Agv continuous positioning system and method based on fiber grating sensing and deep learning and medium

Abstract

The present application relates to the technical field of optical fiber sensing and intelligent positioning, and discloses an AGV continuous positioning system and method based on optical fiber grating sensing and deep learning and a medium. The method comprises the following steps: filtering and processing the multi-channel time sequence response signal of the AGV, cutting to form windowed samples, processing the windowed samples to obtain processed windowed samples, extracting the time sequence difference features of the windowed samples for adaptive modulation to obtain the adaptive modulation time sequence difference features, splicing and fusing the processed windowed samples and the adaptive modulation time sequence difference features to construct a multi-dimensional joint feature representation, and performing statistical standardization processing to serve as a running data set; inputting the running data set into the trained deep learning positioning network to obtain the AGV running continuous positioning result. The method can stably obtain the AGV running response signal in a complex industrial environment, realize AGV continuous positioning, and has good anti-interference performance, positioning accuracy and engineering implementability.

Description

Technical Field

[0001] This invention relates to the field of fiber optic sensing and intelligent positioning technology, specifically to an AGV continuous positioning system, method, and medium based on fiber optic grating sensing and deep learning. Background Technology

[0002] Automated Guided Vehicles (AGVs) are crucial transportation equipment in smart manufacturing, unmanned warehousing, and rail logistics. Their positioning accuracy and continuous positioning capabilities directly impact operational safety, scheduling efficiency, and system stability. Existing AGV positioning methods primarily include radio frequency identification (RFI), vision, QR codes, magnetic nails / strips, and laser-synchronized localization and mapping (SLAM) technologies. Among these, laser SLAM is susceptible to dust, obstructions, and environmental changes, and is complex and computationally expensive. While QR codes, magnetic nails, and magnetic strips are relatively simple to implement, they typically only provide accurate positioning information at discrete locations, failing to meet the requirements for continuous positioning along the entire path, and are prone to error accumulation between reference points.

[0003] Fiber optic sensing technology, with its advantages of strong resistance to electromagnetic interference, high sensitivity, and suitability for long-distance deployment and multiplexing, has promising applications in industrial sensing. In particular, fiber optic grating sensing technology, compared to some scattering-based distributed fiber optic sensing methods, offers a higher signal-to-noise ratio and a clearer positional correspondence, making it more promising for positioning and detection. However, current fiber optic sensing research largely focuses on traffic monitoring, structural health monitoring, vehicle speed recognition, or vibration detection, with limited research on continuous high-precision positioning of AGVs in closed industrial environments. Furthermore, the multi-channel fiber optic response signals generated by AGV operation exhibit significant spatiotemporal coupling and nonlinear characteristics, making it difficult for traditional manual feature extraction methods to fully utilize the effective information within them.

[0004] In recent years, deep learning has shown promising prospects in signal processing, pattern recognition, and fiber optic sensing data interpretation. However, most existing solutions are geared towards tasks such as strain recognition, condition monitoring, or signal demodulation, and lack technical solutions for AGV positioning. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned problems in the prior art and provide an AGV continuous positioning system, method and medium based on fiber optic grating sensing and deep learning. This method can stably acquire AGV operation response signals and realize AGV continuous positioning in complex industrial environments, while taking into account anti-interference, positioning accuracy and engineering feasibility, so as to meet the application requirements of smart factories for highly reliable and highly continuous positioning technology.

[0006] To achieve the above objectives, the present invention provides an AGV continuous positioning system based on fiber Bragg grating sensing and deep learning, comprising: a multi-channel time-series response signal determination module, which uses a fiber Bragg grating sensor to receive the strain response and vibration response of the AGV during operation, and then generates a spectral displacement signal; uses a fiber Bragg grating demodulation instrument to collect the spectral displacement signal; and then demodulates the spectral displacement signal to obtain the multi-channel time-series response signal of the AGV. The dataset construction module performs unified frequency band filtering on the multi-channel time-series response signal of the AGV to obtain a filtered multi-channel continuous time-series signal. The filtered multi-channel continuous time-series signal is then segmented to form multiple windowed samples. Nonlinear amplitude limiting and outlier suppression are applied to each windowed sample to obtain a processed windowed sample. The temporal difference features of each windowed sample are extracted, and each temporal difference feature is adaptively modulated to obtain an adaptively modulated temporal difference feature. Each processed windowed sample is concatenated and fused with the corresponding adaptively modulated temporal difference feature to construct multiple multi-dimensional joint feature representations. Each multi-dimensional joint feature representation is statistically standardized to obtain standardized multi-dimensional joint features, which are then used as the running dataset. The positioning confirmation module inputs the running dataset into the trained deep learning positioning network to obtain continuous positioning results of the AGV operation.

[0007] A second aspect of this invention provides a continuous positioning method for AGVs based on fiber optic grating sensing and deep learning, comprising: The strain and vibration responses of the AGV during operation are received by a fiber Bragg grating sensor, and then a spectral displacement signal is generated. The spectral displacement signal is collected by a fiber Bragg grating demodulator, and then demodulated to obtain the multi-channel timing response signal of the AGV. The multi-channel timing response signal of the AGV is subjected to unified frequency band filtering to obtain a filtered multi-channel continuous timing signal. The filtered multi-channel continuous timing signal is then segmented into multiple windowed samples. Nonlinear amplitude limiting and outlier suppression are applied to each windowed sample to obtain a processed windowed sample. The timing difference features of each windowed sample are extracted and adaptively modulated to obtain an adaptively modulated timing difference feature. Each processed windowed sample is then concatenated and fused with the corresponding adaptively modulated timing difference feature to construct multiple multi-dimensional joint feature representations. Each multi-dimensional joint feature representation is then statistically standardized to obtain standardized multi-dimensional joint features. The standardized multi-dimensional joint features are used as the running dataset. Input the running dataset into the trained deep learning localization network to obtain continuous localization results of AGV operation.

[0008] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0009] The purpose of this invention is to overcome the technical shortcomings of existing AGV positioning methods, such as insufficient continuity, limited anti-interference ability, low positioning accuracy, and difficulty in effectively processing multi-channel fiber optic response signals. This invention proposes a continuous AGV positioning method based on the fusion of fiber optic grating sensing and deep learning to achieve the following technical objectives: 1. Provide a stable AGV signal acquisition system to reliably acquire multi-channel fiber optic response signals caused by AGV movement in real time under industrial scenarios with strong electromagnetic interference and complex environments.

[0010] 2. To address the problems of existing discrete positioning methods being unable to achieve continuous positioning along the entire path, as well as the significant spatiotemporal coupling, prominent nonlinear characteristics, and difficulty in effectively representing multi-channel fiber optic response signals, thereby improving the continuous position perception capability and the ability to extract positioning-related feature information during AGV operation.

[0011] 3. A technical solution is provided that integrates fiber optic grating sensing, signal preprocessing, feature representation and deep learning positioning algorithm to achieve continuous high-precision positioning of AGV in complex industrial environments. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the structure of an AGV continuous positioning system based on fiber optic grating sensing and deep learning. Figure 2 This is a schematic diagram of the original signal preprocessing and sample signal construction process; Figure 3 This is a schematic diagram of dividing a continuous time-series signal into samples using a sliding time window. Figure 4 This is a schematic diagram of the deep learning network framework for AGV positioning; Figure 5 This is a positioning result diagram of an AGV positioning method based on fiber Bragg gratings and deep learning algorithms; Figure 6 This is a schematic diagram of the continuous positioning method for AGVs based on fiber optic grating sensing and deep learning. Detailed Implementation

[0013] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0014] The endpoints and any values ​​of the ranges disclosed herein are not limited to the precise ranges or values, and these ranges or values ​​should be understood to include values ​​close to these ranges or values. For numerical ranges, the endpoint values ​​of the various ranges, the endpoint values ​​of the various ranges and individual point values, and individual point values ​​can be combined with each other to obtain one or more new numerical ranges, which should be considered as specifically disclosed herein.

[0015] Furthermore, the technical solutions provided in the various embodiments of the present invention can be combined with each other, but only if they are feasible to those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0016] Example 1 One such Figure 1 The AGV continuous positioning system based on fiber Bragg grating sensing and deep learning shown includes: a multi-channel time-series response signal determination module, which uses fiber Bragg grating sensors to receive the strain response and vibration response of the AGV during operation, and then generates a spectral displacement signal. The spectral displacement signal is collected using a fiber Bragg grating demodulation instrument, and then demodulated to obtain the multi-channel time-series response signal of the AGV. The dataset construction module performs unified frequency band filtering on the multi-channel time-series response signal of the AGV to obtain a filtered multi-channel continuous time-series signal. The filtered multi-channel continuous time-series signal is then segmented to form multiple windowed samples. Nonlinear amplitude limiting and outlier suppression are applied to each windowed sample to obtain a processed windowed sample. The temporal difference features of each windowed sample are extracted, and each temporal difference feature is adaptively modulated to obtain an adaptively modulated temporal difference feature. Each processed windowed sample is concatenated and fused with the corresponding adaptively modulated temporal difference feature to construct multiple multi-dimensional joint feature representations. Each multi-dimensional joint feature representation is statistically standardized to obtain standardized multi-dimensional joint features, which are then used as the running dataset. The positioning confirmation module inputs the running dataset into the trained deep learning positioning network to obtain continuous positioning results of the AGV operation.

[0017] In this invention, the multi-channel timing response signal determination module is used to complete the physical deployment of the fiber optic grating sensing track and acquire the multi-channel response signals during AGV operation.

[0018] Regarding the signal acquisition operation of the fiber Bragg grating sensor, specifically, based on the pre-designed travel route of the AGV, the laying path of the sensing track is marked on the industrial site ground to match the actual running path of the AGV. Then, the ground is cut and grooved along the predetermined laying path, and the fiber Bragg grating sensor is laid and fixed at the predetermined position in the groove. After completion, epoxy resin or other encapsulation materials are filled into the groove for backfilling and leveling. The groove width is 8cm and the groove depth is 5cm, completing the construction of the AGV positioning and sensing track; the spacing between adjacent sensing units of the fiber Bragg grating sensor is 3m. In this way, during the actual operation of the AGV, as it travels along the preset route, the fiber Bragg grating sensor can receive the influence of the AGV's real-time operation, or in other words, the physical influence of the AGV's movement on it through the ground transmission medium, and generate a corresponding spectral displacement signal. (The fiber Bragg grating sensor is used to capture external vibration and strain responses, which are fed back to itself, manifesting as changes in grating spacing and refractive index. This invention utilizes the fiber Bragg grating sensor to sense the mechanical response acting on the sensing optical cable during AGV operation. The mechanical response includes strain and vibration responses caused by the AGV wheel load, and a spectral displacement signal is generated based on the mechanical response.) The sampling frequency is typically 1000Hz. Then, a fiber Bragg grating demodulation instrument is used to collect the spectral displacement signal and demodulate it (the demodulator is a DAS demodulation system, demodulated using the principle of interference; the demodulated signal shows a dynamic shift in the phase term of the interference signal (due to external influence)). This yields a corresponding multi-channel time-series response signal. This multi-channel time-series response signal can reflect the dynamic characteristics of the AGV's effects on each sensing area at different locations.

[0019] The steps of running the dataset building module are used to convert the acquired multi-channel time-series response signals into standardized input samples that can be directly used by deep learning models. Specific steps are as follows: Figure 2 As shown.

[0020] Furthermore, the multi-channel timing response signal of the AGV undergoes unified frequency band filtering (this invention uses a bandpass filter with a filtering frequency range of 0.025Hz-0.25Hz to extract effective signal components related to AGV positioning and suppress high-frequency noise, environmental interference, and irrelevant disturbance components. Using a unified preprocessing frequency band for signals acquired under different operating conditions improves the consistency of data processing under different speeds, loads, or experimental conditions), resulting in the filtered continuous timing signal. The multi-channel timing response signal of the AGV at the sampling index The response signal is represented as follows: in, Indicates the first A fiber grating sensor at the sampling index The vibration response signal under the following conditions, among which ; Indicates the total number of channels; Indicates the AGV at the sampling index The multi-channel timing response signal; among which... In This indicates transpose.

[0021] For AGVs at the sampling index The multi-channel timing response signals are subjected to unified frequency band filtering: in, Indicates the sampling index The multi-channel continuous-time signal after down-filtering; This represents a filtering operator.

[0022] In this invention, the filtered continuous time-series signal is segmented to form multiple windowed samples, such as... Figure 3 As shown, each windowed sample corresponds to a multi-channel response segment within a local time period. The windowed sample is represented as follows: in, Indicates the first The current sampling index of the windowed sample; Indicates the first A windowed sample; Indicates the width of the time window; Indicates the sampling index The multi-channel continuous timing signal after down-filtering.

[0023] Preferably, to reduce the adverse effects of individual anomalous shocks, local mutations, and amplitude outliers on the results, it is also necessary to perform nonlinear amplitude limiting and outlier suppression processing on each windowed sample to keep the sample amplitude within a preset range, thereby improving the stability of data distribution and the robustness of subsequent calculations. Specifically, the process of performing nonlinear amplitude limiting and outlier suppression processing on each windowed sample to obtain the processed windowed sample includes: Where β represents the amplitude scaling factor; α represents the amplitude constraint threshold; Indicates the first A processed windowed sample. In this embodiment... , .

[0024] Based on this processing, while retaining effective trend information, the interference of weak response samples or abnormal samples on feature expression can be reduced.

[0025] In a preferred embodiment, based on each windowed sample, it is necessary to further extract time-series differential features that reflect the signal change trend and local dynamic characteristics. The time-series differential features include first-order time-series differential features and second-order time-series differential features, which are used to characterize the signal change rate and change acceleration, thereby enhancing the ability to describe the dynamic operation characteristics of the AGV.

[0026] Specifically, the process of extracting the temporal difference features of each windowed sample includes: The first-order time-series difference feature matrix and the second-order time-series difference feature matrix are defined as follows: in, Indicates the sampling index The next first-order temporal difference time characteristics; Indicates the sampling index Second-order temporal difference time characteristics; Indicates the first The first-order temporal difference feature matrix of a windowed sample; Indicates the first The second-order temporal difference feature matrix of windowed samples.

[0027] In this invention, based on the overall response intensity or energy level of the current windowed sample, each temporal difference feature needs to be adaptively modulated to enhance the transient vibration information related to AGV motion. The operation of adaptively modulating each temporal difference feature to obtain the adaptively modulated temporal difference feature includes: Dynamic weighting coefficients of temporal difference features for: in, in, Indicates the first The average energy of a windowed sample; This represents the average energy normalization coefficient; This indicates the lower bound of the dynamic weighting coefficients; This indicates the upper limit of the dynamic weighting coefficients; Indicates the first In the nth windowed sample The first channel in the Normalized time response values ​​at each time sampling point; Indicates the channel index. ; Indicates the index of the time sampling point. ; Indicates the total number of channels; Indicates the width of the time window.

[0028] Adaptive modulation of first-order and second-order time-series difference features is performed using this dynamic weighting coefficient: in, Indicates the first The first-order temporal difference feature matrix after adaptive modulation of windowed samples; Indicates the first The second-order temporal difference feature matrix after adaptive modulation of windowed samples.

[0029] In this invention, each processed windowed sample is spliced ​​and fused with the corresponding adaptively modulated temporal difference feature to construct multiple multidimensional joint feature representations containing original amplitude information and dynamic change information, thereby forming multidimensional input samples for AGV continuous positioning tasks.

[0030] Specifically, the expression for the multidimensional joint feature representation is: ,and ; in, For the first Multidimensional joint feature representation of a windowed sample; Represents the set of all real numbers; Indicates the total number of channels; Indicates the width of the time window.

[0031] In this invention, each multidimensional joint feature representation is statistically standardized to give each feature a uniform data scale and distribution range, thereby reducing the dimensional differences between different channels and feature components and improving computational stability and convergence efficiency.

[0032] In a specific instance of this invention, the method for obtaining standardized multidimensional joint features by statistically standardizing each multidimensional joint feature representation includes: in, This represents the mean of the corresponding windowed sample data; This represents the standard deviation of the corresponding windowed sample data; This represents the standardized multidimensional feature representation.

[0033] Through the above steps, this invention achieves the conversion from multi-channel raw response signals to standardized positioning samples that can be directly used by the model.

[0034] Next, the running dataset is input into the trained deep learning localization network. As the sliding window is continuously updated, the network continuously outputs the localization values ​​corresponding to each time step, thus forming the continuous localization results of the AGV running along the predetermined path.

[0035] In this invention, the historical operation dataset of AGVs is used to train the deep learning localization network. This historical operation dataset is constructed by further combining standardized historical multidimensional joint features with the actual physical locations corresponding to the historical windowed samples (based on the actual physical locations of the AGVs within that time period, corresponding location labels are assigned, thus establishing a one-to-one correspondence between historical windowed samples and location labels. The actual physical locations are obtained through other high-precision positioning methods, such as laser SLAM and reflectors. These actual physical locations are only used during calibration training and not during actual prediction; therefore, the standardized multidimensional joint features are directly used as the operation dataset in actual measurement). This is represented as follows: in, For the first A standardized historical multidimensional feature representation; Indicates the first The actual physical location corresponding to each historical windowed sample; Represents the historical running dataset; This represents the total number of samples in the historical running dataset.

[0036] Specifically, the method for obtaining the standardized historical multidimensional joint features is consistent with the method for obtaining standardized multidimensional joint features, except that it uses the strain response and vibration response of the AGV during historical operation and then performs corresponding processing. The specific operation is as follows: The strain and vibration responses of the AGV during its historical operation are received using fiber Bragg grating sensors, and then historical spectral displacement signals are generated. The historical spectral displacement signals are collected using fiber Bragg grating demodulation instruments, and then demodulated to obtain the historical multi-channel timing response signals of the AGV. The historical multi-channel timing response signals of the AGV are subjected to unified frequency band filtering to obtain filtered historical multi-channel continuous timing signals. The filtered historical multi-channel continuous timing signals are then segmented to form multiple historical windowed samples. Nonlinear amplitude limiting and outlier suppression are applied to each historical windowed sample to obtain processed historical windowed samples. The timing difference features of each historical windowed sample are extracted and adaptively modulated to obtain adaptively modulated historical timing difference features. Each processed historical windowed sample is concatenated and fused with the corresponding adaptively modulated historical timing difference features to construct multiple historical multidimensional joint feature representations. Each historical multidimensional joint feature representation is then statistically standardized to obtain standardized historical multidimensional joint features.

[0037] The historical dataset was used for training, validation, and testing of the deep learning localization network. The training, validation, and test sets were divided into sets of 80%, 10%, and 10%, respectively.

[0038] Historical operational data is input into the deep learning localization network for training. The samples simultaneously contain raw response information, temporal change information, and multi-channel coupling information.

[0039] like Figure 4 As shown, the methods for training a deep learning localization network using historical runtime datasets to obtain the trained deep learning localization network include: Local feature extraction is performed on historical datasets using convolutional neural networks. One-dimensional convolution operations are used to extract local patterns and key response structures from multi-channel signals, resulting in a convolutional feature sequence, expressed as: in, This represents a one-dimensional convolution operation; Indicates batch normalization; Represents the ReLU activation function; This represents the max pooling operation; This represents the result after the first convolutional layer. This represents the result after the second convolutional layer, and it sets the convolutional feature sequence... .

[0040] The convolutional neural network includes convolutional layers, batch normalization layers (a batch in batch normalization refers to a set of samples input to the network each time during model training. This layer normalizes the convolutional output by statistically analyzing the mean and variance of the current batch features, thereby stabilizing the feature distribution, alleviating the problem of constantly changing input distributions of different layers during training, and improving the model's convergence speed and training stability), nonlinear activation layers, and pooling layers. The convolutional feature extraction module consists of two one-dimensional convolutional layers, used to extract local dynamic response features and inter-channel coupling features from multi-channel temporal inputs. First, the input features are processed by the first one-dimensional convolutional layer, with the convolutional kernel sliding along the time dimension to capture short-term local change patterns generated when the AGV passes through different sensing areas. Subsequently, the convolutional outputs are sequentially processed by… The system employs a batch normalization layer and a non-linear activation function. Batch normalization stabilizes feature distribution and improves convergence during model training, while the non-linear activation function enhances the network's ability to express responses to complex non-linear signals. Building upon this, a pooling layer downsamples the first-layer features in the temporal dimension, compressing redundant information while preserving key response features and reducing the computational complexity of subsequent models. The pooled features are then further input into a second one-dimensional convolutional layer to extract higher-level temporal abstract features, and batch normalization and non-linear activation further enhance feature representation. Finally, the high-dimensional feature representation output by the convolutional module provides effective input for subsequent attention mechanisms and LSTM temporal modeling, enhancing feature extraction capabilities, improving training stability, and compressing redundant information. Specifically, the system has two convolutional layers: the first layer has 64 channels and an 11-kernel size; the second layer has 128 channels and a 5-kernel size.

[0041] Next, the convolutional feature sequence is input into the self-attention module to model the correlation between different time positions, enabling the network to adaptively focus on key time segments and feature response regions that are more important to the localization task. This invention employs a multi-head self-attention mechanism (with 4 attention heads) to enhance the convolutional feature sequence, thereby improving its ability to model long-range dependencies and global contextual information, resulting in enhanced features. : Convolutional feature sequence is , This represents the last convolutional feature vector in the convolutional feature sequence; This represents the total length of the convolutional feature sequence; the operation of the self-attention mechanism is as follows: in, Represents the query matrix; Represents the key matrix; Represents a value matrix; Represents the learnable weight matrix; Indicates the feature embedding dimension; In This indicates transpose.

[0042] By residually fusing enhanced features with convolutional feature sequences, global dependency information is introduced while preserving local convolutional features, thereby improving feature representation ability and enhancing the stability of information transmission during network training, resulting in a fused feature sequence. ,Right now ; The fused feature sequence is input into an LSTM (Long Short-Term Memory) network to model the time dependencies during the continuous operation of the AGV. This invention employs a two-layer stacked unidirectional LSTM structure to utilize historical feature information to characterize the current position change trend and meet the requirement that real-time inference in continuous positioning scenarios relies solely on historical information. The hidden layer has a dimension of 128, and the output is a hidden state sequence. The LSTM state update process is as follows: in, Indicates time step Input features; Indicates the current hidden state; This indicates the hidden state of the previous step.

[0043] The hidden state sequences are aggregated to extract high-level temporal features that characterize the overall temporal evolution process, while utilizing the hidden states at the last time step. and global average features Joint representation is performed to consider both the instantaneous state information near the current position and the overall temporal information within the entire window, resulting in high-level temporal features. Specifically, an aggregated feature vector is obtained through feature concatenation, expressed as: in, Indicates the hidden state at the last time step; This represents the global feature obtained by average pooling over the entire sequence; This represents the high-level temporal features obtained after aggregation; Indicates the hidden state at the last time step; Indicates the first The hidden state vector corresponding to each time step .

[0044] Input the high-level time series features into the fully connected regression module, and output the first... AGV position prediction value corresponding to each historical windowed sample This establishes a regression mapping relationship between multidimensional input samples and the actual position of the AGV, expressed as: in, This represents a regression network consisting of multiple fully connected layers; The first AGV position prediction value corresponding to each historical windowed sample With actual physical location Error calculation is performed, and the parameters of the deep learning localization network are optimized using a loss function. The process includes: No. Positioning error of historical windowed samples for: .

[0045] To improve the model's adaptability to continuous localization regression tasks and reduce the impact of outlier errors on the training process, this invention uses Huber loss as the basic loss function: in, This represents the Huber loss function; δ is the switching threshold. Considering the need to further suppress large local errors during continuous positioning, this invention introduces a large error penalty term; Final loss function for: Where τ represents the outlier threshold and λ is the penalty weight coefficient.

[0046] During training, mini-batch samples are input into the network. The learnable parameters in the convolutional layers, self-attention modules, LSTM modules, and fully connected regression modules are iteratively updated using the backpropagation algorithm and optimizer, so that the model output gradually approximates the real AGV position labels. After training, the model parameters with the best performance on the validation set are selected as the final localization model to achieve continuous AGV position prediction.

[0047] The deep learning localization network, after training with the dataset, first enters a convolutional neural network module consisting of two one-dimensional convolutional layers. The first convolutional layer slides along the time dimension to extract short-term local vibration responses and channel coupling features. The output then passes through a batch normalization layer to stabilize the feature distribution, a ReLU nonlinear activation layer to mine nonlinear signal relationships, and a max pooling layer to perform temporal downsampling and compress redundant data. The pooling result is fed into the second convolutional layer to extract higher-order temporal abstract features, and then undergoes batch normalization and activation processing again to finally generate a convolutional feature sequence. Next, the convolutional feature sequence is input into a 4-head multi-head self-attention module. By constructing three sets of learnable weight matrices (query, key, and value), the correlation between features is calculated, and temporal segments crucial for localization are adaptively selected to generate enhanced features. Subsequently, the enhanced features are residually fused with the original convolutional feature sequence, preserving local details while supplementing global long-range dependency information to obtain a fused feature sequence. The fused feature sequence is fed into a two-layer stacked unidirectional LSTM network. The LSTM's internal door control mechanism records the temporal evolution pattern, and historical window features characterize the continuous change trend of the AGV's position. The complete hidden state sequence is calculated iteratively layer by layer. Feature aggregation is performed on all hidden state sequences to extract the instantaneous hidden state at the last time step. Simultaneously, global average pooling is performed on the entire sequence to obtain global temporal features. The two sets of features are concatenated and fused to generate high-level temporal features that consider both instantaneous states and overall window changes. These high-level temporal features are input into a multi-layer fully connected regression module. Multi-layer linear mapping completes the regression operation from features to physical coordinates, outputting the AGV's instantaneous position prediction value corresponding to the current window sample. The acquisition device continuously acquires the AGV's fiber optic vibration signal in real time, updating the time window in a fixed step size. The entire process of signal preprocessing, dataset construction, and network forward inference is repeatedly executed, continuously outputting the real-time position prediction value corresponding to each sliding window. The position prediction results under continuous temporal sequence are concatenated sequentially to obtain the continuous positioning result of the AGV along its entire running track.

[0048] Through the above steps, this invention utilizes the synergistic effect of convolutional feature extraction, self-attention feature enhancement, LSTM temporal modeling, and position regression output to achieve effective parsing of multi-channel fiber optic response signals and complete continuous high-precision positioning of AGVs. The specific positioning results are as follows: Figure 5 As shown.

[0049] The technical effects of this invention are as follows: 1. This invention achieves stable acquisition and real-time demodulation of multi-channel fiber optic response signals by constructing a fiber optic grating sensing track deployed along the AGV's running path. It can adapt to industrial application scenarios with strong electromagnetic interference and complex environments, providing a reliable signal foundation for continuous AGV positioning.

[0050] 2. This invention improves the ability to extract positioning-related feature information by performing signal preprocessing on multi-channel optical fiber response signals, and enhances the stability and effectiveness of signal characterization under complex working conditions.

[0051] 3. This invention proposes a deep learning localization method that integrates CNN, self-attention mechanism and LSTM temporal modeling. It can effectively learn the nonlinear mapping relationship between multi-channel fiber response signals and AGV position, realize continuous high-precision localization during AGV operation, and improve the continuity and accuracy of localization results.

[0052] 4. The present invention has been verified by simulation and actual measurement, and has shown good positioning accuracy, repeatability and robustness. Under repeated operation at the same speed and under different speed and load conditions, the positioning accuracy reaches an average positioning error of no more than 3cm, indicating that the present invention has good engineering feasibility and prospects for promotion and application.

[0053] Example 2 like Figure 6 The AGV continuous positioning method based on fiber optic grating sensing and deep learning shown includes: The strain and vibration responses of the AGV during operation are received by a fiber Bragg grating sensor, and then a spectral displacement signal is generated. The spectral displacement signal is collected by a fiber Bragg grating demodulator, and then demodulated to obtain the multi-channel timing response signal of the AGV. The multi-channel timing response signal of the AGV is subjected to unified frequency band filtering to obtain a filtered multi-channel continuous timing signal. The filtered multi-channel continuous timing signal is then segmented into multiple windowed samples. Nonlinear amplitude limiting and outlier suppression are applied to each windowed sample to obtain a processed windowed sample. The timing difference features of each windowed sample are extracted and adaptively modulated to obtain an adaptively modulated timing difference feature. Each processed windowed sample is then concatenated and fused with the corresponding adaptively modulated timing difference feature to construct multiple multi-dimensional joint feature representations. Each multi-dimensional joint feature representation is then statistically standardized to obtain standardized multi-dimensional joint features. The standardized multi-dimensional joint features are used as the running dataset. Input the running dataset into the trained deep learning localization network to obtain continuous localization results of AGV operation.

[0054] Example 3 A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method as described in Embodiment 2.

[0055] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0056] The preferred embodiments of the present invention have been described in detail above; however, the present invention is not limited thereto. Within the scope of the inventive concept, various simple modifications can be made to the technical solutions of the present invention, including combinations of various technical features in any other suitable manner. These simple modifications and combinations should also be considered as the content disclosed in the present invention and are all within the protection scope of the present invention.

Claims

1. An AGV continuous positioning system based on fiber Bragg grating sensing and deep learning, characterized in that, include: The multi-channel timing response signal determination module uses fiber Bragg grating sensors to receive the strain and vibration responses of the AGV during operation, and then generates a spectral displacement signal. The spectral displacement signal is acquired using a fiber Bragg grating demodulator, and then demodulated to obtain the multi-channel timing response signal of the AGV. The dataset construction module performs unified frequency band filtering on the multi-channel time-series response signal of the AGV to obtain a filtered multi-channel continuous time-series signal. The filtered multi-channel continuous time-series signal is then segmented to form multiple windowed samples. Nonlinear amplitude limiting and outlier suppression are applied to each windowed sample to obtain a processed windowed sample. The temporal difference features of each windowed sample are extracted, and each temporal difference feature is adaptively modulated to obtain an adaptively modulated temporal difference feature. Each processed windowed sample is concatenated and fused with the corresponding adaptively modulated temporal difference feature to construct multiple multi-dimensional joint feature representations. Each multi-dimensional joint feature representation is statistically standardized to obtain standardized multi-dimensional joint features, which are then used as the running dataset. The positioning confirmation module inputs the running dataset into the trained deep learning positioning network to obtain continuous positioning results of the AGV operation.

2. The AGV continuous positioning system based on fiber optic grating sensing and deep learning according to claim 1, characterized in that, The process of performing unified frequency band filtering on the multi-channel timing response signals of AGV to obtain the filtered continuous timing signal includes: The multi-channel timing response signal of the AGV at the sampling index The response signal is represented as follows: in, Indicates the first A fiber grating sensor at the sampling index The vibration response signal under the following conditions, among which ; Indicates the total number of channels; Indicates the AGV at the sampling index The multi-channel timing response signal; among which... In Indicates transpose; For AGVs at the sampling index The multi-channel timing response signals are subjected to unified frequency band filtering: in, Indicates the sampling index The multi-channel continuous-time signal after down-filtering; This represents a filtering operator.

3. The AGV continuous positioning system based on fiber optic grating sensing and deep learning according to claim 1, characterized in that, Windowed samples are represented as: in, Indicates the first The current sampling index of the windowed sample; Indicates the first A windowed sample; Indicates the width of the time window; Indicates the sampling index The multi-channel continuous timing signal after down-filtering.

4. The AGV continuous positioning system based on fiber optic grating sensing and deep learning according to claim 1, characterized in that, The process of performing nonlinear amplitude limiting and outlier suppression on each windowed sample to obtain the processed windowed sample includes: in, Indicates the amplitude scaling factor; Indicates the amplitude constraint threshold; Indicates the first A processed windowed sample.

5. The AGV continuous positioning system based on fiber optic grating sensing and deep learning according to claim 4, characterized in that, The temporal difference features include first-order temporal difference features and second-order temporal difference features; The process of extracting the temporal difference features of each windowed sample includes: The first-order time difference feature matrix and the second-order time difference feature matrix are defined as follows: in, Indicates the sampling index The next first-order temporal difference time characteristics; Indicates the sampling index Second-order temporal difference time characteristics; Indicates the first The first-order temporal difference feature matrix of a windowed sample; Indicates the first The second-order temporal difference feature matrix of windowed samples; The operation of adaptively modulating each temporal difference feature to obtain the adaptively modulated temporal difference feature includes: Dynamic weighting coefficients of temporal difference features for: in, in, Indicates the first The average energy of a windowed sample; This represents the average energy normalization coefficient; This indicates the lower bound of the dynamic weighting coefficients; This indicates the upper limit of the dynamic weighting coefficients; Indicates the first In the nth windowed sample The first channel in the Normalized time response values ​​at each time sampling point; Indicates the channel index. ; Indicates the index of the time sampling point. ; Indicates the total number of channels; Indicates the width of the time window; Adaptive modulation of first-order and second-order time-series difference features is performed using this dynamic weighting coefficient: in, Indicates the first The first-order temporal difference feature matrix after adaptive modulation of windowed samples; Indicates the first The second-order temporal difference feature matrix after adaptive modulation of windowed samples.

6. The AGV continuous positioning system based on fiber optic grating sensing and deep learning according to claim 1, characterized in that, The expression for the multidimensional joint feature representation is: ,and ; in, For the first Multidimensional joint feature representation of a windowed sample; Represents the set of all real numbers; Indicates the total number of channels; Indicates the width of the time window.

7. The AGV continuous positioning system based on fiber optic grating sensing and deep learning according to claim 6, characterized in that, Methods for statistically standardizing each multidimensional joint feature representation to obtain standardized multidimensional joint features include: in, This represents the mean of the corresponding windowed sample data; This represents the standard deviation of the corresponding windowed sample data; Indicates the first A standardized multidimensional feature representation.

8. The AGV continuous positioning system based on fiber optic grating sensing and deep learning according to claim 1, characterized in that, Methods for obtaining trained deep learning localization networks include: The standardized historical multidimensional joint features are combined with the actual physical locations corresponding to the historical windowed samples to construct a historical operation dataset: in, For the first A standardized historical multidimensional feature representation; Indicates the first The actual physical location corresponding to each historical windowed sample; Represents the historical running dataset; This represents the total number of samples in the historical running dataset; Using a convolutional neural network to extract local features from a historical dataset, we obtain a convolutional feature sequence, expressed as: in, This represents a one-dimensional convolution operation; Indicates batch normalization; Represents the ReLU activation function; This represents the max pooling operation; This represents the result after the first convolutional layer. This represents the result after the second convolutional layer, and it sets the convolutional feature sequence... ; The convolutional neural network includes convolutional layers, batch normalization layers, nonlinear activation layers, and pooling layers; Multi-head self-attention mechanism is used to enhance convolutional feature sequences to obtain enhanced features. : Convolutional feature sequence is , This represents the last convolutional feature vector in the convolutional feature sequence; This represents the total length of the convolutional feature sequence; the operation of the self-attention mechanism is as follows: in, Represents the query matrix; Represents the key matrix; Represents a value matrix; Represents the learnable weight matrix; Indicates the feature embedding dimension; In Indicates transpose; The enhanced features and convolutional feature sequences are residually fused to obtain the fused feature sequence. ,Right now ; The fused feature sequence is input into the LSTM, and the output is the hidden state sequence. The LSTM state update process is as follows: in, Indicates time step Input features; Indicates the current hidden state; This indicates the hidden state from the previous step; The hidden state sequences are aggregated to obtain high-level temporal features, expressed as follows: in, Indicates the hidden state at the last time step; This represents the global feature obtained by average pooling over the entire sequence; This represents the high-level temporal features obtained after aggregation; Indicates the hidden state at the last time step; Indicates the first The hidden state vector corresponding to each time step ; Input the high-level time series features into the fully connected regression module, and output the first... AGV position prediction value corresponding to each historical windowed sample The expression is: in, This represents a regression network consisting of multiple fully connected layers; The first AGV position prediction value corresponding to each historical windowed sample With actual physical location Error calculation is performed, and the parameters of the deep learning localization network are optimized using a loss function. The process includes: No. Positioning error of historical windowed samples for: ; We use Huber loss as the basic loss function: in, Represents the Huber loss function; The switching threshold; Final loss function for: in, This represents the threshold for identifying outliers; This is the penalty weighting coefficient.

9. A continuous positioning method for AGVs based on fiber optic grating sensing and deep learning, characterized in that, include: The strain and vibration responses of the AGV during operation are received by a fiber Bragg grating sensor, and then a spectral displacement signal is generated. The spectral displacement signal is collected by a fiber Bragg grating demodulator, and then demodulated to obtain the multi-channel timing response signal of the AGV. The multi-channel timing response signal of the AGV is subjected to unified frequency band filtering to obtain a filtered multi-channel continuous timing signal. The filtered multi-channel continuous timing signal is then segmented into multiple windowed samples. Nonlinear amplitude limiting and outlier suppression are applied to each windowed sample to obtain a processed windowed sample. The timing difference features of each windowed sample are extracted and adaptively modulated to obtain an adaptively modulated timing difference feature. Each processed windowed sample is then concatenated and fused with the corresponding adaptively modulated timing difference feature to construct multiple multi-dimensional joint feature representations. Each multi-dimensional joint feature representation is then statistically standardized to obtain standardized multi-dimensional joint features. The standardized multi-dimensional joint features are used as the running dataset. Input the running dataset into the trained deep learning localization network to obtain continuous localization results of AGV operation.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in claim 9.

Images