Distributed optical fiber temperature measurement positioning calibration method and system based on raman effect

Abstract

The application provides a distributed optical fiber temperature measurement positioning calibration method and system based on Raman effect, relates to the technical field of optical fiber temperature measurement calibration, and comprises the following steps: acquiring a Raman scattered light signal to extract a temperature distribution; performing expansion causal convolution and linear activation to obtain an optical fiber starting point coordinate prediction value; constructing a graph topology structure, constructing an adaptive adjacency matrix based on a temperature gradient, and performing feature extraction; performing lifting wavelet decomposition and adaptive denoising on the graph topology features, and performing registration and calibration after reconstructing a temperature field.

Description

A Distributed Fiber Optic Temperature Measurement and Positioning Calibration Method and System Based on Raman Effect Technical Field

[0001] This invention relates to the field of fiber optic temperature measurement and calibration technology, and in particular to a distributed fiber optic temperature measurement and positioning calibration method and system based on the Raman effect. Background Technology

[0002] Raman scattering fiber optic temperature measurement technology uses optical fiber as a sensing element to measure temperature by analyzing the Raman scattered light signal transmitted along the fiber. It has advantages such as distributed measurement, high precision, long distance, and resistance to electromagnetic interference, and is widely used in temperature monitoring systems in fields such as power, oil, natural gas, tunnels, and construction.

[0003] In practical applications, Raman scattering fiber optic temperature measurement systems suffer from inaccurate positioning due to factors such as fiber optic transmission characteristics and system errors. Traditional fiber optic temperature measurement positioning methods mainly rely on optical path delay to determine the spatial location of the measurement point. However, existing technologies still have shortcomings, including failing to effectively handle fiber optic starting point position offset, being unable to adapt to dynamic changes in fiber optic deployment under complex environments, lacking full utilization of temperature gradient characteristics, failing to accurately capture temperature change characteristics, reducing the system's ability to identify hotspot anomalies and positioning accuracy, and struggling to balance noise suppression and detail preservation. Furthermore, they are prone to losing key thermal feature information in complex temperature fields, affecting the quality of temperature field reconstruction and positioning accuracy. Summary of the Invention

[0004] This invention provides a distributed optical fiber temperature measurement and positioning calibration method and system based on the Raman effect, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of this invention provides a distributed optical fiber temperature measurement and positioning calibration method based on the Raman effect, comprising:

[0006] The Raman scattered light signal distributed along the optical fiber transmission path is obtained, and the temperature distribution along the optical fiber axis is extracted by combining the optical path delay relationship to obtain the original temperature sequence.

[0007] The original temperature sequence is subjected to dilated causal convolution and linear activation to obtain activation features, and then skip connections and channel splicing are performed to obtain the coordinate prediction value of the fiber starting point. The offset between the coordinate prediction value and the theoretical starting position is calculated, and the aligned temperature sequence is calculated based on the offset and the original temperature sequence.

[0008] The sampling points of the aligned temperature sequence are constructed into a graph topology. An adaptive adjacency matrix is ​​constructed based on the magnitude and directionality of the temperature gradient between sampling points. The set of neighboring nodes for each graph node is determined based on the adaptive adjacency matrix. The attention coefficient of each node on the graph topology to the neighboring nodes in the set of neighboring nodes is calculated. The neighboring nodes are weighted and aggregated based on the attention coefficient to obtain the graph topology features.

[0009] The graph topology features are subjected to adaptive lifting wavelet decomposition to obtain high-frequency detail coefficients. Adaptive denoising based on Bayesian risk estimation is applied to the high-frequency detail coefficients. Variational interpolation upsampling is performed by combining lifting wavelet inverse transform reconstruction to obtain the reconstructed temperature field. The reconstructed temperature field and the coordinate prediction value are registered to obtain the calibrated fiber temperature field.

[0010] In one alternative implementation,

[0011] The Raman scattered light signal distributed along the optical fiber transmission path is obtained, and the temperature distribution along the optical fiber axis is extracted by combining the optical path delay relationship to obtain the original temperature sequence, including:

[0012] A pulsed laser is injected into the incident end of an optical fiber, and the backscattered Raman light signal generated along the transmission path of the optical fiber is collected. The backscattered Raman light signal includes a Stokes light intensity component and an anti-Stokes light intensity component. The Stokes light intensity component and the anti-Stokes light intensity component are photoelectrically converted to obtain a Stokes electrical signal and an anti-Stokes electrical signal, respectively.

[0013] Cross-correlation is performed on the Stokes signal and the anti-Stokes signal. The time delay difference between the Stokes signal and the anti-Stokes signal is calculated based on the cross-correlation peak position. Time-domain alignment is performed based on the time delay difference to obtain an aligned Stokes signal and an aligned anti-Stokes signal. The intensity ratio of the aligned anti-Stokes signal and the aligned Stokes signal at each sampling time is calculated, and the temperature value at each sampling time is determined based on the intensity ratio.

[0014] Based on the propagation speed of the pulsed laser in the optical fiber and the time interval between each sampling moment and the pulsed laser injection moment, the axial position coordinates of the optical fiber corresponding to each sampling moment are calculated. The original temperature sequence is obtained by establishing a correspondence between the temperature value and the axial position coordinates of the optical fiber.

[0015] In one alternative implementation,

[0016] The original temperature sequence is subjected to dilated causal convolution and linear activation to obtain activation features. Skip connections and channel splicing are then performed to obtain the predicted coordinates of the fiber optic starting point. The offset between the predicted coordinates and the theoretical starting position is calculated. Based on the offset and the original temperature sequence, an aligned temperature sequence is calculated, including:

[0017] The original temperature sequence is mapped to a state vector to obtain a state vector sequence. The transition weights between adjacent state vectors in the state vector sequence are calculated to obtain a state transition weight matrix. The original temperature sequence is then weighted to obtain a state-weighted temperature sequence. The state-weighted temperature sequence is then subjected to dilated causal convolution to obtain a convolution feature sequence. Linear activation is then performed on the state-weighted temperature sequence to obtain activation features. An activation feature sequence is then constructed based on the activation features.

[0018] The activation feature sequence is passed through skip connections and channel concatenation is performed to obtain a multi-channel feature sequence. The variance of each channel feature in the multi-channel feature sequence is calculated to obtain a channel variance vector. The channel variance vector is then used to perform a weighted summation of each channel feature of the activation feature sequence to obtain a weighted aggregated feature vector.

[0019] A preset causal mask is applied to the weighted aggregated feature vector to obtain a masked feature vector and an attention score matrix is ​​calculated. Based on the attention score matrix, a global temporal representation vector is calculated and a linear transformation is performed to obtain the predicted coordinate value of the fiber optic starting point.

[0020] The fiber axial position coordinates of each sampling point in the original temperature sequence are transformed according to the offset to obtain the corrected axial position. The corrected axial position is then correlated with the temperature value of the corresponding sampling point in the original temperature sequence to obtain the aligned temperature sequence.

[0021] In one alternative implementation,

[0022] The sampling points of the aligned temperature sequence are constructed into a graph topology. An adaptive adjacency matrix is ​​constructed based on the magnitude and directionality of the temperature gradient between sampling points. The set of neighboring nodes for each graph node is determined based on the adaptive adjacency matrix, including:

[0023] Each sampling point in the aligned temperature sequence is mapped to a high-dimensional feature vector to obtain a node feature matrix. The node feature matrix is ​​then subjected to graph structure learning to obtain an initial topological connection relationship. Based on the initial topological connection relationship, the sampling points of the aligned temperature sequence are constructed into a graph topological structure to obtain a graph node set.

[0024] A temperature change sequence is obtained by performing a difference operation on the temperature values ​​of adjacent graph nodes in the graph node set. Variational mode decomposition is then performed on the temperature change sequence, and the envelopes of different mode components are extracted to obtain the temperature gradient amplitude features. The instantaneous phase of each mode component is extracted to obtain the temperature gradient direction features.

[0025] A similarity model is performed between nodes to obtain an amplitude correlation graph for the temperature gradient amplitude features, a directional correlation graph is obtained by performing directional consistency modeling on the temperature gradient direction features, a graph fusion operation is performed on the amplitude correlation graph and the directional correlation graph to obtain a fused correlation graph, and the edge weight matrix in the fused correlation graph is extracted and sparsified to obtain an adaptive adjacency matrix.

[0026] The adaptive adjacency matrix is ​​pruned to obtain a sparse adjacency matrix. A graph sampling algorithm is then performed on the sparse adjacency matrix to obtain sampling results of neighboring nodes. Neighboring importance is evaluated to obtain the neighborhood weight distribution. The set of neighboring nodes for each graph node is determined based on the neighborhood weight distribution.

[0027] In one alternative implementation,

[0028] Calculate the attention coefficient of each node in the graph topology to its neighboring nodes in the set of neighboring nodes, and perform weighted aggregation of the neighboring nodes based on the attention coefficients to obtain the graph topology features, including:

[0029] The node features of each node are extracted from the graph topology and constructed with the node features of the neighboring nodes in the neighboring node set. Causal relationship analysis is performed on the feature pairs to construct a directed acyclic graph to obtain a causal graph. In the causal graph, path tracing is used to identify the confounding variables that simultaneously affect the node pairs to obtain a set of confounding variables. Based on the set of confounding variables, a backdoor adjustment operation is performed on the feature pairs to control the influence of the confounding variables and obtain an adjusted feature pair.

[0030] The neighboring node features in the adjusted feature pair are replaced with counterfactual features to obtain the intervention feature pair. The distance metric between the adjusted feature pair and the intervention feature pair in the feature space is calculated to obtain the causal effect value. The causal effect value and the adjusted feature pair are concatenated to obtain the contrast feature. The contrast feature is mapped to the query vector and the key vector and the dot product similarity is calculated to obtain the attention weight matrix. The attention weight matrix is ​​weighted and summed to obtain the context feature. The correlation strength between the current node and the neighboring nodes is extracted from the context feature to obtain the causal attention score. The attention coefficient of the current node to the neighboring nodes is calculated based on the causal attention score.

[0031] The graph topology features are obtained by weighted summation of the node features of the neighboring nodes in the neighboring node set based on the attention coefficient.

[0032] In one alternative implementation,

[0033] Adaptive lifting wavelet decomposition is performed on the graph topological features to obtain high-frequency detail coefficients. Adaptive denoising based on Bayesian risk estimation is applied to the high-frequency detail coefficients. Variational interpolation upsampling is then performed on the reconstructed temperature field using inverse lifting wavelet transform reconstruction. The reconstructed temperature field and the predicted coordinates are registered to obtain the calibrated fiber temperature field, which includes:

[0034] Multi-scale analysis is performed on the graph topological features to calculate the information entropy and determine the optimal decomposition level. Based on the optimal decomposition level, lifting wavelet multi-level decomposition is performed on the graph topological features to obtain low-frequency approximation coefficients and high-frequency detail coefficients. The mean and variance are extracted from the high-frequency detail coefficients to construct a noise statistical distribution and an adaptive threshold is determined based on the noise statistical distribution. Based on the adaptive threshold, soft threshold shrinkage processing is performed on the high-frequency detail coefficients to obtain denoised high-frequency coefficients.

[0035] The denoised high-frequency coefficients and the low-frequency approximation coefficients are subjected to inverse wavelet transform to obtain the topological features of the denoised map. The spatial coordinates and eigenvalues ​​of the topological features of the denoised map are continuously fitted based on the neural implicit representation algorithm to obtain the implicit field representation. The implicit field representation is then subjected to coordinate sampling and feature mapping at a preset target resolution to obtain the reconstructed temperature field.

[0036] A reference coordinate grid is constructed based on the predicted coordinate values. The similarity measure between the reconstructed temperature field and the reference coordinate grid is calculated based on the differentiable image registration algorithm to obtain the registration loss. The registration loss is optimized by backpropagation to obtain spatial transformation parameters. The reconstructed temperature field is subjected to affine transformation and non-rigid deformation based on the spatial transformation parameters to obtain the registration temperature field. The registration temperature field is filled by bicubic interpolation to obtain the calibration fiber temperature field.

[0037] In one alternative implementation,

[0038] The implicit field representation obtained by continuously fitting the spatial coordinates and eigenvalues ​​of the topological features of the denoised map based on the neural implicit representation algorithm includes:

[0039] Spatial coordinates and corresponding feature values ​​are extracted from the topological features of the denoised image. Based on the preset multi-scale radial basis function, the radial distance between the spatial coordinates is calculated to construct a kernel matrix of multiple scales. The weights of the kernel matrix at each scale are solved to obtain the interpolation weights of the corresponding scale. The interpolation weights of each scale are combined with the feature values ​​to obtain the multi-scale radial basis representation. The multi-scale radial basis representation is fused between scales to obtain the global interpolation representation.

[0040] Based on spatial hash coding, a multi-resolution grid is obtained by dividing the spatial coordinates into spatial grids. A hash index is assigned to each grid vertex in the multi-resolution grid and the corresponding feature vector is stored to construct a hash feature table. The spatial coordinates are mapped to the multi-resolution grid to perform a neighboring vertex query to obtain a neighboring hash index. The corresponding feature vector is extracted from the hash feature table according to the neighboring hash index and trilinear interpolation is performed to obtain local hash features.

[0041] The implicit field representation is obtained by weighting and fusing the global interpolation representation and the local hash feature based on a preset weight.

[0042] A second aspect of the present invention provides a distributed fiber optic temperature measurement and positioning calibration system based on the Raman effect, comprising:

[0043] The first unit is used to acquire the Raman scattered light signal distributed along the optical fiber transmission path, and extract the temperature distribution along the optical fiber axis by combining the optical path delay relationship to obtain the original temperature sequence.

[0044] The second unit is used to perform dilated causal convolution and linear activation on the original temperature sequence to obtain activation features and perform skip connections and channel splicing to obtain the coordinate prediction value of the fiber starting point, calculate the offset between the coordinate prediction value and the theoretical starting position, and calculate the aligned temperature sequence based on the offset and the original temperature sequence.

[0045] The third unit is used to construct a graph topology structure from the sampling points of the aligned temperature sequence, construct an adaptive adjacency matrix based on the magnitude and directionality of the temperature gradient between sampling points, determine the set of neighboring nodes for each graph node based on the adaptive adjacency matrix, calculate the attention coefficient of each node on the graph topology structure to the neighboring nodes in the set of neighboring nodes, and perform weighted aggregation of the neighboring nodes based on the attention coefficient to obtain the graph topology features.

[0046] The fourth unit is used to perform adaptive lifting wavelet decomposition on the graph topology features to obtain high-frequency detail coefficients, apply adaptive denoising based on Bayesian risk estimation to the high-frequency detail coefficients, combine the lifting wavelet inverse transform reconstruction with variational interpolation upsampling to obtain the reconstructed temperature field, and register the reconstructed temperature field and the coordinate prediction value to obtain the calibrated fiber temperature field.

[0047] A third aspect of the present invention provides an electronic device, comprising:

[0048] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0049] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0050] In this invention, the offset of the fiber optic starting point coordinates is calculated by processing the original temperature sequence through dilated causal convolution and linear activation, thus achieving precise calibration of the fiber optic position. This solves the spatial positioning error problem caused by optical path delay and improves the spatial positioning accuracy of the temperature field. By weighted aggregation of neighborhood node information through graph attention mechanism, the local and global spatial features of the temperature field are effectively captured, enhancing the expression of spatial correlation and improving the accuracy of temperature field reconstruction. Fine reconstruction of the temperature field is achieved through variational interpolation upsampling. Finally, the calibrated fiber optic temperature field is obtained through registration, improving the spatial positioning accuracy and temperature detection accuracy of the distributed fiber optic temperature measurement system. Attached Figure Description

[0051] Figure 1 is a schematic flowchart of the distributed optical fiber temperature measurement and positioning calibration method based on the Raman effect according to an embodiment of the present invention;

[0052] Figure 2 is a flowchart of the multi-scale topology analysis temperature field calibration process of the distributed optical fiber temperature measurement and positioning calibration method based on the Raman effect according to an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0055] Figure 1 is a flowchart illustrating the distributed optical fiber temperature measurement and positioning calibration method based on the Raman effect according to an embodiment of the present invention. As shown in Figure 1, the method includes:

[0056] The Raman scattered light signal distributed along the optical fiber transmission path is obtained, and the temperature distribution along the optical fiber axis is extracted by combining the optical path delay relationship to obtain the original temperature sequence.

[0057] The original temperature sequence is subjected to dilated causal convolution and linear activation to obtain activation features, and then skip connections and channel splicing are performed to obtain the coordinate prediction value of the fiber starting point. The offset between the coordinate prediction value and the theoretical starting position is calculated, and the aligned temperature sequence is calculated based on the offset and the original temperature sequence.

[0058] The sampling points of the aligned temperature sequence are constructed into a graph topology. An adaptive adjacency matrix is ​​constructed based on the magnitude and directionality of the temperature gradient between sampling points. The set of neighboring nodes for each graph node is determined based on the adaptive adjacency matrix. The attention coefficient of each node on the graph topology to the neighboring nodes in the set of neighboring nodes is calculated. The neighboring nodes are weighted and aggregated based on the attention coefficient to obtain the graph topology features.

[0059] The graph topology features are subjected to adaptive lifting wavelet decomposition to obtain high-frequency detail coefficients. Adaptive denoising based on Bayesian risk estimation is applied to the high-frequency detail coefficients. Variational interpolation upsampling is performed by combining lifting wavelet inverse transform reconstruction to obtain the reconstructed temperature field. The reconstructed temperature field and the coordinate prediction value are registered to obtain the calibrated fiber temperature field.

[0060] In one alternative implementation,

[0061] The Raman scattered light signal distributed along the optical fiber transmission path is obtained, and the temperature distribution along the optical fiber axis is extracted by combining the optical path delay relationship to obtain the original temperature sequence, including:

[0062] A pulsed laser is injected into the incident end of an optical fiber, and the backscattered Raman light signal generated along the transmission path of the optical fiber is collected. The backscattered Raman light signal includes a Stokes light intensity component and an anti-Stokes light intensity component. The Stokes light intensity component and the anti-Stokes light intensity component are photoelectrically converted to obtain a Stokes electrical signal and an anti-Stokes electrical signal, respectively.

[0063] Cross-correlation is performed on the Stokes signal and the anti-Stokes signal. The time delay difference between the Stokes signal and the anti-Stokes signal is calculated based on the cross-correlation peak position. Time-domain alignment is performed based on the time delay difference to obtain an aligned Stokes signal and an aligned anti-Stokes signal. The intensity ratio of the aligned anti-Stokes signal and the aligned Stokes signal at each sampling time is calculated, and the temperature value at each sampling time is determined based on the intensity ratio.

[0064] Based on the propagation speed of the pulsed laser in the optical fiber and the time interval between each sampling moment and the pulsed laser injection moment, the axial position coordinates of the optical fiber corresponding to each sampling moment are calculated. The original temperature sequence is obtained by establishing a correspondence between the temperature value and the axial position coordinates of the optical fiber.

[0065] A pulsed laser is injected into the optical fiber's input end. This pulsed laser can be a laser source with a wavelength of 1064 nm, a pulse width of 10 nanoseconds, and a repetition frequency of 10 kHz. As the pulsed laser propagates through the fiber, Raman scattering occurs due to the thermal vibrations of the fiber material's molecules. This Raman scattering includes Stokes scattering and anti-Stokes scattering. A highly sensitive photodetector is used to collect the backscattered Raman light signal generated along the fiber's transmission path. This backscattered Raman light signal contains both Stokes and anti-Stokes intensity components. Specifically, a bandpass filter with a wavelength of 1114 nm is used to separate the Stokes intensity component, and a bandpass filter with a wavelength of 1014 nm is used to separate the anti-Stokes intensity component. The two separated light signals are then input to the photodetector for photoelectric conversion, yielding Stokes and anti-Stokes electrical signals, respectively. The photodetector has a response time of 1 nanosecond, effectively capturing weak scattered signals. The converted electrical signals are digitized using an analog-to-digital converter, with a sampling rate set to 250 MHz to ensure sufficient spatial resolution.

[0066] After acquiring the digitized Stokes and anti-Stokes electrical signals, a cross-correlation operation is performed on the two signals to determine the time delay difference. The cross-correlation operation is achieved by calculating the sum of the products of the two signals at different time delays. Specifically, the Stokes and anti-Stokes signals each contain a series of sampling points. The cross-correlation function value is obtained by calculating the sum of the products of corresponding sampling points of the two signals at different offsets. The offset at which the cross-correlation function value is maximized is the time delay difference between the two signals. In a practical application, a fiber optic test case showed that the time delay corresponding to the peak cross-correlation value was 25 sampling points, indicating a time delay difference of 24 sampling intervals between the two signals.

[0067] A time-domain alignment operation is performed based on the calculated time delay difference. The alignment process involves shifting the entire anti-Stokes signal by the number of sampling points corresponding to the time delay, aligning it with the Stokes signal in time. For example, if the time delay is 25 sampling points, the entire anti-Stokes signal is shifted back by 25 points to align the two signals, resulting in an aligned Stokes signal and an aligned anti-Stokes signal.

[0068] After time-domain alignment, the intensity ratio of the aligned anti-Stokes signal to the aligned Stokes signal is calculated at each sampling time. For each sampling point, the corresponding anti-Stokes signal value is divided by the Stokes signal value to obtain the intensity ratio. Due to the quantum mechanical properties of Raman scattering, there is a definite relationship between the intensity ratio and temperature. Based on the experimental calibration curve, the intensity ratio is converted into a temperature value. In fiber optic temperature measurement systems, the calibration curve is usually obtained by placing a section of the fiber in a known temperature environment. For example, an intensity ratio of 0.2 corresponds to a temperature of approximately 20 degrees Celsius, a ratio of 0.3 corresponds to a temperature of approximately 40 degrees Celsius, and a ratio of 0.4 corresponds to a temperature of approximately 60 degrees Celsius.

[0069] Based on the propagation speed of the pulsed laser in the optical fiber and the time interval between each sampling moment and the pulsed laser injection moment, the axial position coordinates of the optical fiber corresponding to each sampling moment are calculated. The propagation speed of light in an optical fiber is approximately two-thirds of the speed of light in a vacuum, or approximately 200 million meters per second. Considering that the optical signal propagates back and forth in the optical fiber, the distance can be calculated using the time interval and the speed of light. For example, if the time of a sampling point is 1 microsecond, its corresponding optical fiber position is approximately 100 meters, because the optical signal propagates back and forth in the optical fiber for 1 microsecond, and the one-way distance is half of the total propagation distance.

[0070] By establishing a correspondence between temperature values ​​and the axial position coordinates of the optical fiber, the original temperature sequence is obtained. In practical applications, the measured temperature data and position coordinates can be paired to form a complete temperature distribution curve. For example, in a certain test case, the temperature was recorded as 25 degrees Celsius at 0 meters, 26 degrees Celsius at 50 meters, 30 degrees Celsius at 100 meters, and 28 degrees Celsius at 150 meters, and so on, to plot the complete temperature distribution curve.

[0071] Because Raman scattering signals are weak, the original temperature sequence usually contains noise. To improve measurement accuracy, a moving average filter can be applied to the original temperature sequence, with a window size of 5 sampling points, to smooth the temperature curve. Specifically, the temperature value of each temperature point and the two points before and after it is averaged to obtain the new temperature value for that point. Furthermore, an iterative adaptive thresholding method is used to identify and correct abnormal temperature points. The initial threshold can be set to three times the standard deviation of the difference between adjacent temperature points. When the temperature difference between a point and its adjacent points exceeds the threshold, it is marked as an abnormal point and replaced with the average value of adjacent normal points. After these processing steps, a smoother and more accurate temperature distribution curve can be obtained, providing a reliable basis for subsequent temperature anomaly monitoring and location.

[0072] In this embodiment, by injecting pulsed laser light at the optical fiber incident end and collecting the backscattered Raman light signal generated along the optical fiber transmission path, the Stokes and anti-Stokes components are photoelectrically converted and cross-correlation is performed to achieve signal time-domain alignment. This effectively eliminates measurement errors caused by differences in signal propagation paths and time drift. By calculating the temperature value based on the aligned signal intensity ratio and combining it with the propagation time to determine the spatial location, high-precision temperature distribution reconstruction along the optical fiber axis is achieved. This effectively suppresses noise interference and asynchronous sampling errors, thereby obtaining distributed temperature monitoring results with more accurate spatial positioning and more sensitive temperature response.

[0073] In one alternative implementation,

[0074] The original temperature sequence is subjected to dilated causal convolution and linear activation to obtain activation features. Skip connections and channel splicing are then performed to obtain the predicted coordinates of the fiber optic starting point. The offset between the predicted coordinates and the theoretical starting position is calculated. Based on the offset and the original temperature sequence, an aligned temperature sequence is calculated, including:

[0075] The original temperature sequence is mapped to a state vector to obtain a state vector sequence. The transition weights between adjacent state vectors in the state vector sequence are calculated to obtain a state transition weight matrix. The original temperature sequence is then weighted to obtain a state-weighted temperature sequence. The state-weighted temperature sequence is then subjected to dilated causal convolution to obtain a convolution feature sequence. Linear activation is then performed on the state-weighted temperature sequence to obtain activation features. An activation feature sequence is then constructed based on the activation features.

[0076] The activation feature sequence is passed through skip connections and channel concatenation is performed to obtain a multi-channel feature sequence. The variance of each channel feature in the multi-channel feature sequence is calculated to obtain a channel variance vector. The channel variance vector is then used to perform a weighted summation of each channel feature of the activation feature sequence to obtain a weighted aggregated feature vector.

[0077] A preset causal mask is applied to the weighted aggregated feature vector to obtain a masked feature vector and an attention score matrix is ​​calculated. Based on the attention score matrix, a global temporal representation vector is calculated and a linear transformation is performed to obtain the predicted coordinate value of the fiber optic starting point.

[0078] The fiber axial position coordinates of each sampling point in the original temperature sequence are transformed according to the offset to obtain the corrected axial position. The corrected axial position is then correlated with the temperature value of the corresponding sampling point in the original temperature sequence to obtain the aligned temperature sequence.

[0079] The original temperature sequence is mapped to a state vector sequence. For each temperature point in the original temperature sequence, the temperature values ​​of the five points before and after it are selected to form a state window. Temperature normalization is then applied to map the temperature values ​​within each window to a range of 0 to 1, forming a five-dimensional state vector. For example, if the temperatures of a certain point and its neighboring points are 25℃, 26℃, 30℃, 28℃, and 27℃, after normalization, the state vector [0.0, 0.2, 1.0, 0.6, 0.4] can be obtained. This completes the mapping from the temperature sequence to the state vector sequence. It should be noted that for the boundary points in the original temperature sequence, namely the first 5 points and the last 5 points, since it is impossible to collect enough forward or backward data points, conventional techniques in the field of signal processing technology can be used to fill the non-existent data points with 0 values ​​or with the nearest point. For example, for the sequence [A,B,C,D,E], the forward data point corresponding to point C can be [0,0,0,A,B] or [A,A,A,A,B]. In actual operation, the filling method can be selected according to the requirements of the scheme.

[0080] The state transition weight matrix is ​​obtained by calculating the transition weights between adjacent state vectors in the state vector sequence. For adjacent state vectors in the sequence, the transition weight is measured by calculating the Euclidean distance between them. The smaller the distance, the smoother the state change, and the larger the weight; conversely, if the distance is large, it indicates possible abnormal fluctuations, and the weight is small. In practical cases, if the Euclidean distance between two adjacent state vectors is 0.2, the transition weight can be set to 0.8; if the distance is 0.5, the weight can be set to 0.5. The complete state transition weight matrix is ​​constructed, and its size is the sequence length minus one.

[0081] The original temperature sequence is weighted using a state transition weight matrix to obtain a state-weighted temperature sequence. For each temperature value in the original temperature sequence, the corresponding state transition weight is multiplied as an adjustment factor to achieve weighted processing of the temperature value. For example, for the original temperature value 30℃, if its corresponding transition weight is 0.8, the weighted temperature value is 24℃. It should be noted that since the matrix size is one less than the sequence length, there may be data in the original temperature sequence that does not have a corresponding state transition weight during the weighting process. Therefore, the weighting process uses either forward mapping or backward mapping. In the forward mapping, the last state transition weight is reused, and in the backward mapping, the first state transition weight is reused. For example, for the original temperature sequence [u,v,w,x,y], the corresponding state transition weights are [U,V,W,X]. If forward mapping is used, then X is the state transition weight corresponding to x and y in the original temperature sequence; if backward mapping is used, then U is the state transition weight corresponding to u and v. The remaining correspondences are w and V, x and W, and y and X, respectively.

[0082] A dilated causal convolution operation is performed on the state-weighted temperature sequence to obtain a convolutional feature sequence. The dilated causal convolution employs a special kernel design with a dilation rate of 2, a kernel size of 3, and 16 output channels. Unlike traditional convolution, the dilated causal convolution increases the receptive field by inserting holes between kernel elements while maintaining causality; that is, the current output depends only on the current and past inputs, not on future inputs. After the convolution operation, the resulting feature sequence contains 16 channels, each capturing a different feature pattern of the temperature sequence.

[0083] Linear activation is applied to the convolutional feature sequence to obtain the activation features. The linear activation function is a rectified linear function with a slope of 0.1, which processes each element in the convolutional feature sequence. When the element value is greater than 0, it remains unchanged; when the element value is less than or equal to 0, it is multiplied by 0.1. After activation processing, a complete activation feature sequence is constructed, containing feature representations of 16 channels.

[0084] The activation feature sequence is passed through skip connections and concatenated by channels to obtain a multi-channel feature sequence. Skip connections directly pass features from earlier layers to subsequent layers. In this embodiment, the original state vector sequence and the activation feature sequence are concatenated along the channel dimension to form a multi-channel feature sequence with 21 channels.

[0085] The variance of each channel feature in the multi-channel feature sequence is calculated to obtain the channel variance vector. For each of the 21 channels, the variance of all its elements is calculated to form a 21-dimensional channel variance vector. The larger the variance value, the richer the information variation contained in that channel; the smaller the variance value, the more stable the information of that channel. In the test case, the variance of each channel may be distributed between 0.01 and 0.5, and the specific value depends on the temperature distribution characteristics.

[0086] A weighted aggregated feature vector is obtained by weighting and summing the features of each channel in the activated feature sequence based on the channel variance vector. The channel variance vector is then converted into weights using softmax, with channels having larger variances receiving higher weights. These weights are then used to weight and sum the features across multiple channels, resulting in a weighted aggregated feature vector with dimensions equal to the sequence length.

[0087] A predefined causal mask is applied to the weighted aggregated feature vector to obtain a masked feature vector, and the attention score matrix is ​​calculated. The causal mask ensures that the current position can only focus on itself and previous positions, maintaining the causality of temporal prediction. The masked feature vector is used to calculate the attention score matrix in the self-attention mechanism, and the matrix size is the sequence length multiplied by the sequence length. This matrix represents the degree of correlation between different positions in the sequence, with values ​​ranging from 0 to 1; larger values ​​indicate stronger correlations.

[0088] The global temporal representation vector is calculated based on the attention score matrix and then linearly transformed to obtain the predicted coordinates of the fiber optic origin. The attention score matrix is ​​multiplied by the mask feature vector to obtain a weighted feature representation, which is then processed by average pooling to obtain the global temporal representation vector. This global temporal representation vector undergoes a linear transformation through a fully connected layer, outputting a scalar value as the predicted coordinates of the fiber optic origin, i.e., the offset. In actual testing, the predicted offset may be -2.5 meters, indicating that the actual fiber optic origin is 2.5 meters forward from the originally assumed origin.

[0089] The corrected axial position is obtained by transforming the fiber axial position coordinates of each sampling point in the original temperature sequence according to the offset. For each sampling point in the original temperature sequence, the predicted offset is subtracted from its fiber axial position coordinates to obtain the corrected axial position. For example, if the original position of a sampling point is 100 meters and the predicted offset is -2.5 meters, the corrected axial position is 102.5 meters. A correspondence is established between the corrected axial position and the temperature value of the corresponding sampling point in the original temperature sequence to obtain the aligned temperature sequence.

[0090] In this embodiment, by mapping the original temperature sequence to a state vector and constructing a state transition weight matrix, dynamic modeling of the temporal variation characteristics of temperature is achieved, which can effectively capture the continuous evolution law of fiber temperature along the axis. By combining dilated causal convolution with linear activation, the temporal receptive field is expanded, which maintains computational efficiency while being sensitive to long-term dependencies, improving the comprehensiveness and temporal correlation of feature extraction. Multi-scale information fusion is achieved by using skip connections and channel splicing mechanisms, and weighted aggregation is performed based on channel variance, which enhances the robustness to abnormal fluctuations and noise interference. Global temporal representation is extracted by causal mask constraints and attention score calculation, which effectively avoids the problem of future information leakage and makes the coordinate offset prediction more physically consistent and temporally reasonable.

[0091] In one alternative implementation,

[0092] The sampling points of the aligned temperature sequence are constructed into a graph topology. An adaptive adjacency matrix is ​​constructed based on the magnitude and directionality of the temperature gradient between sampling points. The set of neighboring nodes for each graph node is determined based on the adaptive adjacency matrix, including:

[0093] Each sampling point in the aligned temperature sequence is mapped to a high-dimensional feature vector to obtain a node feature matrix. The node feature matrix is ​​then subjected to graph structure learning to obtain an initial topological connection relationship. Based on the initial topological connection relationship, the sampling points of the aligned temperature sequence are constructed into a graph topological structure to obtain a graph node set.

[0094] A temperature change sequence is obtained by performing a difference operation on the temperature values ​​of adjacent graph nodes in the graph node set. Variational mode decomposition is then performed on the temperature change sequence, and the envelopes of different mode components are extracted to obtain the temperature gradient amplitude features. The instantaneous phase of each mode component is extracted to obtain the temperature gradient direction features.

[0095] A similarity model is performed between nodes to obtain an amplitude correlation graph for the temperature gradient amplitude features, a directional correlation graph is obtained by performing directional consistency modeling on the temperature gradient direction features, a graph fusion operation is performed on the amplitude correlation graph and the directional correlation graph to obtain a fused correlation graph, and the edge weight matrix in the fused correlation graph is extracted and sparsified to obtain an adaptive adjacency matrix.

[0096] The adaptive adjacency matrix is ​​pruned to obtain a sparse adjacency matrix. A graph sampling algorithm is then performed on the sparse adjacency matrix to obtain sampling results of neighboring nodes. Neighboring importance is evaluated to obtain the neighborhood weight distribution. The set of neighboring nodes for each graph node is determined based on the neighborhood weight distribution.

[0097] For each sampling point in the temperature sequence, its location information, temperature value, and temperature change trend of the five points before and after it are extracted to construct a 16-dimensional feature vector. Specifically, for a sampling point located at 100 meters on the optical fiber with a temperature of 30℃, its feature vector includes information such as a normalized location value of 0.1, a normalized temperature value of 0.6, and temperature difference values ​​before and after it [0.5, 1.0, 2.0, -1.0, -0.5]. The feature vectors of all sampling points are combined to form a node feature matrix, and the matrix dimension is the number of sampling points multiplied by 16.

[0098] Initial topological connections are obtained by performing graph structure learning on the node feature matrices. A similarity-based graph learning method is used to calculate the cosine similarity between node feature vectors and construct a similarity matrix. A similarity threshold of 0.85 is set; a connection is established when the similarity between two nodes is greater than the threshold. For example, if the feature vector similarity between nodes A and B is 0.9, an edge is established between A and B; if the similarity between nodes A and C is 0.7, no connection is established. Based on the calculated similarity matrix and the threshold condition, the initial topological connections between nodes in the graph are determined.

[0099] Based on the initial topological connections, the sampling points of the aligned temperature sequence are constructed into a graph topology, resulting in a set of graph nodes. Each sampling point serves as a node in the graph, and its attributes include the point's temperature value and location information. The edge weights between adjacent nodes are set to the similarity values ​​of the node feature vectors. In the constructed graph structure, the total number of nodes equals the number of sampling points, and the number of edges depends on the density of the initial topological connections. For example, for a temperature sequence containing 1000 sampling points, the constructed graph may contain approximately 3000 edges, with each node connecting an average of 6 adjacent nodes.

[0100] A temperature change sequence is obtained by performing a difference operation on the temperature values ​​of adjacent nodes in the graph node set. For each pair of nodes connected by an edge in the graph, the difference in temperature values ​​is calculated. For example, if the temperature of node A is 28℃ and the temperature of its connected node B is 30℃, then the temperature difference between them is 2℃. The temperature differences on all edges are arranged according to the order of the nodes on the optical fiber to form the temperature change sequence.

[0101] Variational mode decomposition (VMD) was performed on the temperature change sequence, and the envelopes of different modal components were extracted to obtain the temperature gradient magnitude features. The VMD parameters were set to 5 modes, a penalty factor of 2000, and a tolerance of 10. -7 Variational mode decomposition (VMD) decomposes the temperature change sequence into five modal components of different frequencies. A Hilbert transform is calculated for each modal component to obtain its analytic signal amplitude, i.e., its envelope. The envelope constitutes the temperature gradient amplitude characteristic, reflecting the intensity of temperature changes at different scales. In the fiber optic temperature measurement case, the amplitude range of the first mode envelope is 0.1, representing low-frequency temperature changes; the amplitude range of the second mode envelope is 0.15, representing mid-frequency temperature changes.

[0102] The instantaneous phase of each modal component is extracted to obtain the temperature gradient direction feature. The phase information of each modal component's analytical signal is extracted to obtain an instantaneous phase sequence. The instantaneous phase values ​​range from -π to π, reflecting the directional characteristics of temperature changes. The instantaneous phase sequences of all modes are combined to form a complete temperature gradient direction feature.

[0103] A similarity model of the temperature gradient amplitude characteristics between nodes is used to obtain an amplitude correlation graph. The Pearson correlation coefficient of the temperature gradient amplitude characteristics between different nodes is calculated to construct an amplitude correlation matrix. The correlation coefficient ranges from -1 to 1; the closer the value is to 1, the more similar the temperature change amplitudes of the two nodes are. The correlation coefficient is used as the edge weight to construct the amplitude correlation graph. For example, the correlation coefficient between adjacent nodes is typically above 0.7, while the correlation coefficient between distant nodes may be below 0.3.

[0104] A directional consistency model is obtained by modeling the directional features of temperature gradients to obtain a directional association graph. The cosine similarity of the directional features of temperature gradients between nodes is calculated to construct a directional consistency matrix. The similarity value ranges from -1 to 1, with a larger value indicating a more consistent direction of temperature change. The similarity value is used as the edge weight to construct the directional association graph. Near hotspot regions, the directional consistency of adjacent nodes is usually low, with a similarity of around 0.3; while in temperature-stable regions, the directional consistency of adjacent nodes is usually high, with a similarity exceeding 0.9.

[0105] A graph fusion operation is performed on the magnitude correlation graph and the direction correlation graph to obtain a fused correlation graph. A weighted average method is used for graph fusion, with a weight of 0.6 for the magnitude correlation graph and 0.4 for the direction correlation graph. For the edge connecting node A and node B, if the weight in the magnitude correlation graph is 0.8 and the weight in the direction correlation graph is 0.6, then the weight of the fused edge is 0.8 × 0.6 + 0.6 × 0.4 = 0.72. Taking into account both the magnitude and direction characteristics of the temperature gradient, the fused correlation graph is constructed.

[0106] The edge weight matrix is ​​extracted from the fused association graph and sparsified to obtain an adaptive adjacency matrix. A weight threshold of 0.5 is set; when an edge weight is below this threshold, it is set to 0, i.e., the edge is deleted. For the retained edges, the weight values ​​are normalized to the range of 0 to 1 to reduce noisy connections in the graph and retain significant node relationships, thus obtaining the adaptive adjacency matrix.

[0107] A sparse adjacency matrix is ​​obtained by performing graph pruning on the adaptive adjacency matrix. A Top-K pruning strategy is used, where only the K edges with the highest weights are retained for each node, and K is set to 8. For example, for node A, if it is connected to 10 edges, only the top 8 edges by weight are retained, and the rest are pruned.

[0108] A graph sampling algorithm is performed on the sparse adjacency matrix to obtain sampling results of neighboring nodes, and neighborhood importance is evaluated to obtain the neighborhood weight distribution. An attention-based graph sampling method is used to sample the neighborhood of each node. The attention weight is calculated based on the product of the node's feature similarity and the edge weight. For example, for node A and its neighbor node B, if their feature similarity is 0.7 and their edge weight is 0.8, then the attention weight is 0.7 × 0.8 = 0.56. The attention weights of all neighboring nodes are normalized to obtain the neighborhood weight distribution.

[0109] The neighborhood node set for each graph node is determined based on the distribution of neighborhood weights. A weight threshold of 0.1 is set; when the weight of a neighboring node is lower than this threshold, it is removed from the neighborhood set. The remaining neighboring nodes are then sorted by weight from highest to lowest to form the final neighborhood node set.

[0110] In this embodiment, by mapping the sampling points of the aligned temperature sequence to high-dimensional feature vectors and performing graph structure learning, the intrinsic correlation between sampling points can be adaptively mined from the data, realizing an effective transformation from temporal information to spatial topological structure. By combining temperature difference operation and variational mode decomposition, multi-scale temperature gradient amplitude and direction features are extracted, which can simultaneously perceive the intensity and trend direction of temperature changes, enhancing the ability to characterize the dynamic characteristics of complex temperature fields. Based on the fusion construction of amplitude correlation graph and direction correlation graph, the correlation and directional consistency of temperature changes in space are comprehensively reflected, significantly improving the fitting degree of the graph structure to the real heat conduction relationship. By sparsifying the edge weight matrix of the fused correlation graph and performing graph pruning, redundant connections are reduced, improving the computational efficiency and generalization ability of the graph model.

[0111] In one alternative implementation,

[0112] Calculate the attention coefficient of each node in the graph topology to its neighboring nodes in the set of neighboring nodes, and perform weighted aggregation of the neighboring nodes based on the attention coefficients to obtain the graph topology features, including:

[0113] The node features of each node are extracted from the graph topology and constructed with the node features of the neighboring nodes in the neighboring node set. Causal relationship analysis is performed on the feature pairs to construct a directed acyclic graph to obtain a causal graph. In the causal graph, path tracing is used to identify the confounding variables that simultaneously affect the node pairs to obtain a set of confounding variables. Based on the set of confounding variables, a backdoor adjustment operation is performed on the feature pairs to control the influence of the confounding variables and obtain an adjusted feature pair.

[0114] The neighboring node features in the adjusted feature pair are replaced with counterfactual features to obtain the intervention feature pair. The distance metric between the adjusted feature pair and the intervention feature pair in the feature space is calculated to obtain the causal effect value. The causal effect value and the adjusted feature pair are concatenated to obtain the contrast feature. The contrast feature is mapped to the query vector and the key vector and the dot product similarity is calculated to obtain the attention weight matrix. The attention weight matrix is ​​weighted and summed to obtain the context feature. The correlation strength between the current node and the neighboring nodes is extracted from the context feature to obtain the causal attention score. The attention coefficient of the current node to the neighboring nodes is calculated based on the causal attention score.

[0115] The graph topology features are obtained by weighted summation of the node features of the neighboring nodes in the neighboring node set based on the attention coefficient.

[0116] For each sampling point in the fiber optic temperature monitoring system, a 16-dimensional node feature vector corresponding to each sampling point and all node feature vectors in its neighboring node set are extracted. For sampling point A located at 150 meters in the fiber, its feature vector includes a normalized temperature value of 0.65, a location code of 0.15, and temperature change trend features; its neighboring node set includes four nodes B, C, D, and E, located at 145 meters, 148 meters, 152 meters, and 155 meters respectively, and the feature vectors of the neighboring nodes also have 16 dimensions. The feature vector of node A is paired with the feature vectors of each neighboring node to form a feature pair set {(A, B), (A, C), (A, D), (A, E)}, and each feature pair contains two 16-dimensional vectors.

[0117] A causal graph is constructed by performing causal relationship analysis on feature pairs. The PC algorithm is used for causal discovery, with a significance level of 0.05. A conditional independence test is calculated for each pair of feature vectors, and the causal direction between nodes is determined based on the test results. For example, if the temperature change of node A occurs before that of node B and the conditional independence test is satisfied, a directed edge from A to B is established, indicating that A may be the cause of the temperature change in B. In practice, for feature pair (A, B), by analyzing their mutual information under different time windows, it is found that the mutual information value reaches its maximum value of 0.78 when A lags B by three time units, thus determining the causal relationship from B to A. Similarly, other feature pairs are analyzed to construct a complete causal graph containing the causal relationship representation between all nodes.

[0118] In a causal graph, path tracing is used to identify confounding variables that simultaneously affect node pairs, thus obtaining a set of confounding variables. For each feature pair (Xi, Xj), it is checked whether there exists a node Z in the graph that points to both Xi and Xj. If so, Z is a potential confounding variable. A path length threshold of 3 is set, considering only confounding relationships with path lengths not exceeding 3. In the fiber optic temperature measurement scenario, for the feature pair (A, B), path tracing reveals that nodes C and F simultaneously point to A and B, therefore {C, F} is identified as a set of confounding variables. For other feature pairs, the same method is used to identify confounding variables, forming a complete set of confounding variables.

[0119] Adjusted feature pairs are obtained by performing backdoor adjustment operations on feature pairs based on the confusion variable set Z to control the influence of the confusion variables. An inverse probability weighting method is used to implement backdoor adjustment. For each feature pair (Xi, Xj), an inverse probability weight is calculated based on the confusion variable set Z. P(Xi|Z) is estimated by fitting a probability model, and the weight is set to 1 / P(Xi|Z). For adjacent nodes A and B on the optical fiber, if the confusion variable set is {C, F} and the calculated weight is 1.5, then the feature vector of A is multiplied by 1.5 to obtain the adjusted feature vector; a similar process is performed on the feature vector of B. Backdoor adjustment is performed on all feature pairs to control the influence of the confusion variables, resulting in a set of adjusted feature pairs.

[0120] Counterfactual replacement is performed on the neighboring node features in the adjusted feature pairs to obtain the intervention feature pairs. The intervention magnitude is set to 0.2. For each adjusted feature pair (Xi, Xj), the temperature-related feature of Xj is increased by 20% as a counterfactual intervention. For the adjusted feature pair (A, B), if the normalized temperature feature of B is 0.5, then the temperature feature after the counterfactual replacement is 0.6; keeping other features of B unchanged, the intervention feature vector is formed. The original feature vector of A is paired with the intervention feature vector of B to form the intervention feature pair (A, B'); the same operation is performed on all adjusted feature pairs to obtain the complete set of intervention feature pairs.

[0121] The causal effect value is obtained by calculating the distance between the adjustment feature pair and the intervention feature pair in the feature space. Euclidean distance is used to measure the difference between the two feature pairs, calculated as the square root of the sum of the squares of the differences between the two feature vectors. For the adjustment feature pair (A, B) and the intervention feature pair (A, B'), if the Euclidean distance in the feature space is 0.15, this value is taken as the causal effect value of node A to node B. The causal effect values ​​of all feature pairs are calculated in the same way to form a causal effect matrix.

[0122] The causal effect value and the adjusted feature pair are concatenated to obtain the contrast features. For each adjusted feature pair (Xi, Xj) and its corresponding causal effect value eij, eij is added as a new feature dimension to the adjusted feature pair to form the contrast features. In the fiber optic temperature measurement scenario, for the adjusted feature pair (A, B) and the causal effect value 0.15, the three are concatenated to form a 33-dimensional contrast feature vector; the same operation is performed on all feature pairs to construct a complete set of contrast features.

[0123] The contrastive features are mapped to query and key vectors, and the dot product similarity is calculated to obtain the attention weight matrix. Through two independent linear transformation layers, the contrastive features are mapped to query and key vectors respectively, with a transformation matrix dimension of 33×8, resulting in 8-dimensional query and key vectors. The dot product of the query and key vectors is calculated and divided by a scaling factor of 2.83 (the square root of 8) to obtain the attention weight. In practical applications, for feature pairs (A, B), assuming the mapped query vector is [0.5, 0.3, 0.2, 0.1, 0.4, 0.2, 0.3, 0.1] and the key vector is [0.4, 0.5, 0.1, 0.2, 0.3, 0.4, 0.2, 0.3], the dot product yields an attention weight of 0.63; the same calculation is performed on all feature pairs to construct the attention weight matrix.

[0124] The context features are obtained by weighted summation of the attention weight matrix. The attention weights are converted into a probability distribution using the softmax function and then weighted and summed with the feature vectors. For node A and its neighboring nodes B, C, D, and E, if the corresponding attention weights after softmax processing are [0.4, 0.3, 0.2, 0.1], then the context features are calculated as: B feature vector × 0.4 + C feature vector × 0.3 + D feature vector × 0.2 + E feature vector × 0.1. The context features of all nodes are calculated using the same method.

[0125] The causal attention score is obtained by extracting the association strength between the current node and its neighboring nodes from the context features. A linear layer maps the context features to association strengths, with a transformation matrix of dimension 16×4, outputting a 4-dimensional vector corresponding to the association strength between the current node and its four neighboring nodes. In fiber optic temperature measurement applications, for the context features of node A, after linear transformation, the association strengths with neighboring nodes B, C, D, and E are [0.8, 0.6, 0.4, 0.2], which are the causal attention scores. The same operation is performed on all nodes to construct a complete set of causal attention scores.

[0126] The attention coefficients of the current node to its neighboring nodes are calculated based on the causal attention score. The causal attention score is normalized to a probability distribution using the softmax function and used as the attention coefficient. For node A, the causal attention score [0.8, 0.6, 0.4, 0.2] is processed by softmax to obtain the attention coefficients [0.35, 0.28, 0.22, 0.15], representing the degree of attention A pays to its neighboring nodes B, C, D, and E. The attention coefficients of all nodes are calculated using the same method to construct the attention coefficient matrix.

[0127] The graph topological features are obtained by weighted summation of the node features of neighboring nodes in the set of neighboring nodes based on the attention coefficients. For each node, the feature vectors of its neighboring nodes are weighted and summed according to the attention coefficients to generate the graph topological features of that node. In practical applications, for node A and its neighboring nodes B, C, D, and E, and the attention coefficients [0.35, 0.28, 0.22, 0.15], the calculated graph topological features are: B feature vector × 0.35 + C feature vector × 0.28 + D feature vector × 0.22 + E feature vector × 0.15. The graph topological features of all nodes are calculated in the same way to construct a complete set of graph topological features.

[0128] In this embodiment, by introducing causal relationship analysis and backdoor adjustment mechanisms into the graph topology, causal modeling of the relationships between nodes is achieved, which can effectively distinguish between real causal influences and pseudo-correlation relationships. By constructing a directed acyclic graph and identifying a set of confusing variables, potential factors that simultaneously affect the relationships between multiple nodes can be discovered. By eliminating interference through backdoor adjustment, the dependencies between features become more interpretable and robust. By using counterfactual substitution to generate intervention feature pairs and calculate causal effect values, the direct causal influence between nodes can be quantified at the feature level, improving the discriminability and controllability of feature representation. Based on the causal effect-guided attention mechanism, neighboring nodes are weighted and aggregated through causal attention scores, making the attention distribution more consistent with physical laws and causal logic, and reducing the impact of noise and redundant features.

[0129] In one alternative implementation,

[0130] Adaptive lifting wavelet decomposition is performed on the graph topological features to obtain high-frequency detail coefficients. Adaptive denoising based on Bayesian risk estimation is applied to the high-frequency detail coefficients. Variational interpolation upsampling is then performed on the reconstructed temperature field using inverse lifting wavelet transform reconstruction. The reconstructed temperature field and the predicted coordinates are registered to obtain the calibrated fiber temperature field, which includes:

[0131] Multi-scale analysis is performed on the graph topological features to calculate the information entropy and determine the optimal decomposition level. Based on the optimal decomposition level, lifting wavelet multi-level decomposition is performed on the graph topological features to obtain low-frequency approximation coefficients and high-frequency detail coefficients. The mean and variance are extracted from the high-frequency detail coefficients to construct a noise statistical distribution and an adaptive threshold is determined based on the noise statistical distribution. Based on the adaptive threshold, soft threshold shrinkage processing is performed on the high-frequency detail coefficients to obtain denoised high-frequency coefficients.

[0132] The denoised high-frequency coefficients and the low-frequency approximation coefficients are subjected to inverse wavelet transform to obtain the topological features of the denoised map. The spatial coordinates and eigenvalues ​​of the topological features of the denoised map are continuously fitted based on the neural implicit representation algorithm to obtain the implicit field representation. The implicit field representation is then subjected to coordinate sampling and feature mapping at a preset target resolution to obtain the reconstructed temperature field.

[0133] A reference coordinate grid is constructed based on the predicted coordinate values. The similarity measure between the reconstructed temperature field and the reference coordinate grid is calculated based on the differentiable image registration algorithm to obtain the registration loss. The registration loss is optimized by backpropagation to obtain spatial transformation parameters. The reconstructed temperature field is subjected to affine transformation and non-rigid deformation based on the spatial transformation parameters to obtain the registration temperature field. The registration temperature field is filled by bicubic interpolation to obtain the calibration fiber temperature field.

[0134] The acquired graph topological features were analyzed using discrete wavelet transform at multiple scales. Shannon information entropy was calculated for each decomposition level, with the number of decomposition levels increasing from 1 to 5. On a graph topological feature composed of 1000 fiber optic monitoring points, the information entropy after decomposition was 4.82 for level 1, 4.65 for level 2, 4.31 for level 3, 4.28 for level 4, and 4.27 for level 5. Decomposition was stopped when the rate of change of information entropy was less than 0.01; therefore, the optimal number of decomposition levels was determined to be 4.

[0135] Based on the optimal decomposition level, lifting wavelet multi-level decomposition is performed on the graph topological features to obtain low-frequency approximation coefficients and high-frequency detail coefficients. A db4 wavelet basis is used to perform a four-level decomposition of the graph topological features, with the boundary extension method set to symmetric extension during the decomposition process. Taking the temperature feature at the 200-meter mark on the optical fiber as an example, its initial graph topological feature values ​​are [0.72, 0.68, 0.65, 0.70, 0.69, 0.73, 0.71, 0.67, 0.64, 0.69]. After multi-level decomposition, the low-frequency approximation coefficients are [1.25, 1.22, 1.20, 1.23, 1.21], and the first-level high-frequency detail coefficients are [0.05, -0.0]. The low-frequency coefficients are [0.02, -0.02, -0.04, 0.03], the second-level high-frequency detail coefficients are [0.02, -0.01, 0.01, -0.02, 0.01], the third-level high-frequency detail coefficients are [0.01, -0.005, 0.008, -0.006, 0.004], and the fourth-level high-frequency detail coefficients are [0.003, -0.002, 0.001, -0.002, 0.001]. The low-frequency coefficients reflect the overall trend of the temperature field, while the high-frequency coefficients capture the detailed fluctuations and noise within the temperature field.

[0136] The mean and variance of high-frequency detail coefficients are extracted to construct a noise statistical distribution, and adaptive thresholds are determined based on this distribution. Statistical characteristics are calculated for each level of high-frequency detail coefficients to obtain the mean and variance. For the first-level high-frequency detail coefficients, the calculated mean is 0.006 and the variance is 0.0012; for the second-level coefficients, the mean is 0.002 and the variance is 0.0004; for the third-level coefficients, the mean is 0.0008 and the variance is 0.00015; and for the fourth-level coefficients, the mean is 0.0002 and the variance is 0.00005. Assuming the noise follows a Gaussian distribution, adaptive thresholds for each level are determined based on the mean and variance. The threshold calculation uses a standard deviation-based method, with a coefficient set to 3.5, resulting in a first-level threshold of 0.12, a second-level threshold of 0.07, a third-level threshold of 0.04, and a fourth-level threshold of 0.025. The adaptive thresholds effectively distinguish between noise and effective information from the temperature field.

[0137] Denoising high-frequency coefficients are obtained by applying soft thresholding to high-frequency detail coefficients based on adaptive thresholding. The soft thresholding function processes the high-frequency coefficients, setting them to zero when the absolute value of a coefficient is less than the threshold, and subtracting the threshold's positive or negative value when it is greater. Taking the first-level high-frequency detail coefficients [0.05, -0.03, 0.02, -0.04, 0.03] as an example, the corresponding threshold is 0.12. Since all coefficient absolute values ​​are less than the threshold, the denoising high-frequency coefficients after soft thresholding are [0, 0, 0, 0, 0]. For another set of coefficients [0.15, -0.20, 0.08, -0.25, 0.10], the result after soft thresholding is [0.03, -0.08, 0, -0.13, 0]. Similar processing is applied to all levels of high-frequency coefficients to obtain a complete set of denoising high-frequency coefficients.

[0138] The denoised graph topological features are obtained by performing inverse wavelet transform on the denoised high-frequency coefficients and low-frequency approximation coefficients. The processed low-frequency approximation coefficients and the denoised high-frequency detail coefficients at each level are then reconstructed using inverse wavelet transform to reconstruct the graph topological features at the original scale. The inverse transform also uses the db4 wavelet basis and the same extension method as the decomposition process. For the aforementioned example, the reconstructed denoised graph topological features are [0.71, 0.68, 0.65, 0.70, 0.69, 0.73, 0.70, 0.67, 0.64, 0.69], and the noise component is effectively suppressed compared to the original features.

[0139] An implicit field representation is obtained by continuously fitting the spatial coordinates and feature values ​​of the denoised image's topological features using a neural implicit representation algorithm. A multilayer perceptron network is constructed, containing four hidden layers with 256 neurons per layer and the SIREN activation function. The network input is the normalized optical fiber spatial coordinates (x, y), and the output is the temperature feature value at that location. During training, the coordinates of 1000 sampling points and their corresponding denoised image topological feature values ​​constitute the training data. The learning rate is set to 0.0001, the training epochs are 10000, and the batch size is 64. Taking the 200-meter and 201-meter locations on the optical fiber as examples, the normalized coordinates are (0.2, 0) and (0.201, 0), respectively, and the corresponding denoised image topological feature values ​​are 0.71 and 0.70. The trained neural network predicts these two coordinates to be 0.708 and 0.702, respectively, with a fitting error of less than 0.01. The trained neural network can represent discrete temperature features as a continuous implicit field function.

[0140] The implicit field representation is reconstructed by performing coordinate sampling and feature mapping at a preset target resolution. The target resolution is set to five times the original sampling points, meaning four new sampling points are inserted between every two original sampling points. For a 10-kilometer-long optical fiber, the original sampling resolution is 10 meters per point, totaling 1000 points; this is increased to 2 meters per point, resulting in 5000 points. Normalized coordinates for these 5000 points are generated and input into a trained neural network to obtain the temperature feature value corresponding to each coordinate. For example, for adjacent original sampling points 200 meters (0.2, 0) and 210 meters (0.21, 0), the inserted coordinate points are 202 meters (0.202, 0), 204 meters (0.204, 0), 206 meters (0.206, 0), and 208 meters (0.208, 0), and the neural network predicts temperature feature values ​​of 0.705, 0.698, 0.692, and 0.687, respectively. In this way, a higher resolution representation of the continuous temperature field can be generated.

[0141] A reference coordinate grid is constructed based on predicted coordinate values. Given the physical layout of the optical fiber, a standardized reference coordinate grid is established according to the actual installation location. For linearly laid optical fibers, the reference coordinate grid is a one-dimensional grid with equal intervals; for optical fibers laid in a planar serpentine pattern, the reference coordinate grid is a two-dimensional grid. Taking the monitoring of a certain industrial equipment as an example, 10 kilometers of optical fiber are laid in a serpentine pattern within a 10m × 1000m area. The reference coordinate grid is a 500 × 5 two-dimensional grid, with a horizontal interval of 0.02 meters and a vertical interval of 200 meters. Each node in the reference grid represents the standard position where the optical fiber should ideally be located.

[0142] The registration loss is calculated based on the similarity metric between the reconstructed temperature field and the reference coordinate grid using a differentiable image registration algorithm. The reconstructed temperature field is considered as a feature image, and the reference coordinate grid as a target template. Mutual information is used as the similarity metric. In the actual calculation, the reconstructed temperature field is sampled onto the reference grid under the currently estimated transform parameters, and the mutual information between the sampled temperature values ​​and the reference temperature values ​​is calculated. The mutual information calculation uses a density estimation method based on the Parzen window, with a window width of 0.1. For the initial registration state, the calculated mutual information value is 1.25, indicating a low degree of matching between the current reconstructed temperature field and the reference coordinate grid.

[0143] The spatial transformation parameters were obtained by backpropagation optimization of the registration loss. An optimization model for the transformation parameters was constructed, including affine transformation parameters (rotation, scaling, shearing, and translation) and non-rigid deformation field parameters. Gradient descent was used to optimize the transformation parameters with a learning rate of 0.01 and 500 iterations. In each iteration, the gradient of the mutual information with respect to the transformation parameters was calculated, and the parameter values ​​were updated. For example, in a certain iteration, the initial affine parameters were [1.0, 0, 0, 1.0, 0, 0], representing no transformation state; the optimized parameters were [0.98, 0.01, -0.02, 0.99, 2.5, 1.8], representing slight rotation, scaling, and translation. After complete optimization, the mutual information value increased from the initial 1.25 to 3.78, indicating a significant improvement in registration quality.

[0144] The reconstructed temperature field is subjected to affine transformation and non-rigid deformation based on spatial transformation parameters to obtain the registered temperature field. The optimized transformation parameters are applied to perform spatial transformation on the reconstructed temperature field. For the affine transformation, a transformation matrix is ​​constructed and the coordinates are linearly mapped; for the non-rigid deformation, B-spline interpolation is used to calculate the displacement field at each point. Taking the point at the 500-meter mark on the optical fiber as an example, the original reconstructed coordinates are (0.5, 0), and the temperature characteristic value is 0.65. After the affine transformation, the coordinates become (0.49, 0.005), and after non-rigid deformation, the final coordinates are (0.488, 0.007), while keeping the temperature characteristic value unchanged. By performing the same transformation operation on all points, the registered temperature field in the reference coordinate system is obtained.

[0145] The registered temperature field is filled using bicubic interpolation to obtain the calibrated fiber temperature field. Since the transformed point positions may be irregular, temperature values ​​on a regular grid need to be generated through interpolation. A bicubic interpolation algorithm is used to calculate the temperature value at each node of the reference coordinate grid. The interpolation range is the 16 nearest points within a 4×4 area around the node, and the weighting coefficients are calculated using a cubic polynomial kernel function. Taking the point (0.5, 0) in the reference grid as an example, the coordinates and temperature values ​​of the nearest data point after transformation are [(0.488, 0.007, 0.65), (0.492, 0.003, 0.64), (0.485, -0.002, 0.66), (0.495, 0.005, 0.63)]. The calibrated temperature value for this point is calculated to be 0.647 using bicubic interpolation. After completing the interpolation calculations for all reference grid points, the final calibrated fiber temperature field is obtained.

[0146] In this embodiment, by performing multi-scale analysis on the graph topological features and calculating the information entropy to determine the optimal number of decomposition layers, the most suitable signal decomposition scale can be adaptively selected, so that the feature decomposition retains key structural information while avoiding the loss of details caused by over-decomposition. By using lifting wavelet multi-level decomposition to extract low-frequency and high-frequency features, hierarchical separation of temperature field information is achieved. The adaptive threshold set based on the noise statistical distribution effectively improves the denoising accuracy, accurately suppresses noise interference in high-frequency components, and significantly improves the signal-to-noise ratio and stability of the features. The denoised graph topological features obtained by wavelet inverse transform reconstruction can reduce the impact of environmental disturbances and sampling errors while maintaining the continuity of spatial structure. The neural implicit representation algorithm is introduced to continuously fit the spatial coordinates and feature values, so that the temperature field is transformed from discrete point representation to continuous implicit field representation, realizing high-resolution, smooth and differentiable temperature field reconstruction. Combined with the differentiable image registration algorithm, the spatial transformation parameters obtained by similarity measurement and backpropagation optimization are used to achieve global affine correction and local non-rigid alignment of the temperature field, which significantly improves the consistency and registration accuracy of the temperature field in spatial scale.

[0147] Figure 2 is a flowchart of the multi-scale topology analysis temperature field calibration process of the distributed optical fiber temperature measurement and positioning calibration method based on the Raman effect according to an embodiment of the present invention.

[0148] In one alternative implementation,

[0149] The implicit field representation obtained by continuously fitting the spatial coordinates and eigenvalues ​​of the topological features of the denoised map based on the neural implicit representation algorithm includes:

[0150] Spatial coordinates and corresponding feature values ​​are extracted from the topological features of the denoised image. Based on the preset multi-scale radial basis function, the radial distance between the spatial coordinates is calculated to construct a kernel matrix of multiple scales. The weights of the kernel matrix at each scale are solved to obtain the interpolation weights of the corresponding scale. The interpolation weights of each scale are combined with the feature values ​​to obtain the multi-scale radial basis representation. The multi-scale radial basis representation is fused between scales to obtain the global interpolation representation.

[0151] Based on spatial hash coding, a multi-resolution grid is obtained by dividing the spatial coordinates into spatial grids. A hash index is assigned to each grid vertex in the multi-resolution grid and the corresponding feature vector is stored to construct a hash feature table. The spatial coordinates are mapped to the multi-resolution grid to perform a neighboring vertex query to obtain a neighboring hash index. The corresponding feature vector is extracted from the hash feature table according to the neighboring hash index and trilinear interpolation is performed to obtain local hash features.

[0152] The implicit field representation is obtained by weighting and fusing the global interpolation representation and the local hash feature based on a preset weight.

[0153] The denoised image topology features obtained from the distributed fiber optic temperature measurement system contain the spatial location information and temperature feature values ​​of a series of sampling points. For a 10-kilometer-long fiber, a point is collected every 10 meters, resulting in 1000 measurement points. Each measurement point records its normalized spatial coordinates and corresponding temperature feature value. For example, the measurement point located 2000 meters from the beginning of the fiber has normalized spatial coordinates of (0.2, 0.0) and a corresponding temperature feature value of 25.6℃, which is normalized to 0.427; the measurement point located at 4500 meters has normalized spatial coordinates of (0.45, 0.0) and a temperature feature value of 32.8℃, which is normalized to 0.546. By traversing all measurement points, a dataset containing 1000 pairs of spatial coordinates and feature values ​​is constructed as the input for subsequent multi-scale radial basis function interpolation.

[0154] Based on a pre-defined multi-scale radial basis function, the radial distance between spatial coordinates is calculated to construct kernel matrices at multiple scales. Three different scales of radial basis functions are selected as the basic kernel functions: a Gaussian kernel function with a scale parameter of 2.0 is used for large-scale operations, a Gaussian kernel function with a scale parameter of 0.5 is used for medium-scale operations, and a Gaussian kernel function with a scale parameter of 0.1 is used for small-scale operations. For each pair of spatial coordinate points, their Euclidean distance is calculated, and the kernel value is calculated using the kernel function at each scale. For example, for two points with a distance of 0.05, the output value of the large-scale kernel function is 0.9988, the output value of the medium-scale kernel function is 0.9512, and the output value of the small-scale kernel function is 0.7788. By calculating the kernel function values ​​for all point pairs, three 1000×1000 kernel matrices are constructed to represent the correlation between points at the large, medium, and small scales, respectively.

[0155] The interpolation weights for each scale are obtained by solving the weights of the kernel matrix at each scale. For each scale's kernel matrix, a system of linear equations is solved to obtain the interpolation weights. During the solution process, a regularization term is added to the kernel matrix to prevent overfitting, with a regularization coefficient set to 0.001. Taking a large-scale kernel matrix as an example, a system of equations is constructed by multiplying the temperature eigenvalue vector with the kernel matrix, and then iteratively solved using the conjugate gradient method. The upper limit for the number of iterations is set to 1000, and the convergence threshold is set to 10. -6 After 875 iterations, a large-scale interpolation weight vector is obtained, containing 1000 elements, for example, the first 5 elements are [0.0213, -0.0157, 0.0324, -0.0189, 0.0276]. The same method is used to solve for the interpolation weights of the kernel matrix at the medium and small scales, respectively, to obtain the medium-scale weight vector and the small-scale weight vector. For example, the first 5 elements at the medium scale are [0.0417, -0.0233, 0.0519, -0.0341, 0.0382], and the first 5 elements at the small scale are [0.0842, -0.0563, 0.0971, -0.0688, 0.0753].

[0156] The multi-scale radial basis function (RBF) representation is obtained by weighting the interpolation weights and eigenvalues ​​at each scale. For any query coordinate point, the kernel function values ​​at each scale are calculated for both the query point and all sampled points, forming three kernel vectors. These kernel vectors are then multiplied by the interpolation weight vectors at the corresponding scales to obtain the interpolation estimates at the three scales. Taking the query point (0.375, 0.0) as an example, the dot product of the large-scale kernel vector and the large-scale interpolation weights yields an interpolation value of 0.487, the medium-scale value is 0.503, and the small-scale value is 0.512. These three values ​​constitute the multi-scale RBF representation of the query point, representing the temperature feature estimates at different spatial scales. This process is repeated for the gridded set of query points to obtain the complete multi-scale RBF representation.

[0157] A global interpolated representation is obtained by inter-scale fusion of the multi-scale radial basis representation. An adaptive weighting method is used to fuse the interpolation results of the three scales. The weight coefficients are automatically adjusted according to the local temperature gradient: regions with gentle temperature changes are assigned higher weights to larger scales, and regions with drastic temperature changes are assigned higher weights to smaller scales. In specific calculations, the temperature gradient near the query point is estimated. When the gradient value is less than 0.01, the weights for the three scales are set to [0.6, 0.3, 0.1]; when the gradient value is between 0.01 and 0.05, the weights are set to [0.3, 0.5, 0.2]; and when the gradient value is greater than 0.05, the weights are set to [0.1, 0.3, 0.6]. Taking the query point (0.375, 0.0) as an example, the local temperature gradient at this point is 0.023, corresponding to a weight of [0.3, 0.5, 0.2]. The fused interpolated value is 0.3×0.487+0.5×0.503+0.2×0.512=0.499. Perform the same fusion operation on all query points to obtain a global interpolated representation.

[0158] A multi-resolution grid is obtained by dividing spatial coordinates into spatial grids based on spatial hashing encoding. The spatial region for fiber optic temperature monitoring is divided into a multi-resolution grid structure, including three resolution levels: 32×32×1 grid (coarse resolution), 64×64×1 grid (medium resolution), and 128×128×1 grid (fine resolution). At each resolution level, spatial coordinates are mapped to corresponding grid cells. For example, the coordinate point (0.375, 0.0) is located in grid cell (12, 0) in the coarse resolution grid, in grid cell (24, 0) in the medium resolution grid, and in grid cell (48, 0) in the fine resolution grid. By establishing a multi-resolution grid, the characteristics of spatial data can be captured at different scales.

[0159] A hash feature table is constructed by assigning a hash index to each vertex in the multi-resolution mesh and storing the corresponding feature vector. A hash function is used to map the vertex coordinates to integer hash indices. The hash function design combines prime multiplication and modular arithmetic, specifically: for vertex coordinates (i, j, k) and resolution level l, the hash index is ((1022707i+1048573j+1013k)XOR(2048l))MOD2097152. For example, the vertex (12, 0, 0) in the coarse-resolution mesh has a hash index of 823914. A hash table is established, using the hash index as the key to store the corresponding 8-dimensional feature vector. During the initialization phase, all mesh vertices are traversed, their hash indices are calculated, and the initial feature vector of each vertex is obtained through interpolation. For example, the feature vector corresponding to hash index 823914 is [0.492, 0.487, 0.503, 0.512, 0.498, 0.505, 0.488, 0.495]. In this way, a hash feature table containing all grid vertices is constructed.

[0160] Spatial coordinates are mapped to a multi-resolution grid to obtain a neighbor hash index through neighbor vertex lookup. For any query point, the grid cell containing it is determined in each resolution grid, and the eight grid vertices surrounding the point are found. Taking the query point (0.375, 0.0) as an example, in the coarse resolution grid, the point is located within a grid cell surrounded by eight vertices: (12, 0, 0), (12, 1, 0), (13, 0, 0), and (13, 1, 0); in the medium resolution grid, the point is located within a grid cell surrounded by eight vertices: (24, 0, 0); and in the fine resolution grid, the point is located within a grid cell surrounded by eight vertices: (48, 0, 0). Calculate the hash indices of these vertices, for example, the set of nearest neighbor hash indices for [823914, 824936, 846690, 847712, 823914, 824936, 846690, 847712] at coarse resolution.

[0161] Local hash features are obtained by extracting the corresponding feature vector from the hash feature table using the nearest neighbor hash index and performing trilinear interpolation. Trilinear interpolation weights are calculated based on the relative positions of the query point and its eight surrounding grid vertices. For example, the query point (0.375, 0.0) has normalized coordinates of (0.0, 0.0, 0.0) within the coarse-resolution grid cell, with corresponding interpolation weights of [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]. Based on these weights and the feature vector retrieved from the hash table, the coarse-resolution feature representation is calculated using trilinear interpolation as [0.492, 0.487, 0.503, 0.512, 0.498, 0.505, 0.488, 0.495]. The same interpolation operation is performed on the medium-resolution and fine-resolution representations, yielding medium-resolution and fine-resolution feature representations respectively. These three resolution-level feature representations are then concatenated to form a 24-dimensional local hash feature.

[0162] The implicit field representation is obtained by weighting and fusing the global interpolation representation and local hash features based on preset weights. The scalar value obtained from global interpolation is converted into an 8-dimensional vector by repeating the value 8 times. For example, the global interpolation representation of query point (0.375, 0.0) 0.499 is converted to [0.499, 0.499, 0.499, 0.499, 0.499, 0.499, 0.499, 0.499]. The weight of the global interpolation representation is set to 0.4, and the weight of the local hash features is set to 0.6. The 8-dimensional global representation vector and the 24-dimensional local hash features are concatenated to obtain a 32-dimensional feature vector. Each dimension is weighted according to its corresponding weight to obtain the implicit field representation of query point (0.375, 0.0) as a 32-dimensional vector, where the first 8 dimensions are the global interpolation components and the last 24 dimensions are the local hash feature components.

[0163] In this embodiment, by modeling the spatial coordinates and eigenvalues ​​of the denoised image topology features within a multi-scale radial basis function framework, fine-grained characterization of temperature field distribution features at different spatial scales is achieved. By calculating the multi-scale kernel matrix and solving the corresponding interpolation weights, the spatial correlation of temperature changes in both local and global ranges can be adaptively captured. Smoothness and detail fidelity are balanced during the reconstruction process. Combined with a spatial hashing mechanism, high-dimensional spatial information is efficiently compressed and fast feature retrieval is achieved by performing multi-resolution grid partitioning and hash index mapping on the spatial coordinates. Based on local hash feature extraction using nearest-neighbor vertex query and trilinear interpolation, local geometric details and minute temperature change features can be captured with low storage overhead. This enables more accurate reconstruction of the spatial distribution characteristics of complex fiber optic temperature fields, providing stronger modeling support for high-precision temperature monitoring and dynamic temperature field analysis.

[0164] A second aspect of the present invention provides a distributed fiber optic temperature measurement and positioning calibration system based on the Raman effect, comprising:

[0165] The first unit is used to acquire the Raman scattered light signal distributed along the optical fiber transmission path, and extract the temperature distribution along the optical fiber axis by combining the optical path delay relationship to obtain the original temperature sequence.

[0166] The second unit is used to perform dilated causal convolution and linear activation on the original temperature sequence to obtain activation features and perform skip connections and channel splicing to obtain the coordinate prediction value of the fiber starting point, calculate the offset between the coordinate prediction value and the theoretical starting position, and calculate the aligned temperature sequence based on the offset and the original temperature sequence.

[0167] The third unit is used to construct a graph topology structure from the sampling points of the aligned temperature sequence, construct an adaptive adjacency matrix based on the magnitude and directionality of the temperature gradient between sampling points, determine the set of neighboring nodes for each graph node based on the adaptive adjacency matrix, calculate the attention coefficient of each node on the graph topology structure to the neighboring nodes in the set of neighboring nodes, and perform weighted aggregation of the neighboring nodes based on the attention coefficient to obtain the graph topology features.

[0168] The fourth unit is used to perform adaptive lifting wavelet decomposition on the graph topology features to obtain high-frequency detail coefficients, apply adaptive denoising based on Bayesian risk estimation to the high-frequency detail coefficients, combine the lifting wavelet inverse transform reconstruction with variational interpolation upsampling to obtain the reconstructed temperature field, and register the reconstructed temperature field and the coordinate prediction value to obtain the calibrated fiber temperature field.

[0169] A third aspect of the present invention provides an electronic device, comprising:

[0170] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0171] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0172] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0173] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed fiber optic temperature measurement and positioning calibration method based on the Raman effect, characterized in that, include: Raman scattered light signals distributed along the optical fiber transmission path are acquired, and the temperature distribution along the optical fiber axis is extracted by combining the optical path delay relationship to obtain the original temperature sequence. The original temperature sequence is subjected to dilated causal convolution and linear activation to obtain activation features, and skip connections and channel splicing are performed to obtain the coordinate prediction value of the optical fiber starting point. The offset between the coordinate prediction value and the theoretical starting position is calculated, and the aligned temperature sequence is calculated based on the offset and the original temperature sequence. The sampling points of the aligned temperature sequence are constructed into a graph topology. An adaptive adjacency matrix is ​​constructed based on the magnitude and directionality of the temperature gradient between sampling points. The set of neighboring nodes for each graph node is determined based on the adaptive adjacency matrix. The attention coefficient of each node in the graph topology to the neighboring nodes in the set of neighboring nodes is calculated. The neighboring nodes are weighted and aggregated based on the attention coefficient to obtain the graph topology features. The graph topology features are subjected to adaptive lifting wavelet decomposition to obtain high-frequency detail coefficients. Adaptive denoising based on Bayesian risk estimation is applied to the high-frequency detail coefficients. Variational interpolation upsampling is performed in combination with inverse lifting wavelet transform reconstruction to obtain the reconstructed temperature field. The reconstructed temperature field and the coordinate prediction values ​​are registered to obtain the calibrated fiber temperature field.

2. The method according to claim 1, characterized in that, Acquiring Raman scattered light signals distributed along the optical fiber transmission path and extracting the temperature distribution along the optical fiber axis to obtain the original temperature sequence includes: injecting pulsed laser light into the incident end of the optical fiber; acquiring backscattered Raman light signals generated along the optical fiber transmission path, wherein the backscattered Raman light signals include Stokes intensity components and anti-Stokes intensity components; performing photoelectric conversion on the Stokes intensity components and the anti-Stokes intensity components respectively to obtain Stokes electrical signals and anti-Stokes electrical signals; performing cross-correlation calculation on the Stokes electrical signals and the anti-Stokes electrical signals, and determining the temperature sequence based on the cross-correlation peak position. The time delay difference between the Stokes signal and the anti-Stokes signal is calculated, and time-domain alignment is performed based on the time delay difference to obtain an aligned Stokes signal and an aligned anti-Stokes signal. The intensity ratio of the aligned anti-Stokes signal and the aligned Stokes signal at each sampling moment is calculated, and the temperature value at each sampling moment is determined based on the intensity ratio. Based on the propagation speed of the pulsed laser in the optical fiber and the time interval between each sampling moment and the pulsed laser injection moment, the axial position coordinate of the optical fiber corresponding to each sampling moment is calculated. The temperature value and the axial position coordinate of the optical fiber are then correlated to obtain the original temperature sequence.

3. The method according to claim 1, characterized in that, The original temperature sequence is subjected to dilated causal convolution and linear activation to obtain activation features. Skip connections and channel splicing are then performed to obtain the predicted coordinates of the fiber optic starting point. The offset between the predicted coordinates and the theoretical starting position is calculated. Based on the offset and the original temperature sequence, an aligned temperature sequence is calculated, including: mapping the original temperature sequence to state vectors to obtain a state vector sequence; calculating the transition weights between adjacent state vectors in the state vector sequence to obtain a state transition weight matrix; weighting the original temperature sequence to obtain a state-weighted temperature sequence; performing dilated causal convolution on the state-weighted temperature sequence to obtain a convolutional feature sequence; performing linear activation on the state-weighted temperature sequence to obtain activation features; constructing an activation feature sequence based on the activation features; and then... A multi-channel feature sequence is obtained by skipping connections and concatenating channels. The variance of each channel feature in the multi-channel feature sequence is calculated to obtain a channel variance vector. The channel variance vector is used to perform a weighted summation of each channel feature in the activated feature sequence to obtain a weighted aggregated feature vector. A preset causal mask is applied to the weighted aggregated feature vector to obtain a mask feature vector, and an attention score matrix is ​​calculated. A global temporal representation vector is calculated based on the attention score matrix and a linear transformation is performed to obtain the predicted coordinate value of the fiber optic starting point. The fiber axial position coordinates of each sampling point in the original temperature sequence are transformed according to the offset to obtain a corrected axial position. The corrected axial position is then correlated with the temperature value of the corresponding sampling point in the original temperature sequence to obtain an aligned temperature sequence.

4. The method according to claim 1, characterized in that, The sampling points of the aligned temperature sequence are constructed into a graph topology. An adaptive adjacency matrix is ​​constructed based on the magnitude and directionality of the temperature gradient between sampling points. The neighborhood node set of each graph node is determined based on the adaptive adjacency matrix, including: mapping each sampling point in the aligned temperature sequence to a high-dimensional feature vector to obtain a node feature matrix; performing graph structure learning on the node feature matrix to obtain an initial topological connection relationship; constructing a graph topology based on the initial topological connection relationship to obtain a graph node set; performing a difference operation on the temperature values ​​of adjacent graph nodes in the graph node set to obtain a temperature change sequence; and performing variational mode decomposition on the temperature change sequence and extracting the envelopes of different mode components to obtain the temperature gradient magnitude. The system extracts the instantaneous phase of each modal component to obtain the temperature gradient direction feature; it models the inter-node similarity of the temperature gradient amplitude feature to obtain an amplitude correlation graph, models the direction consistency of the temperature gradient direction feature to obtain a direction correlation graph, performs a graph fusion operation on the amplitude correlation graph and the direction correlation graph to obtain a fused correlation graph, extracts the edge weight matrix in the fused correlation graph and performs sparsification to obtain an adaptive adjacency matrix; it performs graph pruning on the adaptive adjacency matrix to obtain a sparse adjacency matrix, performs a graph sampling algorithm on the sparse adjacency matrix to obtain the sampling results of neighboring nodes and performs neighborhood importance evaluation to obtain the neighborhood weight distribution, and selects and determines the set of neighboring nodes for each graph node based on the neighborhood weight distribution.

5. The method according to claim 1, characterized in that, Calculating the attention coefficient of each node in the graph topology to its neighboring nodes in the neighboring node set, and then weighting and aggregating the neighboring nodes based on the attention coefficients to obtain graph topology features includes: extracting node features from the graph topology and constructing feature pairs with the node features of neighboring nodes in the neighboring node set; performing causal relationship analysis on the feature pairs to construct a directed acyclic graph to obtain a causal graph; identifying confounding variables that simultaneously affect the node pairs through path tracing in the causal graph to obtain a set of confounding variables; performing a backdoor adjustment operation on the feature pairs based on the set of confounding variables to control the influence of the confounding variables to obtain adjusted feature pairs; and performing counterfactual substitution on the neighboring node features in the adjusted feature pairs to obtain a... The process involves: 1) Pre-feature pair; 2) Calculating the distance metric between the adjustment feature pair and the intervention feature pair in the feature space to obtain a causal effect value; 3) Concatenating the causal effect value with the adjustment feature pair to obtain a contrast feature; 4) Mapping the contrast feature to a query vector and a key vector and calculating the dot product similarity to obtain an attention weight matrix; 5) Weighted summing of the attention weight matrix to obtain context features; 6) Extracting the correlation strength between the current node and neighboring nodes from the context features to obtain a causal attention score; 7) Calculating the attention coefficient of the current node to its neighboring nodes based on the causal attention score; and 8) Weighted summing of the node features of the neighboring nodes in the set of neighboring nodes based on the attention coefficient to obtain graph topology features.

6. The method according to claim 1, characterized in that, Adaptive lifting wavelet decomposition is performed on the graph topology features to obtain high-frequency detail coefficients. Adaptive denoising based on Bayesian risk estimation is applied to these high-frequency detail coefficients. Variational interpolation upsampling is then performed using inverse lifting wavelet transform reconstruction to obtain a reconstructed temperature field. Registration of the reconstructed temperature field and the predicted coordinates yields a calibrated fiber temperature field. This process includes: performing multi-scale analysis on the graph topology features to calculate information entropy and determine the optimal decomposition level; performing multi-level lifting wavelet decomposition on the graph topology features according to the optimal decomposition level to obtain low-frequency approximation coefficients and high-frequency detail coefficients; extracting the mean and variance from the high-frequency detail coefficients to construct a noise statistical distribution and determining an adaptive threshold based on the noise statistical distribution; and performing soft thresholding on the high-frequency detail coefficients based on the adaptive threshold to obtain denoised high-frequency coefficients. The process involves: performing inverse wavelet transform on the denoised high-frequency coefficients and the low-frequency approximation coefficients to obtain the topological features of the denoised image; continuously fitting the spatial coordinates and eigenvalues ​​of the topological features of the denoised image using a neural implicit representation algorithm to obtain an implicit field representation; performing coordinate sampling and feature mapping on the implicit field representation at a preset target resolution to obtain a reconstructed temperature field; constructing a reference coordinate grid based on the predicted coordinate values; calculating the similarity measure between the reconstructed temperature field and the reference coordinate grid using a differentiable image registration algorithm to obtain a registration loss; optimizing the registration loss through backpropagation to obtain spatial transformation parameters; performing affine transformation and non-rigid deformation on the reconstructed temperature field based on the spatial transformation parameters to obtain a registered temperature field; and performing bicubic interpolation filling on the registered temperature field to obtain a calibration fiber temperature field.

7. The method according to claim 6, characterized in that, The implicit field representation obtained by continuously fitting the spatial coordinates and eigenvalues ​​of the denoised map topological features based on the neural implicit representation algorithm includes: extracting the spatial coordinates and corresponding eigenvalues ​​from the topological features of the denoised map; constructing kernel matrices of multiple scales by calculating the radial distances between the spatial coordinates based on preset multi-scale radial basis functions; solving for the weights of the kernel matrices at each scale to obtain the interpolation weights of the corresponding scale; weighting the interpolation weights of each scale with the eigenvalues ​​to obtain the multi-scale radial basis representation; and performing inter-scale fusion of the multi-scale radial basis representations to obtain the global interpolation representation. Based on spatial hash encoding, a multi-resolution grid is obtained by dividing the spatial coordinates into spatial grids; a hash feature table is constructed by assigning a hash index to each grid vertex in the multi-resolution grid and storing the corresponding eigenvector; the spatial coordinates are mapped to the multi-resolution grid to perform a neighboring vertex query to obtain a neighboring hash index; the corresponding eigenvectors are extracted from the hash feature table based on the neighboring hash index and trilinear interpolation is performed to obtain local hash features; and the implicit field representation is obtained by weighting and fusing the global interpolation representation and the local hash features based on preset weights.

8. A distributed fiber optic temperature measurement and positioning calibration system based on the Raman effect, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire the Raman scattered light signal distributed along the optical fiber transmission path, and extract the temperature distribution along the optical fiber axis by combining the optical path delay relationship to obtain the original temperature sequence. The second unit is used to perform dilated causal convolution and linear activation on the original temperature sequence to obtain activation features and perform skip connections and channel splicing to obtain the coordinate prediction value of the fiber starting point, calculate the offset between the coordinate prediction value and the theoretical starting position, and calculate the aligned temperature sequence based on the offset and the original temperature sequence. The third unit is used to construct a graph topology structure from the sampling points of the aligned temperature sequence, construct an adaptive adjacency matrix based on the magnitude and directionality of the temperature gradient between sampling points, determine the set of neighboring nodes for each graph node based on the adaptive adjacency matrix, calculate the attention coefficient of each node in the graph topology structure to the neighboring nodes in the set of neighboring nodes, and perform weighted aggregation on the neighboring nodes based on the attention coefficient to obtain the graph topology features; the fourth unit is used to perform adaptive lifting wavelet decomposition on the graph topology features to obtain high-frequency detail coefficients, apply adaptive denoising based on Bayesian risk estimation to the high-frequency detail coefficients, combine with inverse lifting wavelet transform reconstruction to perform variational interpolation upsampling to obtain the reconstructed temperature field, and register the reconstructed temperature field and the coordinate prediction value to obtain the calibrated fiber temperature field.

9. An electronic device, characterized in that, include: processor; A memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images