do not explicitly utilize this inter-layer representation consistency information.

[0015] 6. Existing methods lack reliability verification of φ-OTDR phase maps:

[0016] Unlike natural images, φ-OTDR phase maps may exhibit unknown perturbations such as local stripe anomalies, phase texture shifts, spatiotemporal continuity disruptions, or multi-scale vibration mode mismatches. A single scoring function is insufficient to adequately measure their reliability.

[0017] In actual engineering sites, intrusion methods, environmental disturbances, equipment noise, compound events, and unknown abnormal signals change constantly over time. It is difficult to exhaust all possible event types during the training phase. Therefore, how to maintain a high pass rate for known event recognition and effectively reject unknown events or abnormal inputs based on the frozen φ-OTDR closed-set event classifier without changing the original closed-set classification network structure or retraining the classifier remains an urgent problem to be solved. Summary of the Invention

[0018] The purpose of this invention is to design and develop an open set event recognition method that integrates inter-layer consistency, addressing the open set event recognition problem in water monitoring scenarios. This method extracts multi-level evidence from multi-stage features, significantly improving the ability to reject unknown classes.

[0019] The technical solution provided by this invention is as follows:

[0020] A method for identifying open set events with inter-layer consistency includes the following steps:

[0021] Step 1: Based on the φ-OTDR / DAS distributed acoustic sensing system, acquire vibration response signals distributed along the optical fiber;

[0022] Step 2: Perform coherent detection, phase demodulation, and time-space mapping on the vibration response signal to construct a two-dimensional phase event image;

[0023] Step 3: Construct the φ-OTDR open set event recognition model;

[0024] The φ-OTDR open set event recognition model includes an MSRS post-processing module, which includes a multi-layer Mahalanobis distance branch, an inter-layer consistency branch, and an MSP branch.

[0025] Step 4: Input the two-dimensional phase event image into the φ-OTDR open set event recognition model to obtain the final reliability score and perform open set discrimination;

[0026] The open set discrimination satisfies:

[0027] ;

[0028] In the formula, For open sets, For unknown class, For the first The Softmax probability normalized value of each sample. For the first The final reliability score for each sample This is the reliability threshold.

[0029] Preferably, the time-space mapping satisfies:

[0030] ;

[0031] ;

[0032] In the formula, The propagation time of the echo relative to the probe pulse. The speed of light in a vacuum. The effective refractive index of the optical fiber. The position of the optical fiber is given by the coefficient 2, which reflects the round-trip propagation process of the probe light within the fiber. For dynamic phase terms, The strain sensitivity coefficient is related to the fiber parameters and the operating wavelength. This is local strain.

[0033] Preferably, the two-dimensional phase event image includes:

[0034] 224 consecutive time frames are selected from the continuous time-space phase matrix, and 224 spatial positions are selected in the spatial direction to obtain a phase event matrix with a size of 224×224;

[0035] The phase event matrix is ​​plotted as a two-dimensional phase event image in time-space, with its vertical axis corresponding to continuous time frames and its horizontal axis corresponding to the spatial positions arranged sequentially along the sensing optical fiber.

[0036] Preferably, the φ-OTDR open set event recognition model further includes a closed set event classifier, which uses MobileNetV4-Conv-Medium as the event recognition network.

[0037] Preferably, the multi-layer Mahalanobis distance branch satisfies:

[0038] ;

[0039] In the formula, For the first The branch intermediate score of each sample after multi-level weighted fusion Indicates the robustness of the branch level after multi-level weighted fusion. Score standardization For the first The original scores of each sample are obtained through multiple Mahalanobis distance branches.

[0040] Preferably, the inter-layer consistency branch satisfies:

[0041] ;

[0042] In the formula, For the first The original scores of each sample after the inter-stratum consistency branch, , and These are the Euclidean distances between the first feature stage and the second feature stage, the second feature stage and the third feature stage, and the third feature stage and the pre-classification feature level before the classification head, respectively.

[0043] Preferably, the MSP branch satisfies:

[0044] ;

[0045] In the formula, For the first The original scores of each sample after passing through the MSP branch.

[0046] Preferably, step three further includes:

[0047] Robustly evaluate the original scores of the multilevel Mahalanobis distance branch, the inter-level consistency branch, and the original MSP score. Score standardization:

[0048] ;

[0049] In the formula, For the first The original scores of each branch, and , For the first Standardized scores for each branch To verify the samples in the set at the 1st The median of the original scores for each branch. This represents the absolute deviation of the median.

[0050] Preferably, the final reliability score satisfies:

[0051] ;

[0052] In the formula, , and They represent the first The scores were obtained from a robust standardization process using multiple Mahalanobis distance branches, inter-layer consistency branches, and MSP branches. , and These represent the weights of the three types of scores in the final reliability score. This is used to determine the final reliability score.

[0053] Preferably, the range of the reliability threshold satisfies:

[0054] ;

[0055] in, and These are the minimum and maximum values ​​of the final reliability score for the validation set, respectively.

[0056] The beneficial effects of this invention are as follows:

[0057] (1) The present invention designs and develops an open set event recognition method that integrates inter-layer consistency. Without changing the original closed set classifier structure and training process, it can utilize the multi-layer intermediate features, inter-layer representation changes and output probability information of the frozen classifier to construct a unified sample reliability score through post-processing, thereby realizing the acceptance of known classes and the rejection of unknown classes.

[0058] (2) The open set event recognition method designed and developed in this invention is aimed at the open set event recognition problem in water monitoring scenarios. It provides a unified sample reliability metric and a decision basis for known class acceptance / unknown class rejection during the open set testing and deployment stages, significantly improves the ability to reject unknown classes, maintains a high coverage of known samples under a high unknown rejection rate, improves the overall recognition accuracy of open sets, reduces the risk of high confidence misjudgment of unknown events by closed set classifiers, makes full use of the feature information of the intermediate layers of the network, improves the discrimination robustness by utilizing the stability of inter-layer representation evolution, and the post-processing method is convenient for engineering deployment. It does not rely on external unknown sample training, is suitable for long-term online operation of safety monitoring systems, and has good generalization potential. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the distributed acoustic sensing system based on φ-OTDR described in this invention.

[0060] Figure 2 This is a flowchart illustrating the inter-layer consistency open set event identification method of the present invention.

[0061] Figure 3 This is the phase event image described in this invention.

[0062] Figure 4 This is a schematic diagram of the fusion weights of the MSRS post-processing module described in this invention, using three-dimensional scatter plots.

[0063] Figure 5 This is a schematic diagram of the fusion weighted triangular surface of the MSRS post-processing module described in this invention.

[0064] Figure 6 This is the original confusion matrix of the closed-set test set described in this invention.

[0065] Figure 7 This is the normalized confusion matrix of the closed-set test set described in this invention.

[0066] Figure 8 This is a schematic diagram of the ROC curve for a multi-classification task on the closed-set test set described in this invention.

[0067] Figure 9 This is a schematic diagram of the Precision-Recall curve for a multi-classification task on the closed-set test set described in this invention.

[0068] Figure 10 This is the confusion matrix of the hybrid test set described in this invention.

[0069] Figure 11 This is the normalized confusion matrix of the Imagenette dataset described in this invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0071] This invention provides a method for identifying open set events with inter-layer consistency, employing a distributed acoustic sensing (DAS) system based on φ-OTDR. The system structure is as follows: Figure 1 As shown, the output light from a narrow-linewidth laser with a wavelength of 1550.12 nm is first split into two paths by a 90:10 fiber coupler, with approximately 90% used as probe light and the remainder as local oscillator light. The probe light enters an acousto-optic modulator (AOM) driven by an arbitrary waveform generator (AWG), where it is modulated into pulsed probe light and frequency-shifted. It is then amplified by an erbium-doped fiber amplifier (EDFA) and injected into the sensing fiber through an optical circulator. Rayleigh backscattered light from the sensing fiber returns and interferes with the local oscillator light in the optical coupler. This interference is converted into an electrical signal by a balanced photodetector (BPD), and finally acquired by a data acquisition card (DAQ) at a sampling rate of 1 GSa / s.

[0072] The aforementioned device uses coherent detection to acquire vibration response information distributed along the optical fiber, providing raw spatiotemporal data for subsequent phase demodulation, event segment construction, and open set identification. In the controlled experiment, the system pulse repetition frequency was set to 10 kHz, the pulse width to 100 ns, the sensing medium to 10 km of single-mode optical fiber, and the piezoelectric ceramic transducer (PZT) to be connected at a distance of approximately 4.92 km from the input end of the optical fiber, with its vibration frequency provided by a signal generator.

[0073] The method for identifying open set events with inter-layer consistency provided by this invention specifically includes the following steps:

[0074] Step 1: Based on the φ-OTDR / DAS distributed acoustic sensing system, acquire vibration response signals distributed along the optical fiber;

[0075] Step 2: Perform coherent detection, phase demodulation, and time-space mapping on the vibration response signal to construct a two-dimensional phase event image. Specifically:

[0076] In a φ-OTDR system, numerous Rayleigh scattering centers distributed along the optical fiber generate backscattered light from the incident pulse. The system obtains the phase change related to external disturbances through coherent interference and phase demodulation. For distributed acoustic sensing, the time sampling points recorded by the data acquisition card must first be mapped to spatial locations on the optical fiber. When the probe pulse reaches this location and generates backscattering, the scattered light needs to return to the receiver along the original path. Therefore, the echo propagation time includes both forward propagation and backward return, which can be expressed as:

[0077] (1)

[0078] In the formula, The propagation time of the echo relative to the probe pulse. The speed of light in a vacuum. The effective refractive index of the optical fiber. The position of the optical fiber is given by the coefficient 2, which reflects the round-trip propagation process of the probe light within the fiber. This indicates the equivalent propagation delay of light in the optical fiber medium;

[0079] Equation (1) gives the basic mapping relationship between the time axis and the spatial axis in the φ-OTDR signal. That is, under the condition of a fixed sampling rate, different echo times correspond to different positions on the optical fiber. The system can rearrange the continuously acquired backscattered signals into a distributed response that expands along the length of the optical fiber. This relationship is the basis for the subsequent construction of spatiotemporal phase data. Because the event recognition model does not directly process the original electrical signal at a single time point, but needs to extract the disturbance mode in the "time-space" two-dimensional structure. According to Equation (1), in each detection cycle, the echo sampling points correspond to the spatial positions distributed along the optical fiber according to the propagation delay. Multiple consecutive detection cycles constitute the time dimension. Therefore, the original φ-OTDR acquisition data can be physically organized into a "time-space" two-dimensional response matrix, where the time dimension describes the change of the same spatial position with the continuous detection cycle, and the spatial dimension describes the distribution of different positions along the optical fiber at the same time.

[0080] To obtain phase information that is more sensitive to external vibrations, the coherent received signal is quadrature demodulated. The received signal is mixed with mutually orthogonal cosine and sine reference signals, and then low-pass filtered to obtain the baseband in-phase component. and orthogonal components The phase is determined by the four-quadrant arctangent relationship of the two baseband components:

[0081] (2)

[0082] In the formula, To wrap the phase, The quadrant correction integer is determined based on the sign relationship between the I and Q signals and is used to correct the phase quadrant of the arctangent.

[0083] income The packaged phase ranges from [-π, π]. From a physical perspective, the demodulated phase can be expressed as the sum of the static phase background and the dynamic phase change caused by the disturbance:

[0084] (3)

[0085] In the formula, The static phase is determined by factors such as the initial state of the optical fiber, the distribution of scattering centers, installation conditions, and fixed optical path difference, and it usually changes slowly over time. This refers to the dynamic phase term caused by local strain, pressure disturbance, or external vibration; it is the part that more directly reflects disturbance information after the event occurs. This is a static phase term. For the φ-OTDR / DAS event recognition task, the static phase provides the basic phase background for each spatial location, while the dynamic phase reflects the change process of the event in different time frames and different spatial locations. Writing the two in an additive form helps to explain that the phase event image used in this invention is not an ordinary natural image, but a spatiotemporal feature representation transformed from the distributed phase response of optical fiber.

[0086] Under small perturbation conditions, dynamic phase Changes and local strain The changes have a corresponding relationship:

[0087] (4)

[0088] In the formula, The strain sensitivity coefficient is related to the fiber parameters and operating wavelength;

[0089] Therefore, different types of external events will form different stripes, textures, local undulations and continuity features in the time-space phase matrix.

[0090] When constructing event recognition samples, 224 consecutive time frames are selected from a continuous time-space phase matrix, and 224 spatial positions are selected at fixed intervals in the spatial direction, thus obtaining a phase event matrix of size 224×224. The phase event matrix is ​​plotted into an image in time-space. The vertical axis of the image corresponds to the continuous time frames, and the actual time interval between adjacent rows is determined by the system pulse repetition period. The horizontal axis of the image corresponds to the spatial positions arranged sequentially along the sensing fiber, and the actual spatial interval between adjacent columns is determined by the data acquisition sampling interval and the sampling interval. Therefore, the two-dimensional texture in the image actually represents the phase response distribution of the disturbance at different times and different fiber positions.

[0091] To convert the phase event matrix into a grayscale image that can be directly read by a neural network, the wrapped phase is linearly normalized using a fixed physical range [-π, π]. The normalization relationship is as follows:

[0092] (5)

[0093] In the formula, This is the normalized wrap-around phase, and ;

[0094] The normalization result is then mapped to an 8-bit grayscale value:

[0095] (6)

[0096] In the formula, For fiber optic location and transmission time The corresponding 8-bit grayscale value of the pixel can be seen from equations (5) and (6). Grayscale values ​​0 and 255 correspond to the cases where the phase is close to -π and π, respectively, and the intermediate grayscale value corresponds to the phase value between the two.

[0097] In summary, the source of the two-dimensional phase event image can be summarized as follows: the φ-OTDR coherent echo first establishes the spatial position relationship based on the echo propagation delay, and then obtains the wrapped phase distributed along the fiber through orthogonal demodulation; the continuous detection period forms the time dimension, and the fixed-interval fiber positions form the spatial dimension. The two are truncated to form a 224×224 time-space phase matrix. Finally, the phase event image is obtained by normalization of the fixed phase range and 8-bit gray quantization. This representation method not only meets the input size requirements of the subsequent two-dimensional convolutional network, but also retains the main phase response structure of the event perturbation in time and space.

[0098] Step 3, as follows Figure 2 As shown, the φ-OTDR open set event recognition model is constructed, specifically including:

[0099] Step 1: Construct a closed-set event classifier. Train the closed-set event classifier using samples of known classes, and then freeze the trained closed-set event classifier.

[0100] MobileNetV4-Conv-Medium is used as the event recognition network. The event recognition network consists of two parts: a feature extraction module and a classification head. The feature extraction module includes a shallow feature extraction stage, a first feature stage, a second feature stage, and a third feature stage. The classification head includes a feature aggregation head, a pre-classification feature output layer, a linear classifier, and a Softmax probability normalization. After training, the parameters of the closed-set classifier are frozen. The MSRS post-processing module only reads the intermediate features and output probabilities without changing the original network connection relationship of the closed-set classifier.

[0101] The shallow feature extraction stage consists of Stem and FusedIB, which are used to perform initial feature extraction on the 224×224 phase event image. Then, it enters the first feature stage, the second feature stage and the third feature stage according to the predetermined hierarchy of MobileNetV4-Conv-Medium. The three stages output the first stage feature F3, the second stage feature F4 and the third stage feature F5 respectively. F3, F4 and F5 represent the stage features from shallower to deeper, respectively, which are used to characterize phase texture, local spatiotemporal structure and higher-level event patterns, and serve as the multi-layer feature input of the MSRS post-processing module.

[0102] After the third feature stage, the system enters the feature aggregation head, which aggregates spatial features and maps them to fixed-dimensional pre-classification features. The pre-classification features That is, pre-logits, also denoted as ,in, For the first The pre-classification features of each input sample, and the linear classifier for... Perform a linear mapping to obtain Unnormalized score vector That is, logits. For the logits output of a linear classifier, only one Softmax probability normalization is performed to obtain the sample's probability. The relative probability distribution over known categories is calculated as follows:

[0103] (7)

[0104] In the formula, For the first The scores corresponding to each known category For the first The scores corresponding to each known category, and For the classification head pair Sample The output unnormalized score vector, Indicates the first The non-negative response after class score exponential mapping, the denominator pair The responses of the known categories are normalized, so the sum of the probabilities of all categories is 1;

[0105] For the multi-layer representation required for MSRS post-processing, the outputs of the first feature stage, the second feature stage, the third feature stage, and the pre-classification feature level are read respectively. When the stage output is a feature tensor with spatial dimension, it is only sorted into the corresponding hierarchical feature vector during the post-processing reading process.

[0106] During closed-set training, network parameters are learned by minimizing the cross-entropy loss:

[0107] (8)

[0108] In the formula, The number of training samples. Indicates the first Sample True known category labels, Indicates the sample index. Indicates the first Each sample represents the softmax probability corresponding to a known true class, indicating the network's confidence in the correct class of that sample.

[0109] Minimizing the cross-entropy loss will cause samples of the same known class to gradually cluster in the feature space and form separable boundaries between different known classes in the output space.

[0110] Step 2: Introduce the MSRS post-processing module after the closed-set event classifier to perform unknown class rejection during the open-set testing or deployment phase.

[0111] The MSRS post-processing module includes a multi-layer Mahalanobis distance branch, an inter-layer consistency branch, and a maximum softmax probability branch. The multi-layer Mahalanobis distance branch measures the distance between a sample and the known class feature distribution. The inter-layer consistency branch (UCons) characterizes whether the representation of a sample is stable across different layers of the network. The maximum softmax probability (MSP) branch supplements the output layer confidence information. The three branches obtain three types of evidence, which evaluate the reliability of the sample from the perspectives of feature distribution, inter-layer evolution, and output probability, respectively. After robust standardization, the evidence is linearly fused to obtain the final reliability score.

[0112] 1) For the multi-level Mahalanobis distance branch, all samples up to the th The Mahalanobis distance between the centers of each category is defined as:

[0113] (9)

[0114] In the formula, In the first When the original features of the layer are in the first layer The sample to the first The distance of Mahalanobis from the center of each category, For the first The sample obtained by the frozen classifier Layer original features, These represent the pre-classification feature levels after the first feature stage, the second feature stage, the third feature stage, and before the classifier head, respectively. , For the first The first in the layer The feature mean of the nth known category is used to describe the nth... The center position of the training sample in the feature space of this layer. For the first The covariance matrix of a layer is used to describe the degree of dispersion and correlation of each dimension of the features of that layer.

[0115] Equation (9) is used to characterize the degree of deviation of the sample from the statistical distribution of each known category. In combination with the actual processing flow of the program, the minimum value of the distance between each category is first taken for each layer:

[0116] (10)

[0117] In the formula, For the first Layer take each category The minimum distance between samples;

[0118] Then, through negative logarithmic compression transformation, the "smaller the distance, the more reliable" principle is converted into a hierarchical reliability score where "the larger the score, the more reliable".

[0119] (11)

[0120] In the formula, For the first The first layer Original reliability score for each sample;

[0121] This level of scoring is robust as described in this specification. The score standardization method yields hierarchical standardized scores, which are then multi-layered weighted fusions. Finally, the fusion results are subjected to branch-level robust standardization again.

[0122] (12)

[0123] (13)

[0124] (14)

[0125] (15)

[0126] In the formula, for The score after hierarchical robust standardization Indicates the first Layer robustness Score standardization To verify the samples in the set at the 1st The median of the original reliability score for the layer hierarchy. To verify the samples in the set at the 1st The absolute deviation of the original hierarchical reliability score of the layer from its median. For the corresponding robust scaling parameters, For the first The original layer weights corresponding to each feature layer are used to characterize the relative contribution of the features of that layer in the multi-layer Mahalanobis score fusion. For the first The original layer weights corresponding to each feature layer The first layer is obtained by nonnegative normalization of the original layer weights. The normalized weights of each feature layer are such that the sum of the normalized weights of all layers is 1. For the first The branch intermediate score of each sample after multi-level weighted fusion Indicates the robustness of the branch level after multi-level weighted fusion. The scores are standardized, and their central and scaling parameters are predetermined by the corresponding validation reference branch scores. For the first The original scores of each sample were obtained through multi-level Mahalanobis branches and used as Mahalanobis terms in the MSRS three-branch linear fusion.

[0127] 2) The inter-layer consistency branch is used to measure the stability of the representation evolution of samples in the network depth direction. For known-class phase event images, the network has learned a relatively stable feature extraction path during training, and its shallow texture, local structure, and high-level semantic representation usually have a smooth transition relationship. Although unknown samples are also forced to be mapped to the known class space by the network, their feature changes between different layers may be more unstable. Based on this consideration, this invention maps features from different layers to a unified dimension and compares the differences between features of adjacent layers. The smaller the difference between adjacent layers, the more likely the sample is to conform to the known class representation evolution law; the larger the difference, the more likely the sample is to belong to unknown or abnormal input. This branch is also robustly standardized to be fused with other branches at the same scale, specifically including:

[0128] Will The corresponding four-layer vectors are uniformly mapped to a common dimension D. If a certain layer dimension is higher than D, the first D dimensions are truncated; if it is lower than D, zeros are added. The purpose is to enable distance comparison between different layer outputs in the same vector space.

[0129] After unifying the dimensions, perform in-sample robust standardization based on median / MAD:

[0130] (16)

[0131] In the formula, To achieve intermediate features after dimensional unification and robust standardization, To unify operations across dimensions, This represents the absolute deviation of the median. The median in the feature dimension. To prevent stable terms with a denominator of zero.

[0132] Compared to mean and standard deviation, median / MAD is less sensitive to outlier features and is more suitable for in-sample standardization in open set scenarios where there may be out-of-distribution inputs.

[0133] Then, L2 normalization is performed on it to obtain the final features used for inter-layer distance comparison:

[0134] (17)

[0135] In the formula, For the final level features, For L2 normalization;

[0136] L2 normalization can reduce the impact of overall feature magnitude differences on inter-layer distance, allowing subsequent comparisons to focus more on vector direction and relative structure. Therefore, the processing order in the inter-layer consistency branch is clear: original features... →Dimensional unification → Robust standardization intermediate characteristics →L2 normalized features After the above processing is completed, the differences between adjacent layers are processed using the final normalized values. The Euclidean distance is calculated as follows:

[0137] (18)

[0138] (19)

[0139] (20)

[0140] In the formula, , and These are the Euclidean distances between the first feature stage and the second feature stage, the second feature stage and the third feature stage, and the third feature stage and the pre-classification feature level before the classification head, respectively.

[0141] If a sample belongs to a known distribution, its features typically evolve smoothly and gradually along the network depth, with relatively small differences between adjacent layers. Unknown samples are more likely to exhibit unstable cross-layer changes. Since a larger inter-layer distance indicates a less reliable sample, to maintain consistency with the overall scoring semantics of "the larger the score, the more reliable," we directly take the negative average of the three inter-layer distances to obtain the original score of the inter-layer consistency branch:

[0142] (twenty one)

[0143] In the formula, The original score for the interlayer consistency branch is denoted as . The larger the score, the smaller the distance between the three adjacent layers and the more stable the cross-layer representation evolution of the sample.

[0144] 3) The maximum softmax probability branch uses the maximum softmax probability to reflect the confidence level of the classifier's output, which most strongly leans towards the known class. This can be written as:

[0145] (twenty two)

[0146] In the formula, The original score of the MSP branch before the branch-level robust standardization is given. The larger the score, the more the closed-set classifier tends to classify the sample as a known class. This branch is simple to calculate and can supplement the confidence information at the classification head level. However, since the softmax probability is only normalized within the set of known classes, unknown samples may also obtain a high maximum probability. Therefore, MSP cannot be used alone to handle the task of rejecting unknown classes.

[0147] After obtaining the original scores for each branch, MSRS uniformly performs a robust evaluation based on the score distribution of the validation reference samples before final linear fusion. Score standardization:

[0148] (twenty three)

[0149] In the formula, For the first The original scores of each branch, and , For the first Standardized scores for each branch To verify the samples in the set at the 1st The median of the original scores for each branch;

[0150] Taking into account the multi-layer Mahalanobis branch, inter-layer consistency branch, and MSP branch, the final reliability score (MSRS score) is defined as follows:

[0151] (twenty four)

[0152] In the formula, , and The scores obtained by applying the existing robust standardization process of this invention to the multi-level Mahalanobis branches, inter-level consistency branches, and MSP branches are represented. , and These represent the weights of the three types of scores in the final reliability score;

[0153] This linear fusion form has clear interpretability: the multi-layer Mahalanobis branch provides evidence of class-conditional distribution, the inter-layer consistency branch supplements the stability of the network's internal representation evolution, the MSP branch preserves the output layer confidence information, and the weights are not manually adjusted on the test set, but determined by a discrete search on the validation partition.

[0154] The MSRS post-processing module calculates a reliability score for the input samples during the open set testing or deployment phase, and uses this score to determine whether to accept the known class or reject the unknown class.

[0155] Step 4: Input the two-dimensional phase event image into the φ-OTDR open set event recognition model to obtain the MSRS score, and perform threshold selection and open set discrimination:

[0156] The semantics of the MSRS score is "the higher the score, the closer the sample is to the known class distribution." Therefore, after obtaining continuous reliability scores, it is necessary to further set a reliability threshold. The open set discrimination rule divides the samples into acceptable known class samples and unknown class samples that need to be rejected.

[0157] (25)

[0158] In the formula, when When a sample is considered sufficiently close to the known class distribution, the system accepts the class with the highest probability given by the closed-set classifier as the known class prediction; when At this time, although the sample can still be mapped to a known category by the closed-set classifier, its reliability is insufficient, so it is rejected as an unknown class.

[0159] The reliability threshold is adaptively determined through a validation set: the threshold is selected through multiple experiments, and the search range for candidate thresholds is determined by the actual distribution of the final reliability scores on the validation set. in and These are the minimum and maximum values ​​of the final reliability score for the validation set, respectively.

[0160] In this embodiment, under the current robust z-score cutoff parameter setting, the theoretical range of the final reliability score is [-12, 12]. The candidate threshold is composed of the midpoint of adjacent different validation scores and the two ends of the score range. The optimal reliability threshold is selected based on the open set accuracy under each candidate threshold. =-3.572342.

[0161] This step actually separates the category prediction of the closed-set classifier from the rejection decision of the open-set classifier: the closed-set classifier is responsible for providing candidate known categories, and the MSRS is responsible for judging whether the candidate result is trustworthy.

[0162] To verify the performance of the method described in this invention, the following tests were conducted:

[0163] To measure the performance of open sets from an engineering application perspective, this invention mainly uses the following metrics: known sample coverage, unknown sample rejection rate, and open set accuracy. Known sample coverage represents the proportion of known samples accepted by the system under a given threshold out of all known samples; unknown sample rejection rate represents the proportion of unknown samples correctly identified as unknown categories; and open set accuracy is a comprehensive statistical measure of the overall proportion of known samples that are accepted and correctly classified, as well as the proportion of unknown samples that are rejected.

[0164] Furthermore, this invention employs curve-based metrics such as AUROC, AUPR-In, AUPR-Error, and FPR95 to evaluate the overall separation capability of different scoring methods for known and unknown samples. The combined use of curve-based metrics and engineering working point metrics helps to simultaneously examine the model's separation performance and deployment availability.

[0165] The φ-OTDR water monitoring dataset, comprising 37,337 samples, was used as the mixed test set. This set included 32,120 known samples and 5,217 unknown samples. The known samples encompassed seven known categories: 10 kHz, 200 Hz, background noise, model aircraft sound, rowing boat sound, speedboat sound, and impact sound. The known samples from the mixed test set were used as the closed-set test set for both closed-set classification and open-set tests.

[0166] like Figure 3 As shown, the vibration response signal is constructed into a two-dimensional phase event image, where the brightness of the pixels represents the phase magnitude at the corresponding time-space location, and the change of local gray level with time or space reflects the change law of phase response. By using a unified fixed value range of [-π, π] for mapping, the phase-gray level correspondence between different event samples can be kept consistent, avoiding the change of the original phase scale between different samples due to image-by-image adaptive stretching.

[0167] like Figure 4 , 5 As shown, the visualization results of the fused weighted search are presented, using each change... and The correct recognition rate obtained by the weights is plotted as a graph. This graph shows that the system mainly relies on the multi-layer Mahalanobis statistical branches to provide evidence of class conditional distribution. At the same time, it uses inter-layer consistency information to correct the stability of the internal representation evolution of the network and retains a certain output confidence as a supplement. In the experiment of this invention, the optimal peak point was used, the fusion weight ratio was α=0.55, β=0.20, γ=0.25, the judgment threshold was -3.572342, and the recognition accuracy of the open set reached 0.9753.

[0168] Table 1 presents the overall performance metrics for the closed-set test on a closed-set test set containing only 7 known categories:

[0169] Table 1 Overall classification performance indicators of the closed set test set

[0170]

[0171] As can be seen, the closed-set event classifier achieved an accuracy of 97.12%. Furthermore, the mean precision, mean recall, and mean F1 score reached 0.9738, 0.9739, and 0.9738, respectively, with a weighted average F1 score of 0.9712. These results indicate that the model has high overall recognition ability and exhibits good balance among different categories. In particular, the differences in mean precision and weighted average precision are small, indicating that its classification performance does not have obvious category bias and can provide a stable closed-set classification foundation for subsequent open-set rejection experiments.

[0172] Table 2 shows the classification performance of each category under the closed-set test:

[0173] Table 2 Classification Performance of Each Category

[0174]

[0175] As shown in Table 2, the precision, recall, and F1 score of the four categories of 10 kHz, 200 Hz, background noise, and speedboat sound are all close to 0.99 or above, indicating that these events have good separability in the feature space. In contrast, the indices of model aircraft flight noise, rowing boat sound, and knocking sound are slightly lower. Among them, the F1 score of rowing boat sound is 0.9356, which is the lowest among all categories, suggesting that there is some feature overlap between it and adjacent categories.

[0176] like Figure 6 , 7 As shown, the original confusion matrix and normalized confusion matrix of the closed test set are given respectively. Overall, the main diagonal of the matrix is ​​dominant, indicating that most samples can be correctly identified. Further observation of the easily confused categories shows that some samples of model aircraft flight noise are misclassified as rowing sounds and knocking sounds; there is a significant bidirectional confusion between rowing sounds and knocking sounds; there are only a few cross-classifications among the 10 kHz, 200 Hz and background noise categories, which indicates that the boundaries of high-frequency, low-frequency and background noise samples are relatively clear, and the vibration modes of ship-related events are more similar.

[0177] like Figure 8 , 9As shown, the ROC curves and false positive rate-recall curves for the closed-set multi-class classification task are further illustrated. The ROC curves for each category are generally close to the upper left corner, and the AUC is basically maintained above 0.995, indicating that the model has a strong class discrimination ability. The false positive rate-recall curves are also generally close to the upper right corner. The average precision (AP) for model aircraft flight noise is 0.9929, while the APs for rowing sound and knocking sound are 0.9816 and 0.9862, respectively. Although slightly lower than other categories, they are still at a high level. Combining the confusion matrix and the curve-shaped indicators, it can be seen that the closed-set classifier has formed a relatively stable discrimination basis.

[0178] On the mixed test set, without rejection, the model achieves a closed-set accuracy of 83.58%; when only known class ground truth samples are counted, the closed-set accuracy is 97.16%. The average inference time per batch is 0.0060 s. Based on JSON statistics, the φ-OTDR open-set event recognition model achieves an open-set accuracy of 97.44% at the current threshold, meaning that the overall ratio of correctly accepting and classifying known samples and correctly rejecting unknown samples reaches 97.44%.

[0179] like Figure 10 As shown, the confusion matrix of the open set containing the unknown class is given. It can be seen that 5213 out of 5217 samples of the unknown class were correctly rejected, and only 4 were misclassified as knocking, corresponding to a rejection rate of 99.92% for the unknown class. At the same time, the 10 kHz, 200 Hz and background noise classes maintained a high number of diagonal counts, indicating that the rejection mechanism did not significantly damage the recognition results of the known class.

[0180] Table 3 shows a comparison of the engineering metrics for open set recognition using five post-processing methods on a mixed test set, listing representative performance indicators such as known sample coverage, unknown class rejection rate, and open set accuracy.

[0181] Table 3 Comparison of Engineering Indicators for Open Set Recognition (including rejection rate for unknown classes)

[0182]

[0183] Overall, both Energy and Entropy scores can achieve a certain degree of separation between known and unknown samples, indicating the feasibility of constructing post-processing scores based on the output distribution in this task. Energy's unknown class rejection rate and open set accuracy reached 96.17% and 91.75%, respectively, higher than Entropy's 92.41% and 91.07%. This suggests that compared to directly utilizing the predicted distribution entropy, energy scores can more fully characterize the discriminative boundary between known and unknown samples in this task. ViM, as a strong post-processing baseline, demonstrated a high level of performance on engineering workpoint metrics, with known sample coverage and unknown class rejection rates reaching 99.25% and 99.60%, respectively. This indicates that the virtual logit matching strategy can maintain a high pass rate for known samples while suppressing false acceptance of unknown samples. In contrast, OpenMax can also achieve a certain degree of known / unknown separation under the current data configuration, but its known sample coverage and open set accuracy are 88.09% and 87.84%, respectively, which are generally weaker than Energy and ViM. Compared to MSRS, OpenMax typically uses the activation vector output from the last fully connected (FC) layer of a convolutional neural network. However, the intermediate features of a convolutional model often contain richer spatial, textural, and local structural information. Performing distance modeling only on the last vector may lose the feature differences needed for fine-grained classification. Its principle of conservatively weakening the confidence of known classes to identify unknown classes also impairs the accurate discrimination ability of convolutional models on known classes. In contrast, the method described in this invention achieves optimal or near-optimal results on engineering operating point metrics. According to JSON metrics, at the current threshold, the known sample coverage, unknown class rejection rate, and open set accuracy are 99.87%, 99.92%, and 97.44%, respectively. Compared to ViM, the method described in this invention improves the open set accuracy by 1.96 percentage points, the known sample coverage by 0.62 percentage points, and the unknown class rejection rate by 0.32 percentage points; compared to OpenMax, the open set accuracy is improved by 9.60 percentage points. The results show that the method described in this invention, by integrating multi-layered statistical evidence, inter-layer consistency, and output confidence, can still maintain a strong ability to reject unknowns even with a high retention rate of known samples, demonstrating good comprehensive balance and engineering deployment potential.

[0184] To further analyze the contribution of each component branch to the overall open set discrimination performance, this invention conducts ablation comparisons on the multi-layer Mahalanobis branch, inter-layer consistency branch, and MSP branch in the method described in this invention under the same frozen classifier and the same mixed test set settings. The experiments examine three configurations in sequence: single branch, pairwise combination, and complete fusion of three branches. The focus is on the changes in AUROC, FPR95, and open set accuracy to verify the role of each branch in class-conditional statistical modeling, inter-layer representation stability characterization, and output confidence supplementation. It should be noted that the inter-layer consistency branch only focuses on inter-layer consistency in the discrimination principle and cannot be used as an independent discrimination method. It can only be used as an auxiliary to other methods. Therefore, it is not tested separately. The specific results are shown in Table 4.

[0185] Table 4. Results of the component branch ablation experiment of the method described in this invention.

[0186]

[0187] The results from the single branch show that the multi-level Mahalanobis branch, when used alone, achieved an AUROC of 0.9900, an FPR95 of 0.0045, and an open set accuracy of 90.31%, which is better than the 0.9662, 0.2482, and 88.41% achieved when only the MSP branch was used. This indicates that the class conditional statistical distance is the main source of discrimination in the framework of this invention.

[0188] Further observation of the pairwise fusion results reveals that the overall performance of introducing multi-layer Mahalanobis branches is superior to combinations without these branches. Specifically, the Mahalanobis branch + maximum softmax probability branch achieves an AUROC of 0.9975 and an open-set accuracy of 95.27%, while the Mahalanobis branch + inter-layer consistency branch achieves an AUROC of 0.9953 and an open-set accuracy of 94.32%. The combination consisting only of the inter-layer consistency branch and the MSP branch achieves an open-set accuracy of 93.10%. This indicates that the multi-layer Mahalanobis branch provides the main contribution to the separation of known / unknown information, and the MSP branch, as evidence in the output space, further strengthens the final discrimination boundary.

[0189] Comparing the complete and reduced models, the complete method with three branches achieved an AUROC score of 0.999993 and an open set accuracy of 97.44%, making it the best result among all configurations. Compared to the complete model, removing the MSP branch reduced the open set accuracy to 94.32%, and removing the inter-layer consistency branch reduced it to 95.27%. These results indicate that while the MSP branch is not dominant, it provides direct and stable supplementary evidence for the output layer; the inter-layer consistency branch is mainly used to assist in modeling the evolution of inter-layer representations. Overall, the complete fusion method outperforms any single-branch or dual-branch configuration.

[0190] As can be seen from Tables 3 and 4, the advantage of the method described in this invention in the φ-OTDR open set recognition task is not a simple application of the OOD post-processing method for natural images, but is closely related to the physical imaging mechanism and event representation method of this type of phase signal. The φ-OTDR phase map is composed of the dynamic phase response unfolded along the spatial position of the fiber and continuous time frames. The pixel gray level mainly corresponds to the phase fluctuations caused by external disturbances, rather than the color, edge, or object semantics in the natural image. Therefore, unknown events may not necessarily show a significantly low confidence in the final classification, but may appear in the form of local stripe morphological anomalies, spatiotemporal continuity disruption, phase texture shift, or multi-scale vibration mode mismatch. If only softmax, Energy, or the last layer features are relied upon, it may be difficult to fully capture the hierarchical manifestation of these distribution deviations in the shallow, middle, and high layers of the network. The evidence can measure the distance between the sample and the known class statistical distribution at different levels, such as early phase texture, local spatiotemporal structure, and high-level event representation. This allows the rejection judgment to no longer rely solely on a single output confidence level, but to be based on the consistency of multi-scale class conditional distributions. In Table 4, the Mahalanobis branch alone has achieved an AUROC of 0.9900 and an open set accuracy of 90.31%, which is better than the MSP branch alone (0.9662 and 88.41%). This indicates that the class conditional statistical distance is an important source of discrimination for the method described in this invention. The complete structure further achieves an AUROC of 0.9999 and an open set accuracy of 97.44%, indicating that multi-level statistical evidence has a complementary effect with other reliability information.

[0191] The contribution of the inter-layer consistency branch lies in the fact that it is not a traditional multi-layer feature splicing, weighting, or fusion, but rather an explicit characterization of whether the representation evolution of samples is stable in the network depth direction. Traditional multi-layer feature fusion usually emphasizes the complementarity of information from different layers, with the main goal of enhancing classification expressive power. Open set rejection, on the other hand, focuses more on whether samples follow the normal representation transformation path formed by known classes within the network. For known φ-OTDR events, there is usually a relatively stable progressive relationship between phase texture, local vibration structure, and high-level class representation. Even if unknown perturbations, random garbled signals, or unregistered events are locally similar to some known classes in a certain layer of features, they may exhibit abnormal representation jumps between adjacent layers. The inter-layer consistency branch utilizes this cross-layer change pattern to supplement the Mahalanobis branch's judgment on the static feature distribution distance, thereby helping to identify samples that are "locally similar but evolutionarily unstable." As shown in Table 4, after introducing the inter-layer consistency branch, the open set accuracy of the Mahalanobis branch + inter-layer consistency branch is 90.31% higher than that of the Mahalanobis branch alone. The efficiency was improved to 94.32%; the efficiency of the complete structure was further improved to 97.44% compared to the Mahalanobis branch + MSP branch with this branch removed. This result shows that the interlayer consistency branch provides cross-layer stability evidence that is difficult to replace by traditional multilayer feature fusion.

[0192] This strategy is more suitable for φ-OTDR / DAS signals, also because of the differences in data modality and semantic source between phase event images and natural images. Although the φ-OTDR phase map is organized as a two-dimensional image input network, its two coordinate axes have clear temporal and spatial physical meanings. The sample differences mainly come from the phase response formed by the propagation of disturbances along the fiber position and time frame, rather than the object contours, textures, or scene semantics in natural images. DAS scenes are also affected by factors such as Rayleigh fading, background noise, environmental disturbances, equipment status changes, and compound events. Unknown samples often exhibit phase patterns that are locally similar to known classes but have abnormal overall statistical distributions or spatiotemporal evolution patterns. Therefore, an open set strategy suitable for this task should simultaneously focus on whether the sample is close to the feature distribution of known classes and whether it conforms to the cross-layer representation evolution pattern of known classes. The method described in this invention combines multi-layer Mahalanobis statistical evidence, inter-layer consistency evidence, and output confidence evidence, which can perform a more reliable assessment of φ-OTDR phase event images that conforms to the physical signal characteristics without retraining the closed set classifier.

[0193] Therefore, the effectiveness of the method described in this invention does not merely stem from the transfer of the idea of ​​open set recognition in natural images, but rather from its compatibility with the multi-scale phase texture, spatiotemporal continuous structure, and unknown perturbation characteristics of φ-OTDR / DAS signals. The method described in this invention achieves a good balance between known sample coverage, unknown class rejection rate, and open set accuracy. The complete three-branch configuration outperforms any single-branch or two-branch configuration, all of which support the above analysis from an experimental perspective.

[0194] To further observe the performance of this strategy on open sets in other data modalities, this invention conducts supplementary experiments on the Imagenette natural image dataset. It should be noted that Imagenette is a natural image dataset, which differs from φ-OTDR phase difference event images in terms of data modality, imaging mechanism, and category semantics. Therefore, this experiment is only used to examine the generalization performance of the method on natural image open set tasks and cannot replace the main experimental conclusions based on the φ-OTDR water monitoring dataset mentioned above.

[0195] Six classes in Imagenette are used as known classes, and the remaining four classes are used as unknown classes. During the training phase, 5666 samples of known classes are used; during the validation phase, 2328 known samples and 1198 unknown samples are used, and the total number of samples in the mixed test set is 3526.

[0196] The backbone network used in this supplementary experiment was still MobileNetV4-Conv-Medium, with an MSRS threshold of -1.417932. ViM and OpenMax were selected as comparison methods, and all three methods were evaluated on the same mixed test set. The experimental results are shown in Table 5. The evaluation metrics included AUROC, AUPR-In, AUPR-Error, and FPR95. AUROC reflects the model's overall ability to distinguish between in-distribution and out-of-distribution samples; a higher value indicates better discriminative performance. AUPR-In measures the model's recognition performance on in-distribution samples and is more valuable in cases of class imbalance. AUPR-Error evaluates the model's ability to detect misclassified samples; a higher value indicates stronger error detection capability. FPR95 represents the false positive rate when the true positive rate reaches 95%, measuring the degree of misacceptance of out-of-distribution samples; a lower value is better. The normalized open-set confusion matrix of MSRS on this dataset is shown below. Figure 11 As shown.

[0197] Table 5. Recognition results of the Imagenette dataset open set.

[0198]

[0199] like Figure 11As shown, the open set discrimination performance of the method described in this invention in the Imagenette supplementary experiment is further demonstrated at the category level. The matrix is ​​normalized row by row according to the true category. It can be seen that the normalized correct rejection rate of the unknown class is 0.98, which is basically consistent with the rejection rate of 0.9843 for unknown samples in Table 5. The diagonal values ​​of most known classes remain above 0.93, indicating that the method described in this invention still has a relatively stable recognition ability on known classes of natural images.

[0200] Meanwhile, the confusion matrix also revealed the main sources of error in this supplementary experiment. 0.30 samples from class n03000684 were classified as unknown, and 0.11 and 0.07 samples from classes n03394916 and n02102040, respectively, were rejected as "Unknown." Unknown samples were almost never mistakenly accepted as known classes. This phenomenon corresponds to the known sample coverage of approximately 0.9012 achieved by the method described in this invention, indicating that at this operating point, the model error is more often manifested as over-rejection of known samples than misacceptance of unknown samples. Therefore, the method described in this invention exhibits a conservative but beneficial open-set decision-making characteristic on this dataset, which helps reduce the risk of misacceptance of unknown classes.

[0201] As shown in Table 5, on the Imagenette mixed test set, all three methods exhibit a certain ability to separate known and unknown samples. ViM's AUROC is 0.9732, slightly higher than the 0.9682 of the method described in this invention and 0.9615 of OpenMax, indicating that ViM performs well in overall ranking metrics. Looking at the open set discrimination results at a given operating point, the open set accuracy of the method described in this invention is 0.9259, higher than ViM's 0.9232 and OpenMax's 0.9244. Meanwhile, the unknown sample rejection rate of the method described in this invention reaches 0.9843, the highest among the three methods, indicating that under the current threshold strategy, the method described in this invention is more inclined to robustly reject unknown samples.

[0202] It should be noted that the known sample coverage rate of the method described in this invention is 0.9012, which is lower than OpenMax's 0.9167 and only slightly close to ViM's 0.9008. This indicates that in order to improve the rejection rate of unknown samples, this method sacrifices some of the pass rate of known samples. In terms of error types, the total number of erroneous samples in this experiment was 248, of which 230 known samples were rejected as unknown, 2 known samples were accepted but misclassified, and 16 unknown samples were incorrectly accepted as known. It can be seen that the main source of error is not that unknown samples are mistakenly accepted as known, but rather that some known samples are rejected as unknown by the system, indicating that the overall discrimination strategy is conservative.

[0203] This conservative characteristic is somewhat reasonable in security monitoring tasks. In real open environments, misclassifying unknown anomalies or unregistered events as a known category often leads to a high risk of misjudgment based on confidence. In contrast, while ensuring the necessary coverage of known classes, moderately increasing the rejection rate of unknown samples helps reduce the probability of the system misreceiving unknown inputs. Therefore, the Imagenette supplementary experiment illustrates from another dataset perspective that the multi-layer statistical evidence, inter-layer consistency, and output confidence fusion mechanism of the method described in this invention still has a certain degree of effectiveness in natural image open set recognition tasks.

[0204] This invention designs and develops a φ-OTDR open-set event recognition method based on multi-layer Mahalanobis-Softmax reliability scoring, addressing the open-set event recognition problem in aquatic monitoring scenarios. It uses a trained and frozen closed-set event classifier as a foundation, extracting Mahalanobis evidence from multi-stage features and combining inter-layer consistency branches and output confidence branches for robust standardization and linear fusion. This provides a unified sample reliability metric and a decision-making basis for known class acceptance / unknown class rejection during the open-set testing and deployment stages without altering the original classification network structure and training process. It significantly improves the ability to reject unknown classes, maintains high known sample coverage even with high unknown rejection rates, increases the overall accuracy of open-set recognition, reduces the risk of high-confidence misjudgment of unknown events by the closed-set classifier, fully utilizes the feature information of intermediate layers, and improves discrimination robustness through the stability of inter-layer representation evolution. The post-processing method facilitates engineering deployment, does not rely on external unknown samples for training, is suitable for long-term online operation of safety monitoring systems, and has good generalization potential.

[0205] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A method for identifying open set events with inter-layer consistency, characterized in that, Includes the following steps: Step 1: Based on the φ-OTDR / DAS distributed acoustic sensing system, acquire vibration response signals distributed along the optical fiber; Step 2: Perform coherent detection, phase demodulation, and time-space mapping on the vibration response signal to construct a two-dimensional phase event image; Step 3: Construct the φ-OTDR open set event recognition model; The φ-OTDR open set event recognition model includes an MSRS post-processing module, which includes a multi-layer Mahalanobis distance branch, an inter-layer consistency branch, and an MSP branch. Step 4: Input the two-dimensional phase event image into the φ-OTDR open set event recognition model to obtain the final reliability score and perform open set discrimination; The open set discrimination satisfies: ； In the formula, For open sets, For unknown class, For the first The Softmax probability normalized value of each sample. For the first The final reliability score for each sample This is the reliability threshold.

2. The method for identifying open set events with inter-layer consistency according to claim 1, characterized in that, The time-space mapping satisfies: ； ； In the formula, The propagation time of the echo relative to the probe pulse. The speed of light in a vacuum. The effective refractive index of the optical fiber. The position of the optical fiber is given by the coefficient 2, which reflects the round-trip propagation process of the probe light within the fiber. For dynamic phase terms, The strain sensitivity coefficient is related to the fiber parameters and the operating wavelength. This is local strain.

3. The method for identifying open set events with inter-layer consistency according to claim 2, characterized in that, The two-dimensional phase event image includes: 224 consecutive time frames are selected from the continuous time-space phase matrix, and 224 spatial positions are selected in the spatial direction to obtain a phase event matrix with a size of 224×224; The phase event matrix is ​​plotted as a two-dimensional phase event image in time-space, with its vertical axis corresponding to continuous time frames and its horizontal axis corresponding to the spatial positions arranged sequentially along the sensing optical fiber.

4. The method for identifying open set events with inter-layer consistency according to claim 3, characterized in that, The φ-OTDR open set event recognition model also includes a closed set event classifier, which uses MobileNetV4-Conv-Medium as the event recognition network.

5. The method for identifying open set events with inter-layer consistency according to claim 4, characterized in that, The multi-layer Mahalanobis distance branch satisfies: ； In the formula, For the first The branch intermediate score of each sample after multi-level weighted fusion Indicates the robustness of the branch level after multi-level weighted fusion. Score standardization For the first The original scores of each sample are obtained through multiple Mahalanobis distance branches.

6. The method for identifying open set events with inter-layer consistency according to claim 5, characterized in that, The inter-layer consistency branch satisfies: ； In the formula, For the first The original scores of each sample after the inter-stratum consistency branch, , and These are the Euclidean distances between the first feature stage and the second feature stage, the second feature stage and the third feature stage, and the third feature stage and the pre-classification feature level before the classification head, respectively.

7. The method for identifying open set events with inter-layer consistency according to claim 6, characterized in that, The MSP branch satisfies: ； In the formula, For the first The original scores of each sample after passing through the MSP branch.

8. The method for identifying open set events with inter-layer consistency according to claim 7, characterized in that, Step three also includes: Robustly evaluate the original scores of the multilevel Mahalanobis distance branch, the inter-level consistency branch, and the original MSP score. Score standardization: ； In the formula, For the first The original scores of each branch, and , For the first Standardized scores for each branch To verify the samples in the set at the 1st The median of the original scores for each branch. This represents the absolute deviation of the median.

9. The method for identifying open set events with inter-layer consistency according to claim 8, characterized in that, The final reliability score satisfies: ； In the formula, , and They represent the first The scores were obtained from a robust standardization process using multiple Mahalanobis distance branches, inter-layer consistency branches, and MSP branches. , and These represent the weights of the three types of scores in the final reliability score. This is used to determine the final reliability score.

10. The method for identifying open set events with inter-layer consistency according to claim 9, characterized in that, The range of values ​​for the reliability threshold satisfies: ； in, and These are the minimum and maximum values ​​of the final reliability score for the validation set, respectively.

Images