Battery compartment foreign matter identification method and system based on laser scanning and AI identification

Abstract

The invention belongs to the technical field of foreign matter detection and recognition, and provides a battery compartment foreign matter recognition method and system based on laser scanning and AI recognition. A laser scanning module is used for scanning and collecting three-dimensional point cloud data in a battery cabin in a preset period, point cloud preprocessing is carried out on the collected original point cloud data, the collected original point cloud data are converted into a unified cabin coordinate system, point cloud segmentation is carried out, a hypersphere model of normal state point cloud is constructed based on a support vector data description method, and a hypersphere model is constructed. The abnormal state is judged through a dynamic threshold value, foreign matter position tracing is carried out according to personnel positioning data in combination with multi-period scanning data, and the multi-source data alignment precision and trajectory estimation stability are cooperatively improved through dynamic time warping and space-time filtering. The method can effectively extract complex features, improve foreign matter recognition precision, recognize foreign matters in real time, judge the properties of the foreign matters, improve data processing efficiency, automatically adapt to complex and changeable environments and ensure stable and reliable operation under various working conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of foreign object detection and identification, and specifically relates to a battery compartment foreign object identification method and system based on laser scanning and AI identification. Background Art

[0002] Energy storage power stations are systems that store, convert, and release cyclical electrical energy through electrochemical cells or electromagnetic energy storage media. As the core energy storage unit of an energy storage power station, the battery compartment of an energy storage station faces risks such as short circuits and thermal runaway caused by foreign objects such as metal debris, condensed water, and animal intrusion.

[0003] In the current field of foreign object identification in battery compartments of energy storage power stations, the main methods used include traditional manual inspections and simple image recognition technology.

[0004] Manual inspections involve staff periodically entering the battery compartment and inspecting the equipment and surrounding environment one by one using the naked eye and simple tools like flashlights and binoculars. However, traditional manual inspections are limited by human visual resolution. Small foreign objects, such as metal debris less than 1 cm in size and fine fiber strands, are difficult to detect in the complex battery compartment environment. Furthermore, when foreign objects resemble surrounding equipment in color or shape, misidentification or omission can occur. For example, darker plastic fragments falling onto a black battery casing may be difficult for inspectors to detect.

[0005] Simple image recognition technology: A camera is installed inside the battery compartment to capture images of the interior. These images are transmitted to the image analysis system, which then uses basic image processing algorithms, such as edge detection and color recognition, to attempt to identify foreign objects. Simple image recognition technology also has limitations. Due to the large number of devices and complex backgrounds within the battery compartment, existing edge detection and color recognition algorithms are susceptible to noise when processing images, resulting in inaccurate extraction of foreign object edges and color features. This in turn makes it impossible to accurately distinguish between tiny foreign objects or similar objects, increasing the probability of misjudgment and missed detection.

[0006] As the scale of energy storage power stations continues to expand, the number of battery compartments increases, and the amount of data that needs to be monitored is growing exponentially. Manual inspections require a lot of time and manpower, and the inspection cycle is long, making real-time monitoring impossible. For example, a large energy storage power station may contain dozens of battery compartments. A manual inspection may take hours or even days. During this period, if foreign objects enter and cause safety issues, it is difficult to detect and deal with them in time. For simple image recognition technology, due to its relatively simple image analysis algorithm, the processing speed is slow when faced with large amounts of image data. When real-time analysis of continuous video images is required, the existing processing unit may not be able to complete the analysis of all images in a short period of time, resulting in monitoring delays and failure to meet the needs of real-time monitoring.

[0007] The battery compartment environment is complex, with issues like light variations and equipment obstruction. The light intensity and angle inside the battery compartment change during different time periods, which can affect manual inspections and image recognition. For example, in the morning and evening when the light is dim, manual inspections can easily miss foreign objects, and the image recognition system may be unable to accurately identify foreign objects due to image quality degradation caused by insufficient light. Furthermore, the battery compartment is densely populated with equipment and contains numerous obstructions, such as battery racks and cable trays. During manual inspections, some obscured areas may not be fully inspected, and image recognition technology, due to a lack of complete information, struggles to accurately determine whether foreign objects are present, severely impacting recognition effectiveness. Summary of the Invention

[0008] The purpose of the present invention is to overcome the existing defects and provide a battery compartment foreign body identification method and system based on laser scanning and AI recognition.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0010] A battery compartment foreign object recognition method based on laser scanning and AI recognition includes:

[0011] Use the laser scanning module to scan and collect the three-dimensional point cloud data inside the battery compartment at a preset period;

[0012] Perform point cloud preprocessing on the collected original point cloud data and convert it into a unified cabin coordinate system;

[0013] An improved PointNet++ architecture is used for point cloud segmentation. A hypersphere model of the normal point cloud is constructed based on the support vector data description method, and abnormal states are identified using a dynamic threshold.

[0014] The personnel positioning data of the access control system is obtained and combined with multi-period scanning data to trace the location of foreign objects. The multi-source data alignment accuracy and trajectory estimation stability are synergistically improved through dynamic time warping and spatiotemporal filtering.

[0015] Furthermore, when the laser scanning module is used to scan and collect three-dimensional point cloud data inside the battery compartment at a preset period, the scanning path of the laser scanning module is planned based on an improved Hilbert curve pattern. In the densely populated area of ​​the battery rack, the path density is dynamically adjusted according to the local complexity of the point cloud reconstruction model. During the scanning process, the resolution is adaptively adjusted according to the accuracy requirements of the detection task.

[0016] Furthermore, the point cloud preprocessing of the collected original point cloud data and converting it into a unified cabin coordinate system includes:

[0017] Dynamic background modeling: Based on an adaptive Gaussian mixture model, the fixed structures and stationary equipment in the battery compartment are modeled. Long-standing and stable point cloud information is extracted as the baseline background. The probability distribution of each point is continuously updated to eliminate spurious differences caused by lighting changes or minor vibrations.

[0018] Outlier Filtering: A statistical filtering algorithm is used to identify and remove spatially isolated points. The average distance from each point to surrounding points in its local neighborhood is calculated and compared with the overall distribution. Points whose average distance to their neighborhood exceeds a preset threshold are filtered out.

[0019] Data augmentation: Introducing slight perturbations on the original point cloud to improve the model's adaptability to changes in posture and position;

[0020] UWB positioning data processing: A density-based outlier detection method is introduced, and a multi-scale judgment strategy is constructed by integrating a statistical filtering mechanism to handle interpolation errors caused by sudden offset points in UWB trajectories.

[0021] Coordinate system conversion: Based on the pre-calibrated cabin reference points, a global coordinate mapping relationship is established, and spatial alignment of point cloud data is achieved through rotation and translation transformation matrices. Point cloud data collected at different time periods and angles are fused under the same spatial reference.

[0022] Furthermore, the improved PointNet++ architecture is used for point cloud segmentation, a hypersphere model of the normal point cloud is constructed based on the support vector data description method, and abnormal states are identified by dynamic thresholds, including:

[0023] Point cloud segmentation: Using an improved PointNet++ architecture, a deformable convolutional layer is introduced in the encoder to accommodate the non-rigid and local deformation information of the cabin structure. A multi-scale feature extraction pipeline is also constructed to capture spatial details at different scales.

[0024] Foreign body classification: Based on the support vector data description method, a hypersphere model of the normal state point cloud is constructed, which tightly surrounds the normal data in the feature space. The abnormal state is judged by dynamic thresholds, and points outside the threshold are marked as potential foreign bodies.

[0025] Confidence fusion: Analyze the stability of detection results in the time series dimension, use Kalman filtering to smooth the output probability distribution of N consecutive frames, calculate the confidence of the presence of foreign objects in each frame, and automatically trigger an alarm mechanism when the accumulated abnormal probability over time exceeds the set value.

[0026] Furthermore, the improved PointNet++ architecture includes:

[0027] A deformable convolutional layer is introduced in the encoder to adapt to the non-rigid and local deformation information of the cabin structure. At the same time, a multi-scale feature extraction pipeline is constructed, using 1×1×1, 3×3×3, and 5×5×5 convolution kernels in parallel branches to capture spatial details at different scales.

[0028] The decoder stage introduces a dual attention mechanism of space and channel, dynamically adjusts the response strength of each dimension in the feature map, and performs weighted reinforcement on the loss function. The optimized loss function is expressed as

[0029]

[0030] Where S is the predicted point cloud output by the model, is a real point cloud;

[0031] x represents any point in the predicted point cloud S;

[0032] y represents the real point cloud Any point in

[0033] ||xy|| 2 Represents the square of the Euclidean distance between point x and point y, which is used to quantify the position error between the two points;

[0034] Represents the square distance from the predicted point x to the nearest point y in the real point cloud;

[0035] Represents the square distance from the real point y to the nearest point x in the predicted point cloud.

[0036] Furthermore, the personnel positioning data of the access control system is obtained, and the position of the foreign object is traced by combining the multi-period scanning data. Dynamic time warping and spatiotemporal filtering are used to collaboratively improve the multi-source data alignment accuracy and trajectory estimation stability, including:

[0037] Multi-source data collection: obtain personnel location data from the access control system and multi-period scanning data;

[0038] Dynamic Time Warping: Using an event-triggered dynamic time warping method, scanning nodes are triggered when people enter or leave a specific area, enabling nonlinear alignment of the trajectory time series and the laser scanning time series on the event point set. The trajectory node alignment process is constrained by minimizing the time difference cost function.

[0039] Probability density modeling: Based on the spatial interpolation algorithm unit, Kriging interpolation is introduced to perform probability density modeling on discrete point clouds, and an estimation model is constructed for the spatial position of the point cloud;

[0040] Joint Kalman Filtering: For the fusion of sensors with different accuracies, the observation model combines laser point cloud and UWB positioning information, sets point cloud error and UWB error, and performs weighted processing by assigning different observation error covariances. The process noise uses an adaptive covariance matrix to reflect the system uncertainty in positioning and motion prediction, and dynamically adjust the filter response capability.

[0041] Furthermore, in the multi-source data collection, a timestamp calibration method is used to establish a timestamp mapping function, and the historical data used includes personnel trajectory data and laser scanning data within at least the past week.

[0042] Another object of the present invention is to provide a battery compartment foreign object recognition system based on laser scanning and AI recognition, characterized in that the system includes:

[0043] The laser scanning module consists of a linear array laser radar mounted on a programmable rotating bracket, which is used to collect 3D point cloud data of the battery compartment interior at a preset cycle;

[0044] The data processing module is connected to the laser scanning module and is used to perform noise reduction on the original point cloud data and convert it into a unified cabin coordinate system.

[0045] The AI ​​recognition module is connected to the data processing module and contains a trained foreign object detection neural network. It is used to establish a baseline model based on the historical normal state point cloud dataset and identify foreign objects by comparing the real-time point cloud differences.

[0046] The location tracking module is used to obtain personnel location data from the access control system and to trace the location of foreign objects by combining multi-period scanning data;

[0047] The work order interface module is connected to the AI ​​recognition module and the location tracking module to generate a maintenance work order containing foreign object coordinates and trajectory traceability information.

[0048] Furthermore, the data processing module includes: a point cloud pre-processing unit for dynamic background modeling and outlier removal from the laser scanning data, while processing abnormal interference from the UWB positioning data;

[0049] The coordinate system conversion unit is used to convert the original point cloud data into a unified cabin coordinate system.

[0050] Furthermore, the AI ​​recognition module includes:

[0051] The point cloud segmentation unit uses an improved PointNet++ architecture, introducing deformable convolutional layers in the encoder to adapt to the non-rigid and local deformation information of the cabin structure, while also building a multi-scale feature extraction pipeline to capture spatial details at different scales.

[0052] The foreign body classification unit constructs a hypersphere model of the normal state point cloud based on the support vector data description method, tightly surrounds the normal data in the feature space, and uses a dynamic threshold to distinguish abnormal states. Points outside the threshold are marked as potential foreign bodies;

[0053] The confidence fusion unit is used to analyze the stability of detection results in the time series dimension. It uses Kalman filtering to smooth the output probability distribution of N consecutive frames, calculates the confidence of the presence of foreign objects in each frame, and automatically triggers the alarm mechanism when the time-accumulated abnormal probability is greater than the set value.

[0054] In combination with the above technical solutions, the present invention has the following beneficial effects compared with the prior art:

[0055] The present invention uses a laser scanning module to collect high-resolution three-dimensional point cloud data, accurately capturing subtle structures and object information. At the same time, the AI ​​recognition module uses advanced algorithms such as the improved PointNet++ architecture to more effectively extract complex features and improve the accuracy of foreign object recognition.

[0056] The present invention leverages the powerful data analysis and learning capabilities of AI algorithms to rapidly process and analyze large amounts of point cloud data acquired through laser scanning. It can identify foreign objects and determine their properties in real time, significantly improving data processing efficiency.

[0057] The present invention optimizes algorithms and system designs. For example, the dynamic background modeling unit eliminates false differences such as lighting changes. The AI ​​recognition module adopts multiple optimization methods to adapt to complex geometric shapes and environments. The position tracking module uses multiple technologies to achieve multi-source data fusion and precise spatiotemporal positioning. These enable the system to automatically adapt to complex and changing environments and ensure stable and reliable operation under various working conditions.

[0058] The present invention constructs a complete system architecture that includes laser scanning, data processing, AI recognition, location tracking and work order interface. The modules work together to achieve closed-loop management of the foreign object identification process. The location tracking module uses multi-source heterogeneous data fusion technology to accurately trace the location of foreign objects, improve the accuracy of the intersection judgment of abnormal areas and personnel behavior, and generate maintenance work orders containing foreign object coordinates and trajectory traceability information to facilitate subsequent processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0060] Figure 1 This is a flow chart of a battery compartment foreign object recognition method based on laser scanning and AI recognition provided by an embodiment of the present invention;

[0061] Figure 2 This is a schematic diagram of the structure of a battery compartment foreign object recognition system based on laser scanning and AI recognition provided by an embodiment of the present invention;

[0062] Figure 3 is a structural principle diagram of a data processing module provided by an embodiment of the present invention;

[0063] Figure 4 and Figure 5 This is a system signaling diagram of a battery compartment foreign object recognition system based on laser scanning and AI recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0065] Example 1:

[0066] like Figure 1 FIG. 1 is an embodiment of a battery compartment foreign object recognition method and system based on laser scanning and AI recognition provided by the present invention, comprising:

[0067] S1: Use the laser scanning module to scan and collect the three-dimensional point cloud data inside the battery compartment at a preset period;

[0068] S2: Perform point cloud preprocessing on the collected original point cloud data and convert it into a unified cabin coordinate system;

[0069] S3: Uses an improved PointNet++ architecture for point cloud segmentation, constructs a hypersphere model of the normal point cloud based on the support vector data description method, and uses dynamic thresholds to identify abnormal states;

[0070] S4: Obtain personnel positioning data from the access control system, combine it with multi-period scanning data to trace the location of foreign objects, and collaboratively improve the multi-source data alignment accuracy and trajectory estimation stability through dynamic time warping and spatiotemporal filtering.

[0071] As a preference, in step S1, the installation spacing of the laser scanning modules is designed to ensure complete coverage of the cabin space and to avoid blind spots to the greatest extent possible, and is configured using geometric relationships. Specifically, the linear array laser radar is arranged inside the cabin through a programmable rotating bracket, and the installation spacing satisfies the formula

[0072]

[0073] Where D represents the minimum horizontal spacing between adjacent lidars, H is the vertical height of the battery compartment, and θ is the lidar's vertical field of view. This formula, derived from the spatial projection relationship between the radar's conical field of view and the compartment height, ensures that the scanning sectors of adjacent devices intersect or slightly overlap horizontally, eliminating data loss caused by radar blind spots. Scan path planning is based on a modified Hilbert curve model, which offers excellent space-filling performance and path continuity, effectively improving coverage efficiency on complex surfaces within the compartment. In densely populated battery rack areas, the path density is dynamically adjusted based on the local complexity of the point cloud reconstruction model, implementing a high-frequency retracement strategy to enhance point cloud sampling accuracy in critical areas. During scanning, the system adaptively adjusts resolution based on the accuracy requirements of the inspection task. In its basic state, a standard point cloud model is generated using uniformly spaced sampling intervals of 5 mm. Once the detection algorithm identifies a suspected foreign object or unusual structure, the scanning module immediately increases the local resolution to 2 mm. This denser point cloud sampling enhances detail capture, providing higher-quality spatial data for subsequent point cloud processing and foreign object identification.

[0074] It should be noted that the improved Hilbert curve pattern planning is based on the space-filling characteristics of the traditional Hilbert curve and is optimized for the distribution of complex structures within the battery compartment. Specifically, by analyzing the local complexity of the point cloud reconstruction model, the scanning path density is dynamically increased in areas with dense battery racks, and a high-frequency retracement strategy is used to improve the point cloud sampling accuracy in key areas. At the same time, during the scanning process, the resolution is adaptively adjusted according to the accuracy requirements of the detection task. In the basic state, a standard point cloud model is generated with 5mm equal spacing. When suspected foreign objects or abnormal structures are detected, the local resolution is automatically increased to 2mm, and the ability to capture details is enhanced through dense point cloud sampling. After applying this improved method, the coverage efficiency of the complex structural surfaces inside the cabin can be effectively improved, and the data loss caused by the radar blind spot can be eliminated. This provides higher-quality spatial data support for subsequent point cloud processing and foreign object identification, significantly improving the detection accuracy of small foreign objects and adaptability in complex environments.

[0075] Preferably, in step S2, performing point cloud preprocessing on the collected original point cloud data and converting it into a unified cabin coordinate system includes:

[0076] Dynamic background modeling: Based on an adaptive Gaussian mixture model, the fixed structures and stationary equipment in the battery compartment are modeled. Long-standing and stable point cloud information is extracted as the baseline background. The probability distribution of each point is continuously updated to eliminate spurious differences caused by lighting changes or minor vibrations.

[0077] Outlier Filtering: A statistical filtering algorithm is used to identify and remove spatially isolated points. The average distance from each point to surrounding points in its local neighborhood is calculated and compared with the overall distribution. Points whose average distance to their neighborhood exceeds a preset threshold are filtered out.

[0078] Data augmentation: Introducing slight perturbations on the original point cloud to improve the model's adaptability to changes in posture and position;

[0079] UWB positioning data processing: A density-based outlier detection method is introduced, and a multi-scale judgment strategy is constructed by integrating a statistical filtering mechanism to handle interpolation errors caused by sudden offset points in UWB trajectories.

[0080] Coordinate system conversion: Based on the pre-calibrated cabin reference points, a global coordinate mapping relationship is established, and spatial alignment of point cloud data is achieved through rotation and translation transformation matrices. Point cloud data collected at different time periods and angles are fused under the same spatial reference.

[0081] Specifically, point cloud preprocessing of the collected raw point cloud data aims to improve the robustness and accuracy of the subsequent recognition model for the input data. First, the dynamic background modeling unit models the fixed structures and stationary equipment in the battery compartment based on the adaptive Gaussian mixture model, and extracts the long-standing and morphologically stable point cloud information as the reference background for the subsequent accurate detection of dynamic changes. The model gradually eliminates false differences caused by lighting changes or small vibrations by continuously updating the probability distribution state of each point. Secondly, the outlier filtering unit uses a statistical filtering algorithm to identify and eliminate isolated points in space. In the local neighborhood of each point, the average distance to the surrounding points is counted and compared with the overall distribution. The set filtering parameters include the neighborhood radius r = 50 mm and the standard deviation multiple k = 2.0. The filtering condition is that the average distance between a point and its neighborhood exceeds the global mean plus k times the standard deviation, that is,

[0082]

[0083] in is the average neighborhood distance of the i-th point, μ and σ are the average distance and standard deviation of all points, respectively. This strategy effectively removes discrete points caused by laser interference or reflection. Finally, the data augmentation unit improves the model's adaptability to pose and position changes by introducing slight perturbations to the original point cloud. The augmentation method includes random rotations of ±2° and random translations of ±5mm. This method expands the number of training samples while ensuring that the spatial structure remains largely unchanged, thereby improving the generalization ability of the neural network.

[0084] Point cloud preprocessing not only performs refined outlier removal on laser scanning data, but also extends to processing abnormal interference of UWB positioning data. In response to the interpolation error caused by sudden offset points in UWB trajectories, this paper introduces a density-based outlier detection method and integrates the statistical filtering mechanism to construct a multi-scale judgment strategy. In the processing of each UWB positioning point, the system analyzes the distribution of its neighborhood points within the unit time window, calculates the local density index and determines whether it is consistent with the surrounding trajectory structure. Density outliers will be regarded as interference items and filtered out. Density judgment is based on the calculation of the local outlier factor (LOF). The LOF value of each point is defined as the ratio of its reachable density to the average reachable density of the neighborhood. If it satisfies

[0085] LOF i >δ

[0086] Among them, LOF i is the local outlier factor of the i-th UWB positioning point, δ is the set threshold, and the system determines that the point is an outlier. On this basis, the statistical filtering strategy is further superimposed to construct a dynamic window on the time series. The mean μ and standard deviation σ of the trajectory point position change are statistically calculated within the sliding interval of length TTT, and the points that meet the |x i -μ|>βσ|, where x i is the current point's position coordinate, and β is the sensitivity coefficient. This dual filtering mechanism effectively eliminates positioning anomalies caused by transient occlusion, signal reflection, or device interference, ensuring the accuracy and robustness of subsequent trajectory data-based interpolation, alignment, and matching calculations in terms of continuity and stability.

[0087] The coordinate conversion step is implemented through the coordinate system conversion unit. Its core is to convert the raw point cloud data collected by the laser scanning module from the device coordinate system to a unified cabin coordinate system to eliminate errors caused by differences in the spatial positions of different sensors. This process establishes a global coordinate mapping relationship based on pre-calibrated cabin reference points. The point cloud data is spatially aligned through rotation and translation transformation matrices, ensuring that point cloud data collected at different times and angles are integrated under the same spatial reference. This provides a unified spatial reference framework for subsequent point cloud segmentation, foreign object identification, and position tracing, improving the spatiotemporal consistency and analytical accuracy of multi-source data.

[0088] Preferably, in step S3, using the improved PointNet++ architecture to perform point cloud segmentation specifically includes:

[0089] The point cloud segmentation network adopts an improved PointNet++ architecture, introduces a deformable convolution layer in the encoder to adapt to the non-rigid and local deformation information in the cabin structure, and constructs a multi-scale feature extraction path. In the parallel branches, 1×1×1, 3×3×3, and 5×5×5 convolution kernels are used to capture spatial details at different scales, effectively improving the ability to express complex geometric shapes. The decoder stage introduces a dual attention mechanism of space and channel, dynamically adjusts the response strength of each dimension in the feature map, and achieves significant enhancement of key areas. In the loss function design, the FocalLoss idea is used to weightedly strengthen the traditional Chamfer Distance. The optimized loss function is expressed as follows:

[0090]

[0091] Where S is the predicted point cloud output by the model, is the real point cloud, x represents any point in the predicted point cloud S; y represents the real point cloud Any point in ||xy|| 2 Represents the square of the Euclidean distance between point x and point y, which is used to quantify the position error between the two points; Represents the square distance from the predicted point x to the nearest point y in the real point cloud; Represents the squared distance from the actual point y to the nearest point x in the predicted point cloud. This function improves sensitivity to sparse foreign structures while ensuring the accuracy of point cloud geometric alignment.

[0092] As a preferred method, a hypersphere model of a normal state point cloud is constructed based on a support vector data description method, and abnormal states are discriminated by a dynamic threshold, including:

[0093] Based on the support vector data description method, a hypersphere model of the normal state point cloud is constructed, which tightly surrounds the normal data in the feature space and achieves high-precision discrimination of abnormal states through a dynamic threshold α = μ + 3σ, where μ is the average distance from the normal sample point to the center of the hypersphere and σ is its standard deviation. Points outside the threshold will be marked as potential foreign objects. The confidence fusion unit analyzes the stability of the detection results in the time series dimension, uses Kalman filtering to smooth the output probability distribution of N consecutive frames, and calculates the confidence p of the presence of foreign objects in each frame. i , the cumulative abnormal probability P = 1-Π(1-p i ) is greater than 0.95, the alarm mechanism is automatically triggered, and the instantaneous misjudgment is suppressed and the system stability is improved by fusing time domain information.

[0094] Preferably, in step S4, the personnel positioning data of the access control system is obtained, and the position of the foreign object is traced by combining the multi-period scanning data. The multi-source data alignment accuracy and trajectory estimation stability are synergistically improved through dynamic time warping and spatiotemporal filtering, including:

[0095] Fusion of time synchronization and spatial positioning: Using event-triggered dynamic time warping subunits, an event triggering mechanism is introduced on the basis of the traditional sliding window dynamic time warping algorithm, and key behavior nodes are used to enhance the accuracy of time registration. When a clearly marked event such as a person entering or leaving a specific area occurs, the system immediately initiates a forced synchronization strategy, breaking the passive sliding mode under a fixed time step, and achieving precise alignment of trajectory data and laser scanning data at the point of event occurrence, thereby significantly reducing the timing lag error caused by delays. In scenarios where there are differences in the sampling frequencies of multi-source data, trajectory data is recorded at a frequency of 1Hz, while the sampling frequency of laser scanning data is 0.2Hz. The time synchronization processing adopts a sliding window mechanism. The window length for each processing is set to 5 seconds, and the window slides on the time axis at a step size of 1 second. The optimal matching path is constructed in each window, so that data points with inconsistent sampling intervals establish the minimum cost matching path in the time domain. The cost function is based on the time difference. Based on Indicates the timestamp of the i-th trajectory data point, in seconds, corresponding to the position data recorded by the UWB or other positioning system at the i-th moment. Indicates the timestamp of the j-th laser scanning data frame, in seconds, corresponding to a point cloud acquisition completed by the laser radar at the j-th moment; Δt ij This represents the absolute difference in time between the i-th track point and the j-th scan frame. It is used to calculate the time cost of matching track points and scan frames during dynamic time warping. Dynamic planning of the alignment path is performed with the goal of minimizing the total time difference, ensuring that the data at the event trigger point falls stably within the time window corresponding to the adjacent scan frames, thereby improving the accuracy of track and scan fusion.

[0096] The timestamp calibration subunit addresses the inconsistency between the acquisition frequency and timestamp accuracy of laser scanning data and personnel trajectory data by constructing a time mapping function based on regression learning. It uses historical trajectory data to train a prediction model to accurately estimate the position that was not collected at the scanning moment. Under the premise that the UWB positioning system generates trajectory points at a frequency of 1Hz, this subunit extracts the time series information from continuous trajectory segments, constructs the functional relationship between time and position, and uses supervised learning to train a multivariate regression model to accurately estimate the position of a given scanning time point t. s Predict the corresponding personnel location The mapping function is defined as

[0097]

[0098] Where (t i ,P i ) represents the i-th moment in the historical trajectory and its corresponding position, is the scanning time t s The corresponding predicted position, The training sample sequence provided for the model, where t i is the timestamp of the i-th historical moment, P i is the actual position of the person at that moment, and n is the number of trajectory points used for modeling. This data is used to train the function f, enabling it to capture the time series characteristics of person movement. By selecting a regression model with good generalization capabilities, such as support vector regression, LSTM regression network, or polynomial regression for training, the system improves the accuracy of temporal registration without introducing hard synchronization of multiple source devices. This ensures that the prediction results are synchronized and consistent with the laser scanning point cloud data in the spatial dimension, thereby enhancing the accuracy of matching trajectory and environmental information.

[0099] The hierarchical Kalman filter subunit achieves systematic suppression of UWB positioning errors and high-precision modeling of continuous trajectories by constructing a multi-layer filtering architecture. In the first layer, the high spatial resolution of the laser point cloud data is used to correct the high-frequency noise in the UWB original trajectory data. The spatial accuracy of the lidar in the cabin structure scanning reaches ±3 mm. This accuracy is input into the Kalman filter as the measurement update, effectively enhancing the confidence of the position observation and enabling the UWB measurement point to be smoothly corrected under the constraints of the local environment. In the second layer, the state vector is introduced.

[0100]

[0101] Describe the position and velocity of a person in three-dimensional space through the state transfer equation

[0102] x k+1 =F k x k +w k

[0103] The motion evolution of the simulated personnel at continuous time steps, where F k is the state transfer matrix, w k is the process noise, reflecting the system dynamic uncertainty. The observation update adopts the joint measurement model

[0104] z k =H k x k +v k

[0105] where z k is the multi-source measurement value after fusion, H kis the observation matrix, v k In the prediction-update filtering loop, the system dynamic model is used to estimate the occupant's position at the next moment. The measurement model incorporates point cloud correction data and UWB positioning values ​​for error feedback adjustment. By jointly optimizing spatial position and velocity information, continuous modeling of discrete trajectory points is achieved, significantly reducing positioning jumps and drift caused by low-precision sensors. The resulting smooth trajectory provides a stable and reliable data foundation for foreign object detection and trajectory tracing in both time and space.

[0106] The adaptive weight fusion subunit improves the intelligence of multi-source data fusion by introducing an environmental perception mechanism. The system adjusts the fusion weight of laser point cloud data and UWB positioning data in real time according to the complexity of the internal structure of the battery compartment. In view of the spatial heterogeneity distribution, the Bayesian inference model is used to construct the posterior probability distribution of the position estimate at each moment.

[0107] P(P t |Z t )∝P(Z t |P t )P(P t )

[0108] Where P(P t |Z t ) is the posterior probability of the fused position information, P(Z t |P t ) represents the likelihood of the observation value Zt under the current position Pt, P(P t ) is the prior position distribution. In the densely populated area of ​​battery racks, the boundary contours and obstacle distribution provided by the laser point cloud have a high confidence level, and the system automatically increases their weight by dynamically adjusting the parameter λ. lidar ,λ uwb , introducing a weighted formula in fusion position estimation

[0109]

[0110] where λ lidar +λ uwb =1, is the estimated position of the person after fusion, P t lidar is the position estimated based on the laser point cloud data, P t uwb To achieve optimal integration of heterogeneous information based on the estimated location of UWB positioning data, the system also generates a probability heat map in three-dimensional space, expressing the spatial residence probability of people at different times, which serves as input for subsequent interpolation optimization.

[0111] The spatiotemporal database enhancement subunit supplements the environment state variables corresponding to the scanning time on the basis of the original trajectory database and defines the triple (T i ,P i ,E i ), which respectively represent the personnel position, trajectory point and environmental status at the i-th moment. By establishing a "trajectory-point cloud-environment" association map, the contextual association between foreign object detection results and personnel activity paths is realized, which significantly improves the accuracy of the intersection judgment of abnormal areas and personnel behaviors.

[0112] The reverse interpolation optimization subunit constructs a spatial probability density field based on the Kriging interpolation method. Construct an estimation function at

[0113]

[0114] in is the probability density of the point to be estimated, w i is the weight calculated by the spatial covariance matrix, f(x i ) at a known point x i The observation value or position probability density at n is the number of known points used for interpolation. For multi-period trajectory and scanning data, the system uses dynamic time warping to align the trajectory sequence, extract possible related spatiotemporal segments, and generate multiple sets of residual position candidate points based on the estimated probability density function through Monte Carlo simulation to form the final position search area. Where P(p j )>∈, R is the set of possible locations where the remnant may appear, p j is the jth candidate location point, P(p j ) is the candidate position p j The probability of occurrence of ∈ is the set probability threshold. Only candidate points with a value higher than this value are retained to form the search area. m is the number of candidate position points that meet the conditions, which effectively compresses the search space and improves the retrieval efficiency and the operability of residual foreign matter positioning.

[0115] The dynamic time warping unit handles the problem of inconsistent sampling frequencies between laser scanning and personnel trajectory. It uses a dynamic time warping method based on event triggering mechanism to trigger scanning nodes when personnel enter and exit a specific area, so that the trajectory time series T uwb ={t1,t2,...,t n} and laser scanning time series T scan ={s1,s2,...,s m}Achieve nonlinear alignment on the event point set {ek} by minimizing the time difference cost function Constraint trajectory node alignment process, is the personnel trajectory timestamp corresponding to the kth event point, sjk is the laser scanning timestamp corresponding to the kth event point, and D is the total nonlinear alignment error between time series, which is used to measure the degree of synchronization. The spatial interpolation algorithm unit introduces Kriging interpolation to model the probability density of discrete point clouds and constructs an estimated value calculation formula for the spatial position x0 of the point cloud:

[0116]

[0117] where z(x i ) is the i-th known spatial position x i The observed value of λ i The optimal interpolation weight is calculated based on the covariance function, which satisfies the unbiasedness and minimum variance constraints. The joint Kalman filter part introduces motion state modeling on this basis, assuming that the system state vector is X = [x, y, z, v x ,v y ,v z ], the system state transfer equation is X k+1 =AX k +w k , the observation model is Z k =HX k +v k , where A is the state transfer matrix, H is the observation matrix, wk and vk are process noise and observation noise respectively. For different precision sensor fusion, the observation model combines laser point cloud and UWB positioning information. The point cloud error is set to ±3mm and the UWB error is ±100mm. The system performs weighted processing by assigning different observation error covariances. The process noise adopts the adaptive covariance matrix

[0118]

[0119] where σ x =σ y =σ z =10mm is the standard deviation of the process noise of the position state in each axis, reflecting the uncertainty of position prediction, σ v =5mm / s velocity state in each axis. This reflects the disturbance range of the velocity estimate and dynamically adjusts the filter response capability. This filter structure, as the second layer of the hierarchical Kalman filter, forms a cascade with the first layer's point cloud correction results, ultimately outputting a high-confidence continuous trajectory estimate to support reverse reasoning and localization and traceability of the location of leftover foreign objects.

[0120] The timestamp mapping function established by the timestamp calibration subunit of the position tracking module uses historical data including personnel trajectory data and laser scanning data within at least the past week.

[0121] Example 2:

[0122] like Figure 2 As shown, an embodiment of the present invention provides a battery compartment foreign object recognition system based on laser scanning and AI recognition, including:

[0123] The laser scanning module consists of a linear array laser radar mounted on a programmable rotating bracket, which is used to collect 3D point cloud data of the battery compartment interior at a preset cycle;

[0124] The data processing module is connected to the laser scanning module and is used to perform noise reduction on the original point cloud data and convert it into a unified cabin coordinate system.

[0125] The AI ​​recognition module is connected to the data processing module and contains a trained foreign object detection neural network. It is used to establish a baseline model based on the historical normal state point cloud dataset and identify foreign objects by comparing the real-time point cloud differences.

[0126] The location tracking module is used to obtain personnel location data from the access control system and to trace the location of foreign objects by combining multi-period scanning data;

[0127] The work order interface module is connected to the AI ​​recognition module and the location tracking module to generate a maintenance work order containing foreign object coordinates and trajectory traceability information.

[0128] Preferably, the data processing module in the embodiment of the present invention includes: a point cloud pre-processing unit for dynamic background modeling, removing outliers from laser scanning data, and processing abnormal interference from UWB positioning data;

[0129] The coordinate system conversion unit is used to convert the original point cloud data into a unified cabin coordinate system.

[0130] Preferably, the AI ​​recognition module in the embodiment of the present invention includes:

[0131] The point cloud segmentation unit uses an improved PointNet++ architecture, introducing deformable convolutional layers in the encoder to adapt to the non-rigid and local deformation information of the cabin structure, while also building a multi-scale feature extraction pipeline to capture spatial details at different scales.

[0132] The foreign body classification unit constructs a hypersphere model of the normal state point cloud based on the support vector data description method, tightly surrounds the normal data in the feature space, and uses a dynamic threshold to distinguish abnormal states. Points outside the threshold are marked as potential foreign bodies;

[0133] The confidence fusion unit is used to analyze the stability of detection results in the time series dimension. It uses Kalman filtering to smooth the output probability distribution of N consecutive frames, calculates the confidence of the presence of foreign objects in each frame, and automatically triggers the alarm mechanism when the time-accumulated abnormal probability is greater than the set value.

[0134] The various modules in the embodiment of the present invention form an organic whole through reasonable connection methods and collaborative working mechanisms. The laser scanning module is responsible for collecting high-precision three-dimensional point cloud data, providing a data basis for subsequent analysis; the data processing module pre-processes and converts the coordinate system of the original point cloud data so that it can be effectively processed by the AI ​​recognition module; the AI ​​recognition module uses advanced neural network algorithms to accurately identify foreign objects; the position tracking module achieves accurate tracing of the position of foreign objects through multi-source heterogeneous data fusion; the work order interface module converts the identification and tracing results into an operational maintenance work order, realizing closed-loop management of the entire foreign object identification process. This system architecture design gives full play to the advantages of each module, improves the overall performance and reliability of the system, and has obvious differences and advantages compared with existing technologies.

[0135] Key algorithms: In the AI ​​recognition module, the point cloud segmentation network adopts an improved PointNet++ architecture, especially the addition of deformable convolution layers in the encoder part, as well as the introduction of multi-scale feature extraction structure and spatial-channel dual attention weights. Compared with traditional point cloud segmentation algorithms, it can more effectively extract complex features in point cloud data and improve the accuracy and efficiency of foreign object recognition. The foreign object classification model constructs a normal state hypersphere based on the support vector data description (SVDD) and sets a dynamic threshold method. Compared with traditional classification algorithms, it can better adapt to the complex environment and data distribution in the battery compartment of the energy storage power station, and accurately classify and judge foreign objects. In the position tracking module, the event-triggered dynamic time warping sub-unit method adopted by the dynamic time warping unit and the spatiotemporal joint Kalman filter algorithm can effectively solve the problem of time synchronization and spatial positioning accuracy fusion when processing multi-source heterogeneous data, thereby improving the accuracy of foreign object location tracing.

[0136] Data processing flow: Starting from laser scanning data acquisition, through noise reduction and coordinate system conversion in the data processing module, to foreign object identification in the AI ​​recognition module, and then to position tracing in the position tracking module, and finally to the work order interface module to generate a maintenance work order, the entire data processing flow is unique and advantageous. Through a series of preprocessing, feature extraction, model training, and data fusion operations, efficient conversion from raw data to final foreign object identification and processing results is achieved. This data processing flow not only improves the performance and accuracy of the system, but also has good scalability and adaptability. It can meet the foreign object identification needs of battery compartments of energy storage power stations of different sizes and complexities. It is the key difference between this invention and the prior art and should be protected by patents.

[0137] It should be understood that, although the various steps in the flow charts of the various embodiments of the present invention are shown in sequence according to the instructions of the arrows, these steps are not necessarily performed in sequence according to the order indicated by the arrows. Unless otherwise specified in the present invention, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the various embodiments may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0138] Those skilled in the art will appreciate that all or part of the processes in the above-described embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-described methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0139] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A battery compartment foreign object recognition method based on laser scanning and AI recognition, characterized in that: The method comprises: Use the laser scanning module to scan and collect the three-dimensional point cloud data inside the battery compartment at a preset period; Perform point cloud preprocessing on the collected original point cloud data and convert it into a unified cabin coordinate system; An improved PointNet++ architecture is used for point cloud segmentation. A hypersphere model of the normal point cloud is constructed based on the support vector data description method, and abnormal states are identified using a dynamic threshold. The personnel positioning data of the access control system is obtained and combined with multi-period scanning data to trace the location of foreign objects. The multi-source data alignment accuracy and trajectory estimation stability are synergistically improved through dynamic time warping and spatiotemporal filtering.

2. The battery compartment foreign object identification method based on laser scanning and AI recognition according to claim 1 is characterized in that: In the process of using the laser scanning module to scan and collect the three-dimensional point cloud data inside the battery compartment at a preset period, the scanning path of the laser scanning module is planned based on an improved Hilbert curve pattern. In the densely populated area of ​​the battery rack, the path density is dynamically adjusted according to the local complexity of the point cloud reconstruction model. During the scanning process, the resolution is adaptively adjusted according to the accuracy requirements of the detection task.

3. The battery compartment foreign object identification method based on laser scanning and AI recognition according to claim 1 is characterized in that: The point cloud preprocessing of the collected original point cloud data and converting it into a unified cabin coordinate system includes: Dynamic background modeling: Based on an adaptive Gaussian mixture model, the fixed structures and stationary equipment in the battery compartment are modeled. Long-standing and stable point cloud information is extracted as the baseline background. The probability distribution of each point is continuously updated to eliminate spurious differences caused by lighting changes or minor vibrations. Outlier Filtering: A statistical filtering algorithm is used to identify and remove spatially isolated points. The average distance from each point to surrounding points in its local neighborhood is calculated and compared with the overall distribution. Points whose average distance to their neighborhood exceeds a preset threshold are filtered out. Data augmentation: Introducing slight perturbations on the original point cloud to improve the model's adaptability to changes in posture and position; UWB positioning data processing: A density-based outlier detection method is introduced, and a multi-scale judgment strategy is constructed by integrating a statistical filtering mechanism to handle interpolation errors caused by sudden offset points in UWB trajectories. Coordinate system conversion: Based on the pre-calibrated cabin reference points, a global coordinate mapping relationship is established, and spatial alignment of point cloud data is achieved through rotation and translation transformation matrices. Point cloud data collected at different time periods and angles are fused under the same spatial reference.

4. The battery compartment foreign object identification method based on laser scanning and AI recognition according to claim 1 is characterized in that: The improved PointNet++ architecture is used for point cloud segmentation, a hypersphere model of the normal point cloud is constructed based on the support vector data description method, and abnormal states are identified by dynamic thresholds, including: Point cloud segmentation: Using an improved PointNet++ architecture, a deformable convolutional layer is introduced in the encoder to accommodate the non-rigid and local deformation information of the cabin structure. A multi-scale feature extraction pipeline is also constructed to capture spatial details at different scales. Foreign body classification: Based on the support vector data description method, a hypersphere model of the normal state point cloud is constructed, which tightly surrounds the normal data in the feature space. The abnormal state is judged by dynamic thresholds, and points outside the threshold are marked as potential foreign bodies. Confidence fusion: Analyze the stability of detection results in the time series dimension, use Kalman filtering to smooth the output probability distribution of N consecutive frames, calculate the confidence of the presence of foreign objects in each frame, and automatically trigger an alarm mechanism when the accumulated abnormal probability over time exceeds the set value.

5. The battery compartment foreign object recognition method based on laser scanning and AI recognition according to claim 4 is characterized in that: The improved PointNet++ architecture includes: A deformable convolutional layer is introduced in the encoder to adapt to the non-rigid and local deformation information of the cabin structure. At the same time, a multi-scale feature extraction pipeline is constructed, using 1×1×1, 3×3×3, and 5×5×5 convolution kernels in parallel branches to capture spatial details at different scales. The decoder stage introduces a dual attention mechanism of space and channel, dynamically adjusts the response strength of each dimension in the feature map, and performs weighted reinforcement on the loss function. The optimized loss function is expressed as Where S is the predicted point cloud output by the model, is a real point cloud; x represents any point in the predicted point cloud S; y represents the real point cloud Any point in ||xy|| 2 Represents the square of the Euclidean distance between point x and point y, which is used to quantify the position error between the two points; Represents the square distance from the predicted point x to the nearest point y in the real point cloud; Represents the square distance from the real point y to the nearest point x in the predicted point cloud.

6. The battery compartment foreign object identification method based on laser scanning and AI recognition according to claim 1 is characterized in that: The method of obtaining personnel positioning data from the access control system and tracing the location of foreign objects by combining multi-period scanning data, and collaboratively improving the multi-source data alignment accuracy and trajectory estimation stability through dynamic time warping and spatiotemporal filtering, includes: Multi-source data collection: obtain personnel location data from the access control system and multi-period scanning data; Dynamic Time Warping: Using an event-triggered dynamic time warping method, scanning nodes are triggered when people enter or leave a specific area, enabling nonlinear alignment of the trajectory time series and the laser scanning time series on the event point set. The trajectory node alignment process is constrained by minimizing the time difference cost function. Probability density modeling: Based on the spatial interpolation algorithm unit, Kriging interpolation is introduced to perform probability density modeling on discrete point clouds, and an estimation model is constructed for the spatial position of the point cloud; Joint Kalman Filtering: For the fusion of sensors with different accuracies, the observation model combines laser point cloud and UWB positioning information, sets point cloud error and UWB error, and performs weighted processing by assigning different observation error covariances. The process noise uses an adaptive covariance matrix to reflect the system uncertainty in positioning and motion prediction, and dynamically adjust the filter response capability.

7. The battery compartment foreign object identification method based on laser scanning and AI recognition according to claim 6 is characterized in that: In the multi-source data collection, a timestamp calibration method is used to establish a timestamp mapping function, and the historical data used includes personnel trajectory data and laser scanning data within at least the past week.

8. A battery compartment foreign object recognition system based on laser scanning and AI recognition, characterized in that: The system comprises: The laser scanning module consists of a linear array laser radar mounted on a programmable rotating bracket, which is used to collect 3D point cloud data of the battery compartment interior at a preset cycle; The data processing module is connected to the laser scanning module and is used to perform noise reduction on the original point cloud data and convert it into a unified cabin coordinate system. The AI ​​recognition module is connected to the data processing module and contains a trained foreign object detection neural network. It is used to establish a baseline model based on the historical normal state point cloud dataset and identify foreign objects by comparing the real-time point cloud differences. The location tracking module is used to obtain personnel location data from the access control system and to trace the location of foreign objects by combining multi-period scanning data; The work order interface module is connected to the AI ​​recognition module and the location tracking module to generate a maintenance work order containing foreign object coordinates and trajectory traceability information.

9. The battery compartment foreign object recognition system based on laser scanning and AI recognition according to claim 8 is characterized in that: The data processing module includes: a point cloud pre-processing unit for dynamic background modeling, removing outliers from laser scanning data, and processing abnormal interference from UWB positioning data; The coordinate system conversion unit is used to convert the original point cloud data into a unified cabin coordinate system.

10. The battery compartment foreign object recognition system based on laser scanning and AI recognition according to claim 8, characterized in that: The AI ​​recognition module includes: The point cloud segmentation unit uses an improved PointNet++ architecture, introducing deformable convolutional layers in the encoder to adapt to the non-rigid and local deformation information of the cabin structure, while also building a multi-scale feature extraction pipeline to capture spatial details at different scales. The foreign body classification unit constructs a hypersphere model of the normal state point cloud based on the support vector data description method, tightly surrounds the normal data in the feature space, and uses a dynamic threshold to distinguish abnormal states. Points outside the threshold are marked as potential foreign bodies; The confidence fusion unit is used to analyze the stability of detection results in the time series dimension. It uses Kalman filtering to smooth the output probability distribution of N consecutive frames, calculates the confidence of the presence of foreign objects in each frame, and automatically triggers the alarm mechanism when the time-accumulated abnormal probability is greater than the set value.

Images