A multi-parameter demodulation method for optical fiber speckle based on parallel deep learning

Abstract

The present invention discloses a multi-parameter demodulation method for optical fiber speckle based on parallel deep learning, which relates to the field of optical fiber sensing. The method comprises the following steps: applying a combined perturbation to a multimode optical fiber, synchronously collecting speckle images and corresponding curvature, vibration intensity, and position coordinate labels, and constructing a multi-parameter mapping dataset; dynamically dividing the dataset into subsets based on computing power, and independently training a feature extraction subnetwork with each subset to generate a pre-trained weight set; constructing a parallel neural network architecture, loading and freezing the pre-trained weight set, and training a fusion classification network to obtain an optimized fusion classification network; inputting the speckle image to be measured into the optimized fusion classification network, and outputting the corresponding parameter demodulation results. Through the parallel deep learning architecture and the reuse mechanism of pre-trained weights, this method achieves high-precision synchronous demodulation of curvature, vibration, and position coordinates in optical fiber speckle, combining strong anti-interference performance with efficient computing power utilization.

Description

Technical Field

[0001] The present invention relates to the field of optical fiber sensing, and more particularly to a multi-parameter demodulation method for optical fiber speckle based on parallel deep learning. Background Art

[0002] Fiber optic sensing technology, especially multi-parameter sensing based on multimode fiber speckle pattern analysis, has attracted widespread attention in fields such as structural health monitoring and industrial process control due to its high sensitivity, resistance to electromagnetic interference and distributed sensing potential.

[0003] Currently, traditional fiber optic speckle sensing methods typically focus on detecting a single parameter, making them ineffective in complex scenarios involving multiple disturbances acting simultaneously or intertwined in real-world applications. Existing technologies often face severe interference from cross-sensitivity when handling multi-parameter demodulation, resulting in a significant decrease in demodulation accuracy. Furthermore, methods that serially process different parameters or rely on complex physical models are inefficient and cannot meet the demands of applications with demanding real-time performance, limiting the technology's practical application.

[0004] Furthermore, traditional methods based on speckle image correlation analysis or simple machine learning models are extremely sensitive to the inherent random noise and environmental interference of speckles, resulting in poor robustness. These methods struggle to effectively capture the complex, multi-scale features inherent in speckle patterns, which are associated with variations in various physical quantities. Recognition accuracy plummets, especially in noisy environments or when faced with unseen data, and generalization capabilities are insufficient. Existing model architectures exhibit significant limitations in terms of comprehensive feature extraction and resistance to overfitting.

[0005] Furthermore, to achieve high accuracy, some existing methods significantly increase computational complexity and parameter count. This not only makes training and inference time-consuming and labor-intensive, but also places stringent demands on hardware computing power. It also makes lightweight deployment difficult in resource-constrained embedded systems or industrial field monitoring equipment. Traditional methods lack effective computational load optimization mechanisms, making it difficult to balance high accuracy with high efficiency, and high computing power with low power consumption.

[0006] Therefore, how to design a multi-parameter demodulation method for fiber speckle based on parallel deep learning that can efficiently and synchronously demodulate multiple key sensing parameters in multimode fiber speckle and have good generalization ability and low computing power resource consumption is an urgent problem that needs to be solved by technical personnel in this field. Summary of the Invention

[0007] In view of this, the present invention provides a multi-parameter demodulation method for optical fiber speckle based on parallel deep learning. By constructing a parallel deep learning architecture and a pre-trained weight reuse mechanism, high-precision synchronous recognition of bending curvature, vibration intensity and position coordinates is achieved, the demodulation efficiency and anti-interference performance are improved, and a lightweight deployment solution is provided for embedded devices.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A multi-parameter demodulation method for optical fiber speckle based on parallel deep learning includes the following steps:

[0010] S1. By applying combined perturbations to the multimode optical fiber, the speckle image and the corresponding curvature, vibration intensity, and position coordinate labels are collected synchronously to construct a multi-parameter mapping dataset;

[0011] S2. Dynamically divide the data set into subsets based on computing power, and use each subset to independently train a feature extraction subnetwork to generate a pre-trained weight set;

[0012] S3. Build a parallel neural network architecture, load and freeze the pre-trained weight set, perform fusion classification network training, and obtain an optimized fusion classification network;

[0013] S4. Input the speckle image to be measured into the optimized fusion classification network, and output the corresponding parameter demodulation result.

[0014] Preferably, in S1, applying the combined disturbance includes:

[0015] The bending curvature disturbance is applied by the optical fiber curvature parameter regulator, the vibration gear disturbance is applied by the vibration intensity controller, and the disturbance coordinate position is located by the optical fiber fixing device.

[0016] Preferably, in S2, the feature extraction subnetwork includes:

[0017] Pyramid Attention Module: Parallel 1×1, 3×3, and 7×7 convolutional layers, each with 256 output channels, extract local details, mid-scale, and global context features, respectively.

[0018] Feature fusion module: concatenates the three 256-channel features output by PAM into a 768-channel feature map, and reduces the dimension to 256 channels through 1×1 convolution;

[0019] Classification output module: The residual block, global average pooling layer and linear classification layer are connected in sequence to output the single parameter classification result.

[0020] Preferably, in the pyramid attention module, the step sizes of the 1×1 convolution layer, the 3×3 convolution layer, and the 7×7 convolution layer are all 1, and the 1×1 convolution layer is padded with 0, the 3×3 convolution layer is padded with 1, and the 7×7 convolution layer is padded with 3.

[0021] Preferably, the residual block includes a 7×7 depth convolution layer, a LayerNorm normalization layer, a linear projection layer, a GELU activation function, and a random depth dropout layer connected in sequence to form a gradient stabilization path.

[0022] Preferably, the random depth drop layer randomly skips the calculation path of the current residual block with a preset probability p, so that the input data is directly passed to the next layer.

[0023] Preferably, in S3, the fusion classification network includes:

[0024] Weight loading module: loads pre-trained weight sets and freezes parameters;

[0025] Dynamic sharding module: divides the input batch into n equal-sized shards according to the number n;

[0026] Feature extraction module: contains n computing units that reuse the S2 subnet structure. Each unit independently processes a slice and outputs a 256-dimensional feature vector.

[0027] Adaptive merging layer: concatenate n 256-dimensional feature vectors into 256×n-dimensional fusion features;

[0028] Fully connected classifier: Synchronously outputs three types of labels based on fused features.

[0029] Preferably, the feature extraction module satisfies:

[0030] The network structure of each computing unit is consistent with the feature extraction subnetwork in S2, and the parameters are initialized by the pre-trained weight set.

[0031] Preferably, in S3, the fusion classification network training adopts a dynamic optimization strategy:

[0032] In the initial stage, the learning rate is 0.001 and the batch size is 32;

[0033] When the training loss and validation loss reach a plateau, reduce the learning rate to 0.0001 and the batch size to 16;

[0034] The training is terminated when the loss function fluctuation is less than 5%.

[0035] Preferably, said S3 also includes model validity verification;

[0036] An independent test set was used as input to optimize the model, and the confusion matrix and F1 score of the three types of parameters were calculated. The model was considered valid when the curvature recognition F1 score was greater than 95%, the vibration intensity was greater than 92%, and the position coordinates were greater than 90%.

[0037] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0038] 1. This method effectively addresses the issue of multi-parameter cross-sensitivity by constructing a parallel neural network architecture, dynamically loading and freezing pre-trained subnetwork weights, and integrating multi-source features with an adaptive fusion layer. A sharding mechanism is employed to process multimodal data in parallel, achieving simultaneous output of curvature, vibration intensity, and position coordinates. This significantly improves the demodulation efficiency and overall performance of multimode fiber speckle sensing while maintaining high recognition accuracy.

[0039] 2. The feature extraction subnet integrates a pyramid attention module with a residual structure, capturing local details and global features through multi-scale convolutions and introducing a random depth dropout layer to mitigate overfitting. Independent subsets are used in the pre-training phase to learn differentiated features, providing diverse initialization for the parallel network. During the fusion training phase, parameters are frozen and a dynamic optimization strategy is employed to improve the model's adaptability to noise and unseen data.

[0040] 3. Computational load balancing is achieved through modular design. Standard convolution replaces complex attention mechanisms to reduce parameter count, while dynamic sharding is used to parallelize input data and maximize GPU utilization. The adaptive merging layer automatically adjusts the fusion dimension to avoid redundant computation, significantly reducing hardware computing power requirements. This facilitates deployment in resource-constrained embedded systems, such as industrial field monitoring equipment, and improves the feasibility of technology implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0042] Figure 1 A flowchart of a multi-parameter demodulation method for optical fiber speckle based on parallel deep learning provided by an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of the feature extraction subnetwork structure provided by an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of the fusion classification network structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] like Figure 1 As shown, this embodiment provides a method for multi-parameter demodulation of optical fiber speckle based on parallel deep learning, comprising the following steps:

[0047] S1. By applying combined perturbations to the multimode optical fiber, the speckle image and the corresponding curvature, vibration intensity, and position coordinate labels are collected synchronously to construct a multi-parameter mapping dataset;

[0048] S2. Dynamically divide the data set into subsets based on computing power, and use each subset to independently train a feature extraction subnetwork to generate a pre-trained weight set;

[0049] S3. Build a parallel neural network architecture, load and freeze the pre-trained weight set, perform fusion classification network training, and obtain an optimized fusion classification network;

[0050] S4. Input the speckle image to be measured into the optimized fusion classification network, and output the corresponding parameter demodulation result.

[0051] By constructing a parallel neural network architecture and loading frozen pre-trained subnetwork weights, high-precision synchronous demodulation of multiple parameters such as curvature, vibration intensity, and position coordinates in optical fiber speckle is achieved. At the same time, the combination of the pyramid attention module and dynamic optimization strategy significantly enhances the model's anti-interference ability and generalization. Its modular design and dynamic sharding mechanism optimize computing resource utilization, providing an efficient solution for embedded deployment.

[0052] The following is a further detailed description of each step of the above method;

[0053] In this embodiment S1, a multi-parameter mapping data set is constructed by applying a combined perturbation to a multimode optical fiber and synchronously collecting a speckle image and corresponding curvature, vibration intensity, and position coordinate labels;

[0054] The combined disturbances include:

[0055] The bending curvature disturbance is applied by the optical fiber curvature parameter regulator, the vibration gear disturbance is applied by the vibration intensity controller, and the disturbance coordinate position is located by the optical fiber fixing device.

[0056] In this step, the combined perturbation is implemented collaboratively through three devices: a fiber curvature parameter regulator controls the fiber's bend radius through a precision mechanical structure, with each curvature value calibrated in real time by a laser ranging sensor; a vibration intensity controller generates standard vibration waveforms with preset gears (low, medium, and high) based on a piezoelectric ceramic actuator, with the acceleration amplitude verified by a MEMS sensor; and a fiber fixture precisely locates the perturbation point coordinates using a three-dimensional mobile platform and gridded positioning holes. These three devices are synchronously triggered by a central controller, ensuring that the perturbation is applied simultaneously to the CCD at the moment of exposure, achieving strict spatiotemporal alignment of the speckle image with the curvature, vibration intensity, and position coordinate labels.

[0057] It realizes high-fidelity data acquisition of multi-parameter coupled disturbances: the mechanical-electronic control joint system accurately reproduces the real-world complex working conditions, and the synchronous trigger mechanism eliminates timing errors, providing a multi-parameter mapping data set with strict physical correspondence for model training, solving the data distortion problem caused by disturbance separation experiments.

[0058] In this embodiment S2, the data set is dynamically divided into subsets based on computing power, and a feature extraction subnetwork is independently trained using each subset to generate a pre-trained weight set;

[0059] like Figure 2 As shown, the feature extraction subnetwork includes:

[0060] Pyramid Attention Module: Parallel 1×1, 3×3, and 7×7 convolutional layers, each with 256 output channels, extract local details, mid-scale, and global context features, respectively.

[0061] Feature fusion module: concatenates the three 256-channel features output by PAM into a 768-channel feature map, and reduces the dimension to 256 channels through 1×1 convolution;

[0062] Classification output module: The residual block, global average pooling layer and linear classification layer are connected in sequence to output the single parameter classification result.

[0063] Furthermore, in the pyramid attention module, the stride of the 1×1 convolution layer, 3×3 convolution layer, and 7×7 convolution layer are all 1, and the 1×1 convolution layer is padded with 0, the 3×3 convolution layer is padded with 1, and the 7×7 convolution layer is padded with 3.

[0064] Furthermore, the residual block includes a 7×7 depthwise convolutional layer, a LayerNorm normalization layer, a linear projection layer, a GELU activation function, and a random depth dropout layer connected in sequence to form a gradient stabilization path.

[0065] Furthermore, the random depth dropout layer randomly skips the calculation path of the current residual block with a preset probability p, so that the input data is directly passed to the next layer.

[0066] During pre-training, the feature extraction subnet utilizes a pyramid attention module, using three convolution kernels (1×1, 3×3, and 7×7) in parallel to capture pixel-level details, local patterns, and global distribution characteristics of the speckle image. This information is then fused with multi-scale information through concatenation and 1×1 convolution to enhance feature representation. Combined with an optimized residual block design, this further reduces computational complexity and enhances the model's adaptability to speckle data distribution. A random depth dropout mechanism is also introduced to dynamically skip some residual paths to improve training stability.

[0067] To enhance feature diversity, the pre-training process dynamically partitions mutually exclusive data subsets based on GPU memory, forcing each subnetwork to learn differentiated feature representations. This strategy, combined with multi-scale convolution, effectively overcomes the limitations of a single receptive field and enhances the model's robustness to noise and unseen perturbations. Furthermore, the generated diverse pre-trained weights serve as prior knowledge, providing a multi-dimensional feature foundation for subsequent parallel fusion networks.

[0068] In this embodiment S3, a parallel neural network architecture is constructed, the pre-trained weight set is loaded and frozen, and a fusion classification network is trained to obtain an optimized fusion classification network;

[0069] like Figure 3 As shown, the fusion classification network includes:

[0070] Weight loading module: loads pre-trained weight sets and freezes parameters;

[0071] Dynamic sharding module: divides the input batch into n equal-sized shards according to the number n;

[0072] Feature extraction module: contains n computing units that reuse the S2 subnet structure. Each unit independently processes a slice and outputs a 256-dimensional feature vector.

[0073] Adaptive merging layer: concatenate n 256-dimensional feature vectors into 256×n-dimensional fusion features;

[0074] Fully connected classifier: Synchronously outputs three types of labels based on fused features.

[0075] Furthermore, the feature extraction module satisfies:

[0076] The network structure of each computing unit is consistent with the feature extraction subnetwork in S2, and the parameters are initialized by the pre-trained weight set.

[0077] Furthermore, dynamic optimization strategy is adopted for fusion classification network training:

[0078] In the initial stage, the learning rate is 0.001 and the batch size is 32;

[0079] When the training loss and validation loss reach a plateau, reduce the learning rate to 0.0001 and the batch size to 16;

[0080] The training is terminated when the loss function fluctuation is less than 5%.

[0081] Furthermore, S3 also includes model validity verification;

[0082] An independent test set was used as input to optimize the model, and the confusion matrix and F1 score of the three types of parameters were calculated. The model was considered valid when the curvature recognition F1 score was greater than 95%, the vibration intensity was greater than 92%, and the position coordinates were greater than 90%.

[0083] The parallel neural network it constructs utilizes a weight reuse and dynamic computation architecture: by freezing the weights of pre-trained subnetworks to preserve feature extraction capabilities, the input data is dynamically split into n pieces, processed in parallel by n independent computing units. The 256-dimensional features output by each unit are concatenated and fed into a fully connected layer for simultaneous three-parameter classification. Training utilizes a two-stage dynamic optimization approach: an initial high learning rate accelerates convergence, while when loss plateaus, the learning rate and batch size are simultaneously reduced to fine-tune parameters. Training is terminated when loss fluctuations fall below 5%. Model validity is verified using an independent test set, requiring simultaneous compliance with curvature, vibration, and position recognition accuracy.

[0084] Achieve collaborative optimization of multi-parameter high-precision demodulation and computational efficiency: weight reuse inherits the diversified feature priors of the subnet, and the sharding parallel mechanism increases the inference speed by nearly n times; the dynamic optimization strategy avoids local optimality, and the fusion feature layer fully explores cross-scale correlations, so that the recognition accuracy of the three types of parameters exceeds the threshold of industrial-grade applications and meets the real-time monitoring needs of complex scenarios.

[0085] In this embodiment, S4, the speckle image to be measured is input into the optimized fusion classification network, and the corresponding parameter demodulation result is output. Specifically, this step is further explained in conjunction with specific application scenarios;

[0086] 1) Field data collection and input preprocessing;

[0087] After deploying multimode fiber optic sensors at key points on the bridge, high-speed CCD cameras capture speckle images in real time when vehicles pass through, causing micro-vibrations in the structure. These images are normalized and resized before being fed into an optimized fusion classification network. Before input, illumination consistency correction is performed: by calculating the brightness variance of the background speckle pattern, pixel value shifts caused by ambient light fluctuations are dynamically compensated to ensure that the input data conforms to the training set distribution.

[0088] 2) Parallel network dynamic reasoning process;

[0089] The fusion classification network activates a dynamic sharding mechanism: a single speckle image is divided into n non-overlapping 256×256 sub-images based on the available GPU memory. Each sub-image is assigned to an independent computational unit, which loads frozen pre-trained weights for feature extraction. The pyramid attention module simultaneously analyzes the sub-image's local deformation texture (1×1 convolution), vibration patch boundaries (3×3 convolution), and global strain distribution (7×7 convolution), outputting a 256-dimensional feature vector. The 16 feature vectors are concatenated into a 4096-dimensional fused feature vector via an adaptive merging layer. The fully connected classifier then simultaneously outputs three sets of demodulation results:

[0090] 3) Result output and engineering decision-making;

[0091] The demodulation results are generated and transmitted to the monitoring platform within 50ms. Automatically linked to the BIM model, the coordinates of the abnormal location are highlighted in the 3D bridge diagram, and a curvature limit alarm and vibration spectrum are superimposed. Based on multi-parameter coupling analysis, engineers determined that the continued increase in curvature at that point, accompanied by intermittent moderate vibration, indicated local instability caused by loose bolts and immediately triggered a drone inspection command. The entire process achieves end-to-end demodulation, from speckle image to maintenance decision, with a response speed three times faster than traditional single-parameter serial demodulation.

[0092] This scenario verifies the engineering practicality of the method: through millisecond-level response of parallel architecture, joint interpretation of multi-scale features, and strict spatiotemporal alignment of outputs, it meets the needs of rapid location and quantitative assessment of complex faults in infrastructure monitoring.

[0093] The multi-parameter demodulation method for optical fiber speckle based on parallel deep learning provided in this embodiment achieves high-precision synchronous perception of multimode optical fiber speckle through a parallel deep learning architecture and a multi-parameter coupled demodulation mechanism. It integrates dynamic sharding processing, feature diversity pre-training, and an adaptive optimization strategy to construct an intelligent demodulation system that balances computational efficiency and interference resistance. This provides a reliable solution for real-time diagnosis of complex disturbances in fields such as industrial monitoring and infrastructure safety, and promotes the leapfrog development of optical fiber sensing technology towards intelligent multi-parameter collaborative demodulation.

[0094] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. References to the same or similar parts between the various embodiments are sufficient. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For relevant parts, refer to the method description.

[0095] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-parameter demodulation method for optical fiber speckle based on parallel deep learning, characterized in that: The following steps are involved: S1. By applying combined perturbations to the multimode optical fiber, the speckle image and the corresponding curvature, vibration intensity, and position coordinate labels are collected synchronously to construct a multi-parameter mapping dataset; S2. Dynamically divide the data set into subsets based on computing power, and use each subset to independently train a feature extraction subnetwork to generate a pre-trained weight set; The feature extraction subnetwork includes: a pyramid attention module: a 1×1 convolutional layer, a 3×3 convolutional layer, and a 7×7 convolutional layer are set in parallel, each with 256 output channels, respectively extracting local details, medium-scale, and global context features; a feature fusion module: splicing the three 256-channel features output by the PAM into a 768-channel feature map, and reducing the dimension to 256 channels through 1×1 convolution; a classification output module: a residual block, a global average pooling layer, and a linear classification layer connected in sequence, outputting a single-parameter classification result; S3. Build a parallel neural network architecture, load and freeze the pre-trained weight set, perform fusion classification network training, and obtain an optimized fusion classification network; the fusion classification network includes: a weight loading module: loads the pre-trained weight set and freezes the parameters; a dynamic sharding module: divides the input batch into n equal-sized shards according to the number n; a feature extraction module: includes n computing units that reuse the S2 subnet structure, each unit independently processes a shard and outputs a 256-dimensional feature vector; an adaptive merging layer: splices n 256-dimensional feature vectors into a 256×n-dimensional fusion feature; a fully connected classifier: synchronously outputs three types of labels based on the fusion feature; S4. Input the speckle image to be measured into the optimized fusion classification network, and output the corresponding parameter demodulation result.

2. The method for multi-parameter demodulation of optical fiber speckle based on parallel deep learning according to claim 1, characterized in that: In S1, applying the combined disturbance includes: The bending curvature disturbance is applied by the optical fiber curvature parameter regulator, the vibration gear disturbance is applied by the vibration intensity controller, and the disturbance coordinate position is located by the optical fiber fixing device.

3. The method for multi-parameter demodulation of optical fiber speckle based on parallel deep learning according to claim 1, characterized in that: In the pyramid attention module, the stride of the 1×1 convolution layer, the 3×3 convolution layer, and the 7×7 convolution layer are all 1, and the padding of the 1×1 convolution layer is 0, the padding of the 3×3 convolution layer is 1, and the padding of the 7×7 convolution layer is 3.

4. The method for multi-parameter demodulation of optical fiber speckle based on parallel deep learning according to claim 1, characterized in that: The residual block includes a 7×7 depth convolution layer, a LayerNorm normalization layer, a linear projection layer, a GELU activation function, and a random depth dropout layer connected in sequence to form a gradient stabilization path.

5. The method for multi-parameter demodulation of optical fiber speckle based on parallel deep learning according to claim 4, characterized in that: The random depth drop layer randomly skips the calculation path of the current residual block with a preset probability p, so that the input data is directly passed to the next layer.

6. The method for multi-parameter demodulation of optical fiber speckle based on parallel deep learning according to claim 1, characterized in that: The feature extraction module satisfies: The network structure of each computing unit is consistent with the feature extraction subnetwork in S2, and the parameters are initialized by the pre-trained weight set.

7. The method for multi-parameter demodulation of optical fiber speckle based on parallel deep learning according to claim 1, characterized in that: In S3, a dynamic optimization strategy is used for fusion classification network training: In the initial stage, the learning rate is 0.001 and the batch size is 32; When the training loss and validation loss reach a plateau, reduce the learning rate to 0.0001 and the batch size to 16; The training is terminated when the loss function fluctuation is less than 5%.

8. The method for multi-parameter demodulation of optical fiber speckle based on parallel deep learning according to claim 1, characterized in that: Said S3 also includes model validity verification; An independent test set was used as input to optimize the model, and the confusion matrix and F1 score of the three types of parameters were calculated. The model was considered valid when the curvature recognition F1 score was greater than 95%, the vibration intensity was greater than 92%, and the position coordinates were greater than 90%.

Images