A Method and System for Measuring the Orientation of Underground Pipelines Based on Spatial Coordinate Correction

Abstract

This invention relates to the field of underground pipeline detection and measurement, and provides a method and system for measuring the direction of underground pipelines based on spatial coordinate correction. The method includes: acquiring multi-source underground pipeline detection data and encoding it into a pulse sequence; processing the pulse sequence through an attenuation processing unit containing first-level and second-level attenuation processing units to obtain fused feature data, suppressing transient pulse clusters and smoothing residual fluctuations; based on the fused feature data, performing graph reasoning through a plasticity knowledge graph containing synaptic connection edges to output pipeline category labels and spatial location confidence scores; and generating an adjustment signal based on the spatial location confidence scores to adjust the weight gain of the sensing channels. This invention achieves adaptive switching and enhancement of the sensing channels, solves the problem of direction positioning deviation caused by the loss of signals from a single sensor, and effectively suppresses nonlinear terrain disturbances.

Description

Technical Field

[0001] This application relates to the field of underground pipeline detection and measurement, and in particular to a method and system for measuring the direction of underground pipelines based on spatial coordinate correction. Background Technology

[0002] Currently, the core technologies for as-built surveying of urban underground pipelines mainly rely on electromagnetic induction and ground-penetrating radar (GPR). Electromagnetic induction determines the planar position and burial depth of metallic pipelines by applying an alternating electromagnetic field to the target pipeline and detecting the secondary field. Ground-penetrating radar, on the other hand, identifies non-metallic pipelines and underground cavities by emitting high-frequency electromagnetic pulses and receiving reflected signals. Simultaneously, real-time kinematic (RTK) positioning and inertial measurement units (IMUs) are used to assist in acquiring spatial coordinates and attitude data, and Kalman filtering is typically employed to perform loosely or tightly coupled fusion processing of multi-sensor data.

[0003] However, in actual as-built surveying operations, when encountering non-metallic pipeline sections or pipeline intersections, the received signal of the electromagnetic induction instrument suddenly attenuates or even disappears, forcing the measurement process to be interrupted. Operators can only rely on personal experience to guess the pipeline direction, resulting in significant positioning errors. At the same time, multiple high-amplitude reflection anomalies often appear in the ground-penetrating radar profile images, making it difficult to effectively distinguish the target pipeline from interfering objects such as the stratum interface, leading to frequent misjudgments. Based on the above phenomena, it was found that this phenomenon mainly occurs because sensors based on a single physical principle cannot cover all pipeline material types and burial environments. Furthermore, the existing multi-source data fusion framework based on linear Kalman filtering lacks a mechanism to adaptively select the optimal sensing channel. In areas where electromagnetic signals attenuate, it cannot automatically and smoothly switch to other effective detection modes, resulting in signal loss for non-metallic pipelines or intersecting pipeline sections, causing pipeline direction positioning errors. Summary of the Invention

[0004] To address the limitations of existing technologies, such as the single physical principle sensor's limited sensing dimension and lack of adaptive fusion mechanisms leading to signal loss in non-metallic or intersecting pipeline segments and resulting in deviations in pipeline routing, this application proposes an underground pipeline routing measurement method and system based on spatial coordinate correction. This method encodes multi-source detection data into pulse sequences for adaptive fusion and combines them with a plasticity knowledge graph for graph reasoning and feedback adjustment, thereby achieving blind-zone-free continuous and accurate positioning of mixed metal and non-metal pipelines.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for measuring the orientation of underground pipelines based on spatial coordinate correction includes: acquiring multi-source underground pipeline detection data and encoding it into pulse sequences; processing the pulse sequences through an attenuation processing unit to obtain fused feature data, wherein the attenuation processing unit includes at least a first-level attenuation processing unit and a second-level attenuation processing unit, the first-level attenuation processing unit being configured to suppress transient pulse clusters exceeding a frequency threshold, and the second-level attenuation processing unit being configured to smooth residual fluctuations; based on the fused feature data, performing graph inference through a plasticity knowledge graph to output pipeline category labels and spatial location confidence scores, wherein the plasticity knowledge graph contains synaptic connection edges, and the graph inference includes weight propagation of spatial location confidence scores based on synaptic connection edges; and generating adjustment signals for the corresponding sensing channels based on the spatial location confidence scores, and feeding back to adjust the weight gain of the sensing channels corresponding to the multi-source underground pipeline detection data.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the construction process of the attenuation processing unit specifically includes: configuring a recurrent spiking network among N spiking neurons with a first attenuation time constant to obtain a first-level attenuation processing unit. The first attenuation time constant is used to control the membrane potential integration window length of the basal dendritic module for the input pulse sequence, and the recurrent spiking network is used to realize short-term pulse pattern memorization. At the output of the first-level attenuation processing unit, M spiking neurons with a second attenuation time constant are coupled to form a second-level attenuation processing unit. The second attenuation time constant is greater than the first attenuation time constant and is used to control the integration smoothing window length of the apical dendritic module for the input pulse sequence. Pulse sequence samples with different degrees of surface undulation are acquired, and the synaptic connection weights in the first-level and second-level attenuation processing units are jointly trained with the goal of minimizing the difference between the output pulse sequence of the second-level attenuation processing unit and the target smoothed pulse sequence. The trained first-level and second-level attenuation processing units are then deployed in series to form the attenuation processing unit.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the process of processing the pulse sequence through an attenuation processing unit to obtain fused feature data includes: a first-stage attenuation processing unit performing membrane potential integration and resetting on the pulse sequence, suppressing transient pulse clusters, and outputting an intermediate sequence; a second-stage attenuation processing unit performing long-time-window integration on the intermediate sequence, smoothing residual fluctuations, and outputting a smoothed sequence; and decoding the smoothed sequence to obtain the fused feature data.

[0008] In conjunction with the first aspect mentioned above, one possible implementation of the construction process of the plastic knowledge graph specifically includes: acquiring historical and standardized data containing pipeline attributes, pipeline spatial relationships, geological environment information, and multi-source detection signal characteristics as knowledge sources. Based on the knowledge sources, a plastic knowledge graph containing nodes and synaptic connection edges is constructed. Nodes include pipeline nodes representing underground physical entities, feature nodes representing detection signal characteristics, and spatial constraint nodes representing spatial topological constraints. Synaptic connection edges connect nodes and include feature-to-pipeline mapping edges representing the mapping relationship between features and pipeline categories, as well as spatial constraint edges representing the spatial logical relationship between pipelines. Based on historical detection data samples, with the goal of maximizing the prediction accuracy of feature-to-pipeline mapping edges in associating feature nodes with corresponding pipeline nodes, the initial weights of feature-to-pipeline mapping edges are trained in a supervised manner. During pipeline detection applications, the weights of feature-to-pipeline mapping edges are dynamically updated online based on the pipeline category labels and spatial location confidence scores output by graph reasoning.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves performing graph reasoning using a plastic knowledge graph to output pipeline category labels and spatial location confidence scores. Specifically, this includes: the plastic knowledge graph receiving fused feature data and activating feature nodes whose similarity to the fused feature data exceeds a preset threshold. Through feature-to-pipeline mapping edges, the confidence scores of the activated feature nodes are propagated to the connected pipeline nodes, and the overall confidence score of each pipeline node is calculated. Based on the overall confidence scores, the pipeline nodes are sorted, and at least one pipeline node with an overall confidence score exceeding the confidence threshold is selected as a candidate pipeline node. Based on spatial constraint edges, the spatial relationships between the candidate pipeline nodes are validated for consistency. The category attributes and overall confidence scores of the validated candidate pipeline nodes are output as pipeline category labels and spatial location confidence scores, respectively. For candidate pipeline nodes that fail validation, the overall confidence score of the corresponding pipeline node is reduced according to the type of constraint violated, and the selection and validation steps are re-executed.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the process of acquiring multi-source underground pipeline detection data and encoding it into a pulse sequence specifically includes: acquiring multi-source underground pipeline detection data, which includes electromagnetic induction detection signals, ground-penetrating radar reflection waveform signals, inertial measurement data, and spatial coordinate positioning data. Responding to the voltage amplitude characteristics of the electromagnetic induction detection signals, a frequency and pulse conversion coding unit is activated, mapping the voltage amplitude to the firing frequency of the corresponding spiking neurons per unit time, generating a first pulse sequence. Responding to the waveform amplitude distribution and temporal structure characteristics of the ground-penetrating radar reflection waveform signals, a delay and pulse conversion coding unit is activated, allocating corresponding pulse firing delays based on the amplitude of sampling points in the waveform, generating a second pulse sequence. Responding to the motion state change rate characteristics of the inertial measurement data, an event-driven conversion coding unit is activated, triggering a pulse event only when the change in acceleration or angular velocity exceeds a dynamic threshold, generating a third pulse sequence. Responding to the digital quantization characteristics of the spatial coordinate positioning data, a phase coding unit is activated, mapping the binary bits of the coordinate values ​​to pulses with specific phases, generating a fourth pulse sequence. The first, second, third, and fourth pulse sequences are synchronized and aligned in time to output a pulse sequence under a unified time reference.

[0011] In conjunction with the first aspect mentioned above, one possible implementation includes, after outputting the pipeline category label and spatial location confidence score, a knowledge graph-guided cross-modal feature point matching and 3D reconstruction process. Specifically, this includes: extracting electromagnetic induction direction feature points, ground-penetrating radar reflection feature points, and spatial trajectory feature points from multi-source underground pipeline detection data. Based on spatiotemporal proximity, the electromagnetic induction direction feature points, ground-penetrating radar reflection feature points, and spatial trajectory feature points are paired to generate a cross-modal candidate matching pair set. In response to each candidate matching pair in the cross-modal candidate matching pair set, at least one spatial constraint edge associated with the pipeline node corresponding to the pipeline category label is extracted from the plasticity knowledge graph. Based on the spatial constraint edge, the geometric relationship of the feature points associated with the candidate matching pair in 3D space is verified for consistency. The candidate matching pairs that pass the verification are confirmed as valid cross-modal matching point pairs. Based on all valid cross-modal matching point pairs, a 3D spatial location model of the underground pipeline is reconstructed.

[0012] Secondly, a system for measuring the orientation of underground pipelines based on spatial coordinate correction is provided, comprising: a multi-source data encoding unit for acquiring multi-source underground pipeline detection data and encoding it into pulse sequences; an attenuation processing unit for processing the pulse sequences to obtain fused feature data, wherein the attenuation processing unit includes at least a first-level attenuation processing unit and a second-level attenuation processing unit, the first-level attenuation processing unit being configured to suppress transient pulse clusters exceeding a frequency threshold, and the second-level attenuation processing unit being configured to smooth residual fluctuations; a plasticity knowledge graph inference unit for performing graph inference based on the fused feature data through a plasticity knowledge graph, outputting pipeline category labels and spatial location confidence scores, wherein the plasticity knowledge graph contains synaptic connection edges, and the graph inference includes weight propagation of spatial location confidence scores based on synaptic connection edges; and a feedback adjustment unit for generating adjustment signals for the corresponding sensing channels based on the spatial location confidence scores and providing feedback adjustment to the weight gain of the sensing channels corresponding to the multi-source underground pipeline detection data.

[0013] In conjunction with the second aspect mentioned above, in one possible implementation, the underground pipeline alignment measurement system belongs to a heterogeneous computing platform. This platform includes: a front-end sensing and encoding module, integrating hardware interfaces for an electromagnetic inductor, ground-penetrating radar, inertial measurement unit, and real-time dynamic positioning receiver, and containing a multi-source data encoding unit for receiving raw detection data in real time and performing pulse encoding; a core fusion and inference computing module, including an attenuation processing unit and a plastic knowledge graph inference unit, for performing pulse sequence fusion processing and graph inference computing; and a back-end control and feedback module, including a feedback adjustment unit and integrating a display unit and a human-machine interface, for generating feedback adjustment commands based on the inference results, controlling the entire system workflow, and outputting visualized results.

[0014] In conjunction with the first aspect mentioned above, in one possible implementation, the plastic knowledge graph reasoning unit further includes: a remote knowledge injection interface, used to receive updated pipeline attributes, spatial relationship rules, and typical signal feature data, and incrementally update the nodes and edges of the local plastic knowledge graph; and an incremental learning engine, used to dynamically optimize the weights of the feature-pipeline mapping edges online based on the pipeline category labels and spatial location confidence scores output by graph reasoning.

[0015] This application provides a method and system for measuring the alignment of underground pipelines based on spatial coordinate correction. It unifies the encoding of multi-source heterogeneous detection data (electromagnetic induction, ground-penetrating radar, inertial measurement, and spatial positioning) into pulse sequences, breaking the limitations of single-physical-principle sensors and enabling information-level fusion of heterogeneous data in a unified pulse domain. Specifically, when the electromagnetic induction signal attenuates in non-metallic pipeline sections, the pulse-coded information from the ground-penetrating radar channel can automatically obtain higher weight gain through graph reasoning and feedback adjustment mechanisms of a plasticity knowledge graph, achieving adaptive and smooth switching of the sensing channel, thus ensuring continuous tracking and accurate positioning of the pipeline alignment. Simultaneously, an attenuation processing mechanism including first-stage and second-stage attenuation processing units is introduced. The first-stage attenuation processing unit utilizes a cyclic pulse connection network and a first attenuation time constant to rapidly suppress transient high-density pulse clusters, avoiding the tailing effect caused by nonlinear terrain disturbances. The second-stage attenuation processing unit smooths residual fluctuations through long-time-window integration, replacing the recursive smoothing process of traditional linear Kalman filtering, significantly improving the accuracy and robustness of elevation measurements in complex surface environments. Furthermore, the plasticity knowledge graph encodes historical exploration experience and spatial topological constraints into a graph structure. It realizes graph reasoning through weight propagation and consistency verification of synaptic connection edges, providing traceable prior knowledge guidance for pipeline interpretation, eliminating subjective errors introduced by manual annotation alignment, and improving the credibility and standardization of as-built measurement results. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the method for measuring the direction of underground pipelines based on spatial coordinate correction, provided in an embodiment of this application. Figure 2 A schematic diagram of the attenuation processing unit construction and data processing flow in the underground pipeline routing measurement method based on spatial coordinate correction provided in the embodiments of this application; Figure 3 A schematic diagram of the plasticity knowledge graph construction and reasoning process in the underground pipeline routing measurement method based on spatial coordinate correction provided in the embodiments of this application; Figure 4 A schematic diagram of the pulse coding process for multi-source underground pipeline detection data in the underground pipeline routing measurement method based on spatial coordinate correction provided in the embodiments of this application; Figure 5 A schematic diagram of the cross-modal feature point extraction and matching process in the underground pipeline routing measurement method based on spatial coordinate correction provided in the embodiments of this application; Figure 6 A schematic diagram of the three-dimensional reconstruction process of underground pipelines in the underground pipeline routing measurement method based on spatial coordinate correction provided in the embodiments of this application; Figure 7A schematic diagram of the system structure of the underground pipeline routing measurement system based on spatial coordinate correction provided in the embodiments of this application. Detailed Implementation

[0017] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0018] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0019] Example 1: like Figure 1 As shown in the figure, this embodiment provides a method for measuring the direction of underground pipelines based on spatial coordinate correction. This method constructs a complete closed-loop logic from the perception of heterogeneous data at the bottom layer to the reasoning of knowledge at the top layer, and then to the adaptive adjustment of the perception channel at the bottom layer. Specifically, it includes the following four core steps.

[0020] Step 100: Obtain multi-source underground pipeline detection data and encode it as a pulse sequence.

[0021] Multi-source underground pipeline detection data refers to a collection of heterogeneous detection signals acquired by sensors based on different physical principles, exhibiting fundamental differences in data format, sampling frequency, and physical dimensions. Multi-source underground pipeline detection data includes not only electromagnetic induction detection signals and ground-penetrating radar reflection waveform signals, but can also encompass inertial measurement data, spatial coordinate positioning data, acoustic detection signals, and even infrared thermal imaging data, as long as they can reflect the physical properties or spatial location characteristics of underground pipelines.

[0022] A pulse sequence is a time-series signal stream consisting of a series of discrete pulse events with precise timestamps. It carries information through the pulse's firing time, frequency, or phase, rather than relying on continuous analog amplitudes.

[0023] Unifying heterogeneous, multi-source underground pipeline detection data into pulse sequences essentially transforms continuous or discrete analog signals from different physical domains into a unified stream of discrete events within the pulse time domain through specific encoding mapping rules. For example, the continuous voltage amplitude of electromagnetic induction can be mapped to the pulse firing frequency, and the waveform amplitude of ground-penetrating radar can be mapped to the pulse firing delay. This allows heterogeneous data that could not be directly weighted and fused at the numerical level to be aligned and integrated on a unified pulse time axis.

[0024] This overcomes the information loss and alignment difficulties inherent in traditional multi-sensor fusion, which requires projecting all data into the same numerical space (such as the state space of a Kalman filter). Pulse domain fusion preserves the physical characteristics of the original signals from each sensor, avoiding distortion introduced by forced numerical conversion. Simultaneously, it provides a native data interface for subsequent event-driven neuromorphic computing, fundamentally solving the problem of heterogeneous data being "unalignable and difficult to fuse."

[0025] Step 200: Process the pulse sequence through the attenuation processing unit to obtain fused feature data. The attenuation processing unit includes at least a first-level attenuation processing unit and a second-level attenuation processing unit. The first-level attenuation processing unit is configured to suppress transient pulse clusters that exceed the frequency threshold, and the second-level attenuation processing unit is configured to smooth residual fluctuations.

[0026] The attenuation processing unit refers to an algorithm module or hardware logic unit that uses the dendritic computing mechanism of biological neurons to perform nonlinear filtering and feature integration on the temporal dynamic characteristics of the input pulse sequence.

[0027] Fusion feature data refers to the low-dimensional dense vector or pulse statistical features extracted after filtering and integration by the attenuation processing unit. These features can simultaneously represent effective information from multi-source detection signals and suppress environmental disturbance noise. For the macroscopic function of two-stage attenuation, suppressing transient pulse clusters exceeding the frequency threshold means identifying and rapidly attenuating those abnormally high-frequency pulse clusters that are densely emitted within an extremely short time window, such as the pulse bursts from inertial sensors caused by violent shaking of the measuring wheel. Smoothing residual fluctuations refers to performing integral smoothing over a time window on the low-frequency small-amplitude oscillations that still exist after the first stage of rapid suppression, in order to output stable feature representations.

[0028] In some implementations, when a mixed pulse sequence containing transient pulse clusters and normal pulses is input to the attenuation processing unit, the first-stage attenuation processing unit uses its shorter time constant to generate rapid saturation and leakage effects on the high-frequency dense input during the membrane potential integration process, thereby "counteracting" the continuous influence of the transient pulse clusters on the membrane potential and outputting only an intermediate sequence reflecting the normal signal. Subsequently, the second-stage attenuation processing unit receives this intermediate sequence, uses its longer time constant to perform long-window integration, smooths the residual small fluctuations, and finally outputs a stable smooth sequence and decodes it into fused feature data.

[0029] This effectively eliminates the inherent defects of traditional Kalman filtering and its derivative linear filters in handling non-Gaussian and nonlinear terrain disturbances (traditional Kalman filtering, based on linear state equations and Gaussian noise assumptions, not only fails to converge quickly when encountering transient spike noise caused by road covers or potholes, but also produces a "tailing effect" that deviates from the true value at multiple subsequent sampling points). Through a two-stage attenuation mechanism, the first stage of rapid nonlinear suppression directly "blocks" the propagation of transient disturbances at the neuron level, and the second stage of smoothing ensures output stability. This achieves a brain-like processing pipeline of rapid identification, rapid suppression, and fine smoothing, fundamentally eliminating the tailing effect and significantly improving the robustness of elevation and attitude measurements in complex terrain environments.

[0030] Step 300: Based on the fused feature data, perform graph reasoning through the plasticity knowledge graph to output pipeline category labels and spatial location confidence scores. The plasticity knowledge graph contains synaptic connection edges, and the graph reasoning includes the propagation of spatial location confidence scores based on the weights of the synaptic connection edges.

[0031] Plasticity knowledge graphs refer to graph data structures and computing engines that encode prior knowledge, spatial topological constraints, and signal feature mapping relationships in the field of underground pipeline detection into a graph structure, and whose internal connection weights can be dynamically adjusted based on inference results and online data. The overall structure is not a traditional static triplet knowledge base, but rather a dynamic graph that simulates the plasticity of biological neural synapses.

[0032] Synaptic connection edges are weighted directed edges that connect different nodes (such as feature nodes and pipeline nodes) in a graph. Their weight values ​​characterize the confidence level of the association between nodes or the rigidity of the constraint, and the weight can be enhanced or weakened during the learning process.

[0033] Graph reasoning refers to the computational process of propagating confidence, aggregating and verifying spatial constraints along the graph structure based on the activation state of nodes and the weights of synaptic connection edges in the graph.

[0034] Pipeline category labels refer to the semantic classification results of the detection target's material, purpose, and other attributes (such as "DN300 ductile iron water supply pipe" or "PE gas pipe").

[0035] Spatial location confidence refers to a quantitative evaluation index of the reliability of the system's output pipeline spatial location and category determination results, which is usually a probability value between 0 and 1.

[0036] In some implementations, after the fused feature data is input into the plasticity knowledge graph, feature nodes with a similarity to the feature exceeding a preset threshold can be activated first. Then, the initial confidence of the activated feature nodes is propagated to the connected pipeline nodes according to the weight ratio of the edges through synaptic connection edges, so that the comprehensive confidence of each pipeline node can be calculated.

[0037] Candidate pipeline nodes are then selected based on their overall confidence scores. Spatial constraints (such as pipeline spacing and intersection angle specifications) encoded in the graph are used to verify the consistency of spatial logic between candidate nodes. Only nodes that pass the verification can have their pipeline category label and spatial location confidence score finally output.

[0038] This approach replaces traditional data-driven black-box classification, which relies on manual visual interpretation or lacks physical logic, with knowledge-guided graph reasoning. It effectively avoids misjudgments caused by multiple similar pseudo-anomaly reflections in ground-penetrating radar profiles, a problem that arises when traditional interpretation is challenged. By explicitly encoding prior knowledge such as industry standards and historical experience into spatial constraint edges through a malleable knowledge graph, logical consistency verification is enforced during reasoning. This effectively eliminates pseudo-anomaly features (such as stratigraphic interface reflections) that do not conform to the topological laws of pipeline space, providing an interpretable reasoning chain, significantly reducing the misjudgment rate, and eliminating errors caused by subjective human intervention.

[0039] Step 400: Based on the spatial location confidence, generate the adjustment signal for the corresponding sensing channel and feed back to adjust the weight gain of the sensing channel corresponding to the multi-source underground pipeline detection data.

[0040] Among them, the sensing channel refers to the complete data processing link from data encoding to feature extraction for specific types of multi-source underground pipeline detection data (such as electromagnetic induction channel and ground-penetrating radar channel).

[0041] Regulation signals are control instructions generated from high-level inference results and used to guide the lower-level perception network to adjust its parameters or resource allocation.

[0042] Weight gain refers to the amplification factor applied to the synaptic connection weights or pulse firing sensitivity in the sensing channel, which determines the proportion of influence of the channel's output information in subsequent fusion.

[0043] In some implementations, a regulating signal can be generated based on a high-confidence pipeline category label (e.g., determined to be a "non-metallic PE gas pipe") and fed back to the underlying multi-source data encoding and fusion network. Specifically, if the target is determined to be a non-metallic pipeline, the regulating signal will reduce the weight gain of the electromagnetic induction sensing channel (because electromagnetic methods are ineffective for non-metallic materials) while increasing the weight gain of the ground-penetrating radar sensing channel. This allows subsequent input detection data to automatically assign higher synaptic weights to the ground-penetrating radar pulse sequence during pulse fusion, thereby achieving adaptive switching and enhancement of the sensing channels at the signal level.

[0044] This effectively constructs a feedback loop from high-level semantic cognition to low-level signal perception, completely solving the problem in traditional solutions where the system cannot automatically find an alternative signal source when the electromagnetic signal attenuates and disappears in non-metallic pipe sections, and can only interrupt or guess blindly. By adjusting the weight gain through feedback, the system can automatically "shift its attention" to another effective channel (such as ground-penetrating radar) when it is confirmed that a certain channel is invalid, just like a biological attention mechanism. This achieves seamless and smooth switching of perception modes, ensuring the continuity and blind-spot-free coverage of pipeline tracking.

[0045] Based on the above technical solution, a complete technical closed loop is achieved through the underground pipeline routing measurement method based on spatial coordinate correction. This loop includes pulse coding to unify heterogeneous data, two-stage attenuation to suppress nonlinear disturbances, knowledge graph-guided interpretable reasoning, and finally adaptive channel switching via feedback closed loop. This not only breaks through the perception limitations of a single sensor and enables continuous blind-zone-free tracking of mixed metal and non-metal pipelines, but also eliminates the trailing effect of traditional linear filtering and the subjective error of manual interpretation, thus improving the accuracy, robustness, and automation level of underground pipeline spatial coordinate measurement in complex surface environments.

[0046] Example 2: like Figure 2 As shown, based on step 200 of Embodiment 1, this embodiment further elaborates on the internal construction mechanism and execution data flow of the attenuation processing unit, providing specific algorithm sinking and neural mechanism mapping for the functionally limited attenuation processing unit.

[0047] The construction process of the attenuation processing unit specifically includes: Step 210: Configure a recurrent pulse connection network between spiking neurons containing a first decay time constant to obtain a first-level decay processing unit. The first decay time constant is used to control the length of the membrane potential integration window of the basal dendritic module for the input pulse sequence. The recurrent pulse connection network is used to realize short-term pulse pattern memory.

[0048] The first decay time constant is a core time parameter that determines the leakage rate of the spiking neuron's membrane potential. The smaller its value, the faster the neuron forgets historical inputs, and the shorter the integration window. For example, when dealing with severe ground turbulence, the first decay time constant can be set to about 5 milliseconds to ensure that transient disturbances are not remembered for a long time.

[0049] The basal dendrite module refers to the algorithmic subunit that simulates the computational function of the basal dendrites of biological neurons. It is responsible for receiving the raw pulse input from the bottom layer and using the rapid decay characteristics to perform preliminary filtering of high-frequency noise.

[0050] A recurrent spiking network is a feedback synaptic connection structure established between spiking neurons at the same level, which allows the spiking pulses fired by a neuron to be used as input and fed back to itself or other neurons at the same level, thereby maintaining a specific spiking pattern for a short period of time.

[0051] In terms of operational principle, the mathematical calculation mechanism of the basal dendritic module follows a leakage integral-firing model: when the input pulse sequence arrives at the basal dendritic module, each pulse injects a certain amount of charge into the membrane potential of the corresponding neuron through synaptic weights. Due to the extremely small first decay time constant, the membrane potential rapidly leaks and decays when no new pulse is received. Simultaneously, the recurrent pulse connection network allows the neuron's output pulse to be used as input feedback again through recurrent synapses after firing, forming a short-term high-level maintenance of the membrane potential, i.e., short-term pulse pattern memory.

[0052] When a transient pulse cluster input exceeding the frequency threshold is encountered, the membrane potential will rapidly integrate to the firing threshold in a very short time and trigger pulse reset. Due to the counteracting effect of rapid decay and cyclic feedback, the continuous excitation effect caused by the transient pulse cluster is quickly suppressed, and a long-term accumulation of membrane potential cannot be formed.

[0053] It should be understood that although this embodiment preferably uses a leakage integral-emission model and a time constant of 5 milliseconds, in other embodiments, the first decay time constant can also be set to other values ​​in the range of 2 to 10 milliseconds according to the sensor sampling frequency and terrain disturbance characteristics, as long as it can meet the functional requirement of rapidly suppressing transient pulse clusters.

[0054] This step directly "counters" transient high-frequency noise at the neuron level through the rapid decay and cyclic memory mechanism of basal dendrites. This addresses the problem of traditional linear Kalman filters, which, when dealing with non-Gaussian impulse noise, force abnormal spikes into the state estimate because their state equation must smooth all observations, causing tailing deviations at subsequent sampling points. The basal dendritic module, with its extremely short time constant and cyclic reset, allows transient perturbations to be rapidly digested and forgotten within the neuron, fundamentally blocking the downstream propagation path of noise.

[0055] Step 211: At the output of the first-level attenuation processing unit, M spiking neurons containing the second attenuation time constant are coupled to form a second-level attenuation processing unit. The second attenuation time constant is greater than the first attenuation time constant and is used to control the length of the integral smoothing window of the top dendrite module for the input pulse sequence.

[0056] The second decay time constant refers to the time parameter that controls the leakage rate of the membrane potential of the second-level neuron. Its value is greater than that of the first decay time constant, which means that the neuron has a longer memory time for the input information and a wider integration window. For example, in a scenario of smoothing residual elevation fluctuations, the second decay time constant can be set to about 50 milliseconds.

[0057] The apical dendrite module refers to the algorithmic subunit that simulates the function of the apical dendrites of biological neurons. It is responsible for receiving the intermediate pulse sequence after filtering by the first-level basal dendrites and using the long time window integral characteristic to perform fine smoothing modulation on the residual small fluctuations.

[0058] In terms of operational principle, the apical dendritic module also operates based on the membrane potential integration mechanism of spiking neurons. However, due to its second decay time constant, which is much larger than the first decay time constant, its membrane potential leakage rate is extremely slow. When the pulses of the intermediate sequence are input into the apical dendritic module, the charge injected by each pulse remains on the membrane potential for a long time and accumulates slowly. This is equivalent to performing a moving average integration of the pulse firing frequency over a relatively long time window. This long time window integration ensures that even if there are low-frequency, small-amplitude pulse fluctuations caused by minor topographic undulations in the intermediate sequence, they will be mutually canceled and smoothed out during the long accumulation process, thus outputting a stable pulse firing pattern.

[0059] This step utilizes the long-window smoothing mechanism of the apical dendrite module to compensate for the potential signal over-truncation caused by the rapid suppression of basal dendrites. This allows the basal dendrites to quickly suppress transient spikes, while the apical dendrites handle the fine smoothing of residual glitches. The two, connected in series, form a complete nonlinear filtering pipeline, completely replacing the cumbersome and tail-prone recursive smoothing process in traditional Kalman filtering.

[0060] Step 212: Obtain pulse sequence samples with different degrees of surface undulation. With the goal of minimizing the difference between the output pulse sequence of the second-level attenuation processing unit and the target smooth pulse sequence, jointly train the synaptic connection weights in the first-level attenuation processing unit and the second-level attenuation processing unit.

[0061] Joint training refers to the process of simultaneously optimizing the parameter learning of the synaptic connection weights within the two layers by treating the first and second level attenuation processing units as an end-to-end integrated network and using a backpropagation algorithm. The "target smooth pulse sequence" referred to in this application is an ideal, undisturbed pulse sequence label representing the actual terrain elevation or pipeline attitude, generated by manual annotation or a high-precision reference system.

[0062] In terms of operational principle, the joint training process employs a backpropagation algorithm based on alternative gradients. First, it collects measured pulse sequence samples containing various surface undulations (such as flat roads, manhole cover protrusions, pothole depressions, etc.) and obtains the corresponding target smooth pulse sequences as supervision labels. Then, the samples are input into the constructed two-stage cascaded network. Forward propagation then yields the smoothed sequence output by the second-stage attenuation processing unit. The difference in loss function between the output smoothed sequence and the target smoothed pulse sequence can then be calculated, for example, using mean squared error loss.

[0063] Since the firing process of a spiking neuron is a non-differentiable discrete event, an alternative gradient function is introduced during training to transform the discrete threshold of spiking into a continuously differentiable approximate gradient. This allows the error signal to propagate back through the discrete firing point to all synaptic connection weights in both the first and second levels, and gradient updates are performed with the goal of minimizing this difference. Therefore, iterative joint training with a large number of samples is required. The first-level basal dendritic module learns the optimal transient suppression recurrent weights, and the second-level apical dendritic module learns the optimal long-time-window smoothing integral weights. Together, they achieve the optimal filtering state.

[0064] This joint training design avoids the local optimum problem that may be caused by hierarchical independent optimization, and ensures the deep coupling and tacit cooperation of the two-level mechanisms of fast suppression and fine smoothing at the parameter level, so that the entire attenuation processing unit has the strongest adaptive robustness when facing unknown and complex terrain disturbances.

[0065] Step 213: Deploy the first-level attenuation processing unit and the second-level attenuation processing unit in series after training to form an attenuation processing unit.

[0066] In terms of operational principle, serial deployment means that the data flow strictly follows a unidirectional transmission path, first passing through the first-level basal dendritic module for filtering, and then smoothly entering the second-level apical dendritic module. The output of the first-level processing unit is directly coupled to the input of the second-level processing unit, without any other numerical domain conversion or buffering blocking in between, ensuring the continuity and real-time performance of pulse domain processing.

[0067] Its cascaded deployment forms a non-bypassable neuromorphic nonlinear filtering pipeline. Any input pulse must first undergo transient counter-current testing by the basal dendrites, and then undergo long-term smoothing by the apical dendrites, thereby ensuring that the final output fused feature data is absolutely free of tailing effects and high-frequency spikes.

[0068] The process of processing the pulse sequence through the attenuation processing unit to obtain fused feature data includes: Step 220: The first-stage attenuation processing unit integrates and resets the membrane potential of the pulse sequence, suppresses transient pulse clusters, and outputs the intermediate sequence.

[0069] Among them, membrane potential integration and reset refers to the core dynamic process of spiking neurons: the neuron continuously accumulates the charge brought by the input pulse, causing the membrane potential to rise. When the membrane potential exceeds the firing threshold, the neuron immediately fires an output pulse and resets the membrane potential to the resting potential level in preparation for the next round of integration.

[0070] The intermediate sequence refers to the pulse event stream that has been rapidly suppressed by the first-stage attenuation processing unit, filtering out transient high-frequency pulse clusters but may still contain small low-frequency fluctuations.

[0071] In some implementations, when a mixed pulse sequence containing transient pulse clusters is input to the first-level decay processing unit, neurons in the basal dendritic module begin membrane potential integration. Because the transient pulse clusters arrive densely within a very short time, the membrane potential rapidly rises to the firing threshold, triggering pulse firing and immediately resetting the membrane potential.

[0072] After the reset, due to the extremely small first decay time constant, the membrane potential leaks and falls back rapidly, so that subsequent transient pulses must start integration again and cannot use the residual potential of the preceding pulses to form a continuous superposition.

[0073] This frequent integration-reset cycle effectively offsets and suppresses the continuous impact of transient pulse clusters, retaining only those normal pulse signals that conform to the normal frequency distribution to form the intermediate sequence output.

[0074] This allows the rapid reset and leakage of membrane potential to achieve immediate isolation and removal of transient abnormal events, avoiding the continuous tailing deviation caused by outliers being included in the state estimation matrix in traditional Kalman filtering.

[0075] Step 221: The second-stage attenuation processing unit performs long-time window integration on the intermediate sequence, smooths the residual fluctuations, and outputs a smoothed sequence.

[0076] Among them, long-window integration refers to the calculation process of accumulating and averaging the influence of the input pulse on the membrane potential over a long time span by using a large decay time constant.

[0077] A smooth sequence refers to an ideal pulse event stream that, after undergoing a second-stage attenuation process, has a stable frequency, smooth fluctuations, and is free of transient spikes and high-frequency glitches.

[0078] In some implementations, after the intermediate sequence is input into the second-level decay processing unit, the neurons of the top dendrite module retain the charge injection of each input pulse for a long time due to their large second decay time constant.

[0079] During long-window integration, slight fluctuations in pulse firing frequency caused by minor topographic undulations in the intermediate sequence are smoothed out over a long period of membrane potential accumulation. For example, slightly higher frequency pulse inputs and slightly lower frequency pulse inputs within several adjacent time windows neutralize each other under the membrane potential accumulation effect of long-window integration, allowing the neuron's membrane potential to maintain a steady upward trajectory, thereby outputting a smooth sequence with a uniform and stable firing frequency.

[0080] This long-window integration mechanism can achieve fine-grained smoothing of signals without introducing hysteresis, preserving the low-frequency effective variation trend of real terrain elevation or pipeline attitude, while completely filtering out residual random fluctuation noise.

[0081] Step 222: Decode the smooth sequence to obtain fused feature data.

[0082] Decoding refers to the statistical mapping process of converting a discrete smooth sequence in the pulse domain back into a continuous eigenvector in the numerical domain.

[0083] In some implementations, the decoding process uses a frequency statistical decoding scheme, which involves counting the number of pulses fired or the firing frequency of each spiking neuron in the second-level attenuation processing unit within a preset sliding time window, and mapping the frequency value to the corresponding dimension component of the floating-point feature vector.

[0084] For example, by using a 200-millisecond sliding window with a 50-millisecond step, the firing frequency of each of the 256 neurons is counted, thereby generating a 128-dimensional fusion feature vector, i.e., fusion feature data.

[0085] It should be understood that, in addition to frequency statistical decoding, other statistical mapping methods such as time-weighted decoding or phase-average decoding can also be used in other embodiments, as long as they can accurately extract the effective information carried in the smooth sequence.

[0086] This decoding process can seamlessly connect the results of neuromorphic pulse computation to subsequent traditional numerical computation modules (such as knowledge graph reasoning), realizing cross-domain bridging between neuromorphic computation and symbolic logic reasoning, and ensuring the closed-loop data link of the overall technical solution.

[0087] Based on the above technical solution, a robust underlying algorithm is provided through the attenuation processing unit. This allows for the introduction of the basal and parietal dendrite division mechanism of multi-compartment spiking neurons, and, in conjunction with joint training and cascaded deployment, constructs a two-stage nonlinear filtering pipeline with fast hedging suppression and long-term window fine smoothing. This completely eliminates the unavoidable tailing effect and over-smoothing problem of traditional Kalman filtering when dealing with non-Gaussian transient disturbances, thus improving the root mean square error accuracy of pipeline elevation and attitude measurements in complex terrain environments.

[0088] Example 3: like Figure 3 As shown, based on step 300 of Embodiment 1, this embodiment further elaborates on the internal construction mechanism of the plastic knowledge graph and the specific execution logic of graph reasoning, providing a graph structure underlying implementation and reasoning logic binding for the functionally limited plastic knowledge graph.

[0089] The construction process of a plastic knowledge graph specifically includes: Step 310: Obtain historical and standard data containing pipeline attributes, pipeline spatial relationships, geological environment information, and multi-source detection signal characteristics as a source of knowledge.

[0090] The knowledge source refers to the multi-dimensional dataset used to construct the prior topology and initial weights of the graph, specifically including: Pipeline attributes include pipe diameter, material, burial depth range, and ownership unit; The spatial relationships of pipelines include parallel spacing, intersection angles, and hierarchical order; Geological environmental information includes stratigraphic type, water content, and dielectric constant; The characteristics of multi-source detection signals include the peak intensity of electromagnetic induction signals, attenuation curve parameters, and the depth and radius of curvature of the hyperbolic reflection vertex of ground-penetrating radar.

[0091] The above data can be obtained from historical databases of as-built surveys, industry standards and specifications (such as the "Technical Specification for Detection of Urban Underground Pipelines"), physical model knowledge, and on-site excavation verification records.

[0092] It should be understood that although the above four knowledge sources are preferably listed in this embodiment, in other embodiments, knowledge sources may also cover acoustic detection features, infrared thermal imaging features, and even real-time updated geographic information system data, as long as they can provide prior constraints in the field of pipeline detection.

[0093] In terms of operating principle, the system first performs structured cleaning and semantic alignment on the acquired historical and standard data, transforms the standard clauses described in natural language (such as "the net distance between water supply pipes and gas pipes when they are laid in parallel shall not be less than 1.5 times the pipe diameter") into numerical constraint parameters that can be parsed by computers, and extracts the signal waveform features in the historical detection records as feature vector templates, thereby forming the original material library for constructing the map.

[0094] Step 311: Based on the knowledge source, construct a plastic knowledge graph containing nodes and synaptic connection edges. The nodes include pipeline nodes representing underground physical entities, feature nodes representing the characteristics of detection signals, and spatial constraint nodes representing spatial topological constraints. The synaptic connection edges connect the nodes and include feature-to-pipeline mapping edges representing the mapping relationship between features and pipeline categories, as well as spatial constraint edges representing the spatial logical relationship between pipelines.

[0095] Among them, pipeline nodes refer to semantic anchor points in the map that represent physical entities of underground pipelines. Their attributes include pipe diameter, material, burial depth range and confidence level, for example {node ID:P001, type: water supply, material: ductile iron, pipe diameter:DN300, burial depth range:[1.2m,1.8m]}.

[0096] Feature nodes are semantic units that represent typical patterns or anomalous reflections extracted from multi-source detection signals, including ground-penetrating radar feature nodes and electromagnetic induction feature nodes.

[0097] Spatial constraint nodes are virtual nodes that represent the spatial topological relationship rules that pipelines must follow. Their attributes include minimum safe distance, cross angle range, and hierarchical order.

[0098] Feature-to-pipeline mapping edges refer to weighted directed edges connecting feature nodes and pipeline nodes. Their weight values ​​characterize the confidence that a certain type of detection signal feature belongs to a certain type of pipeline target.

[0099] Spatial constraint edges refer to weighted edges that connect pipeline nodes and spatial constraint nodes. Their weights characterize the rigidity of the constraint. For example, the constraint that the spacing between parallel pipelines must not be less than 1.5 times the pipe diameter is a hard constraint (weight 1.0), while the constraint that the recommended intersection angle is 60 to 90 degrees is a soft constraint (weight 0.7).

[0100] It should be understood that the node types in the map are not limited to the three types mentioned above. In other embodiments, geological environment nodes that characterize the physical properties of the underground medium can also be introduced to further enrich the constraint dimensions.

[0101] In some implementations, the system maps the cleaned knowledge material into a topological structure of the graph: each pipeline type is instantiated as a pipeline node, each typical signal pattern is instantiated as a feature node, and each specification clause is instantiated as a spatial constraint node.

[0102] Subsequently, feature-to-pipeline mapping edges are established based on physical correspondences (for example, the initial weight of mapping the strong peak feature of electromagnetic induction signal to metal pipeline nodes is set to 0.85, and the weight of mapping to non-metallic pipeline nodes is set to 0.05). Spatial constraint edges are established according to specification clauses, thereby forming a network knowledge graph containing prior logic.

[0103] This effectively encodes implicit knowledge, which was originally scattered in normative texts and historical archives, into a graph structure topology and weight matrix that can be traversed and computed by computers. This makes prior knowledge no longer a static textual reference, but an executable logic that can directly participate in dynamic reasoning and computation.

[0104] Step 312: Based on historical detection data samples, with the goal of maximizing the prediction accuracy of the feature-pipeline mapping edge for the association between feature nodes and corresponding pipeline nodes, supervised training is performed on the initial weights of the feature-pipeline mapping edge.

[0105] Supervised training refers to the process of optimizing the initial weights of the feature-pipeline mapping edges in the graph using backpropagation algorithms on historical detection data samples with real pipeline category labels.

[0106] In some implementations, the system collects a large number of historical detection samples with confirmed pipeline attributes, and inputs the extracted feature vectors into the map to activate the corresponding feature nodes.

[0107] Initially, due to unoptimized weights, the calculated overall confidence score of activated feature nodes, after propagating to pipeline nodes through mapping edges, may not match the true label.

[0108] With the goal of maximizing prediction accuracy, the cross-entropy loss between prediction confidence and true label is calculated. Then, the weight values ​​of the feature-pipeline mapping edges are updated through the gradient descent algorithm, so that the weight of the mapping edges that correctly point to the true pipeline category gradually increases, while the weight of the mapping edges that point to the wrong category gradually decreases.

[0109] This effectively avoids the subjective bias and inaccuracy that may result from manually setting weights. By using a data-driven approach, the graph can automatically learn the signal-material mapping relationship that best conforms to the actual physical laws, thereby improving the accuracy of the initial inference.

[0110] Step 313: During the pipeline detection application, the weights of the feature-pipeline mapping edges are dynamically updated online based on the pipeline category labels and spatial location confidence scores output by graph reasoning.

[0111] Online dynamic updating refers to a brain-like plasticity mechanism in which the graph automatically adjusts the weights of synaptic connections based on the co-occurrence relationship between the current inference results and input features during real-time detection operations, without the need for offline retraining. Specifically, this mechanism simulates the Hebbian learning principle of biological neural synapses, where the connection weights between synchronously firing neurons are enhanced, while the weights of asynchronously firing neurons are weakened.

[0112] In some implementations, when graph inference outputs a high-confidence pipeline category label, the system determines that a correct causal relationship exists between the currently active feature node and the output pipeline node. At this point, the system, following the Hebbian update rule, multiplies the weight of the feature-to-pipeline mapping edge connecting the feature node and the pipeline node by an enhancement factor greater than 1 (e.g., 1.05), while multiplying the weight of the mapping edge connecting the feature node to other inactive or low-confidence pipeline nodes by a decay factor less than 1 (e.g., 0.95). As the probing operation continues, the weight distribution of the graph continuously evolves towards a direction that aligns with the actual pipeline distribution in the current region.

[0113] This endows the knowledge graph with dynamic plasticity, becoming more accurate with use, completely breaking through the limitation of traditional static triplet knowledge bases that remain fixed once built. It enables it to address the challenges of complex and ever-changing real-world operating conditions, where differences in pipeline material distribution and signal characteristics exist in different areas. Static knowledge bases cannot adapt to such local variations. Through an effective online Hebbian update mechanism, the graph can continuously self-optimize as data accumulates, always maintaining the highest level of timeliness and adaptability to the current detection environment.

[0114] The process of performing graph reasoning through a plasticity knowledge graph and outputting pipeline category labels and spatial location confidence scores specifically includes: Step 320: Plastic knowledge graph reception. The plastic knowledge graph is received, and feature nodes whose similarity to the fused feature data exceeds a preset threshold are activated.

[0115] Activation refers to the operation of marking a node as valid and assigning it an initial confidence level when the matching degree between the input data and the attribute template of a node in the graph reaches a certain standard.

[0116] The preset threshold refers to the critical similarity value for determining the validity of a match; for example, it can be set to 0.7.

[0117] In some implementations, the fused feature data output by the attenuation processing unit includes electromagnetic induction sub-features and ground-penetrating radar sub-features. The system then calculates the cosine similarity between these sub-feature vectors and the attribute templates of each feature node in the map. When the cosine similarity of a feature node (e.g., a "ground-penetrating radar hyperbolic reflection feature node") exceeds 0.7, the node is activated, and its initial confidence is set to this similarity value; if the similarity is below 0.7, the node remains dormant and does not participate in subsequent inference.

[0118] By using a similarity threshold, a large number of redundant nodes that are irrelevant to the current input are filtered out in the early stages of reasoning, which greatly reduces the search space and improves the computational efficiency and focus of graph reasoning.

[0119] Step 321: By mapping the features to the pipeline nodes, propagate the confidence of the activated feature nodes to the connected pipeline nodes, and calculate the overall confidence of each pipeline node.

[0120] Confidence propagation refers to a graph computation operation that transfers the confidence of a source node to a target node along the direction of the synaptic connection edge, according to the weight of the edge.

[0121] The overall confidence level refers to the final confidence level value obtained by aggregating the confidence levels received by a pipeline node from all connected active feature nodes.

[0122] In some implementations, the formula for calculating the overall confidence level of any pipeline node in the graph is: ,in To assess the overall confidence level, Let j be the weight of the edge mapping the j-th feature to the pipeline. Let be the initial confidence level of the j-th feature node. Let be the cosine similarity between the j-th feature node and the fused feature.

[0123] This formula allows for the weighted aggregation of the confidence scores of all activated feature nodes to the pipeline node, enabling those pipeline nodes that receive support from multiple strong mapping edges to stand out and achieve a high overall confidence score.

[0124] This allows the confidence propagation mechanism to simulate the joint verification logic of multi-source evidence. The weak features of a single sensor may only give pipeline nodes extremely low confidence, but when electromagnetic induction and ground-penetrating radar features point to the same pipeline node at the same time, their combined confidence will be enhanced nonlinearly, thereby realizing knowledge-level fusion verification of multimodal information.

[0125] Step 322: Sort the pipeline nodes based on the overall confidence level, and select at least one pipeline node whose overall confidence level exceeds the confidence threshold as a candidate pipeline node.

[0126] Among them, candidate pipeline nodes refer to the set of pipeline nodes that rank high in overall confidence and exceed the minimum confidence threshold after confidence propagation, and are the results of potential target pipeline determination.

[0127] In some implementations, the system sorts all pipeline nodes by their overall confidence scores in descending order and sets a confidence threshold (e.g., 0.5). The top K pipeline nodes (e.g., K=5) with an overall confidence score greater than 0.5 are selected as candidate pipeline nodes. If the overall confidence score of all nodes is below 0.5, the current input features are deemed insufficient to identify the pipeline, and an unknown label is output.

[0128] This achieves the effect of ensuring that candidate nodes entering the final verification stage have basic evidence support through dual screening of sorting and thresholding, thus avoiding interference from low-confidence noisy nodes on subsequent spatial constraint verification.

[0129] Step 323: Based on spatial constraint edges, perform consistency verification on the spatial relationships between candidate pipeline nodes, and output the category attributes and comprehensive confidence scores of the verified candidate pipeline nodes as pipeline category labels and spatial location confidence scores, respectively.

[0130] Consistency verification refers to the logical judgment process of checking whether the spatial relationship between candidate pipeline nodes violates industry common sense and physical laws by using the pipeline space topology specifications encoded in the graph.

[0131] In some implementations, the system extracts the spatial constraint edges connecting candidate pipeline nodes and obtains the spatial coordinate data of the current detection area. For example, if the candidate set contains both "DN300 water supply pipe" and "DN200 gas pipe" nodes, and the spatial constraint edge stipulates that "the parallel distance between the water supply pipe and the gas pipe shall not be less than 1.5 meters," the system will verify the inferred planar projection distance between the two pipes based on the current detection. If the inferred distance is 2.0 meters, satisfying the constraint, the verification passes, and the system outputs the category attributes of these two nodes (such as "DN300 ductile iron water supply pipe") as pipeline category labels, and their comprehensive confidence score (such as 0.92) as spatial location confidence score.

[0132] This allows spatial constraint verification to introduce a mandatory review of macroscopic physical logic at the end of the inference chain, eliminating a key line of defense against false anomalies. Ground-penetrating radar profiles often show stratigraphic interface reflections or construction debris reflections that resemble the hyperbolic shape of pipelines. These false anomalies often achieve high similarity activation and confidence propagation at the feature level. However, if a stratigraphic interface is misidentified as a pipeline, its inferred spatial location often leads to violations of specification constraints regarding the spacing between it and surrounding confirmed pipelines (e.g., inferring two pipelines to intersect and overlap, with a spacing of 0). In this case, consistency verification will forcibly reject the misidentification, thus providing an interpretable inference chain and significantly reducing the misidentification rate in complex environments.

[0133] Step 324: For candidate pipeline nodes that fail verification, reduce the overall confidence of the corresponding pipeline nodes according to the type of constraint violated, and re-execute the selection and verification steps.

[0134] In some implementations, if the consistency verification fails, the system deducts the overall confidence score of the pipeline node that violates the constraint based on the type and rigidity of the violated constraint (i.e., the weight of the spatial constraint edge). For example, the confidence score of a node that violates a hard constraint (weight 1.0) is directly multiplied by a penalty coefficient of 0.1, and the confidence score of a node that violates a soft constraint (weight 0.7) is multiplied by a penalty coefficient of 0.5.

[0135] After the weighting is reduced, the system re-sorts and selects pipeline nodes. Nodes that were originally ranked lower but did not violate constraints may be promoted to new candidate nodes, and then consistency verification is performed again until a set of pipeline nodes that fully conforms to the spatial logic is selected.

[0136] Its weight reduction and reselection mechanism can form a self-correcting closed loop for graph reasoning, so that the reasoning process is no longer a single forward judgment, but has the ability to iteratively optimize based on global constraints. This ensures that the final output results are logically consistent in both local feature matching and global spatial topology, and completely eliminates the misjudgment phenomenon of "high confidence but logically absurd" caused by the lack of posterior error correction ability in traditional pure data-driven classifiers.

[0137] Based on the above technical solution, by constructing a dynamic graph containing pipeline nodes, feature nodes, and spatial constraint nodes, and introducing a Hebbian online learning and spatial constraint consistency verification iteration mechanism, the static prior knowledge base is upgraded into a malleable reasoning engine with self-evolution and logical error correction capabilities. This not only effectively eliminates misjudgments caused by false anomaly reflections from ground-penetrating radar through spatial constraint verification, but also provides an interpretable reasoning chain, which can be dynamically updated online, allowing the system performance to continuously improve with data accumulation and maintaining the timeliness and regional adaptability of the detection knowledge.

[0138] Example 4: like Figure 4 As shown, based on step 100 of Embodiment 1, this embodiment further elaborates on the specific mapping algorithm for encoding multi-source underground pipeline detection data into pulse sequences, providing customized encoding strategies for sensors with different physical principles.

[0139] Step 401: Obtain multi-source underground pipeline detection data, which includes electromagnetic induction detection signals, ground-penetrating radar reflection waveform signals, inertial measurement data, and spatial coordinate positioning data.

[0140] Among them, the electromagnetic induction detection signal refers to the continuous analog voltage signal output by the receiving coil of the electromagnetic induction instrument. The continuous analog voltage signal can reflect the secondary field strength of underground metal pipelines.

[0141] The reflected waveform signal of ground-penetrating radar refers to the high-frequency electromagnetic pulse reflection sequence received by the ground-penetrating radar antenna. The high-frequency electromagnetic pulse reflection sequence is usually presented as an A-Scan waveform, reflecting the difference in dielectric constant of the underground medium.

[0142] Inertial measurement data refers to the six-axis discrete time series signal output by the inertial measurement unit. The six-axis discrete time series signal can reflect the three-dimensional attitude and acceleration of the measurement vehicle.

[0143] Spatial coordinate positioning data refers to the absolute coordinate value stream output by a real-time dynamic positioning receiver. The absolute coordinate value stream includes longitude, latitude, and geodetic height information.

[0144] It should be understood that although the above four data types are preferably listed in this embodiment, in other embodiments, multi-source underground pipeline detection data can also cover acoustic detection signals, infrared thermal imaging data, and even semantic feature data from underground pipeline attribute databases, as long as they can provide physical or logical clues for pipeline detection.

[0145] In some implementations, the four types of heterogeneous data streams mentioned above are received in real time through the hardware interface of the front-end sensing and encoding modules. Since the sampling frequencies of various sensors differ significantly (e.g., electromagnetic induction device 1kHz, ground-penetrating radar and IMU 200Hz, RTK 20Hz), it is necessary to allocate an independent buffer queue for each data stream and stamp a unified hardware clock on the timeline to provide raw material for subsequent pulse coding and time alignment.

[0146] By explicitly listing and accepting four of the most representative heterogeneous physical signals, it breaks the limitation of traditional fusion algorithms that can only handle homogeneous numerical matrices, laying the broadest data foundation for subsequent cross-physical-dimensional information-level fusion in the pulse domain.

[0147] Step 402: In response to the voltage amplitude characteristics of the electromagnetic induction detection signal, activate the frequency and pulse conversion coding unit, map the voltage amplitude to the firing frequency of the corresponding spiking neuron per unit time, and generate the first pulse sequence.

[0148] Among them, the frequency and pulse conversion coding unit refers to the algorithm module that adopts the Rate Coding mechanism. Its core logic is to linearly or non-linearly map the continuous amplitude of the input signal to the number of firings of spiking neurons within a fixed time window.

[0149] The first pulse sequence refers to a stream of discrete pulse events output by an electromagnetic induction signal encoder, characterized by the intensity of the secondary field signal based on the pulse emission density.

[0150] In terms of operating principle, the essential characteristic of electromagnetic induction detection signals is that the voltage amplitude reflects the distance and size of the metal pipeline. The frequency and pulse conversion encoding unit maps the voltage range of 0 to 100mV to a pulse firing frequency of 0 to 500Hz.

[0151] Specifically, within each 100ms encoding time window, the encoding unit calculates the target firing frequency based on the voltage amplitude of the current sampling point and fires pulses uniformly or non-uniformly with a fixed pulse width of 2ms. For example, a voltage amplitude of 50mV will be mapped to a firing frequency of 250Hz, that is, 25 pulses will be fired within 100ms.

[0152] This allows frequency coding to perfectly preserve the core physical characteristics of electromagnetic induction signals, namely signal strength. It addresses the technical problem that traditional numerical fusion often directly quantizes voltage amplitudes as floating-point numbers, making it susceptible to noise interference and lacking temporal dynamics. Furthermore, by converting amplitudes into pulse density through frequency coding, it not only provides a natural statistical averaging smoothing effect on transient amplitude spikes, but the generated first pulse sequence can also directly drive the membrane potential integration of subsequent spiking neural networks, achieving a native connection between signal characteristics and neural computing paradigms.

[0153] Step 403: In response to the waveform amplitude distribution and temporal structure characteristics of the ground-penetrating radar reflected waveform signal, activate the delay and pulse conversion coding unit, allocate the corresponding pulse firing delay based on the amplitude of the sampling points in the waveform, and generate the second pulse sequence.

[0154] Among them, the delay and pulse conversion coding unit refers to the algorithm module that adopts the temporal coding mechanism. Its core logic is to determine whether the pulse is emitted earlier or later based on the relative magnitude of the amplitude of each sampling point in the input waveform.

[0155] The second pulse sequence refers to the discrete pulse event stream output by the ground-penetrating radar encoder, which characterizes the timing structure and amplitude distribution of the radar waveform by the precise time delay of pulse emission.

[0156] In terms of operating principle, the essential characteristics of ground-penetrating radar signals are the temporal structure and amplitude fluctuations of hyperbolic reflections in the waveform, with temporal information being more critical than simple frequency information.

[0157] The delay and pulse conversion coding unit sorts the 512 sampling points in each A-Scan waveform according to the amplitude. The larger the amplitude of the peak, the shorter the pulse firing delay.

[0158] The time delay range is set to 0 to 50 ms. The sampling point with the highest amplitude emits a pulse at 0 ms, and the sampling point with the lowest amplitude emits a pulse at 50 ms.

[0159] This time delay allocation allows the spatial morphology of the original waveform to be accurately projected into a density distribution in the pulse time domain. This enables time delay coding to concisely preserve the spatial topological features of the ground-penetrating radar waveform. If frequency coding were also used for the radar signal, the crucial temporal sequence within the waveform would be lost, making it impossible for subsequent networks to recognize the hyperbolic reflection features. Time delay coding, however, transforms the amplitude of the waveform into temporal sequence, allowing the spatial structure of the waveform to be losslessly replicated in the pulse domain. This provides irreplaceable temporal clues for subsequent map inference and identification of hyperbolic features of non-metallic pipelines.

[0160] Step 404: In response to the motion state change rate characteristics of the inertial measurement data, activate the event-driven conversion coding unit, trigger a pulse event only when the change in acceleration or angular velocity exceeds the dynamic threshold, and generate a third pulse sequence.

[0161] Among them, the event-driven conversion coding unit refers to the algorithm module that draws on the principle of neuromorphic visual event camera. Its core logic is to output pulses only when the input signal changes significantly, and not respond when it changes statically or slowly.

[0162] The third pulse sequence refers to an extremely sparse asynchronous pulse event stream output by the IMU encoder that carries only attitude change information, with each pulse accompanied by a channel identifier and polarity label.

[0163] In some implementations, inertial measurement data outputs a large amount of redundant static attitude data when the measuring wheel is moving smoothly. However, when encountering a manhole cover protrusion or a pothole, drastic changes in acceleration and angular velocity occur. At this point, the event-driven conversion encoding unit presets dynamic thresholds (e.g., pitch and roll angle changes of 0.5 degrees, yaw angle changes of 0.2 degrees). A pulse event is triggered only when the change in adjacent sampling points exceeds this threshold. The pulse records not only a timestamp but also the channel identifier (accelerometer channels 0-2, gyroscope channels 3-5) and polarity (positive or negative change). If the change does not exceed the threshold, no pulse is emitted.

[0164] Event-driven coding can be used to compress IMU data into an extremely sparse pulse stream containing only key mutation events. Computational resources are activated only during attitude mutations, which greatly reduces the computational load and estimation power consumption of the subsequent attenuation processing unit, making it suitable for the low-power real-time computing needs of portable field measurement equipment.

[0165] Step 405: In response to the digital quantization characteristics of the spatial coordinate positioning data, activate the phase encoding unit to map the binary bits of the coordinate values ​​into pulses with specific phases, and generate the fourth pulse sequence.

[0166] The phase coding unit refers to the algorithm module that uses the phase coding mechanism. Its core logic is to map the binary bits of the input digital quantization value to a specific phase offset in the pulse oscillation period.

[0167] The fourth pulse sequence refers to a stream of discrete pulse events output by an RTK encoder, in which the pulse phase precisely represents the absolute spatial coordinate values.

[0168] In some implementations, spatial coordinate positioning data (longitude, latitude, geodetic height) are high-precision floating-point numbers, whose numerical characteristics need to be encoded losslessly and accurately. The phase encoding unit can first convert each coordinate floating-point number into a 32-bit binary representation and assign a reference oscillation period to each bit. If the bit is 1, a pulse is emitted at the 0-degree phase of the oscillation period; if the bit is 0, a pulse is emitted at the 180-degree phase or no pulse is emitted.

[0169] This mapping allows a coordinate value to be expanded into a sequence of pulses with precise phase characteristics, enabling the receiver to reconstruct the original value without loss simply by detecting the phase of the pulses.

[0170] This allows phase encoding to utilize the discrete polymorphism of pulse phase to achieve precise binary-level mapping of floating-point numbers, ensuring that the most critical reference data for pipeline measurement—spatial coordinates—does not lose any effective accuracy during pulse domain conversion, thus providing reliable spatial anchors for subsequent 3D reconstruction.

[0171] Step 406: Synchronize and align the first pulse sequence, the second pulse sequence, the third pulse sequence, and the fourth pulse sequence in time, and output a pulse sequence under a unified time reference.

[0172] Time synchronization alignment refers to the process of interpolating, aligning, and merging four asynchronous pulse sequences with different sampling rates on a unified time axis.

[0173] A pulse sequence under a unified time reference refers to a mixed pulse stream in which four pulse events share the same time coordinate system after alignment, and can be synchronously received and processed in parallel by subsequent attenuation processing units.

[0174] In some implementations, since the input sampling rates and coding time windows of the four types of coding units are different, the generated four pulse sequences are asynchronous on the time axis. In this case, the timestamp of the highest sampling rate (e.g., 1kHz) can be used as a reference to perform time interpolation resampling on the other three pulse sequences.

[0175] For the first pulse sequence of frequency encoding, maintain its original time window density; For the time-delay coded second pulse sequence, the time delay segment corresponding to its A-Scan waveform is mapped to the corresponding interval of the reference time axis; For event-driven third pulse sequences, their asynchronous event stamps are directly aligned to the nearest reference time scale; For the phase-encoded fourth pulse sequence, its binary phase segments are mapped to a continuous interval of the reference time axis.

[0176] Ultimately, the four pulses are interwoven and merged on a unified time axis to form a single, unified pulse sequence output.

[0177] Based on the above technical solutions, by tailoring four encoding strategies—frequency, time delay, event-driven, and phase—for four types of heterogeneous data, namely electromagnetic induction, ground-penetrating radar, inertial measurement, and spatial positioning, we not only preserve the core physical characteristics (intensity, timing, abrupt changes, and accuracy) of the original signals from each sensor to the greatest extent possible, avoiding feature loss and distortion caused by traditional numerical quantization, but also achieve sparse computation with extremely low power consumption through event-driven encoding. Through time synchronization alignment, we break down the fusion barriers of multi-source heterogeneous data, providing the highest quality original pulse data source for subsequent brain-like fusion and knowledge graph reasoning.

[0178] Example 5: like Figure 5-6 As shown, after outputting pipeline category labels and spatial location confidence in step 300 of Embodiment 1, this embodiment further expands the cross-modal feature point matching and 3D reconstruction process guided by knowledge graph, establishing a complete data link from feature extraction to 3D spatial model for pipeline route measurement.

[0179] Step 501: Extract electromagnetic induction direction feature points, ground-penetrating radar reflection feature points, and spatial trajectory feature points from the multi-source underground pipeline detection data.

[0180] Among them, electromagnetic induction direction feature points refer to discrete spatial anchor points extracted from electromagnetic induction detection signals that can characterize abrupt changes in the horizontal direction of the pipeline or extreme values ​​of the signal, such as signal peak points and direction change points (location points where the rate of change of direction angle exceeds 15 degrees / meter).

[0181] Ground-penetrating radar reflection feature points refer to the vertices and amplitude maxima points that can be identified from ground-penetrating radar profile images to characterize the hyperbolic reflection morphology of underground targets. These points usually correspond to the top burial depth of pipelines.

[0182] Spatial trajectory feature points refer to the mileage markers and curvature extrema extracted from the fusion trajectory of RTK and IMU, representing the key nodes of the measurement vehicle's motion in three-dimensional space.

[0183] In terms of operational principle, feature extraction algorithms are performed on the three channels of raw detection data respectively: For electromagnetic induction data, peak detection and differential operators are used to identify the locations of abrupt changes in direction; For ground-penetrating radar data, a hyperbola fitting algorithm is used to extract the depth and horizontal offset of the reflection vertex; For RTK trajectory data, curvature calculation and sliding window detection are used to extract trajectory inflection points.

[0184] All extracted feature points are assigned a unified timestamp and preliminary spatial coordinates, forming three types of discrete feature point sets.

[0185] Step 502: Based on spatiotemporal proximity, combine the electromagnetic induction direction feature points, ground-penetrating radar reflection feature points, and spatial trajectory feature points in pairs to generate a cross-modal candidate matching pair set.

[0186] Spatiotemporal proximity refers to the degree of closeness between two feature points of different modalities in terms of timestamps and spatial coordinates. That is, they are recorded at the same or similar times and the distance between them in the horizontal projection or depth direction is less than a preset proximity threshold.

[0187] A cross-modal candidate matching pair set refers to the initial matching relationship set formed by pairing feature points from different sensors according to the principle of spatiotemporal proximity. For example, an electromagnetic induction direction feature point and a ground-penetrating radar reflection feature point are combined into a candidate pair.

[0188] In terms of operation, the system sets temporal proximity thresholds (e.g., 100 milliseconds) and spatial proximity thresholds (e.g., horizontal distance 0.5 meters, depth distance 0.3 meters). Then, it iterates through three sets of feature points, calculating the time difference and spatial distance between any two feature points from different modalities. If both are less than the corresponding proximity threshold, the two points are combined into a candidate matching pair. For example, when the measuring wheel passes a certain point, electromagnetic induction records a point of sudden change in direction, and ground-penetrating radar records a hyperbola vertex at the same moment; if their horizontal coordinates differ by only 0.1 meters, they are combined into a candidate pair. All candidate pairs that meet the conditions are aggregated to form a cross-modal candidate matching pair set.

[0189] Step 503: In response to each candidate matching pair in the cross-modal candidate matching pair set, extract at least one spatial constraint edge associated with the pipeline node corresponding to the pipeline category label within the plasticity knowledge graph.

[0190] In terms of operational principle, once the system confirms the pipeline category label of the current detection area, the physical properties and regulatory requirements of the target pipeline are clearly defined. The system can then retrieve the pipeline node in the plasticity knowledge graph and traverse along its spatial constraint edges to extract all relevant spatial topology rules. For example, if the label is a water supply pipe, the system extracts the minimum clearance constraint edges and intersection angle constraint edges between it and drainage pipes and gas pipes. The rigid or soft weights and parameters carried by these constraint edges will serve as the absolute standard for subsequent geometric consistency verification.

[0191] This allows prior industry standards and historical topological logic to be explicitly introduced into the matching and verification process, so that cross-modal alignment no longer relies on the vague experience in the operator's mind, but on numerical constraints that can be precisely executed by the computer, fundamentally eliminating the subjective arbitrariness and errors brought about by manual annotation alignment.

[0192] Step 504: Based on spatial constraint edges, verify the consistency of the geometric relationship between the feature points associated with the candidate matching pairs in three-dimensional space.

[0193] Consistency verification refers to the process of checking whether the relative positional relationship between two feature points in a candidate matching pair in three-dimensional space conforms to the physical logic and specification requirements of pipeline laying by using the geometric parameters (such as spacing, angle, and hierarchical order) specified by the extracted spatial constraint edges.

[0194] In terms of operational principle, the two feature points in a candidate matching pair are restored to a three-dimensional spatial coordinate system, and the horizontal distance, vertical height difference, and connecting angle between them are calculated. The calculation results can then be compared with the extracted spatial constraint edge parameters. For example, if a candidate pair associates an electromagnetic induction feature point with a ground-penetrating radar feature point as the intersection of a water supply pipe and a gas pipe, but the calculation shows that the vertical height difference between the two is only 0.1 meters (almost overlapping), while the spatial constraint edge stipulates that the vertical clearance between the water supply pipe and the gas pipe must be ≥0.5 meters when they intersect, then the geometric relationship of this candidate pair violates the hard constraint, and the consistency verification fails. Conversely, if the calculated distance and angle both fall within the tolerance range allowed by the constraint edge, then the verification passes.

[0195] This spatial constraint verification is the core mechanism for eliminating false anomaly mismatches. In actual exploration, ground-penetrating radar often captures reflections from stratigraphic interfaces or construction debris. These false features may be very close to real electromagnetic induction feature points in terms of spatiotemporal proximity, easily forming false candidate pairs. However, the spatial relationships of pipelines implied by these false features often violate basic laying specifications (such as overlapping pipes or zero spacing). Consistency verification can accurately eliminate these false matching pairs by utilizing the rigid barriers of physical logic, achieving knowledge-guided automatic error correction.

[0196] Step 505: Confirm the verified candidate matching pairs as valid cross-modal matching point pairs.

[0197] Among them, a valid cross-modal matching point pair refers to a feature point association combination that, after spatial constraint edge consistency verification, has a geometric relationship that conforms to pipeline laying logic and industry standards, and is ultimately recognized by the system as a true reflection of the physical entity of the same underground pipeline or the spatial relationship of related pipelines.

[0198] In terms of operational principle, after performing the verification step 504 on each candidate matching pair in the set, the candidate pairs that pass the verification are retained, given a "valid" status label, and their associated pipeline category label and overall confidence score are inherited to the matching point pair. Candidate pairs that fail the verification are discarded or temporarily stored with reduced weight. Ultimately, all valid cross-modal matching point pairs constitute a clean and reliable multi-source feature association network, where each matching point pair represents a consensus confirmation of the same spatial physical structure from different sensor perspectives.

[0199] Step 506: Based on all valid cross-modal matching point pairs, reconstruct the three-dimensional spatial location model of the underground pipeline.

[0200] In terms of operational principle, all valid cross-modal matching point pairs are arranged in time-stamp order, and interpolation fitting is performed using the three-dimensional spatial coordinates they carry. Specifically, firstly, on the horizontal plane, electromagnetic induction direction feature points and spatial trajectory feature points are fitted to generate the horizontal projection trend line of the pipeline. Then, in the vertical direction, depth information of ground-penetrating radar reflection feature points and RTK elevation data are combined to fit and generate a vertical profile trend line. Finally, the curves of the horizontal and vertical dimensions are merged to generate a continuous and smooth three-dimensional spatial curve, that is, the three-dimensional spatial location model of the underground pipeline.

[0201] By using effective cross-modal matching point pairs as 3D skeleton nodes, the electromagnetic induction horizontal positioning, the ground-penetrating radar depth detection, and the RTK absolute coordinates are automatically and seamlessly stitched into an integrated 3D model. This not only improves the integrity and accuracy of the reconstruction, but also allows the output model to be directly imported into GIS and BIM platforms, meeting the stringent requirements of as-built surveying for the standardization and timeliness of data results.

[0202] Based on the above technical solution, by introducing spatial constraint edges of a plastic knowledge graph as a verification criterion for cross-modal feature point matching, the time-consuming and subjectively error-prone traditional method of relying on manual annotation and alignment is completely replaced, achieving automatic and accurate association of multi-source heterogeneous data in three-dimensional space. Simultaneously, based on the verified effective matching point pairs, a three-dimensional spatial location model is reconstructed, ensuring the integrity and logical consistency of the pipeline's horizontal orientation, vertical burial depth, and three-dimensional curves. This upgrades the as-built survey results from fragmented two-dimensional drawings into highly reliable and traceable three-dimensional digital assets.

[0203] Example 6: like Figure 7 As shown, this embodiment provides an underground pipeline routing measurement system based on spatial coordinate correction. The system constructs a closed-loop architecture of four core units corresponding to the method process, which is used to support the reliable deployment and real-time operation of the aforementioned method process at the hardware and software levels.

[0204] The underground pipeline routing measurement system specifically includes a multi-source data encoding unit, an attenuation processing unit, a plasticity knowledge graph reasoning unit, and a feedback adjustment unit.

[0205] The multi-source data encoding unit is used to acquire multi-source underground pipeline detection data and encode it into pulse sequences. This unit not only includes hardware interface circuits for interfacing with external sensors such as electromagnetic induction instruments, ground-penetrating radar, inertial measurement units, and real-time dynamic positioning receivers, but also incorporates encoding algorithm logic such as frequency conversion, time delay conversion, event-driven conversion, and phase conversion.

[0206] During operation, the multi-source data encoding unit collects heterogeneous detection data in real time. Through its respective encoding subunit, it maps the voltage amplitude, waveform timing, attitude change and coordinate values ​​into the first to fourth pulse sequences, and outputs a unified pulse sequence after completing the timestamp synchronization and alignment internally.

[0207] It should be understood that although this embodiment preferably deploys the multi-source data encoding unit as a front-end independent module close to the sensor to reduce data transmission latency, in other embodiments, the unit can also be used as a software virtual module within the core computing platform, as long as it can perform the unified encoding and alignment function of heterogeneous data to the pulse domain.

[0208] An attenuation processing unit is used to process the pulse sequence to obtain fused feature data. The attenuation processing unit includes at least a first-level attenuation processing unit and a second-level attenuation processing unit. The first-level attenuation processing unit is configured to suppress transient pulse clusters exceeding a frequency threshold, and the second-level attenuation processing unit is configured to smooth residual fluctuations.

[0209] This unit implements a two-stage cascaded processing logic for basal dendrites and apical dendrites at the hardware or software level: The first-stage attenuation processing unit receives the pulse sequence output by the multi-source data encoding unit and uses its equipped fast attenuation time constant and cyclic pulse connection network to perform rapid membrane potential integration and reset offset on the transient high-density pulse clusters caused by terrain abrupt changes, suppressing their downstream propagation; The second-stage attenuation processing unit is coupled to the first-stage output and uses a longer attenuation time constant to perform long-time window integration on the intermediate sequence to smooth the residual low-frequency small fluctuations, and finally outputs low-noise, high signal-to-noise ratio fused feature data through frequency statistical decoding.

[0210] In the system architecture, the attenuation processing unit completely replaces the traditional attitude compensation module based on linear Kalman filtering, fundamentally eliminating the hardware computational root cause of the trailing effect.

[0211] The plasticity knowledge graph reasoning unit performs graph reasoning based on fused feature data and plasticity knowledge graph, outputting pipeline category labels and spatial location confidence scores. The plasticity knowledge graph contains synaptic connection edges, and the graph reasoning includes weight propagation of spatial location confidence scores based on synaptic connection edges.

[0212] Internally, this unit maintains a dynamic graph data structure containing pipeline nodes, feature nodes, and spatial constraint nodes, and is equipped with an inference execution engine for node activation, confidence propagation, and spatial constraint consistency verification. Upon receiving the fused feature data output from the attenuation processing unit, this unit first activates feature nodes with similarity exceeding a threshold. Then, it propagates the confidence to pipeline nodes through feature-pipeline mapping edges, calculates the comprehensive confidence, and sorts and selects candidate nodes. Finally, it uses spatial constraint edges to perform consistency verification and iterative error correction on the physical logic between candidate nodes, ultimately outputting pipeline category labels and spatial location confidence scores that have passed logical self-consistency verification.

[0213] It should be understood that the plastic knowledge graph reasoning unit not only performs static graph traversal calculations, but the weights of its internal synaptic connection edges can also be dynamically updated online based on the reasoning results, giving the unit the plasticity characteristic of continuous evolution with data accumulation.

[0214] The feedback adjustment unit generates adjustment signals for the corresponding sensing channels based on spatial location confidence levels and adjusts the weight gain of the sensing channels corresponding to the multi-source underground pipeline detection data. Specifically, when the plasticity knowledge graph reasoning unit outputs a high-confidence pipeline category label, the feedback adjustment unit generates an adjustment signal based on the physical attribute logic implied by that label. For example, if the target is determined to be a non-metallic pipeline, the feedback adjustment unit generates control commands to reduce the weight gain of the electromagnetic induction sensing channel and increase the weight gain of the ground-penetrating radar sensing channel. This adjustment signal is fed back along the feedback link to the multi-source data encoding unit and its subsequent synaptic weight connection layer, dynamically modifying the pulse firing sensitivity or synaptic connection weight matrix of the corresponding channel, thereby achieving adaptive reallocation of sensing resources and computational attention at the system's underlying layer.

[0215] Regarding the connection relationships and data flow of the four core units mentioned above, the output of the multi-source data encoding unit is connected to the input of the attenuation processing unit via a high-speed pulse data bus, pushing the pulse sequence under a unified time base to the attenuation processing unit in real time. The output of the attenuation processing unit is connected to the input of the plastic knowledge graph inference unit via a feature vector interface, transmitting the filtered and smoothed fused feature data to the inference engine. Furthermore, the output of the plastic knowledge graph inference unit is divided into two paths: one outputs the pipeline category label and spatial location confidence to the system's backend display and reconstruction module, and the other transmits the confidence and label semantics to the input of the feedback adjustment unit. Finally, the output of the feedback adjustment unit is connected in reverse to the weight configuration interfaces of the multi-source data encoding unit and the attenuation processing unit via a parameter configuration bus, forming a complete closed-loop control link.

[0216] This system-level closed-loop architecture breaks away from the loosely coupled paradigm of traditional pipeline detection systems, where each sensor module operates independently and data is only stitched offline at the end. It enables the four main units to construct a brain-like closed loop of perception, cognition, and action through the bidirectional interweaving of pulse data flow and regulation control flow. This allows the system to proactively adjust its underlying perception strategies based on high-level inference conclusions during operation, achieving adaptive and smooth switching of perception channels and real-time suppression of nonlinear disturbances. This provides robust hardware and software deployment support for the continuous, blind-zone-free operation of the methodology.

[0217] Example 7: To more clearly illustrate the operational effects and commercial value of the technical solutions provided in the various embodiments of the present invention in actual engineering projects, the following uses the underground pipeline completion measurement in a renovation project of an old urban community as an example to provide a detailed description of the application scenario.

[0218] It should be understood that this embodiment is illustrative only and not restrictive, and is intended to demonstrate how the system of the present invention addresses typical technical challenges in complex surface environments, and does not mean that the present invention is only applicable to this specific scenario.

[0219] In this renovation scenario of an old residential community, the underground pipeline distribution is extremely complex, including overlapping sections of metal water supply pipes and non-metallic PE gas pipes. Furthermore, the ground surface exhibits severe undulations due to years of disrepair, such as protruding manhole covers and sunken potholes. When the measuring wheel travels along the pipeline route, the system faces three core technical challenges: electromagnetic signal interruptions in non-metallic pipe sections, jagged elevation fluctuations caused by sudden ground changes, and multiple false anomaly reflections in the ground-penetrating radar profile. This embodiment will demonstrate how the system, through the core mechanisms of embodiments 1 to 5 mentioned above, addresses these challenges one by one and outputs an accurate three-dimensional route model.

[0220] Step 901: When the measuring wheel travels to the non-metallic PE gas pipe section, the electromagnetic induction signal chain breaks and the sensing channel adaptively switches: When the measuring wheel travels from the metallic water supply pipe section to the non-metallic PE gas pipe section, the intensity of the secondary field signal detected by the electromagnetic induction receiver coil drops sharply to near zero. Traditional detection equipment would then display a signal chain break, forcing the operator to interrupt the measurement and blindly guess the direction. However, this invention can acquire the electromagnetic induction detection signal of the broken chain in real time through the multi-source data encoding unit. Due to the extremely low voltage amplitude, the frequency and the frequency of the first pulse sequence output by the pulse conversion encoding unit are almost zero. At the same time, the ground-penetrating radar reflection waveform signal still stably receives the hyperbolic reflection of the PE pipe, and the time delay and pulse conversion encoding unit normally output the second pulse sequence. Inertial measurement data and spatial coordinate positioning data are also output as the third and fourth pulse sequences through the event-driven conversion encoding unit and the phase encoding unit, respectively. After time synchronization and alignment, the four pulse sequences are input to the attenuation processing unit, and after transient suppression by the basal dendrite module and long-time-window smoothing by the top dendrite module, stable fused feature data is output.

[0221] The fused feature data was then input into the plasticity knowledge graph reasoning unit. Since the activation similarity of the electromagnetic induction feature nodes was extremely low and failed to exceed the preset threshold, only the ground-penetrating radar feature nodes were activated with high similarity. The activated ground-penetrating radar feature nodes propagated their confidence to the PE gas pipeline nodes through feature-to-pipeline mapping edges. The calculated overall confidence of the PE gas pipeline nodes exceeded the confidence threshold and passed the consistency verification of the spatial constraint edges. The system ultimately output the pipeline category label as non-metallic PE gas pipe, with a spatial location confidence of 0.89.

[0222] Based on the high-confidence non-metallic pipeline tag, the feedback adjustment unit generates a corresponding adjustment signal and adjusts the weight gain of the sensing channels corresponding to the multi-source underground pipeline detection data. Specifically, the system significantly reduces the weight gain of the electromagnetic induction sensing channel while significantly increasing the weight gain of the ground-penetrating radar sensing channel. This allows subsequent input detection data to automatically assign a higher weight to the ground-penetrating radar pulse sequence during pulse fusion and map inference. This feedback closed-loop mechanism enables the system to automatically and smoothly switch to the ground-penetrating radar-dominated detection mode without manual intervention in electromagnetic signal disconnection sections, achieving adaptive and seamless connection of sensing channels and reducing the horizontal deviation of non-metallic pipeline sections.

[0223] Step 902: The measuring wheel travels to the section where the manhole cover protrudes and the pothole is recessed. Transient pulse cluster suppression and elevation smoothing output: When the measuring wheel passes over the manhole cover protrusion, the inertial measurement unit collects drastic changes in acceleration and angular velocity. The event-driven conversion encoding unit detects that the change exceeds the dynamic threshold, triggering a high-frequency pulse event and generating a third pulse sequence containing transient pulse clusters. If traditional Kalman filtering is used to process this type of non-Gaussian pulse noise, the filtering result will continuously deviate from the true elevation value within several sampling points after the outlier appears, producing obvious tailing effects and sawtooth oscillations.

[0224] In the system of this invention, a hybrid pulse sequence input attenuation processing unit is included, comprising transient pulse clusters. The basal dendritic module of the first-stage attenuation processing unit, with its extremely small first attenuation time constant and cyclic pulse connection network, rapidly integrates and resets the membrane potential of the transient high-density pulse clusters. When the high-frequency pulse clusters arrive densely, the membrane potential rapidly rises to the firing threshold and is immediately reset. Due to the rapid attenuation characteristic, subsequent transient pulses cannot utilize the residual potential of the preceding pulses to form a continuous superposition; the transient perturbation is directly counteracted and suppressed at the neuron level, and the output is an intermediate sequence containing only normal signals. Subsequently, the apical dendritic module of the second-stage attenuation processing unit, with its larger second attenuation time constant, performs long-window integration on the intermediate sequence, smoothing residual minor topographical fluctuations, outputting a frequency-stable smooth sequence, and decoding it into fused feature data.

[0225] This dual-stage cascaded pipeline, which combines rapid suppression of basal dendrites with fine smoothing of apical dendrites over a long time window, completely eliminates the tailing effect of traditional Kalman filtering, enabling the system to output smooth and accurate elevation curves even on drastically undulating road surfaces.

[0226] Step 903: Ground penetrating radar encounters multiple false anomaly reflections. Spatial constraint verification eliminates misjudgments and ensures accurate 3D reconstruction: In the underground of old residential areas, due to the presence of stratigraphic interfaces, construction waste, and underground cavities, ground penetrating radar profile images often show multiple high-amplitude false anomaly reflections that resemble the hyperbolic shape of pipelines. Traditional methods relying on manual visual interpretation or purely data-driven classification are prone to misjudging stratigraphic interface reflections as pipeline reflections, leading to incorrect burial depth labeling.

[0227] In this invention's system, when fused feature data is input into the plasticity knowledge graph reasoning unit, not only are genuine PE gas pipeline feature nodes activated, but some pseudo-anomaly feature nodes with similar morphology to geological interfaces may also be activated due to their cosine similarity exceeding a preset threshold. These nodes then propagate to incorrect pipeline nodes through mapping edges, forming a candidate pipeline node set. However, in the consistency verification stage based on spatial constraint edges, the system extracts the spatial constraint edges associated with the candidate pipeline nodes for logical review. For example, if a pseudo-anomaly reflection is candidate-identified as a water supply pipe node, but its inferred burial depth is only 0.1 meters away from the confirmed drainage pipe node, violating the hard spatial constraint edge encoded in the knowledge graph that the vertical clearance between water supply and drainage pipe intersections must be greater than 0.5 meters, the consistency verification forcefully rejects this misjudgment. The system reduces the overall confidence of the corresponding pipeline node according to the type of constraint violation and re-executes the selection and verification steps. Ultimately, only genuine PE gas pipeline nodes that pass the spatial topology logical review are retained, outputting accurate pipeline category labels and spatial location confidence.

[0228] After outputting labels and confidence scores, the system further executes a knowledge graph-guided cross-modal feature point matching and 3D reconstruction process. Electromagnetic induction direction feature points, ground-penetrating radar reflection feature points, and spatial trajectory feature points are extracted, and a set of cross-modal candidate matching pairs is generated based on spatiotemporal proximity. Subsequently, spatial constraint edges associated with PE gas pipeline nodes within the plasticity knowledge graph are extracted, and the geometric relationships of the feature points associated with the candidate matching pairs in 3D space are verified for consistency. False matching pairs composed of pseudo-anomaly feature points are eliminated, and valid cross-modal matching point pairs are confirmed. Finally, based on all valid cross-modal matching point pairs, combined with the actual reflection depth of the ground-penetrating radar and the elevation coordinates of the RTK, a continuous and smooth 3D spatial location model of the underground pipeline is reconstructed.

[0229] Through the complete operation of steps 901 to 903 above, a feedback loop of pulse coding and graph inference enables adaptive and seamless switching of the sensing channel when the signal chain of a non-metallic pipeline segment is broken, eliminating the tens of centimeter-level deviation in orientation caused by manual guessing. Simultaneously, the dual-stage attenuation mechanism of basal and apical dendrites completely suppresses transient pulse noise and tailing effects caused by ground abrupt changes, improving the RMSE of elevation measurements. Furthermore, through spatial constraint consistency verification of the plasticity knowledge graph, false anomalies such as stratigraphic interfaces are automatically eliminated, providing an interpretable inference chain. Finally, based on the verified cross-modal matching point pairs, a high-precision three-dimensional pipeline orientation model is automatically reconstructed, reducing the traditional manual alignment and interpretation process, which takes several hours, to real-time calculation output in tens of seconds, improving the accuracy, robustness, and automation level of underground pipeline as-built measurements in complex surface environments.

[0230] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention, such as adaptive adjustments to the pulse coding strategy, resetting the decay time constant, expanding the types of knowledge graph nodes and edges, or equivalent replacements for heterogeneous computing platform deployment architectures, should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for measuring the routing of underground pipelines based on spatial coordinate correction, characterized in that, include: Acquire multi-source underground pipeline detection data and encode them into pulse sequences; The pulse sequence is processed by an attenuation processing unit to obtain fused feature data. The attenuation processing unit includes at least a first-level attenuation processing unit and a second-level attenuation processing unit. The first-level attenuation processing unit is configured to suppress transient pulse clusters that exceed a frequency threshold, and the second-level attenuation processing unit is configured to smooth residual fluctuations. Based on the fused feature data, graph reasoning is performed through a plasticity knowledge graph to output pipeline category labels and spatial location confidence scores, wherein the plasticity knowledge graph contains synaptic connection edges, and the graph reasoning includes propagating the spatial location confidence scores based on the weights of the synaptic connection edges; Based on the spatial location confidence level, an adjustment signal is generated for the corresponding sensing channel, and the weight gain of the sensing channel corresponding to the multi-source underground pipeline detection data is adjusted accordingly.

2. The method for measuring the routing of underground pipelines based on spatial coordinate correction according to claim 1, characterized in that, The construction process of the attenuation processing unit specifically includes: A recurrent pulse connection network is configured between spiking neurons containing a first decay time constant to obtain a first-level decay processing unit. The first decay time constant is used to control the membrane potential integration window length of the basal dendritic module for the input pulse sequence. The recurrent pulse connection network is used to realize short-term pulse pattern memory. At the output of the first-stage attenuation processing unit, a spiking neuron containing a second attenuation time constant is coupled to form a second-stage attenuation processing unit. The second attenuation time constant is greater than the first attenuation time constant and is used to control the length of the integral smoothing window of the top dendrite module for the input pulse sequence. Pulse sequence samples with different degrees of surface undulation are obtained. With the goal of minimizing the difference between the output pulse sequence of the second-level attenuation processing unit and the target smooth pulse sequence, the synaptic connection weights in the first-level attenuation processing unit and the second-level attenuation processing unit are jointly trained. The first-level attenuation processing unit and the second-level attenuation processing unit after training are deployed in series to form the attenuation processing unit.

3. The method for measuring the routing of underground pipelines based on spatial coordinate correction according to claim 2, characterized in that, The process of processing the pulse sequence through the attenuation processing unit to obtain fused feature data includes: The first-stage attenuation processing unit performs membrane potential integration and reset on the pulse sequence, suppresses transient pulse clusters, and outputs an intermediate sequence. The second-stage attenuation processing unit performs long-time window integration on the intermediate sequence, smooths the residual fluctuations, and outputs a smoothed sequence. The smoothed sequence is decoded to obtain the fused feature data.

4. The method for measuring the routing of underground pipelines based on spatial coordinate correction according to claim 1, characterized in that, The construction process of the plasticity knowledge graph specifically includes: Acquire historical and standardized data, including pipeline attributes, pipeline spatial relationships, geological environment information, and multi-source detection signal characteristics, as a source of knowledge; Based on the aforementioned knowledge sources, a plastic knowledge graph containing nodes and synaptic connection edges is constructed. The nodes include pipeline nodes representing underground physical entities, feature nodes representing detection signal characteristics, and spatial constraint nodes representing spatial topological constraints. The synaptic connection edges connect the nodes and include feature-to-pipeline mapping edges representing the mapping relationship between features and pipeline categories, as well as spatial constraint edges representing the spatial logical relationship between pipelines. Based on historical detection data samples, with the goal of maximizing the prediction accuracy of the feature-pipeline mapping edge for the association between feature nodes and corresponding pipeline nodes, the initial weights of the feature-pipeline mapping edge are trained in a supervised manner. During pipeline detection applications, the weights of the features and pipeline mapping edges are dynamically updated online based on the pipeline category labels and spatial location confidence scores output by the graph inference.

5. The method for measuring the routing of underground pipelines based on spatial coordinate correction according to claim 4, characterized in that, The process of performing graph reasoning through a plasticity knowledge graph and outputting pipeline category labels and spatial location confidence scores specifically includes: The plastic knowledge graph receives the fused feature data and activates feature nodes whose similarity to the fused feature data exceeds a preset threshold. By using the feature-pipeline mapping edges, the confidence of the activated feature nodes is propagated to the connected pipeline nodes, and the overall confidence of each pipeline node is calculated. The pipeline nodes are sorted based on the comprehensive confidence score, and at least one pipeline node whose comprehensive confidence score exceeds the confidence threshold is selected as a candidate pipeline node. Based on the spatial constraint edges, the spatial relationship between the candidate pipeline nodes is validated for consistency. The category attributes of the validated candidate pipeline nodes and the comprehensive confidence scores are output as pipeline category labels and spatial location confidence scores, respectively. For candidate pipeline nodes that fail verification, reduce the overall confidence level of the corresponding pipeline node according to the type of constraint violated, and re-execute the selection and verification steps.

6. The method for measuring the routing of underground pipelines based on spatial coordinate correction according to claim 1, characterized in that, The process of acquiring multi-source underground pipeline detection data and encoding it into a pulse sequence specifically includes: Acquire multi-source underground pipeline detection data, which includes electromagnetic induction detection signals, ground-penetrating radar reflection waveform signals, inertial measurement data, and spatial coordinate positioning data; In response to the voltage amplitude characteristics of the electromagnetic induction detection signal, the frequency and pulse conversion coding unit is activated to map the voltage amplitude to the firing frequency of the corresponding spiking neuron per unit time, thereby generating the first pulse sequence. In response to the waveform amplitude distribution and temporal structure characteristics of the ground-penetrating radar reflected waveform signal, the delay and pulse conversion coding unit is activated, and a corresponding pulse emission delay is allocated based on the amplitude of the sampling points in the waveform to generate a second pulse sequence; In response to the motion state change rate characteristics of the inertial measurement data, the event-driven conversion encoding unit is activated, triggering a pulse event only when the change in acceleration or angular velocity exceeds a dynamic threshold, generating a third pulse sequence; In response to the digital quantization characteristics of the spatial coordinate positioning data, the phase encoding unit is activated to map the binary bits of the coordinate values ​​into pulses with a specific phase, generating a fourth pulse sequence; The first pulse sequence, the second pulse sequence, the third pulse sequence, and the fourth pulse sequence are synchronized and aligned in time to output a pulse sequence under a unified time reference.

7. The method for measuring the routing of underground pipelines based on spatial coordinate correction according to claim 1 or 5, characterized in that, After outputting the pipeline category label and spatial location confidence score, the process also includes a knowledge graph-guided cross-modal feature point matching and 3D reconstruction, specifically including: Extract electromagnetic induction direction feature points, ground-penetrating radar reflection feature points, and spatial trajectory feature points from the multi-source underground pipeline detection data; Based on spatiotemporal proximity, the electromagnetic induction direction feature points, ground-penetrating radar reflection feature points, and spatial trajectory feature points are combined in pairs to generate a cross-modal candidate matching pair set. In response to each candidate matching pair in the cross-modal candidate matching pair set, extract at least one spatial constraint edge associated with the pipeline node corresponding to the pipeline category label within the plasticity knowledge graph; Based on the spatial constraint edges, the geometric relationship of the feature points associated with the candidate matching pairs in three-dimensional space is verified for consistency. The candidate matching pairs that pass the verification are confirmed as valid cross-modal matching point pairs; Based on all valid cross-modal matching point pairs, a three-dimensional spatial location model of the underground pipeline is reconstructed.

8. A system for measuring the routing of underground pipelines based on spatial coordinate correction, characterized in that, The underground pipeline routing measurement system, applied to the spatial coordinate correction-based method according to any one of claims 1-7, specifically includes: The multi-source data encoding unit is used to acquire multi-source underground pipeline detection data and encode it into a pulse sequence; An attenuation processing unit is used to process the pulse sequence to obtain fused feature data. The attenuation processing unit includes at least a first-level attenuation processing unit and a second-level attenuation processing unit. The first-level attenuation processing unit is configured to suppress transient pulse clusters that exceed a frequency threshold, and the second-level attenuation processing unit is configured to smooth residual fluctuations. The plasticity knowledge graph reasoning unit performs graph reasoning based on the fused feature data through the plasticity knowledge graph, and outputs pipeline category labels and spatial location confidence scores, wherein the plasticity knowledge graph contains synaptic connection edges, and the graph reasoning includes propagating the spatial location confidence scores based on the weights of the synaptic connection edges; The feedback adjustment unit is used to generate an adjustment signal for the corresponding sensing channel based on the spatial location confidence level, and to adjust the weight gain of the sensing channel corresponding to the multi-source underground pipeline detection data.

9. The underground pipeline routing measurement system based on spatial coordinate correction according to claim 8, characterized in that, The underground pipeline routing measurement system is deployed on a heterogeneous computing platform, which includes: The front-end sensing and coding module integrates the hardware interfaces of an electromagnetic inductor, ground-penetrating radar, inertial measurement unit, and real-time dynamic positioning receiver, and includes the multi-source data coding unit, which is used to receive raw detection data in real time and complete pulse coding. The core fusion and reasoning computation module includes the attenuation processing unit and the plastic knowledge graph reasoning unit, which are used to perform pulse sequence fusion processing and graph reasoning computation. The backend control and feedback module includes the feedback adjustment unit and integrates a display unit and a human-computer interaction interface. It is used to generate feedback adjustment instructions based on the reasoning results, control the entire system workflow, and output visual results.

10. The underground pipeline routing measurement system based on spatial coordinate correction according to claim 8, characterized in that, The plastic knowledge graph reasoning unit further includes: The remote knowledge injection interface is used to receive updated pipeline attributes, spatial relationship rules, and typical signal characteristic data, and to update the nodes and edges of the local plasticity knowledge graph in an incremental manner. An incremental learning engine is used to dynamically optimize the weights of the feature-pipeline mapping edges online based on the pipeline category labels and spatial location confidence scores output by the graph inference.

Images