A grid-based inventory and data synchronization method based on artificial intelligence

Abstract

The application discloses a kind of based on artificial intelligence's grid check and data synchronization method, comprising the following steps: to be checked area is carried out grid space division, obtains grid position coding set, generates adaptive grid division result;Image data and auxiliary check data are collected;Image data and corresponding grid position coding in adaptive grid division result are bound;Using improved RepVGG model, the structured image input sample set is identified, and the output asset identification result set;Asset identification result and auxiliary check data are fused and checked, and the fusion check result is obtained;The fusion check result is cached in edge device local, and the synchronization check data set is obtained;Synchronization check data set is aggregated, and global check result is generated and visual check report is output.The application uses adaptive grid division and improved RepVGG, realizes grid check and data synchronization.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and data processing technology, and in particular relates to an artificial intelligence-based grid-based inventory and data synchronization method. Background Technology

[0002] Current inventory processes often involve manually scanning barcodes or using RFID technology to read tags for recording, supplemented by image acquisition equipment to generate inventory lists. These methods typically rely on fixed area division, dividing the area to be inventoried into several static grid units, and then identifying assets within each grid. Due to the lack of dynamic adjustment, changes in asset distribution density can easily lead to unreasonable division, insufficient coverage, or duplicate counting. Existing image recognition methods are mostly general models that only classify or detect image features, without incorporating location coding to establish a correspondence between images and grid positions, making it difficult to reflect the precise spatial distribution of assets.

[0003] Regarding data transmission, existing inventory systems mostly employ centralized uploading, which is prone to transmission interruptions or data loss when the network environment is unstable. Furthermore, the lack of mechanisms for resuming interrupted uploads and deduplication / merging leads to incomplete or redundant results. Simultaneously, the lack of version control for inventory data makes traceability and verification difficult. As for result display, current outputs are mostly in list or static report format, lacking visualization of missing data distribution, anomaly location, and asset status, thus failing to provide real-time support for management.

[0004] Therefore, how to provide an AI-based grid-based inventory and data synchronization method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a grid-based inventory and data synchronization method based on artificial intelligence.

[0006] The technical solution of the present invention is as follows:

[0007] An AI-based grid-based inventory and data synchronization method includes the following steps:

[0008] The area to be investigated is divided into grids to obtain a set of grid location codes. Based on historical investigation records and current asset distribution density, an adaptive grid division result is generated.

[0009] Image data and auxiliary investigation data are collected by edge acquisition devices deployed in each adaptive grid cell;

[0010] The image data is bound to the corresponding grid position codes in the adaptive grid partitioning results to construct a set of structured image input samples;

[0011] An improved RepVGG model is used to identify a set of structured image input samples. A location encoding fusion path is introduced during the training phase, and all branches are reparameterized during the deployment phase to output a set of asset recognition results.

[0012] The asset identification results set and auxiliary inventory data are merged and verified, and the confidence weights are configured based on the historical abnormal pattern library to obtain the merged inventory results;

[0013] The merged investigation results are cached locally on the edge device, and a data synchronization operation is performed based on the current network status to obtain a synchronized investigation data set;

[0014] The system aggregates the synchronous investigation data set to generate global investigation results and outputs a visual investigation report.

[0015] Furthermore, the specific method for dividing the area to be investigated into a grid to obtain a set of grid location codes, and generating adaptive grid division results based on historical investigation records and current asset distribution density includes:

[0016] Obtain the physical layout data of the area to be investigated, perform initial grid division of the area based on the physical layout data, and assign a unique grid location code to each grid cell to form a set of grid location codes;

[0017] Generate an asset identification heat map and an asset distribution density map for the current period based on historical inventory records;

[0018] Based on the asset identification heat map and asset distribution density map, the probability area of ​​asset clustering in the next period is predicted, and the boundary range of each grid cell in the initial grid set is dynamically adjusted to generate the adjusted grid set.

[0019] Each grid cell in the adjusted grid set is reassigned a grid position code to form a new set of grid position codes. The adjusted grid set is then combined with the new set of grid position codes to form an adaptive grid partitioning result.

[0020] Furthermore, the edge acquisition device includes a camera, a handheld terminal, and an unmanned mobile platform; the auxiliary investigation data includes RFID signals, barcode data, and sensor data.

[0021] Furthermore, the specific method for binding image data with the corresponding grid position codes in the adaptive grid partitioning results to construct a structured image input sample set includes:

[0022] Based on the deployment coordinates and shooting direction of the edge acquisition device, establish the correspondence between image data and grid position codes;

[0023] Each image data is bound to the corresponding grid position code in the adaptive grid partitioning result to form an image-position pair set;

[0024] Each pair of elements in the image-location pair set is encapsulated according to a unified data structure to form a structured image input sample unit;

[0025] All structured image input sample units are organized and aggregated according to the grid position encoding order to form a structured image input sample set.

[0026] Furthermore, the specific method for using the improved RepVGG model to identify a set of structured image input samples, introducing a location encoding fusion path during the training phase, reparameterizing all branches during the deployment phase, and outputting a set of asset recognition results includes:

[0027] Each sample unit in the structured image input sample set is input into the improved RepVGG model, wherein the sample unit includes an image tensor and a position encoding vector;

[0028] During the training phase of the improved RepVGG model, a 3x3 convolution branch, a 1x1 convolution branch, an identity mapping branch, and a positional encoding fusion branch are used to extract features from the image tensor and the positional encoding vector to obtain a fused feature map.

[0029] The fused feature maps are input to the output layer to generate prediction results, and the loss function is calculated by combining the labeled results.

[0030] The parameters of the improved RepVGG model are updated based on the loss function until the training converges, resulting in an improved RepVGG model with location awareness.

[0031] After training convergence, the 3x3 convolution branch, 1x1 convolution branch, identity mapping branch and position encoding fusion branch are subjected to structural reparameterization to obtain the gridded clearing and identification model.

[0032] The gridded inventory and identification model is invoked in the edge device to infer the structured image input sample set and output the asset identification result set of all grid cells.

[0033] Furthermore, in the training phase of the improved RepVGG model, the specific method for extracting features from the image tensor and positional encoding vector using a 3x3 convolution branch, a 1x1 convolution branch, an identity mapping branch, and a positional encoding fusion branch to obtain the fused feature map includes:

[0034] The image tensor is input into a 3x3 convolution branch to extract local spatial perception features of the image, resulting in a local perception feature map.

[0035] The image tensor is input into the one-to-one convolution branch, and convolution operation and batch normalization are performed sequentially to obtain the channel compressed feature map.

[0036] The image tensor is input into the identity mapping branch, and after batch normalization, the residual information feature map is obtained.

[0037] The position encoding vector is input into the position encoding fusion branch, and a fully connected operation is performed on the position encoding vector to map the position encoding vector into a channel vector;

[0038] Broadcast the channel vectors in the spatial dimension to generate a three-dimensional position tensor. Then, use a one-to-one convolution to fuse the three-dimensional position tensor and the image tensor channel by channel to output a spatially enhanced feature map.

[0039] The local perception feature map, channel compression feature map, residual information feature map, and spatial location enhancement feature map are summed element-wise on the same spatial and channel dimensions to obtain a fused feature map.

[0040] Furthermore, the specific method for performing structural reparameterization processing on the 3x3 convolution branch, 1x1 convolution branch, identity mapping branch, and position encoding fusion branch after training convergence to obtain the gridded inspection and identification model includes:

[0041] By sequentially reading the convolutional weight parameters and batch normalization parameters corresponding to each branch in the improved RepVGG model, the original parameter set of the branch is obtained.

[0042] Based on the original parameter set of each branch, the convolution weights of each branch are combined with the batch normalization parameters to obtain the equivalent convolution weight and bias set.

[0043] Based on the equivalent convolution weights and bias set, zero-padded expansion operation is performed on the equivalent convolution weights of the one-to-one convolution branch to obtain the one-to-one convolution branch expanded convolution kernel;

[0044] Zero-padded expansion is performed on the equivalent convolution weights of the position encoding fusion branch to obtain the expanded convolution kernel of the position encoding fusion branch;

[0045] The equivalent convolution weights of the identity mapping branch are converted into a 3x3 convolution kernel form to obtain the extended convolution kernel of the identity mapping branch.

[0046] The equivalent convolution weights of the 3x3 convolution branch are directly used as the convolution kernels of the 3x3 convolution branch. The extended convolution kernels of the 1x1 convolution branch, the extended convolution kernels of the position encoding fusion branch, the extended convolution kernels of the identity mapping branch, and the extended convolution kernels of the 3x3 convolution branch are combined to form an extended convolution kernel set.

[0047] Based on the extended set of convolutional kernels and the equivalent set of convolutional weights and biases, the convolutional kernels of the four branches are summed element by element to obtain a single 3x3 convolutional kernel. The biases of each branch are numerically summed to obtain a single bias.

[0048] By replacing all the multi-branch structures of the corresponding layer in the improved RepVGG model with a single 3x3 convolution kernel and a single bias, a network topology with a single convolution structure is obtained.

[0049] The network topology of a single convolutional structure is saved in a format that can be deployed on edge devices to form a gridded investigation and identification model.

[0050] Furthermore, the specific method for fusing and verifying the asset identification result set and auxiliary inventory data, and obtaining the fused inventory result based on the historical anomaly pattern library and the assigned confidence weights, includes:

[0051] Each asset identification result in the asset identification result set is matched one by one with the corresponding record in the auxiliary inventory data set to generate a set of matching pairs;

[0052] Based on the historical anomaly pattern library, reliability statistics of multi-source data during the investigation process are extracted, and a confidence weight vector is formed according to the accuracy and anomaly rate distribution of various types of data in historical records.

[0053] For each matching pair in the matching pair set, the asset identification result, RFID signal, barcode data and sensor data are numerically processed according to the confidence weight vector to obtain the data consistency score of the matching pair;

[0054] The data consistency score is compared with a preset threshold. When the data consistency score is less than the preset threshold, a corresponding check anomaly marker is generated and assigned a value of one. Otherwise, it is assigned a value of zero, forming a check anomaly marker set.

[0055] The asset identification result set, the auxiliary inventory data set, and the inventory anomaly marker set are merged to form a merged inventory result.

[0056] Furthermore, the specific method for caching the fused investigation results locally on the edge device and performing edge-cloud collaborative data synchronization operations based on the current network status to obtain the synchronized investigation data set includes:

[0057] The merged investigation results are stored in the local cache of the edge device, forming a local cache set;

[0058] A synchronization strategy function is constructed based on the current network status of the edge device. When the calculation result of the synchronization strategy function is higher than the preset threshold, the full transmission mode is triggered. When the calculation result of the synchronization strategy function is lower than the preset threshold, the breakpoint resume transmission strategy is executed.

[0059] When executing the breakpoint resume strategy, the local cache set is divided into blocks and numbered according to the generated timestamp and the corresponding grid location code, forming multiple block sets;

[0060] During the process of executing block numbering and transmitting to the cloud, the block set is deduplicated and merged to generate a deduplicated and merged result set;

[0061] After the transmission is completed, a version control mechanism is established for the deduplication and merging result set according to the synchronization order, forming a deduplication and merging result set with version identifiers;

[0062] The deduplication and merging result set with version identifiers is uploaded to the cloud platform to obtain a synchronized data set.

[0063] Furthermore, the specific methods for aggregating the synchronous inventory data set, generating global inventory results, and outputting a visualized inventory report include:

[0064] The synchronous inventory data set is aggregated according to the grid location code to create a grid location index table;

[0065] Based on the grid location index table, data entries belonging to the same grid location code are merged to obtain a preliminary aggregation result set;

[0066] Multi-dimensional statistical processing is performed on the preliminary aggregated result set to calculate the global investigation results;

[0067] Based on the overall inventory results, a visual inventory report is generated, which includes a missing distribution map, an anomaly grid location map, and an asset status map.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] This invention introduces an adaptive grid partitioning mechanism into the area to be inventoried, enabling dynamic adjustment based on asset distribution density and historical inventory records. This allows grid cells to better match the actual asset distribution. In this process, each grid cell is assigned a unique grid location code and bound to image data, thus establishing a correspondence between the image and the spatial index. This solves the problems of insufficient recognition coverage or duplicate counting caused by fixed partitioning methods in existing technologies, improving the spatial positioning accuracy of the inventory results.

[0070] This invention utilizes an improved RepVGG model to recognize structured image input samples. During the training phase, a multi-branch convolutional structure and a positional encoding fusion path enhance the model's sensitivity to grid spatial locations. In the deployment phase, structural reparameterization unifies the multi-branch convolutional kernels into a single 3x3 convolutional kernel, balancing recognition accuracy and computational efficiency. By combining auxiliary data and a confidence-weighted mechanism based on a historical anomaly pattern database, this invention can fuse and verify multi-source data, effectively reducing recognition bias and generating more reliable fusion and verification results.

[0071] This invention introduces an edge-cloud collaborative mechanism for data synchronization, supporting breakpoint resumption, deduplication and merging, and version control, ensuring the integrity and consistency of data transmission even under fluctuating network conditions. Finally, through the aggregation and processing of synchronized inventory data, a missing data distribution map, anomaly grid location map, and asset status map are generated as a visualized inventory report, providing an intuitive display of the inventory results. This method guarantees the accuracy of inventory data, the reliability of transmission, and the visualization of results, overcoming the shortcomings of existing technologies in grid division, identification capabilities, data synchronization, and result display. Attached Figure Description

[0072] Figure 1 This is a flowchart of a grid-based inventory and data synchronization method based on artificial intelligence proposed in this invention;

[0073] Figure 2 This is a schematic diagram illustrating the binding of adaptive grid division and grid position encoding in an artificial intelligence-based gridded inventory and data synchronization method proposed in this invention.

[0074] Figure 3 This is a schematic diagram of the improved RepVGG model in the gridded inventory and data synchronization method based on artificial intelligence proposed in this invention. Detailed Implementation

[0075] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0076] This invention provides an artificial intelligence-based grid-based inventory and data synchronization method, such as... Figures 1-3 As shown, it includes the following steps:

[0077] The area to be investigated is divided into grids to obtain a set of grid location codes. Based on historical investigation records and current asset distribution density, an adaptive grid division result containing multiple adaptive grid cells and their corresponding grid location codes is generated.

[0078] Image data and auxiliary investigation data are collected by edge acquisition devices deployed in each adaptive grid cell;

[0079] The image data is bound to the corresponding grid position codes in the adaptive grid partitioning results to construct a set of structured image input samples;

[0080] An improved RepVGG model is used to identify a set of structured image input samples. A location encoding fusion path is introduced during the training phase, and all branches are reparameterized during the deployment phase to output asset recognition results.

[0081] The asset identification results and auxiliary inventory data are integrated and verified, and the reliability weights are configured based on the historical abnormal pattern library to obtain the integrated inventory results;

[0082] The merged investigation results are cached locally on the edge device to form a local cache set, and a data synchronization operation is performed based on the current network status to obtain a synchronized investigation data set;

[0083] The system aggregates the synchronous investigation data set to generate global investigation results and outputs a visual investigation report.

[0084] This invention achieves precise correspondence between assets and spatial locations through adaptive grid partitioning combined with image and location encoding binding. An improved RepVGG model enhances recognition accuracy and computational efficiency, while multi-source data fusion verification ensures result reliability. Combined with an edge-cloud collaborative data synchronization mechanism, it enables breakpoint resumption, deduplication and merging, and version control, ensuring data transmission integrity and traceability. The resulting visualized inventory report intuitively reflects missing and abnormal distributions, improving inventory efficiency and the real-time nature of management decisions.

[0085] In this embodiment, the area to be investigated is divided into grids to obtain a set of grid location codes. Based on historical investigation records and the current asset distribution density, an adaptive grid division result containing multiple adaptive grid cells and their corresponding grid location codes is generated, specifically including:

[0086] Obtain the physical layout data of the area to be investigated. The physical layout data includes structural diagram information, shelf location and arrangement information, and grid granularity configuration parameters. The grid granularity configuration parameters include minimum side length, maximum side length, and boundary change limit value.

[0087] Based on the acquired physical layout data, the area to be investigated is initially divided into grids according to preset spatial division rules. The preset spatial division rules include using the area boundary as a reference and combining the shelf arrangement direction and the size of the storage location to constrain the side length range of the division unit. Under the premise of meeting the minimum and maximum side length constraints, the entire area is divided into several independent rectangular grid units to form an initial grid set. Each grid unit is then assigned a unique grid position code to generate a grid position code set, which is used to identify the unique position of each grid unit in space.

[0088] Based on historical inventory records, the asset identification situation of each location in the inventory area in previous cycles is statistically analyzed to generate an asset identification heat distribution map indexed by regional coordinates. The value corresponding to each coordinate point represents the average number of asset identifications per unit time at that location. At the same time, the asset distribution of each location in the current cycle is obtained, and an asset distribution density map indexed by regional coordinates is calculated to reflect the degree of asset concentration in different areas in the current inventory cycle.

[0089] Based on the asset identification heat map and asset distribution density map, the degree of asset aggregation in each grid unit in the area to be investigated is comprehensively analyzed to determine the high-density asset areas that may appear in the next cycle. The boundary range of each grid unit in the initial grid set is dynamically adjusted according to the preset spatial density response rules so that the grid side length in the high-density area approaches the minimum side length and the grid side length in the low-density area approaches the maximum side length.

[0090] At the same time, the side length change of each grid cell in adjacent periods is limited to no more than the boundary change limit value, thereby generating an adjusted grid set to ensure that the grid division maintains the stability and continuity of the overall division structure while dynamically adapting to changes in asset distribution.

[0091] Each grid cell in the adjusted grid set is reassigned a corresponding grid position code to form a new grid position code set. The adjusted grid set is then combined with the new grid position code set to form a spatial index structure for subsequent binding of image data and execution of investigation operations. The spatial index structure is the adaptive grid partitioning result.

[0092] This invention introduces an adaptive grid partitioning mechanism, so that the division of the inventory area no longer relies on fixed boundaries, but rather combines physical layout, shelf arrangement, and grid granularity parameters for initial partitioning. Based on this, it dynamically adjusts the grid according to the asset identification heat map generated from historical inventory records and the asset distribution density map of the current period. Through spatial density response rules, high-density areas can be divided into smaller grids to improve identification accuracy, while low-density areas can be divided into larger grids to reduce redundant calculations, and the range of side length changes is limited to ensure overall structural stability. This method, while ensuring the rationality and continuity of the partitioning, generates a new spatial index structure, providing an accurate basis for binding image data with location encoding.

[0093] In this embodiment, the acquisition of image data and auxiliary investigation data through edge acquisition devices deployed within each adaptive grid cell specifically includes:

[0094] Edge acquisition devices are deployed within each grid cell of the adaptive grid partitioning result. These edge acquisition devices include fixed cameras, handheld terminals, and unmanned mobile platforms.

[0095] The edge acquisition device is used to acquire images of assets within the adaptive grid cell. The acquired image data is represented as a matrix, where each element of the matrix corresponds to the pixel value of the image at the row and column coordinates, reflecting the light intensity or color information at that location.

[0096] The auxiliary inventory data is synchronously acquired within the adaptive grid cell using edge acquisition devices. The auxiliary inventory data includes RFID signals corresponding to multiple assets acquired through the RFID module, barcode data corresponding to multiple assets acquired through the barcode scanning module, and sensor data acquired through the environmental monitoring module.

[0097] In this embodiment, binding image data with the corresponding grid position codes in the adaptive grid partitioning results to construct a structured image input sample set specifically includes:

[0098] Based on the deployment coordinates and shooting direction of the edge acquisition device, the covering grid cell number of the shooting area corresponding to each image data in the adaptive grid division result is obtained, and the correspondence between image data and grid position coding is established;

[0099] Each image data is bound to the corresponding grid position code in the adaptive grid partitioning result, so that the image data and the grid position code form a one-to-one image-position pair. The image data is used to represent the visual information collected in the grid cell, and the grid position code is used to identify the unique spatial location of the image. All image-position pairs are combined in sequence to form an image-position pair set.

[0100] Each pair of elements in the image-location pair set is encapsulated according to a unified data structure, so that the image data is represented in tensor form, and the tensor contains dimensional information such as the number of channels, height and width.

[0101] Simultaneously, the corresponding grid position codes are represented as vectors, and the image tensor and position code vectors are combined into structured image input sample units. All sample units are organized and assembled according to the order of grid position codes to form a complete set of structured image input samples.

[0102] In this embodiment, an improved RepVGG model is used to identify a set of structured image input samples. A location encoding fusion path is introduced during the training phase, and all branches are reparameterized during the deployment phase. The output asset recognition results specifically include:

[0103] Each sample unit in the structured image input sample set is input into the improved RepVGG model. The sample unit includes an image tensor and a location encoding vector.

[0104] During the training phase of the improved RepVGG model, a 3x3 convolution branch, a 1x1 convolution branch, an identity mapping branch, and a positional encoding fusion branch are used to extract features from the image tensor and the positional encoding vector to obtain a fused feature map.

[0105] The fused feature map is input into the output layer to generate prediction results. The loss function is calculated by combining the annotation results, and the parameters of the improved RepVGG model are updated based on the loss function until the training converges, resulting in an improved RepVGG model with location awareness.

[0106] After training convergence, the 3x3 convolution branch, 1x1 convolution branch, identity mapping branch and position encoding fusion branch are subjected to structural reparameterization to obtain the gridded clearing and identification model.

[0107] The gridded inventory and identification model is invoked in the edge device to infer the structured image input sample set and output the asset identification result set of all grid cells.

[0108] In this embodiment, during the training phase of the improved RepVGG model, a 3x3 convolution branch, a 1x1 convolution branch, an identity mapping branch, and a positional encoding fusion branch are used to extract features from the image tensor and the positional encoding vector to obtain the fused feature map. Specifically, this includes:

[0109] The image tensor is input into the 3x3 convolution branch, and the 3x3 convolution operation and batch normalization are performed sequentially to extract the local spatial perception features of the image, and obtain the local perception feature map that is consistent with the image tensor in the dimensions of channel number, height and width, which is used to characterize the feature information of the image tensor in the spatial neighborhood.

[0110] The image tensor is input into the 1x1 convolution branch, and convolution operation and batch normalization are performed in sequence to obtain the channel compressed feature map. The channel compressed feature map completes information aggregation in the channel dimension and expands in the spatial dimension through zero padding so that it maintains the same height and width as the local perceptual feature map output by the 3x3 convolution branch.

[0111] The image tensor is input into the identity mapping branch, the image tensor is directly passed and processed by batch normalization to obtain the residual information feature map;

[0112] The position encoding vector is input into the position encoding fusion branch, and a fully connected operation is performed on the position encoding vector to map the position encoding vector from the original encoding dimension to a channel vector consistent with the number of convolution channels. Then, the channel vector is broadcast and expanded in the spatial dimension to generate a three-dimensional position tensor consistent with the height and width of the image tensor.

[0113] One-to-one convolution is used to fuse the 3D position tensor and the image tensor channel by channel, and batch normalization is performed on the fused result to output a spatial position enhancement feature map containing spatial position information.

[0114] The local perception feature map, channel compression feature map, residual information feature map, and spatial location enhancement feature map are summed element-wise on the same spatial and channel dimensions to obtain a fused feature map.

[0115] This invention introduces a multi-branch feature extraction mechanism during the training phase of an improved RepVGG model. It comprehensively utilizes a 3x3 convolution branch to acquire local spatially perceptual features, a 1x1 convolution branch to compress channel information, and an identity mapping branch to maintain the continuity of the original features. Simultaneously, in the positional encoding fusion branch, positional codes are mapped and expanded into a 3D tensor, which is then fused with image features to output an enhanced feature map containing spatial location information. Finally, the fused feature map is obtained by element-wise summing the results from the multiple branches, enabling the model to maintain a compact structure while possessing spatial sensitivity.

[0116] In this embodiment, after training convergence, the 3x3 convolution branch, 1x1 convolution branch, identity mapping branch, and position encoding fusion branch undergo structural reparameterization to obtain the gridded clearing and identification model, specifically including:

[0117] In the multi-branch structure of the improved RepVGG model, the convolution weight parameters and batch normalization parameters corresponding to the 3x3 convolution branch, 1x1 convolution branch, identity mapping branch and position encoding fusion branch are read sequentially. The batch normalization parameters include scaling factor, offset, mean, variance and numerical stability constant. The convolution weight parameters and batch normalization parameters of each branch are recorded in a unified manner to obtain the original parameter set of the branch.

[0118] Based on the original parameter set of the branches, batch normalization fusion processing is performed on the 3x3 convolution branch, the 1x1 convolution branch, the identity mapping branch and the position encoding fusion branch respectively. The convolution weights of each branch are combined with the batch normalization parameters to obtain the corresponding equivalent convolution weights and equivalent biases, forming a set of equivalent convolution weights and biases.

[0119] The equivalent convolution weights are obtained by multiplying the original convolution weights by the ratio of the batch normalized scaling factor to the standardized denominator, and the equivalent biases are obtained by subtracting the product of the mean and the scaling factor after standardization from the batch normalized bias.

[0120] Based on the equivalent convolution weights and bias set, the equivalent convolution weights of the one-to-one convolution branch are expanded by zero-value padding at the center of the convolution kernel to obtain a one-to-one convolution branch extended convolution kernel with the same size as the three-to-three convolution kernel.

[0121] Perform the same zero-padded expansion operation on the equivalent convolution weights of the position encoding fusion branch to obtain the position encoding fusion branch expanded convolution kernel;

[0122] The equivalent convolution weights of the identity mapping branch are converted into a 3x3 convolution kernel form, and the center element position of each channel is assigned a value of 1, while the other positions are assigned a value of 0, thus obtaining the extended convolution kernel of the identity mapping branch.

[0123] The equivalent convolution weights of the 3x3 convolution branch are directly used as the convolution kernels of the 3x3 convolution branch. The extended convolution kernels of the 1x1 convolution branch, the extended convolution kernels of the position encoding fusion branch, the extended convolution kernels of the identity mapping branch, and the extended convolution kernels of the 3x3 convolution branch are combined to form an extended convolution kernel set.

[0124] Based on the extended convolution kernel set, the convolution kernels of the four branches are summed element by element. The convolution kernels of the 3x3 convolution branch, the extended convolution kernels of the 1x1 convolution branch, the extended convolution kernels of the identity mapping branch, and the extended convolution kernels of the position encoding fusion branch are added at the corresponding positions to obtain a single 3x3 convolution kernel.

[0125] Numerical summation is performed on the biases of the 3x3 convolution branch, the 1x1 convolution branch, the identity mapping branch, and the position encoding fusion branch in the equivalent convolution weight and bias set to obtain a single bias.

[0126] The location-encoded fusion branch extended convolutional kernel retains spatial location encoding information after merging, enabling a single 3x3 convolutional kernel to remain sensitive to grid location during the deployment phase;

[0127] By replacing all the multi-branch structures of the corresponding layer in the improved RepVGG model with a single 3x3 convolution kernel and a single bias, a network topology with a single convolution structure is obtained.

[0128] The network topology of a single convolutional structure is saved in a format that can be deployed on edge devices to form a gridded investigation and identification model.

[0129] This invention performs structural reparameterization on the multi-branch structure of the improved RepVGG model after training convergence. By fusing the convolution weights of the 3x3 convolution branch, 1x1 convolution branch, identity mapping branch, and positional encoding fusion branch with batch normalization parameters, equivalent convolution weights and biases are obtained. These are then uniformly converted into 3x3 convolution kernels and summed to ultimately form a single 3x3 convolution kernel and a single bias. During this process, the spatial location information of the positional encoding fusion branch is preserved, ensuring that the generated single 3x3 convolution kernel retains grid position sensitivity during deployment. Simplifying the complex multi-branch structure into a single convolution structure not only reduces computational resource consumption and model runtime latency but also ensures recognition accuracy and positional awareness, providing an efficient and reliable model foundation for rapid deployment on edge devices.

[0130] In this embodiment, the asset identification results and auxiliary inventory data are fused and verified. Based on the historical abnormal pattern database, confidence weights are allocated to obtain the fused inventory results, which specifically include:

[0131] Each asset identification result in the asset identification result set is matched one by one with the corresponding record in the auxiliary inventory data set. Based on the unique identification information of the asset under the grid location code, the image recognition result is associated with RFID signals, barcode data and sensor data to generate a set of matching pairs. Each matching pair consists of the asset identification result and its relevant information in the auxiliary inventory data.

[0132] Based on the historical anomaly pattern library, reliability statistics of multi-source data during the investigation process are extracted. According to the distribution of accuracy and anomaly rates of various types of data in historical records, the confidence weights of image recognition results, RFID signals, barcode data, and sensor data are determined to form a confidence weight vector. The confidence weight vector includes the confidence weights of image recognition results, RFID signals, barcode data, and sensor data.

[0133] For each matching pair in the matching pair set, the asset identification result, RFID signal, barcode data and sensor data are numerically processed according to the confidence weight vector, the weighted score corresponding to each data source is calculated, and the weighted scores are accumulated to obtain the data consistency score of the matching pair.

[0134] The data consistency score is compared with a preset threshold. When the data consistency score is less than the preset threshold, a corresponding check anomaly marker is generated and its value is set to one. When the data consistency score is greater than or equal to the preset threshold, a corresponding check anomaly marker is generated and its value is set to zero, thus forming a check anomaly marker set.

[0135] The asset identification result set, the auxiliary inventory data set, and the inventory anomaly mark set are merged. Using each grid cell as an index, the corresponding asset identification result, auxiliary inventory data, and inventory anomaly mark are combined into a unified data entry, thereby forming a merged inventory result set.

[0136] In this embodiment, the fused investigation results are cached locally on the edge device to form a local cache set. Then, based on the current network status, an edge-cloud collaborative data synchronization operation is performed to obtain the synchronized investigation data set, which specifically includes:

[0137] The fused investigation results are stored in the local cache of the edge device. Each data entry in the set is indexed and managed according to grid location code and timestamp to form a local cache set.

[0138] A synchronization strategy function is constructed based on the current network status of the edge device. The network status includes network bandwidth and transmission latency. When the calculation result of the synchronization strategy function is higher than a preset threshold, the full transmission mode is triggered, and the local cache set is synchronized to the cloud platform. When the calculation result of the synchronization strategy function is lower than the preset threshold, the breakpoint resume transmission strategy is executed.

[0139] When executing the breakpoint resume strategy, the local cache set is divided into blocks according to the generated timestamp and the corresponding grid position code to form multiple block sets. During the data synchronization process, the blocks that have been successfully transmitted are recorded. When the transmission is interrupted, the current breakpoint position pointer is saved locally so that the unfinished block data can be transmitted from the breakpoint position after the network is restored, thereby realizing breakpoint resume.

[0140] During the process of executing block numbering and transmitting to the cloud, the block set is deduplicated and merged. Block records with the same grid location code and earlier timestamps are removed, and only the data entries with the latest timestamp under the corresponding grid location code are retained. Based on this, the deduplicated and merged result set is recombined.

[0141] After the transmission is completed, a version control mechanism is established for the deduplication and merging result set according to the synchronization order. A unique version number is generated for each synchronization, and the version number is bound to the corresponding synchronization block. When a new synchronization operation occurs, the version number is automatically incremented, thereby forming a deduplication and merging result set with version identification, so as to ensure that the cleanup data has traceability and historical version management capabilities in subsequent use.

[0142] The deduplication and merging result set with version numbers is uploaded to the cloud platform to form a synchronous inventory data set. The synchronous inventory data set includes asset identification results, auxiliary inventory data, inventory anomaly markers and their corresponding version numbers.

[0143] In this implementation, the synchronous inventory data set is aggregated to generate a global inventory result and output a visual inventory report, specifically including:

[0144] The synchronous inventory data set is aggregated according to the grid location code to create a grid location index table;

[0145] Based on the grid location index table, data entries belonging to the same grid location code are merged to obtain a preliminary aggregation result set;

[0146] Multi-dimensional statistical processing is performed on the preliminary aggregation result set to calculate the global inventory result. The global inventory result includes the total number of assets obtained by summing the asset identification results in each grid cell; the abnormal grid ratio obtained by statistically analyzing the number of grid cells with inventory anomaly markers to the total number of grid cells; and the missing rate obtained by comparing the asset identification results with the auxiliary inventory data to determine the number of missing assets and in the form of the ratio of the number of missing assets to the total number of assets.

[0147] A visual inventory report is generated based on the global inventory results. The distribution of missing assets is plotted as a missing asset distribution map, the location of grid cells with anomaly markers is plotted as an anomaly grid location map, and the total number of assets, the anomaly ratio, and the missing asset rate are plotted as an asset status map.

[0148] To verify the feasibility of this invention in practice, it was applied to the fixed asset inventory and data synchronization management scenario of a large enterprise. The enterprise's office areas and computer rooms are distributed across multiple buildings, and the types of fixed assets include office computers, printers, network cabinets, electrical equipment, and laboratory instruments. The number of assets is enormous and dynamically changes with departmental adjustments and equipment replacements. Traditional methods relying on manual inventory, form registration, or barcode scanning are time-consuming, inaccurate, and suffer from delayed information updates, leading to discrepancies between asset records and actual assets and delays in management decisions.

[0149] In the application process, an initial grid division is first generated based on the physical layout data of the office area and computer room. This grid is then dynamically adjusted based on historical asset inventory records and the current asset distribution density to obtain an adaptive grid set. Each grid cell is assigned a unique grid location code. Subsequently, cameras, handheld terminals, and unmanned inspection devices are deployed in each grid cell to collect asset images in real time, while also incorporating RFID tag, barcode information, and sensor status data. After all image data is bound to the location code, a structured input sample set is formed, which is then identified by an improved RepVGG model. This model enhances spatial location sensitivity through a location code fusion branch during the training phase and simplifies to a single convolutional kernel structure through structural reparameterization during the deployment phase, achieving a balance between recognition accuracy and computational efficiency. The recognition results are fused and verified with auxiliary inventory data on edge devices, and confidence weights are calculated using a historical anomaly pattern library to identify potential asset anomalies, such as missing tags, location offsets, or registration inconsistencies.

[0150] During the data synchronization phase, the merged inventory results are stored locally on edge devices and synchronized to the cloud using breakpoint resume and deduplication merging strategies based on network conditions. A version control mechanism is established to ensure data traceability and update integrity. The cloud aggregates the synchronized data to generate asset distribution maps, anomaly location maps, and asset health status maps, providing managers with intuitive and real-time visual reports.

[0151] Table 1. Comparison of test results between the method of this invention and the traditional method.

[0152] Test metrics Test scale (pieces) Results of the method of the present invention Results of traditional methods Average time spent on inventory checks 25000 2.1 hours 5.4 hours Asset identification accuracy 25000 96.8% 83.2% Data transmission integrity 25000 100% 96.3% Anomaly detection accuracy 25000 97% 70% Asset status positioning accuracy 25000 95.7% 77.9% Traceable version management support 25000 support Not supported

[0153] As can be seen from Table 1 above, the method of this invention significantly improves upon traditional methods in several key performance indicators of fixed asset inventory. Firstly, in terms of inventory efficiency, the method of this invention takes an average of only 2.1 hours when the total number of assets is 25,000, while the traditional manual method requires 5.4 hours, more than doubling the efficiency. Secondly, in terms of identification accuracy, the method of this invention achieves 96.8%, significantly higher than the 83.2% of the traditional method, indicating that the introduction of location coding binding and the improved RepVGG model effectively reduces missed detections and false positives. Regarding data transmission integrity, this invention achieves 100% integrity through breakpoint resumption and deduplication merging mechanisms, while the traditional method suffers from approximately 3.7% data loss when the network is unstable, demonstrating significantly higher reliability.

[0154] In anomaly identification, this invention can accurately identify 97% of historical anomaly patterns, higher than the 70% of traditional methods, demonstrating that its fusion verification and confidence weighting mechanism can more effectively detect asset anomalies. Regarding asset status positioning accuracy, this invention achieves 95.7%, while traditional methods only reach 77.9%, proving the advantages of gridded spatial indexing. Finally, this invention supports version management and historical traceability, capabilities lacking in traditional methods. In summary, this invention demonstrates excellent performance in efficiency, accuracy, reliability, and traceability, providing strong technical support for large-scale fixed asset management.

[0155] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for grid-based inventory and data synchronization based on artificial intelligence, characterized in that, Includes the following steps: The area to be investigated is divided into grids to obtain a set of grid location codes. Based on historical investigation records and current asset distribution density, an adaptive grid division result is generated. Image data and auxiliary investigation data are collected by edge acquisition devices deployed in each adaptive grid cell; The image data is bound to the corresponding grid position codes in the adaptive grid partitioning results to construct a set of structured image input samples; An improved RepVGG model is used to identify a set of structured image input samples. A location encoding fusion path is introduced during the training phase, and all branches are reparameterized during the deployment phase to output a set of asset recognition results. The asset identification results set and auxiliary inventory data are merged and verified, and the confidence weights are configured based on the historical abnormal pattern library to obtain the merged inventory results; The merged investigation results are cached locally on the edge device, and a data synchronization operation is performed based on the current network status to obtain a synchronized investigation data set; The system aggregates the synchronous investigation data set to generate global investigation results and outputs a visual investigation report.

2. The artificial intelligence based gridification and data synchronization method as claimed in claim 1, wherein, The specific method for dividing the area to be investigated into a grid to obtain a set of grid location codes, and generating adaptive grid division results based on historical investigation records and current asset distribution density includes: Obtain the physical layout data of the area to be investigated, perform initial grid division of the area based on the physical layout data, and assign a unique grid location code to each grid cell to form a set of grid location codes; Generate an asset identification heat map and an asset distribution density map for the current period based on historical inventory records; Based on the asset identification heat map and asset distribution density map, the probability area of ​​asset clustering in the next period is predicted, and the boundary range of each grid cell in the initial grid set is dynamically adjusted to generate the adjusted grid set. Each grid cell in the adjusted grid set is reassigned a grid position code to form a new set of grid position codes. The adjusted grid set is then combined with the new set of grid position codes to form an adaptive grid partitioning result.

3. The artificial intelligence based gridification and data synchronization method as claimed in claim 1, wherein, The edge acquisition device includes a camera, a handheld terminal, and an unmanned mobile platform; the auxiliary investigation data includes RFID signals, barcode data, and sensor data.

4. The artificial intelligence based gridification and data synchronization method as claimed in claim 1, wherein, The specific method for binding image data with the corresponding grid position codes in the adaptive grid partitioning results to construct a structured image input sample set includes: Based on the deployment coordinates and shooting direction of the edge acquisition device, establish the correspondence between image data and grid position codes; Each image data is bound to the corresponding grid position code in the adaptive grid partitioning result to form an image-position pair set; Each pair of elements in the image-location pair set is encapsulated according to a unified data structure to form a structured image input sample unit; All structured image input sample units are organized and aggregated according to the grid position encoding order to form a structured image input sample set.

5. The artificial intelligence based gridification and data synchronization method as claimed in claim 1, wherein, The method comprises the following steps: inputting each sample unit in the structured image input sample set into the improved RepVGG model, wherein the sample unit comprises an image tensor and a position encoding vector; in the training stage of the improved RepVGG model, extracting features of the image tensor and the position encoding vector by using a three-by-three convolution branch, a one-by-one convolution branch, an identity mapping branch and a position encoding fusion branch, and obtaining a fusion feature map; inputting the fusion feature map into an output layer to generate a prediction result, and calculating a loss function in combination with a labeled result; updating parameters of the improved RepVGG model based on the loss function until training converges, and obtaining the improved RepVGG model with position perception capability; after the training converges, performing structural reparameterization processing on the three-by-three convolution branch, the one-by-one convolution branch, the identity mapping branch and the position encoding fusion branch, and obtaining a grid-based inspection identification model; calling the grid-based inspection identification model in an edge device to perform reasoning on the structured image input sample set, and outputting an asset identification result set of all grid units.

6. The artificial intelligence based gridification and data synchronization method as claimed in claim 5, wherein, The method comprises the following steps: inputting the image tensor into the three-by-three convolution branch to extract local spatial perception features of the image, and obtaining a local perception feature map; inputting the image tensor into the one-by-one convolution branch to sequentially perform convolution operation and batch normalization processing, and obtaining a channel compression feature map; inputting the image tensor into the identity mapping branch to perform batch normalization processing, and obtaining a residual information feature map; inputting the position encoding vector into the position encoding fusion branch to perform full connection operation on the position encoding vector, and mapping the position encoding vector into a channel vector; broadcasting and expanding the channel vector in a spatial dimension to generate a three-dimensional position tensor, performing channel-by-channel fusion on the three-dimensional position tensor and the image tensor by using the one-by-one convolution, and outputting a spatial position enhancement feature map; performing element-by-element addition processing on the local perception feature map, the channel compression feature map, the residual information feature map and the spatial position enhancement feature map in the same spatial dimension and channel dimension, and obtaining the fusion feature map.

7. The artificial intelligence based gridification and data synchronization method as claimed in claim 5, wherein, The method comprises the following steps: reading, in sequence, convolution weight parameters and batch normalization parameters corresponding to each branch in the improved RepVGG model to obtain a branch original parameter set; based on the branch original parameter set, combining the convolution weight and the batch normalization parameter of each branch to obtain an equivalent convolution weight and bias set; Based on the equivalent convolution weight and bias set, a zero value padding expansion operation is performed on the equivalent convolution weight of the one multiplication convolution branch to obtain an one multiplication convolution branch expansion convolution kernel; A zero value padding expansion operation is performed on the equivalent convolution weight of the position encoding fusion branch to obtain a position encoding fusion branch expansion convolution kernel; The equivalent convolution weight of the identity mapping branch is converted into a three-by-three convolution kernel form to obtain an identity mapping branch expansion convolution kernel; The equivalent convolution weight of the three-by-three convolution branch is directly taken as a three-by-three convolution branch kernel, and the one multiplication convolution branch expansion convolution kernel, the position encoding fusion branch expansion convolution kernel, the identity mapping branch expansion convolution kernel, and the three-by-three convolution branch kernel are collectively composed into an expansion convolution kernel set Based on the expansion convolution kernel set and the equivalent convolution weight and bias set, the convolution kernels of the four branches are processed element by element to obtain a single three-by-three convolution kernel, and the biases of the branches are numerically added to obtain a single bias; The single three-by-three convolution kernel and the single bias replace all multi-branch structures of the corresponding layer of the improved RepVGG model to obtain a network topology with single convolution structure; The network topology with single convolution structure is saved in a format that can be deployed on an edge device to form a grid-based inspection identification model.

8. The artificial intelligence based gridification and data synchronization method as claimed in claim 1, wherein, The specific method of fusing and verifying the asset identification result set and the auxiliary inspection data based on the historical abnormal pattern library includes: Each asset identification result in the asset identification result set is matched with the corresponding record in the auxiliary inspection data set one by one to generate a matching pair set; Based on the historical abnormal pattern library, the reliability statistical information of multi-source data in the inspection process is extracted, and a confidence weight vector is formed according to the correct rate and abnormal rate distribution of each type of data in the historical record; For each matching pair in the matching pair set, the asset identification result, the RFID signal, the barcode data, and the sensor data are numerically processed according to the confidence weight vector to obtain a data consistency score of the matching pair; The data consistency score is compared with a preset threshold, and when the data consistency score is less than the preset threshold, a corresponding inspection anomaly marker is generated and assigned a value of one, otherwise a value of zero, forming an inspection anomaly marker set; The asset identification result set, the auxiliary inspection data set, and the inspection anomaly marker set are fused to form a fusion inspection result.

9. The artificial intelligence based gridification and data synchronization method as claimed in claim 1, wherein, The specific method of caching the fusion inspection result locally on the edge device and performing edge-cloud collaborative data synchronization operation based on the current network state to obtain a synchronized inspection data set includes: The fusion inspection result is stored in the local cache area of the edge device to form a local cache set; A synchronization strategy function is constructed according to the current network state of the edge device, and when the calculation result of the synchronization strategy function is higher than a preset threshold, a full-amount transmission mode is triggered, and when the calculation result of the synchronization strategy function is lower than a preset threshold, a breakpoint resume transmission strategy is executed; When the breakpoint resume transmission strategy is executed, the local cache set is divided into multiple block sets according to the generated time stamp and the corresponding grid position encoding; In the process of performing block numbering and transmitting to the cloud, the block set is de-duplicated and merged to generate a de-duplicated and merged result set; After transmission is completed, a version control mechanism is established for the de-duplicated and merged result set according to a synchronization order to form a de-duplicated and merged result set with a version identifier; The de-duplicated and merged result set with the version identifier is uploaded to the cloud platform to obtain a synchronization check data set.

10. The artificial intelligence based gridification and data synchronization method as claimed in claim 1, wherein, The specific method of the aggregation processing of the synchronization check data set, the generation of a global check result, and the output of a visual check report includes: The synchronization check data set is aggregated according to a grid location code to establish a grid location index table; Based on the grid location index table, data entries belonging to the same grid location code are merged to obtain a preliminary aggregation result set; The preliminary aggregation result set is subjected to multi-dimensional statistical processing to calculate a global check result; Based on the global check result, a visual check report is generated, and the visual check report includes a missing distribution graph, an abnormal grid positioning graph, and an asset situation graph.

Images