Multi-sensor fusion navigation method of indoor wheeled inspection robot of booster station room

Abstract

The application discloses a kind of multi-sensor fusion navigation method and system of indoor wheeled inspection robot of booster station, method includes: synchronous acquisition laser radar, vision camera, inertial measurement unit, odometer data, real-time detection electromagnetic interference intensity and grading, corresponding window anti-interference filter is executed to disturbed sensor, and sensor credibility mark is generated based on residual error.Based on laser point cloud, construct geometric grid map, extract cabinet and passage semantics by visual semantic segmentation, map electromagnetic interference intensity to generate semantic fusion map.Rely on extended Kalman filter fusion positioning, combine credibility mark and observation covariance Dynamic down-regulation disturbed sensor weight, with shortest path, interference accumulation minimum as target planning navigation path.After reaching target area, identify visual mark and combine inertial measurement unit to correct pose, drive robot to berth and inspect.Provide reliable technical support for power field station operation and maintenance intelligentization.

Description

Technical Field

[0001] This invention belongs to the field of intelligent inspection technology, and more specifically, relates to a multi-sensor fusion navigation method and system for an indoor wheeled inspection robot for a booster station. Background Technology

[0002] With the rapid development of intelligent operation and maintenance systems for power systems, indoor substations, as core hubs for voltage conversion and power transmission, require routine inspections of their equipment operation status as a crucial link in ensuring the safety and stability of the power grid. Traditional manual inspection methods are constrained by multiple factors, including the enclosed environment, high-voltage electromagnetic risks, and time limitations, resulting in inherent drawbacks such as low inspection efficiency, high rate of missed hazard detection, and difficulty in ensuring personnel safety. Wheeled inspection robots are gradually becoming the core equipment to replace manual labor in completing routine inspections of indoor substations, and navigation and positioning technology is the core foundation for these robots to achieve autonomous operation.

[0003] Current indoor inspection robots mostly employ navigation schemes using single sensors or fixed-weight multi-sensor fusion, which exhibit significant adaptability deficiencies under the unique operating conditions of substations. Substations are densely packed with high-voltage power distribution equipment, resulting in high-intensity, non-uniformly distributed alternating electromagnetic interference throughout the area. Conventional sensors such as LiDAR, vision cameras, and inertial measurement units are susceptible to electromagnetic radiation, leading to increased ranging noise, increased image noise, inertial data drift, and odometer deviation. The failure of a single sensor directly causes navigation interruption. Fixed-weight fusion schemes cannot dynamically adjust the trust level for disturbed sensors, causing positioning accuracy to drop sharply with increasing interference intensity, making the robot prone to positioning inaccuracies and path deviations.

[0004] Existing navigation map construction schemes mostly only cover spatial geometric information, failing to incorporate the semantic attributes and electromagnetic environment characteristics of the substation scenario. Path planning only optimizes path length, unable to proactively avoid areas with high electromagnetic interference. When the robot passes through areas with strong radiation, the reliability of sensor data decreases significantly, resulting in insufficient navigation safety margin. Furthermore, long-distance global navigation suffers from cumulative positioning errors. When the robot reaches the vicinity of the target equipment, it cannot meet the millimeter-level docking accuracy required for close-range inspection, leading to a decline in the quality of equipment status data acquisition and compromising the effectiveness of inspection operations.

[0005] The aforementioned technical deficiencies directly lead to low operational reliability and high failure rate of indoor substation inspection robots, severely restricting the large-scale implementation of intelligent operation and maintenance of power stations. There is an urgent need for targeted navigation technology solutions to overcome the bottleneck of scenario adaptation, comprehensively improve the anti-interference capability and operational accuracy of inspection robots, and provide technical support for the safe and stable operation of core hubs of the power system. Summary of the Invention

[0006] This invention aims to address the pain points of existing inspection robots in the strong electromagnetic interference environment of substations, such as weak navigation anti-interference capability, insufficient positioning accuracy, lack of path risk avoidance, and substandard end-point docking accuracy. It proposes a multi-sensor fusion navigation scheme adapted to special working conditions, improves the operational reliability and accuracy of inspection robots, and supports the large-scale implementation of intelligent operation and maintenance of power stations.

[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation, comprising: S1. Synchronously acquire data from lidar, vision camera, inertial measurement unit and wheel odometer, and detect the current electromagnetic interference intensity in real time; classify the interference level according to the intensity, perform anti-interference filtering with a corresponding filtering window length on the interfered sensor, and generate a sensor reliability degradation label based on the filtering residual; S2. Construct an indoor geometric grid map based on lidar point cloud, use visual camera images to perform semantic segmentation to extract the semantics of equipment cabinets and inspection channels, and map the electromagnetic interference intensity of each area to the geometric grid map to generate a semantic fusion map. S3. Based on the extended Kalman filter framework, multi-sensor data is fused for localization. The fusion weight of the corresponding sensor is dynamically reduced according to the confidence degradation flag and observation covariance to obtain the real-time pose. The navigation path is generated by searching in the semantic fusion map with the path length and the cumulative intensity of electromagnetic interference in the traversed area as the joint optimization objective. S4. When the robot travels along the navigation path to the preset area of ​​the target equipment cabinet, it uses a vision camera to identify the preset visual markings on the cabinet, combines the data from the inertial measurement unit to perform visual inertial relative pose calculation and closed-loop correction, and drives the robot to stop at the target pose to perform inspection.

[0008] Furthermore, in step S1, real-time detection of the current electromagnetic interference intensity specifically includes: A three-dimensional magnetic field sensor array is arranged on the wheeled inspection robot body to collect spatial three-axis magnetic field components; the magnetic field gradient change rate and high-frequency noise variance of the three-axis magnetic field components are calculated; the magnetic field gradient change rate is compared with a preset static magnetic field threshold, and the high-frequency noise variance is compared with a preset dynamic noise threshold; the electromagnetic interference intensity value of the current environment is calculated by combining the magnetic field gradient change rate and the high-frequency noise variance.

[0009] Furthermore, in step S1, generating a sensor reliability degradation flag based on the filter residual specifically includes: Calculate the filter residual between the output data of each sensor after anti-interference filtering and the filter prediction value; When the filtered residual exceeds a preset residual threshold for multiple consecutive sampling periods, the cumulative deviation of the filtered residual from the residual threshold is calculated. Based on the preset range in which the cumulative deviation is located, a corresponding level of confidence downgrade identifier is generated; wherein, the higher the level of the confidence downgrade identifier, the more severe the distortion of the sensor data caused by electromagnetic interference, and the lower the initial prior weight in the subsequent fusion positioning.

[0010] Further, in step S2, the electromagnetic interference intensity of each region is mapped to the geometric grid map to generate a semantic fusion map, specifically including: Using the historical inspection trajectory points of the wheeled inspection robot as the sampling center, the electromagnetic interference intensity value and spatial coordinates corresponding to each sampling center point are extracted. Spatial interpolation is performed on the discrete electromagnetic interference intensity values ​​to construct an initial electromagnetic interference intensity distribution field covering the booster station room; The spatial outline of the cabinet is determined by extracting semantic tags of the equipment cabinet. The spatial outline of the cabinet is used as the electromagnetic shielding constraint boundary. The electromagnetic shadow area on the side of the cabinet facing away from the electromagnetic radiation source is identified in the initial electromagnetic interference intensity distribution field. The interference intensity value in the electromagnetic shadow area is attenuated and corrected to obtain the corrected continuous electromagnetic interference intensity probability distribution field. The mean and variance of the continuous electromagnetic interference intensity probability distribution field are mapped to the corresponding grids of the geometric grid map as electromagnetic interference attributes. At the same time, the semantic labels of the equipment cabinet and the inspection channel are projected to the corresponding three-dimensional point cloud grids to complete multi-dimensional fusion.

[0011] Furthermore, in step S3, dynamically reducing the fusion weights of the corresponding sensors based on the confidence degradation flag and observation covariance specifically includes: In the update phase of the extended Kalman filter framework, an adaptive attenuation factor is constructed; the adaptive attenuation factor is negatively correlated with the level of the confidence degradation indicator and negatively correlated with the trace of the current observation covariance of the sensor. The adaptive attenuation factor is used to amplify and correct the observation noise covariance matrix of the corresponding sensor, so as to effectively reduce the fusion weight of the sensor in the Kalman gain calculation. When the fusion weight of the lidar or vision camera is detected to drop below a preset failure threshold, a short-term dead reckoning compensation mechanism is triggered. This mechanism uses high-frequency sampled inertial measurement unit data and wheel odometer data for tight coupling integration to maintain the continuity of positioning output.

[0012] Furthermore, the specific calculation method for the adaptive attenuation factor is as follows: In the formula, This is a preset level decay coefficient used to control the penalty applied to weights by credibility downgrades; The pre-defined covariance decay coefficient is used to control the penalty applied to the weights by observation uncertainty. Degrade the credibility level of the identifier at the current moment; The original observation noise covariance matrix, The traces are recorded as ; It is a natural exponential function.

[0013] Furthermore, in step S3, the navigation path is generated by searching the semantic fusion map with the goal of minimizing both path length and the cumulative intensity of electromagnetic interference in the traversed areas. This specifically includes: Construct a comprehensive cost function that includes a distance cost term and an electromagnetic interference penalty cost term; For each candidate access grid in the semantic fusion map, if the average electromagnetic interference intensity it maps exceeds a preset safe access threshold, an infinite penalty cost is assigned to mark it as a no-entry zone; if it does not exceed the threshold, a non-linear electromagnetic interference penalty cost is calculated based on the average electromagnetic interference intensity. A heuristic search algorithm is adopted, with the goal of minimizing the comprehensive cost function, to search and generate the globally optimal navigation path from the current pose to the pose of the target equipment cabinet in the set of passable grids corresponding to the semantics of the inspection channel.

[0014] Furthermore, in step S4, a visual camera is used to identify preset visual markings on the cabinet, and visual-inertial relative pose calculation and closed-loop correction are performed in conjunction with inertial measurement unit data. Specifically, this includes: The preset visual signage is an active light-emitting infrared signage or a reflective signage with a specific coded pattern that is affixed to the surface of the target equipment cabinet. When the visual camera recognizes the preset visual mark, it extracts the pixel feature points of the mark and combines them with the high-frequency angular velocity and acceleration data output by the inertial measurement unit to construct a visual inertial nonlinear optimization model. With the optimization objective of minimizing feature point reprojection error and inertial pre-integration error, the six-degree-of-freedom relative pose of the wheeled inspection robot relative to the target equipment cabinet is solved; The pose deviation is calculated based on the six-degree-of-freedom relative pose and the preset target docking pose. When the pose deviation is less than the preset docking accuracy threshold, the docking is determined to be completed and the chassis is locked to perform the inspection task.

[0015] As a second aspect of the present invention, a multi-sensor fusion navigation system for an indoor wheeled inspection robot in a substation is also provided, comprising: The multi-source acquisition and reliability assessment unit is used to simultaneously acquire data from lidar, vision camera, inertial measurement unit and wheel odometer, and detect the current electromagnetic interference intensity in real time; according to the intensity, the interference level is divided, anti-interference filtering with a corresponding filtering window length is performed on the interfered sensor, and a sensor reliability degradation label is generated based on the filtering residual; The semantic fusion map construction unit is used to construct an indoor geometric grid map based on LiDAR point cloud, extract the semantics of equipment cabinets and inspection channels by semantic segmentation using visual camera images, and map the electromagnetic interference intensity of each area to the geometric grid map to generate a semantic fusion map. The low-interference path planning unit is used for localization based on the fusion of multi-sensor data using the extended Kalman filter framework. It dynamically reduces the fusion weight of the corresponding sensor according to the confidence degradation flag and observation covariance to obtain the real-time pose. The navigation path is generated by searching in the semantic fusion map with the path length and the minimum cumulative intensity of electromagnetic interference in the traversed area as the joint optimization objective. The visual inertial correction unit is used to identify preset visual markings on the cabinet using a visual camera when the robot travels along the navigation path to the preset area of ​​the target equipment cabinet. It then combines the data from the inertial measurement unit to perform visual inertial relative pose calculation and closed-loop correction, driving the robot to stop at the target pose to perform inspection.

[0016] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor, according to any one of the claims, a multi-sensor fusion navigation method for an indoor wheeled inspection robot in a booster station.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation, as described in this invention, simultaneously collects multi-source sensor data from lidar, vision cameras, inertial measurement units, and wheeled odometers. It also relies on a three-dimensional magnetic field sensor array to monitor the intensity of electromagnetic interference across the entire area in real time. By matching filter windows of corresponding lengths according to interference level classification standards, differentiated anti-interference filtering is performed on disturbed sensors. Finally, based on the residual statistical results of the filtered output value and the predicted value, a reliability downgrade label for the corresponding sensor is generated. This method can perform preliminary purification of raw sensor data in strong electromagnetic interference environments, pre-quantify the data reliability of each sensor, and prevent distorted data from directly entering subsequent calculation stages. It reduces the impact of electromagnetic interference on the navigation system from the source of perception, solving the problems of traditional solutions that cannot identify sensor failure risks in advance and that disturbed data directly participates in calculations, leading to positioning deviations. This provides a reliable input foundation for subsequent positioning and navigation stages.

[0018] 2. The multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation, as described in this invention, constructs an indoor spatial geometric grid map using LiDAR point cloud data. It then extracts semantic labels for equipment cabinets and inspection channels by combining semantic segmentation results from visual images. Finally, it maps global electromagnetic interference intensity data to corresponding grids after spatial interpolation and shielding correction. This integrates geometric space, scene semantics, and electromagnetic risk information to generate a semantic fusion map. Global path search is then performed with the joint objective of minimizing path length and cumulative electromagnetic interference intensity. This method can construct a dedicated navigation map adapted to the substation scenario, allowing path planning to simultaneously consider traffic efficiency and electromagnetic safety. It actively guides the robot to avoid high-interference areas, solving the problem that traditional maps only contain geometric information and path planning cannot avoid electromagnetic risks. While ensuring inspection efficiency, it improves the anti-interference margin of the navigation process and guarantees the stability of sensor data during robot operation.

[0019] 3. The multi-sensor fusion navigation method for the indoor wheeled inspection robot of the substation of this invention completes the fusion positioning calculation of multi-sensor data through an extended Kalman filter framework. It dynamically adjusts the fusion weights of each sensor by combining the pre-generated sensor confidence identifiers with real-time observation covariance, reducing the impact of disturbed sensors on the positioning results. After the robot arrives at the area surrounding the target cabinet, it identifies the cabinet's preset visual markers and completes relative pose calculation and closed-loop correction by combining inertial measurement unit data, driving the robot to stop at the target working pose. This method can achieve interference-resistant dynamic fusion positioning, avoiding navigation interruption caused by the failure of a single sensor, and eliminating the cumulative positioning error of long-distance navigation. It solves the problems of poor anti-interference performance and inability of traditional fixed-weight fusion to meet the end-point docking accuracy requirements, ensuring both the continuity of long-distance navigation and meeting the high-precision docking requirements of close-range inspection. Attached Figure Description

[0020] Figure 1 This is a flowchart of a multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of an indoor wheeled inspection robot according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the system units in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] Example 1 Please refer to Figure 1 This embodiment 1 provides a multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation, including: S1. Synchronously acquire data from lidar, vision camera, inertial measurement unit and wheel odometer, and detect the current electromagnetic interference intensity in real time; classify the interference level according to the intensity, perform anti-interference filtering with a corresponding filtering window length on the interfered sensor, and generate a sensor reliability degradation label based on the filtering residual; S2. Construct an indoor geometric grid map based on lidar point cloud, use visual camera images to perform semantic segmentation to extract the semantics of equipment cabinets and inspection channels, and map the electromagnetic interference intensity of each area to the geometric grid map to generate a semantic fusion map. S3. Based on the extended Kalman filter framework, multi-sensor data is fused for localization. The fusion weight of the corresponding sensor is dynamically reduced according to the confidence degradation flag and observation covariance to obtain the real-time pose. The navigation path is generated by searching in the semantic fusion map with the path length and the cumulative intensity of electromagnetic interference in the traversed area as the joint optimization objective. S4. When the robot travels along the navigation path to the preset area of ​​the target equipment cabinet, it uses a vision camera to identify the preset visual markings on the cabinet, combines the data from the inertial measurement unit to perform visual inertial relative pose calculation and closed-loop correction, and drives the robot to stop at the target pose to perform inspection.

[0023] Please refer to Figure 2 This embodiment 1 further elaborates on the above steps.

[0024] (1) Multi-source data collection and credibility assessment In this embodiment, when the wheeled inspection robot performs inspection tasks indoors at the substation, the high-intensity, non-uniformly distributed alternating electromagnetic interference throughout the indoor area makes a single sensor highly susceptible to electromagnetic radiation, leading to problems such as increased ranging noise, increased image noise, or inertial data drift. Therefore, it is necessary to simultaneously collect data from multiple heterogeneous sensors to build a redundant sensing foundation. Specifically, this involves simultaneously collecting point cloud data generated by LiDAR scanning the indoor environment, indoor two-dimensional image data captured by a vision camera, six-axis attitude data (including high-frequency angular velocity and linear acceleration) output by the inertial measurement unit (IMU), and wheel speed data fed back by the wheeled odometer.

[0025] While simultaneously acquiring the aforementioned multi-source navigation data, the system also needs to perform electromagnetic environment benchmark monitoring in parallel to assess whether these sensor data are contaminated by electromagnetic interference in the current environment. Specifically, a three-dimensional magnetic field sensor array deployed on the robot body acquires the spatial three-axis magnetic field components of the current environment in real time. For the acquired three-axis magnetic field components, the magnetic field gradient change rate is obtained by calculating the ratio of the difference in magnetic field components at adjacent sampling times to the sampling time interval; simultaneously, high-pass filtering is performed on the three-axis magnetic field components to extract high-frequency fluctuation signals, and the high-frequency noise variance of these high-frequency fluctuation signals is calculated. Subsequently, the magnetic field gradient change rate is compared with a preset static magnetic field threshold, and the high-frequency noise variance is compared with a preset dynamic noise threshold. By weighted summing the degree of exceedance of both, the electromagnetic interference intensity value of the current environment is calculated comprehensively.

[0026] After obtaining the electromagnetic interference intensity value, it is divided into three interference levels: low, medium, and high, based on the magnitude of the value. For sensors susceptible to electromagnetic interference, such as lidar and vision cameras, an anti-interference filter window of corresponding length is allocated according to the current interference level. Specifically, the higher the interference level, the longer the allocated sliding filter window is to enhance the smoothing and suppression of high-frequency electromagnetic noise; the lower the interference level, the shorter the filter window is used to retain more original data details. By performing anti-interference filtering with the corresponding window length, the smoothed observation data of each sensor is output.

[0027] After completing the anti-interference filtering, the filter residual between the output data of each sensor after anti-interference filtering and the filter prediction value calculated based on the state at the previous moment is further calculated. To prevent misjudgment caused by occasional noise, a continuous judgment mechanism is set: when the filter residual of a certain sensor exceeds a preset residual threshold for multiple consecutive sampling periods (e.g., 5 consecutive sampling periods), the cumulative sum of the portion of the filter residual exceeding the residual threshold is calculated and used as the cumulative deviation. This cumulative deviation can objectively reflect the degree of continuous deviation of sensor data caused by electromagnetic interference over a period of time.

[0028] Finally, based on the preset interval where the calculated cumulative deviation falls, a corresponding level of confidence degradation label is generated. The preset interval is divided into three increasing numerical ranges, corresponding to Level 1, Level 2, and Level 3 confidence degradation labels, respectively. The larger the value of the interval where the cumulative deviation falls, the higher the level of the generated degradation label, indicating a more severe degree of distortion in the sensor data caused by electromagnetic interference. In the subsequent extended Kalman filter fusion positioning process, according to the level of this degradation label, the initial prior weight of the corresponding sensor observation noise covariance matrix is ​​dynamically reduced by a preset attenuation ratio, thereby effectively suppressing the negative impact of the interfered sensor on the global positioning accuracy.

[0029] (2) Semantic fusion map construction Based on the 3D point cloud data generated from the LiDAR scanning of the indoor environment of the booster station, statistical filtering is performed on the 3D point cloud data to remove outlier noise points, resulting in a filtered effective point cloud. A normal distribution transformation registration algorithm is used to align multiple frames of effective point clouds to a unified global coordinate system, and then stitching them together to generate a global 3D point cloud map. To construct a geometric grid map suitable for navigation, the global 3D point cloud map is projected vertically onto a 2D horizontal plane, which is then divided into 2D grids with fixed resolution. The height distribution characteristics of the point clouds falling within each 2D grid are statistically analyzed, and each 2D grid is marked as an obstacle grid or a passable grid based on these height distribution characteristics, thereby generating an indoor geometric grid map representing the physical structure of the indoor space.

[0030] Based on the acquisition of the indoor physical space structure, and using two-dimensional images of the substation interior captured by a visual camera, the two-dimensional images are input into a pre-trained convolutional neural network semantic segmentation model. The convolutional neural network semantic segmentation model extracts multi-scale spatial features from the two-dimensional image through an encoder, and upsamples and fuses these features through a decoder, outputting a pixel-level semantic probability map with the same resolution as the two-dimensional image. For each pixel in the pixel-level semantic probability map, the category label with the highest probability is selected to generate a semantic segmentation result map containing semantic masks for equipment cabinets and inspection channels, thus completing the semantic extraction of equipment cabinets and inspection channels at the image pixel level.

[0031] To achieve spatial unification of geometric structure and semantic information, the intrinsic and extrinsic parameter matrices of the visual camera are obtained. Combined with depth information provided by a depth sensor or depth projection from a LiDAR point cloud, the pixels in the semantic segmentation result image, representing the semantic masks of the equipment cabinet and inspection channels, are back-projected into 3D space. The spatial distance between the back-projected 3D semantic points and the data points in the global 3D point cloud map is calculated. 3D semantic points with a distance less than a preset fusion threshold are assigned category labels to their corresponding global 3D point cloud map data points, generating a 3D semantic point cloud carrying semantic labels for both the equipment cabinet and inspection channels.

[0032] After acquiring the 3D semantic point cloud of the equipment cabinet and the inspection channel semantic tags, the electromagnetic interference intensity value and spatial coordinates corresponding to each sampling center point are extracted, using the historical inspection trajectory points of the wheeled inspection robot as the sampling center. Let the first... The three-dimensional spatial coordinates of the sampling center points are: Its electromagnetic interference intensity value is The total number of sampling center points is The inverse distance weighting method is used to spatially interpolate discrete electromagnetic interference intensity values ​​for any spatial location to be estimated within the substation. Calculate its initial electromagnetic interference intensity value. The calculation formula is In the formula, For the first Each sampling center point is paired with a location The spatial weight is calculated as follows: ; Represents the Euclidean distance norm. The preset distance attenuation index is used to control the rate at which spatial distance affects the attenuation of interference intensity.

[0033] Considering the significant physical shielding effect of metal equipment cabinets within the substation on electromagnetic waves, pure mathematical interpolation would smooth out the sharp drop in intensity caused by this physical shielding, leading to a severe overestimation of interference intensity within the shadowed area. Therefore, the extracted semantic tags of the equipment cabinets are used to determine their spatial contours, which are then used as electromagnetic shielding constraints. This allows for the identification of the electromagnetic shadowed area on the side of the cabinet facing away from the electromagnetic radiation source within the initial electromagnetic interference intensity distribution field. Let the spatial coordinates of the main electromagnetic radiation source be... For any location to be estimated in space , build from point to If the ray vector intersects with the three-dimensional bounding box of any equipment cabinet's spatial contour, then the position is determined. The system is located within the electromagnetic shadow region. The initial interference intensity value within the electromagnetic shadow region is attenuated and corrected to obtain the corrected continuous electromagnetic interference intensity value. The calculation formula is In the formula, The preset material attenuation coefficient reflects the shielding effectiveness of the metal cabinet against electromagnetic waves of a specific frequency band. The equivalent penetration thickness of the ray vector as it passes through the interior of the equipment cabinet's spatial contour.

[0034] Meanwhile, considering the significant time-varying and random nature of the electromagnetic environment within the substation due to the start-up and shutdown of high-voltage equipment and load fluctuations, a single spatial sampling cannot accurately reflect the true interference level. Therefore, a probabilistic model is performed on the corrected continuous electromagnetic interference intensity values ​​to obtain the probability distribution field of continuous electromagnetic interference intensity. This model is applied to any target grid in the geometric grid map. Extract all spatial coordinates that fall within the target grid during the historical inspection cycle. Historical trajectory sampling points within a three-dimensional spatial range are used to obtain continuous electromagnetic interference intensity values ​​after spatial interpolation and attenuation correction. These values ​​are then combined in chronological order to form the target grid. Interference intensity observation sequence ,in This represents the total number of samples in the observed sequence.

[0035] Assuming the observation sequence follows a Gaussian distribution, the target grid is calculated based on the interference intensity observation sequence. mean electromagnetic interference intensity With electromagnetic interference intensity variance The calculation formulas are as follows: as well as In the formula, Represents the target grid The first in the interference intensity observation sequence One corrected interference strength value.

[0036] Through the above probabilistic modeling process, the calculated mean electromagnetic interference intensity is... and electromagnetic interference intensity variance Target raster mapped to a geometric raster map As an electromagnetic interference attribute, the semantic tags of the equipment cabinet and the inspection channel are projected onto the corresponding 3D point cloud grid to complete the multi-dimensional fusion and generate a semantic fusion map.

[0037] (3) Low-interference path planning Furthermore, localization is achieved by fusing multi-sensor data based on an extended Kalman filter framework, where the current data to be fused is... One sensor Obtain its credibility downgrade flag level at the current moment. and the original observation noise covariance matrix .calculate The traces are recorded as , construct the first Adaptive attenuation factor for each sensor Its calculation formula is In the formula, This is the preset level attenuation coefficient; This is the preset covariance attenuation coefficient; It is a natural exponential function. Utilizing the aforementioned adaptive decay factor... For the original observation noise covariance matrix After amplification correction, the corrected observation noise covariance matrix is ​​obtained. The calculation formula is In the update phase of the extended Kalman filter framework, the Kalman gain... The calculation formula is In the formula Let be the prior state covariance matrix. To observe the Jacobian matrix, we will... Substitute into the formula to reduce the Kalman gain Unlike traditional methods that rely solely on posterior covariance for weight adjustment, this embodiment introduces a confidence downgrade indicator level. The reason is that strong electromagnetic interference within the booster station can cause distortion of the lidar point cloud or a surge in noise in the visual image. This interference is sudden, and relying solely on covariance... Feedback adjustment introduces significant statistical lag, causing erroneous observations to contaminate the filtering results before covariance convergence; this can be addressed by using pre-extracted... With the posterior By jointly constructing an attenuation factor, a dual suppression combining feedforward and feedback can be achieved, suppressing the fusion weights of contaminated sensors in advance at the moment of interference. Subsequently, sensor data is acquired. Actual observed values and combined with observation function Calculate the observation residuals Using the calculated Kalman gain and observation residuals The prior state is corrected, and the posterior state estimate is calculated. And update the posterior state covariance matrix. In the formula It is the identity matrix. Finally, from the posterior state estimate... The three-dimensional position and three-dimensional pose are extracted to obtain the robot's real-time pose.

[0038] During the process of dynamically adjusting the fusion weights, the adaptive attenuation factor will be... As an effective fusion weighting indicator, a preset failure threshold is set. In each filter update cycle, the corresponding LiDAR or vision camera is detected in real time. Is it less than When detected When the observation update of the shielded sensor fails, a short-term dead reckoning compensation mechanism is triggered. This mechanism utilizes only the angular velocity data output from the high-frequency sampled inertial measurement unit (IMU) and the linear velocity data output from the wheel odometer, performing tightly coupled integration using a pre-integration technique. The pre-integration process accumulates the high-frequency IMU data between two low-frequency filter update times in a local coordinate system to calculate the relative displacement increment. Relative velocity increment and relative rotation quaternion increment The aforementioned increments are then superimposed onto the prior pose from the previous moment. The reason for using a tightly coupled inertial measurement unit (IMU) and wheeled odometer instead of a single sensor for calculation is that the indoor floor of the substation is often covered with metal anti-slip textures or cable trench covers. Relying solely on the wheeled odometer is prone to loss of mileage due to wheel slippage or suspension, while relying solely on the IMU will result in severe zero-bias drift in strong electromagnetic environments. The tight coupling of the two can utilize the non-holonomic constraints of the wheeled odometer to suppress the heading and altitude drift of the IMU, while utilizing the high-frequency characteristics of the IMU to compensate for the instantaneous velocity changes of the wheeled odometer when crossing cable trenches, thereby maintaining high-precision pose continuity during external sensor failures.

[0039] After acquiring the real-time pose, the navigation path planning stage begins, where a comprehensive cost function is constructed, including distance cost and electromagnetic interference penalty cost. Its calculation formula is In the formula, The preset distance weighting coefficient, These are preset electromagnetic interference weighting coefficients, which are used to normalize the corresponding cost terms. It is the sum of the Euclidean distances between the center points of all adjacent grid cells on the path; It is the sum of the electromagnetic interference penalty values ​​for all grids passed along the path.

[0040] For each candidate access grid in the semantic fusion map Extract the mean electromagnetic interference intensity of its mapping. and compare it with the preset safe passage threshold. Compare. If Then assign the candidate pass grid. An infinite penalty cost is imposed and marked as a no-go zone; if Then calculate the nonlinear electromagnetic interference penalty cost. Its calculation formula is In the formula, The preset penalty baseline coefficient, The nonlinear growth coefficient is preset. The reason for using a nonlinear exponential calculation formula instead of the traditional linear cost is that the electromagnetic interference in the booster station has a significant cumulative effect on the robot's internal bus communication and sensor accuracy, and a step error rate surge. If a linear cost is used, the planned path will often run close to the edge of high interference equipment, causing the robot to crash or go out of control due to interference accumulation during long-term operation. The nonlinear exponential cost can make the cost increase exponentially when the interference intensity is close to the safety threshold, forcing the planning algorithm to actively maintain a sufficient physical safety margin between the path and the high interference source.

[0041] Furthermore, a heuristic search algorithm is employed to search and generate a globally optimal navigation path within the semantic fusion map. The algorithm starts at the grid cell containing the robot's current pose and ends at the grid cell containing the target equipment cabinet's pose, expanding nodes within the set of passable grid cells corresponding to the semantics of the inspection channel. For any currently expanded node during the search process... Calculate its comprehensive evaluation function In the formula, From the starting point to the node The actual cumulative comprehensive value; For the node The heuristic cost estimate for reaching the destination is calculated by multiplying the pure Euclidean distance between the current node and the destination by the base cost per unit distance. The algorithm first expands the range in each iteration. The node with the smallest value is used to trace back to the endpoint, and the resulting sequence of nodes is the globally optimal navigation path.

[0042] (4) Visual inertial correction When the wheeled inspection robot travels along the navigation path to the preset area of ​​the target equipment cabinet, the vision camera begins searching for actively emitting infrared markers or reflective markers with specific coded patterns affixed to the surface of the target equipment cabinet as preset visual markers. When the vision camera recognizes the preset visual marker, it extracts the pixel feature points of the marker. ,in For the first The two-dimensional pixel coordinates of each feature point on the image plane, combined with the high-frequency angular velocity output by the inertial measurement unit. and linear acceleration A visual inertial nonlinear optimization model was constructed.

[0043] To solve for the robot's spatial state, a general objective function is constructed with the joint optimization objective of minimizing feature point reprojection error and inertial pre-integration error. In the formula, The total number of observed pixel feature points. For the first Reprojection error of each feature point This is the intrinsic parameter matrix of the visual camera. Let be the homogeneous transformation matrix from the robot's body coordinate system to the world coordinate system. For the first The known three-dimensional spatial coordinates of each feature point in the labeled coordinate system For camera projection function, The visual observation noise covariance matrix; This represents the total number of inertial pre-integral constraints. For the first Inertial pre-integration error between adjacent keyframes This represents the actual pre-integrated measurement value of the inertial measurement unit between adjacent keyframes. This is the theoretical pre-integral prediction calculated based on the current state estimate. Let be the inertial measurement noise covariance matrix. The overall objective function is iteratively optimized using a nonlinear least squares method. The six-degree-of-freedom relative pose of the wheeled inspection robot relative to the target equipment cabinet is obtained by solving the problem. .

[0044] After obtaining the relative pose, based on the six-degree-of-freedom relative pose... docking position with the preset target Calculate pose deviation From the aforementioned pose deviation Decompose the translation deviation vector from the middle and rotational deviation angle When the translation deviation vector The Euclidean norm is less than the preset translation and docking accuracy threshold. And the rotational deviation angle The absolute value is less than the preset rotational docking accuracy threshold. When the time is right, the vehicle is deemed to have completed docking and the chassis is locked to perform an inspection.

[0045] Example 2 Please refer to Figure 3 This embodiment 2 provides a multi-sensor fusion navigation system for an indoor wheeled inspection robot in a substation, including: The multi-source acquisition and reliability assessment unit is used to simultaneously acquire data from lidar, vision camera, inertial measurement unit and wheel odometer, and detect the current electromagnetic interference intensity in real time; according to the intensity, the interference level is divided, anti-interference filtering with a corresponding filtering window length is performed on the interfered sensor, and a sensor reliability degradation label is generated based on the filtering residual; The semantic fusion map construction unit is used to construct an indoor geometric grid map based on LiDAR point cloud, extract the semantics of equipment cabinets and inspection channels by semantic segmentation using visual camera images, and map the electromagnetic interference intensity of each area to the geometric grid map to generate a semantic fusion map. The low-interference path planning unit is used for localization based on the fusion of multi-sensor data using the extended Kalman filter framework. It dynamically reduces the fusion weight of the corresponding sensor according to the confidence degradation flag and observation covariance to obtain the real-time pose. The navigation path is generated by searching in the semantic fusion map with the path length and the minimum cumulative intensity of electromagnetic interference in the traversed area as the joint optimization objective. The visual inertial correction unit is used to identify preset visual markings on the cabinet using a visual camera when the robot travels along the navigation path to the preset area of ​​the target equipment cabinet. It then combines the data from the inertial measurement unit to perform visual inertial relative pose calculation and closed-loop correction, driving the robot to stop at the target pose to perform inspection.

[0046] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement any step of a multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation.

[0047] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0048] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0049] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation, characterized in that, include: S1. Synchronously acquire data from lidar, vision camera, inertial measurement unit and wheel odometer, and detect the current electromagnetic interference intensity in real time; classify the interference level according to the intensity, perform anti-interference filtering with a corresponding filtering window length on the interfered sensor, and generate a sensor reliability degradation label based on the filtering residual; S2. Construct an indoor geometric grid map based on lidar point cloud, use visual camera images to perform semantic segmentation to extract the semantics of equipment cabinets and inspection channels, and map the electromagnetic interference intensity of each area to the geometric grid map to generate a semantic fusion map. S3. Based on the extended Kalman filter framework, multi-sensor data is fused for localization. The fusion weight of the corresponding sensor is dynamically reduced according to the confidence degradation flag and observation covariance to obtain the real-time pose. The navigation path is generated by searching in the semantic fusion map with the path length and the cumulative intensity of electromagnetic interference in the traversed area as the joint optimization objective. S4. When the robot travels along the navigation path to the preset area of ​​the target equipment cabinet, it uses a vision camera to identify the preset visual markings on the cabinet, combines the data from the inertial measurement unit to perform visual inertial relative pose calculation and closed-loop correction, and drives the robot to stop at the target pose to perform inspection.

2. The multi-sensor fusion navigation method for an indoor wheeled inspection robot in a booster station according to claim 1, characterized in that, In step S1, the real-time detection of the current electromagnetic interference intensity specifically includes: A three-dimensional magnetic field sensor array is arranged on the wheeled inspection robot body to collect spatial three-axis magnetic field components; the magnetic field gradient change rate and high-frequency noise variance of the three-axis magnetic field components are calculated; the magnetic field gradient change rate is compared with a preset static magnetic field threshold, and the high-frequency noise variance is compared with a preset dynamic noise threshold; the electromagnetic interference intensity value of the current environment is calculated by combining the magnetic field gradient change rate and the high-frequency noise variance.

3. The multi-sensor fusion navigation method for an indoor wheeled inspection robot in a booster station according to claim 1, characterized in that, In step S1, the generation of a sensor reliability degradation flag based on the filter residual specifically includes: Calculate the filter residual between the output data of each sensor after anti-interference filtering and the filter prediction value; When the filtered residual exceeds a preset residual threshold for multiple consecutive sampling periods, the cumulative deviation of the filtered residual from the residual threshold is calculated. Based on the preset range in which the cumulative deviation is located, a corresponding level of confidence downgrade identifier is generated; wherein, the higher the level of the confidence downgrade identifier, the more severe the distortion of the sensor data caused by electromagnetic interference, and the lower the initial prior weight in the subsequent fusion positioning.

4. The multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation according to claim 1, characterized in that, In step S2, the electromagnetic interference intensity of each region is mapped to the geometric grid map to generate a semantic fusion map, specifically including: Using the historical inspection trajectory points of the wheeled inspection robot as the sampling center, the electromagnetic interference intensity value and spatial coordinates corresponding to each sampling center point are extracted. Spatial interpolation is performed on the discrete electromagnetic interference intensity values ​​to construct an initial electromagnetic interference intensity distribution field covering the booster station room; The spatial outline of the cabinet is determined by extracting semantic tags of the equipment cabinet. The spatial outline of the cabinet is used as the electromagnetic shielding constraint boundary. The electromagnetic shadow area on the side of the cabinet facing away from the electromagnetic radiation source is identified in the initial electromagnetic interference intensity distribution field. The interference intensity value in the electromagnetic shadow area is attenuated and corrected to obtain the corrected continuous electromagnetic interference intensity probability distribution field. The mean and variance of the continuous electromagnetic interference intensity probability distribution field are mapped to the corresponding grids of the geometric grid map as electromagnetic interference attributes. At the same time, the semantic labels of the equipment cabinet and the inspection channel are projected to the corresponding three-dimensional point cloud grids to complete multi-dimensional fusion.

5. The multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation according to claim 1, characterized in that, In step S3, the fusion weights of the corresponding sensors are dynamically reduced based on the confidence downgrade flag and the observation covariance, specifically including: In the update phase of the extended Kalman filter framework, an adaptive attenuation factor is constructed; the adaptive attenuation factor is negatively correlated with the level of the confidence degradation indicator and negatively correlated with the trace of the current observation covariance of the sensor. The adaptive attenuation factor is used to amplify and correct the observation noise covariance matrix of the corresponding sensor, so as to effectively reduce the fusion weight of the sensor in the Kalman gain calculation. When the fusion weight of the lidar or vision camera is detected to drop below a preset failure threshold, a short-term dead reckoning compensation mechanism is triggered. This mechanism uses high-frequency sampled inertial measurement unit data and wheel odometer data for tight coupling integration to maintain the continuity of positioning output.

6. The multi-sensor fusion navigation method for an indoor wheeled inspection robot in a booster station according to claim 5, characterized in that, The specific calculation method for the adaptive attenuation factor is as follows: In the formula, This is a preset level decay coefficient used to control the penalty applied to weights by credibility downgrades; The pre-defined covariance decay coefficient is used to control the penalty applied to the weights by observation uncertainty. Degrade the credibility level of the identifier at the current moment; The original observation noise covariance matrix, The traces are recorded as ; It is a natural exponential function.

7. The multi-sensor fusion navigation method for an indoor wheeled inspection robot in a booster station according to claim 1, characterized in that, In step S3, the joint optimization objective is to minimize the path length and the cumulative intensity of electromagnetic interference in the traversed areas. A navigation path is then searched and generated within the semantic fusion map, specifically including: Construct a comprehensive cost function that includes a distance cost term and an electromagnetic interference penalty cost term; For each candidate access grid in the semantic fusion map, if the average electromagnetic interference intensity it maps exceeds a preset safe access threshold, an infinite penalty cost is assigned to mark it as a no-entry zone; if it does not exceed the threshold, a non-linear electromagnetic interference penalty cost is calculated based on the average electromagnetic interference intensity. A heuristic search algorithm is adopted, with the goal of minimizing the comprehensive cost function, to search and generate the globally optimal navigation path from the current pose to the pose of the target equipment cabinet in the set of passable grids corresponding to the semantics of the inspection channel.

8. The multi-sensor fusion navigation method for an indoor wheeled inspection robot in a substation according to claim 1, characterized in that, In step S4, a vision camera is used to identify preset visual markings on the cabinet, and visual-inertial relative pose calculation and closed-loop correction are performed in conjunction with data from the inertial measurement unit. Specifically, this includes: The preset visual signage is an active light-emitting infrared signage or a reflective signage with a specific coded pattern that is affixed to the surface of the target equipment cabinet. When the visual camera recognizes the preset visual mark, it extracts the pixel feature points of the mark and combines them with the high-frequency angular velocity and acceleration data output by the inertial measurement unit to construct a visual inertial nonlinear optimization model. With the optimization objective of minimizing feature point reprojection error and inertial pre-integration error, the six-degree-of-freedom relative pose of the wheeled inspection robot relative to the target equipment cabinet is solved; The pose deviation is calculated based on the six-degree-of-freedom relative pose and the preset target docking pose. When the pose deviation is less than the preset docking accuracy threshold, the docking is determined to be completed and the chassis is locked to perform the inspection task.

9. A multi-sensor fusion navigation system for an indoor wheeled inspection robot in a substation, characterized in that, include: The multi-source acquisition and reliability assessment unit is used to simultaneously acquire data from lidar, vision camera, inertial measurement unit and wheel odometer, and detect the current electromagnetic interference intensity in real time; according to the intensity, the interference level is divided, anti-interference filtering with a corresponding filtering window length is performed on the interfered sensor, and a sensor reliability degradation label is generated based on the filtering residual; The semantic fusion map construction unit is used to construct an indoor geometric grid map based on LiDAR point cloud, extract the semantics of equipment cabinets and inspection channels by semantic segmentation using visual camera images, and map the electromagnetic interference intensity of each area to the geometric grid map to generate a semantic fusion map. The low-interference path planning unit is used for localization based on the fusion of multi-sensor data using the extended Kalman filter framework. It dynamically reduces the fusion weight of the corresponding sensor according to the confidence degradation flag and observation covariance to obtain the real-time pose. The navigation path is generated by searching in the semantic fusion map with the path length and the minimum cumulative intensity of electromagnetic interference in the traversed area as the joint optimization objective. The visual inertial correction unit is used to identify preset visual markings on the cabinet using a visual camera when the robot travels along the navigation path to the preset area of ​​the target equipment cabinet. It then combines the data from the inertial measurement unit to perform visual inertial relative pose calculation and closed-loop correction, driving the robot to stop at the target pose to perform inspection.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor according to any one of claims 1-8, which describes a multi-sensor fusion navigation method for an indoor wheeled inspection robot in a booster station.

Images