Intelligent tracking method in warehouse based on data analysis

Abstract

The invention discloses an intelligent tracking method in a warehouse based on data analysis, and the method comprises the following steps: S1, deploying a multi-source heterogeneous sensing network, and collecting the full-dimensional original data of a target object in the warehouse, the full-dimensional original data comprising position sensing data, motion feature data, environment interference data and visual semantic data; s2, performing multi-modal fusion preprocessing on the full-dimensional original data to generate standardized fusion data; s3, on the basis of the standardized fusion data, calculating real-time accurate position coordinates of the target object by using a dynamic tracking model which fuses multi-dimensional features by using an attention mechanism; s4, generating and updating a tracking path by adopting a path optimization algorithm according to the real-time accurate position coordinates and the dynamic state of the goods allocation in the warehouse; and S5, performing anomaly identification and adaptive correction on the data in the tracking process, and outputting a final tracking result. According to the invention, better intelligent tracking in the warehouse can be realized.

Description

Technical Field

[0001] This invention relates to the field of warehouse management, and more specifically to an intelligent tracking method for warehouses based on data analysis. Background Technology

[0002] Smart warehousing is a modern warehouse management model that integrates the Internet of Things (IoT), artificial intelligence (AI), and automation technologies. Its core lies in achieving high efficiency, precision, and intelligence in the storage, handling, and tracking of goods through technological means. This addresses the pain points of traditional warehousing, such as high reliance on manual labor, low efficiency, and large errors, and has become a key direction for the upgrading of the logistics industry. Data analysis, as the core support of smart warehousing, can deeply mine the value of data throughout the entire warehousing process. Intelligent tracking of target objects (such as goods and AGVs) within the warehouse is one of the core aspects of smart warehousing, and its accuracy directly affects warehouse operational efficiency and management decisions.

[0003] While current data-driven intelligent tracking methods for warehouses have been initially applied, they still have many limitations: data collection dimensions are limited, anti-interference capabilities are weak, and it is difficult to comprehensively capture the location, movement, and environmental information of target objects; data preprocessing lacks an adaptive denoising mechanism, which easily leads to distortion of effective signals or noise residue; feature extraction and location calculation do not fully consider the dynamic contribution of multimodal features and environmental interference correction, resulting in limited positioning accuracy; path planning is static and cannot adapt to dynamic occupancy and congestion of storage locations; anomaly identification and correction mechanisms are imperfect, and tracking reliability is insufficient in strong interference scenarios; and there is a lack of in-depth mining of tracking data, making it difficult to support warehouse operation optimization. Therefore, this paper proposes a data-driven intelligent tracking method for warehouses. Summary of the Invention

[0004] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps: S1: Deploy a multi-source heterogeneous sensing network to collect full-dimensional raw data of target objects in the warehouse. The full-dimensional raw data includes location sensing data, motion feature data, environmental interference data, and visual semantic data. S2: Perform multimodal fusion preprocessing on the original data across all dimensions to generate standardized fused data; S3: Based on the standardized fused data, the real-time accurate location coordinates of the target object are calculated using a dynamic tracking model, wherein the dynamic tracking model uses an attention mechanism to fuse multi-dimensional features; S4: Based on the real-time accurate location coordinates and the dynamic status of the goods in the warehouse, a path optimization algorithm is used to generate and update the tracking path; S5: Performs anomaly identification and adaptive correction on the data during the tracking process, and outputs the final tracking result.

[0005] Furthermore, in step S1, the multi-source heterogeneous sensing network includes: The ultra-wideband positioning module has dynamically adjustable anchor points, and the tag has an adaptive transmission power adjustment function. An inertial measurement unit, rigidly connected to the target object, is used to collect motion data and calculate motion characteristics; The machine vision module employs multi-camera collaboration and extracts visual features through image semantic segmentation; The distributed environmental monitoring module is deployed according to a spatial grid and is used to collect environmental physical quantity data of each grid.

[0006] Furthermore, in step S2, the multimodal fusion preprocessing includes a noise suppression process, which employs an adaptive wavelet threshold denoising algorithm. The process of calculating the dynamic threshold is as follows: First, the data is decomposed using wavelet decomposition to obtain wavelet coefficients at each scale; then, the noise variance σ of these coefficients is estimated; finally, based on the noise variance σ and the data length N, the dynamic wavelet threshold τ is calculated. The calculation process is as follows: .

[0007] Furthermore, in step S3, the features used by the dynamic tracking model are extracted from the standardized fused data through a feature extraction network; the feature extraction network is a hybrid architecture of convolutional neural network and long short-term memory network. The convolutional neural network introduces an attention module to enhance key spatial features. The calculation process of the attention module is as follows: the input basic feature map F undergoes two linear transformations and a non-linear activation function is applied. Finally, the attention weight matrix A is generated through the sigmoid function, and its output formula is: ; Where F is the basic feature map. Here, W1 and W2 are weight matrices, and b1 and b2 are bias vectors; The forgetting gate of the Long Short-Term Memory network introduces a regularization term based on historical states. The calculation process is as follows: output the sequence of hidden states from the first k time steps. The summation is multiplied by the regularization coefficient λ, then normalized using the softmax function, and finally multiplied by the original forget gate output ft to obtain the regularized forget gate output. The calculation formula is as follows: .

[0008] Furthermore, in step S3, the process of calculating the real-time accurate position coordinates includes: The first step is to extract the location feature vector X, motion trend vector V, environmental interference vector D, and visual matching vector M from the standardized fused data; The second step is to dynamically generate corresponding attention weight coefficients based on the current features through a trainable weight learning model, including position feature weight α, motion feature weight β, environment correction weight γ, and visual assistance weight δ. The third step is feature fusion: the feature vector is multiplied by its corresponding attention weight coefficient, then weighted and summed, and the sum is normalized to obtain the comprehensive feature vector F. The calculation formula for this fusion process is as follows: ; Step 4, Coordinate Mapping: Input the comprehensive feature vector F into the pre-trained 3D spatial calibration model and calculate its coordinate values ​​on the x, y, and z axes respectively. The mapping process is as follows: multiply the components of F on the corresponding axes by the spatial calibration coefficient, then multiply by the correction term composed of the environmental correction weight γ and the environmental disturbance vector components, and finally add the origin reference value of the coordinate axes. The mapping formulas are as follows: ; ; ; Where Fx, Fy, Fz and Dx, Dy, Dz are the components of vectors F and D in the three coordinate axes, respectively, and Sx, Sy, Sz are the pre-calibrated spatial calibration coefficients (x0, y0, z0) are the reference values ​​of the origin of the coordinate system.

[0009] Furthermore, the path optimization algorithm includes path evaluation and dynamic triggering steps: The path evaluation uses an improved A* algorithm to calculate the evaluation function value f(n) from the starting node to any node n to be evaluated in the search graph consisting of path nodes within the warehouse. The evaluation function f(n) consists of three parts: The first part, g(n), represents the actual path cost from the starting node to the current node n, which is obtained by accumulating the weights of the edges along the path. The second part, h(n), is the heuristic estimated cost from the current node n to the target node, calculated using Euclidean distance or Manhattan distance. The third part is the congestion cost term, which is obtained by multiplying the real-time congestion coefficient C(n) of the path segment where node n is located by a preset congestion impact weight ω; the real-time congestion coefficient C(n) is calculated based on the current occupancy status data of the associated cargo spaces of the path segment, and the higher the cargo space occupancy rate, the larger the value of C(n); the calculation formula of the evaluation function f(n) of the improved A* algorithm is as follows: ; The dynamic triggering step uses the rate of change in cargo space occupancy, the rate of change in movement speed, and the rate of change in environmental disturbance as dynamic adjustment factors. Among them, the rate of change in storage space occupancy The calculation process is as follows: Take the number of occupied storage spaces within the path planning area at the current time t. The number of occupied storage spaces compared to the previous time t-1 The absolute value of the difference is then divided by the total number of storage locations in that area. The calculation process is as follows: ; Rate of change of velocity The calculation process is as follows: take the absolute value of the difference between the target object's velocity V(t) at the current time t and the velocity V(t-1) at the previous time t-1, and then divide it by the historical average velocity Vavg. Environmental disturbance change rate The calculation process is as follows: Take the absolute value of the difference between the environmental disturbance factor evaluation value D(t) at the current time t and the evaluation value D(t-1) at the previous time t-1, and then divide it by the historical average environmental disturbance factor evaluation value Davg. When any of the dynamic adjustment factors exceeds its preset threshold, the replanning of the tracking path is triggered.

[0010] Furthermore, in step S5, the anomaly identification employs a two-layer mechanism: The first layer of identification is based on the dynamic 3σ criterion. Its calculation process is as follows: Within a sliding time window, the mean μ and standard deviation σ of the data are calculated, and the adaptive coefficient k(t) that changes with time is obtained. The mean, standard deviation, and adaptive coefficient are combined to calculate the dynamic anomaly detection threshold θ. The calculation process is as follows: ; The second layer of recognition is based on a graph neural network model for semantic anomaly detection. Its calculation process includes constructing a feature graph and calculating the correlation between nodes: using data samples as graph nodes, the semantic correlation Sij between any two samples i and j is calculated; this calculation first yields the cosine similarity between the feature vectors of the two samples. Then multiply by an exponential decay term with time distance dij as the parameter. Finally, this result is normalized by dividing by the sum of the products of the similarity between sample i and all other samples k and the attenuation term. The formula for calculating semantic relevance is: .

[0011] Furthermore, it also includes step S6: constructing a trajectory knowledge graph of the target object; Graph embedding algorithms are used to transform graphs into low-dimensional vector representations; Based on the low-dimensional vector, motion patterns are analyzed through association rule mining. The calculation process for the support of the association rule X→Y is as follows: The formula for calculating the number of samples that contain both events X and Y is: count(X∪Y), divided by the total number of samples count(U). ; Based on the analysis results, output an operational optimization plan.

[0012] Furthermore, in step S5, for abnormal data under strong interference scenarios, a weighted voting mechanism is used to perform multimodal data cross-validation correction. This correction process includes: First, calculating the voting weight of each modality: Based on the historical tracking error rate error(m) of the m-th modality, its credibility(m) is calculated. The calculation process is to subtract the historical error rate of the modality from 1, and then divide by the sum of (1 - historical error rates) of all modalities to achieve weight normalization. The formula for calculating credibility is: ; Then, weighted fusion is performed: the independent correction result Rm of the m-th modality data is multiplied by its credibility(m), and the product results of all M modalities are summed to obtain the final correction result R, which is calculated using the following formula: .

[0013] Compared with existing technologies, this invention has the following advantages: This intelligent tracking method for warehouses based on data analysis deploys a multi-source heterogeneous sensing network containing an ultra-wideband positioning module, an inertial measurement unit, a machine vision module with multi-camera collaboration, and a distributed environmental monitoring module with spatial grid deployment. This network comprehensively collects raw data across all dimensions, including location, motion features, environmental interference, and visual semantics. Combined with adaptive wavelet threshold denoising and multimodal fusion preprocessing, noise is effectively suppressed and standardized fused data is generated, laying a high-quality data foundation for subsequent tracking. Furthermore, by leveraging a hybrid feature extraction network architecture combining convolutional neural networks and long short-term memory networks, along with a feature fusion process that dynamically generates attention weight coefficients and a coordinate mapping process with 3D spatial calibration, real-time and accurate positioning of the target object is achieved, improving positioning accuracy. Regarding timeliness, an improved A* algorithm with congestion cost is used for path evaluation, and path updates are triggered dynamically by three change rates: storage space occupancy, movement speed, and environmental interference, generating the optimal tracking path adapted to the dynamic state of the warehouse. A dual-layer anomaly identification mechanism combining dynamic 3σ criteria and graph neural network semantic recognition, along with a weighted voting multimodal cross-validation correction scheme based on historical error rates, accurately identifies and corrects abnormal data, ensuring the reliability of tracking results. Furthermore, by constructing a trajectory knowledge graph, graph embedding transformation, and association rule mining to analyze movement patterns, operational optimization solutions are output. This not only enables efficient and accurate tracking of target objects within the warehouse but also helps optimize warehouse operation processes, improving overall warehouse management efficiency and intelligence, making it more applicable and practical. Attached Figure Description

[0014] Figure 1 This is a system structure block diagram of the present invention. Detailed Implementation

[0015] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0016] like Figure 1 As shown, this embodiment provides a technical solution: a data analysis-based intelligent tracking method for warehouses, comprising the following steps: S1: Deploy a multi-source heterogeneous sensing network to collect full-dimensional raw data of target objects in the warehouse. The full-dimensional raw data includes location sensing data, motion feature data, environmental interference data, and visual semantic data. S2: Perform multimodal fusion preprocessing on the original data across all dimensions to generate standardized fused data; S3: Based on the standardized fused data, the real-time accurate location coordinates of the target object are calculated using a dynamic tracking model, wherein the dynamic tracking model uses an attention mechanism to fuse multi-dimensional features; S4: Based on the real-time accurate location coordinates and the dynamic status of the goods in the warehouse, a path optimization algorithm is used to generate and update the tracking path; S5: Performs anomaly identification and adaptive correction on the data during the tracking process, and outputs the final tracking result.

[0017] In step S1, the multi-source heterogeneous sensing network includes: The ultra-wideband positioning module has dynamically adjustable anchor points, and the tag has an adaptive transmission power adjustment function. An inertial measurement unit, rigidly connected to the target object, is used to collect motion data and calculate motion characteristics; The machine vision module employs multi-camera collaboration and extracts visual features through image semantic segmentation; The distributed environmental monitoring module is deployed according to a spatial grid and is used to collect environmental physical quantity data of each grid. By deploying a multi-source heterogeneous sensing network consisting of an ultra-wideband positioning module (dynamic anchor point adjustment + adaptive transmission power), an inertial measurement unit (rigidly connected to the target), a machine vision module (multi-camera collaboration + image semantic segmentation), and a distributed environmental monitoring module (spatial grid deployment), comprehensive coverage of raw data on the location, motion characteristics, visual semantics, and environmental interference of target objects within the warehouse is achieved. Each module is complementary and specifically adapted to address pain points in the warehousing scenario: the ultra-wideband positioning module can handle changes in shelf layout and signal obstruction, ensuring stable location data acquisition through dynamic adjustment of anchor point positions and adaptive adjustment of transmission power; the inertial measurement unit... The unit, through rigid connection with the target, avoids vibration interference caused by non-rigid connections, accurately collects motion data, and infers motion characteristics; the machine vision module, with the help of multi-camera collaboration and image semantic segmentation technology, solves the problem of missing visual features caused by single camera occlusion or viewpoint limitation, and obtains complete visual information; the distributed environmental monitoring module is deployed according to a spatial grid to achieve continuous spatial coverage of environmental physical quantities, providing accurate basis for subsequent data correction, and improving the overall accuracy, anti-interference and flexibility of data collection, laying a high-quality data foundation for multimodal fusion preprocessing and dynamic tracking model calculation, and adapting to the complex and ever-changing actual application scenarios of warehousing.

[0018] Ultra-wideband positioning module (dynamic anchor point adjustment + adaptive transmission power): Assuming the warehouse operation area shelf layout is adjusted, the original anchor point A has initial coordinates of (5,3,4)m. After the new shelf is added, the signal propagation between the anchor point and the AGV tag is blocked. The anchor point needs to be dynamically adjusted to an unobstructed position of (6,3,4)m.

[0019] Known ultra-wideband signal frequency (wavelength c is the speed of light The system is preset to receive signal-to-noise ratio ≥20dB, free-space fundamental propagation loss L=2dB, and tag initial transmit power. .

[0020] Propagation model: The formula for calculating the received power of a signal in free space is as follows: , where d is the distance between the anchor point and the label; Before adjustment: Anchor point A and the initial distance of the label Substituting into the formula, we get ; Shelf obstruction increases transmission loss Actual received power The signal-to-noise ratio is insufficient. Adjusted status: New anchor point coordinates (6,3,4)m, distance from label The tag adjusts its transmission power using an adaptive transmission power adjustment function. Substituting into the formula, we get It meets the requirement of a signal-to-noise ratio of ≥20dB to ensure stable acquisition of location-aware data.

[0021] Inertial measurement unit (rigidly connected to the target): An inertial measurement unit (IMU) is rigidly fixed to the bottom of the warehouse AGV with bolts. A sampling frequency of 10Hz is used to collect velocity data at five consecutive time points to deduce motion characteristics. .

[0022] Acceleration calculation formula: ; Calculation results: part ; part ; part ; It accurately calculates the motion trend of the AGV, which is "uniform acceleration-deceleration". Comparative verification: If the connection is not rigid, vibration interference will cause the velocity measurement value at time t2 to deviate by [value missing]. The corresponding calculated value of a2 is The deviation reached 40%, highlighting the role of rigid connection in ensuring the accuracy of motion data acquisition.

[0023] Machine vision module (multi-camera collaboration + image semantic segmentation): Three high-definition cameras were deployed in the warehouse storage area at coordinates (0,0,5)m, (10,0,5)m, and (5,8,5)m. The target goods were cubic cardboard boxes with a side length of 0.4m (with QR codes on the surface), which were partially obscured by the shelves.

[0024] Calibration parameters: The camera pixel-actual size calibration ratio is 10 pixels / mm, or 100 pixels / m; 2. Single-camera data extraction: Due to occlusion, camera 1 only recognized 1 / 3 of the goods' outline. Through semantic segmentation, the outline pixels were extracted as 400 pixels long and 320 pixels wide, which translates to an actual size of 0.4m long and 0.32m wide. Camera 2 was farther away, and the QR code pixel ratio was small. The extracted outline pixels were 400 pixels long, 400 pixels wide, and 380 pixels high, which translates to an actual size of 0.4m × 0.4m × 0.38m. Camera 3 was unobstructed and had a vertical viewing angle. The extracted outline pixels were 400 pixels long, 400 pixels wide, and 400 pixels high, which translates to an actual size of 0.4m × 0.4m × 0.4m. At the same time, the QR code clarity feature value was extracted as 0.98. 3. Collaborative Fusion: By stitching together the semantic segmentation results from the three cameras and removing biased data caused by occlusion, a complete visual feature vector of the goods, M=(0.4,0.4,0.4,0.98,0.85), is obtained (length, width, height, QR code clarity, and surface texture clarity, respectively), avoiding the problem of feature loss from a single camera.

[0025] Distributed environmental monitoring module (spatial grid deployment): The 20m×15m×6m storage operation area is divided into a 2m×2m×2m spatial grid (a total of 10×8×3=240 grids). An environmental monitoring sensor is deployed at the center of each grid to collect the electromagnetic interference intensity (x, y, z axis components, with a value range of 0-1, and the larger the value, the stronger the interference).

[0026] Basic data: The AGV travels through grids G1 (center coordinates (2,2,2)m) and G2 (center coordinates (4,2,2)m). The sensor collects the interference vector D1=(0.12,0.10,0.08) for G1 and the interference vector D2=(0.15,0.13,0.09) for G2. Interpolation calculation: When the AGV travels to the midpoint between G1 and G2 (coordinates (3,2,2)m), the interference vector at that point is calculated using linear interpolation. The formula is as follows: (d1 and d2 are the distances from the midpoint to G1 and G2, respectively, both 1m), substituting them, we get... Similarly Finally, the midpoint interference vector D=(0.135,0.115,0.085) is obtained, achieving spatial continuous coverage of environmental interference data.

[0027] In step S2, the multimodal fusion preprocessing includes a noise suppression process, which employs an adaptive wavelet threshold denoising algorithm. The process of calculating the dynamic threshold is as follows: First, the data is decomposed using wavelet decomposition to obtain wavelet coefficients at each scale; then, the noise variance σ of these coefficients is estimated; finally, based on the noise variance σ and the data length N, the dynamic wavelet threshold τ is calculated. The calculation process is as follows: ; By employing an adaptive wavelet threshold denoising algorithm in multimodal fusion preprocessing, the threshold τ can be dynamically calculated based on the noise variance σ and data length N of the original data across all dimensions. This achieves precise adaptive noise suppression. Compared to fixed threshold denoising, it avoids noise residue caused by an excessively small fixed threshold and prevents distortion of the effective signal caused by an excessively large fixed threshold. It can retain the effective features in multimodal data such as position, motion, environment, and vision to the greatest extent, generating standardized, high-quality fused data. This provides reliable data support for feature extraction and precise positioning in subsequent dynamic tracking models, improving the stability and positioning accuracy of target tracking within warehouses.

[0028] Using the distance data from the AGV to the anchor point collected by the ultra-wideband positioning module of the warehouse AGV as the processing object, the original data containing Gaussian white noise was simulated and processed by an adaptive wavelet threshold denoising algorithm to verify the denoising effect.

[0029] Raw data acquisition and noise simulation: The ultra-wideband positioning module acquires the actual distance from the AGV to the anchor point at a sampling frequency of 10Hz (constantly). A total of N=10 sampling points were collected. Gaussian white noise (mean) was added to the raw data due to electromagnetic interference. noise variance Noisy raw data .

[0030] Noise variance σ calculation: Step 1, extract the noise sequence: Since the true distance is constant, the noise sequence... (k=1,2,...,10), we calculate n=[0.05,-0.08,0.13,-0.12,0.07,-0.05,0.10,-0.15,0.03,-0.10]; The second step is to calculate the noise variance: Substituting the data yields .

[0031] Calculation of dynamic threshold τ: According to the formula Substitute With N=10, we get .

[0032] Wavelet denoising: Wavelet decomposition: The db4 wavelet (commonly used for denoising warehouse data) was selected for the original data. Perform a single-level wavelet decomposition to obtain the low-frequency coefficients. (Effective signal dominance) and high-frequency coefficients (Noise-dominated); High-frequency coefficient thresholding: Soft thresholding is used (to avoid signal abrupt changes caused by hard thresholding), and the processing rules are as follows: Substitute ,right Coefficients whose absolute value is less than τ are set to 0, and coefficients whose absolute value is greater than τ are subtracted from τ. Wavelet Reconstruction: Based on Processed and Wavelet reconstruction is performed to obtain the denoised data. .

[0033] Denoising effect verification: Calculate the mean square error (MSE) before and after denoising: MSE before noise reduction: ; MSE after noise reduction: ; The comparison shows that the MSE is reduced after denoising, effectively suppressing noise without distorting the effective signal. If a fixed threshold is used (e.g.) This will result in some residual noise ( If a fixed threshold is used This will result in the loss of effective signal details (overly smoothed data after noise reduction). This highlights the advantages of adaptive dynamic thresholds.

[0034] In step S3, the features used by the dynamic tracking model are extracted from the standardized fused data through a feature extraction network; the feature extraction network is a hybrid architecture of convolutional neural network and long short-term memory network. The convolutional neural network introduces an attention module to enhance key spatial features. The calculation process of the attention module is as follows: the input basic feature map F undergoes two linear transformations and a non-linear activation function is applied. Finally, the attention weight matrix A is generated through the sigmoid function, and its output formula is: ; Where F is the basic feature map. Here, W1 and W2 are weight matrices, and b1 and b2 are bias vectors; The forgetting gate of the Long Short-Term Memory network introduces a regularization term based on historical states. The calculation process is as follows: output the sequence of hidden states from the first k time steps. The summation is multiplied by the regularization coefficient λ, then normalized using the softmax function, and finally multiplied by the original forget gate output ft to obtain the regularized forget gate output. The calculation formula is as follows: ; By employing a hybrid architecture of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) as the feature extraction network, this approach leverages the powerful spatial feature extraction capabilities of CNN to accurately capture key spatial information such as the location distribution and visual contours of target objects. Simultaneously, it utilizes the temporal modeling advantages of LSTM to deeply explore the temporal correlation patterns of features such as motion speed and trajectory changes. Furthermore, an attention module is introduced into the CNN to dynamically generate weight matrices, enhancing the representation strength of key spatial features. A regularization term based on historical states is introduced into the LSTM forget gate to optimize memory performance through historical information normalization. This effectively avoids the limitations of single-network architectures in spatial or temporal feature extraction, reduces redundant information interference and the risk of model overfitting, and significantly improves the representation accuracy and stability of multi-dimensional features. This provides high-quality feature support for the efficient fusion of multimodal features and accurate calculation of real-time position coordinates in dynamic tracking models, thereby enhancing the accuracy and robustness of target tracking within warehouses.

[0035] In the AGV warehouse tracking scenario, the AGV moves within a 10m×10m×6m warehouse area. The standardized fused data has been preprocessed by adaptive wavelet threshold denoising. The feature extraction network is a hybrid architecture of 3-layer CNN (including attention module) and 2-layer LSTM. Feature extraction is performed on the standardized fused data of the AGV at time t to verify the actual effect of the attention module and LSTM regularization term.

[0036] CNN Attention Module (Enhancing Key Spatial Features): Basic parameter settings: The input basic feature map F is a 4×4×3 three-dimensional tensor (height×width×number of channels, with channels corresponding to position, visual, and environmental features respectively), which is flattened into a 48-dimensional vector F. flat =[0.21,0.32,0.15,0.28,0.30,0.18,0.25,0.29,0.16,0.33,0.31,0.17,0.22,0.27,0.14,0.26,0.34,0.19,0.23,0.24,0.13,0.20,0.35,0.12,0.29,0.28,0.11,0.27,0.36,0.10,0.24,0.23,0.09,0.22,0.37,0.08,0.21,0.25,0.07,0.19,0.38,0.06,0.18,0.26,0.05,0.17,0.39,0.04]; weight matrix (Mapping 48-dimensional features to 24-dimensional features). (Mapping 24-dimensional features back to 48-dimensional); Bias vector b1 = [0.12, 0.21, 0.13, 0.32, 0.24, 0.11, 0.15, 0.26, 0.17, 0.33, 0.25, 0.16, 0.18, 0.27, 0.14, 0.31, 0.23, 0.19, 0.20, 0.29, 0.10, 0.34, 0.22, 0.09] T , b2=[0.05,0.03,0.04,0.06,0.02,0.07,0.08,0.01,0.09,0.04,0.03,0 .05,0.06,0.02,0.07,0.01,0.08,0.03,0.09,0.02,0.04,0.06,0.01,0. 07,0.03,0.05,0.08,0.02,0.06,0.04,0.09,0.01,0.07,0.03,0.05,0.0 8,0.02,0.06,0.04,0.09,0.01,0.07,0.03,0.05,0.08,0.02,0.06,0.04] T ; Sigmoid function The tanh function is the hyperbolic tangent activation function.

[0037] Linear transformation and activation computation: First, calculate The 24-dimensional intermediate vector M1 = [1.23, 0.85, 0.92, 1.18, 1.05, 0.98, 1.02, 0.89, 1.11, 0.95, 1.03, 0.87, 1.15, 0.91, 1.07, 0.83, 1.09, 0.94, 1.12, 0.88, 1.01, 0.96, 1.17, 0.81] is obtained through matrix-vector multiplication. T ; After activation with tanh, the output is T=[0.86,0.76,0.78,0.83,0.78,0.76,0.77,0.71,0.81,0.75,0.78,0.70,0.82,0.72,0.79,0.68,0.80,0.74,0.81,0.71,0.77,0.75,0.83,0.67]. T .

[0038] Attention weight matrix generation: calculation The 48-dimensional intermediate vector M2 = [0.95, 0.73, 0.84, 0.91, 0.76, 0.88, 0.89, 0.70, 0.92, 0.74, 0.82, 0.71, 0.90, 0.75, 0.86, 0.69, 0.87, 0.72, 0.93, 0.77, 0.81, 0.78] is obtained through matrix-vector multiplication. ,0.94,0.68,0.85,0.79,0.67,0.83,0.96,0.66,0.80,0.76,0.65,0.79,0.97,0.64,0.78,0.84,0.63,0.77,0.98,0.62,0.76,0.85,0.61,0.75,0.99,0.60] T ; Attention weight matrix generated by Sigmoid activation =[0.72,0.68,0.70,0.71,0.68,0.70,0.71,0.67,0.71,0.68,0.70,0.67,0.71,0.68,0.70,0.67,0.70,0.68,0.72,0.69,0.70,0.69,0.72,0.67,0.70,0.69,0.66,0.70,0.72,0.66,0.70,0.72,0.66,0.70,0.68,0.66,0.69,0.72,0.65,0.69,0.70,0.65,0.69,0.72,0.65,0.68,0.70,0.65,0.68,0.72,0.65].

[0039] Key feature enhancement effect: In weight matrix A, the weights of dimensions 10-15 corresponding to the current position feature of the AGV are 0.70-0.72, and the weights of dimensions 25-30 corresponding to the visual contour feature are 0.70-0.72, both higher than the 0.65-0.67 of the environmental interference feature channel, achieving accurate enhancement of key spatial features; if the attention module is not introduced, all feature weights are 1.0, and key features such as position and vision will be diluted by redundant environmental interference features, increasing the error of subsequent positioning calculation by about 23%, highlighting the role of the attention module in improving the accuracy of feature representation.

[0040] LSTM forgetting gate regularization (optimizing sequential memory): Basic parameter settings: k=3 (first 3 time steps), regularization coefficient The hidden states in the first three time steps are as follows: (Corresponding to position, motion, and visual feature components). , Original forget gate output (The three components correspond to the forgetting weights of the three types of features, respectively.) softmax function .

[0041] Historical State Summation and Normalization: Calculate the sum of the hidden states in the first three time steps. ; Multiply by regularization coefficient have to ; Calculated after softmax normalization: The result after normalization is .

[0042] Forget gate output after regularization: According to the formula Calculated .

[0043] Temporal memory optimization effect: After regularization, the forgetting gate contributes more to the positional features that have a high contribution to historical states ( The mid-position component (0.63) retains more weight to avoid excessive forgetting of effective time-series information; if no regularization term is introduced, When historical states are subject to noise interference (such as...) When the location component is abnormally low (0.12), the original forgetting gate will incorrectly forget valid temporal information, resulting in a decrease in the continuity of temporal features and an increase in subsequent localization errors of approximately 18%. After introducing a regularization term, the stability of the temporal feature representation is significantly improved by summing historical states and smoothing out noise through normalization.

[0044] In step S3, the process of calculating the real-time accurate position coordinates includes: The first step is to extract the location feature vector X, motion trend vector V, environmental interference vector D, and visual matching vector M from the standardized fused data; The second step is to dynamically generate corresponding attention weight coefficients based on the current features through a trainable weight learning model, including position feature weight α, motion feature weight β, environment correction weight γ, and visual assistance weight δ. The third step is feature fusion: the feature vector is multiplied by its corresponding attention weight coefficient, then weighted and summed, and the sum is normalized to obtain the comprehensive feature vector F. The calculation formula for this fusion process is as follows: ; Step 4, Coordinate Mapping: Input the comprehensive feature vector F into the pre-trained 3D spatial calibration model and calculate its coordinate values ​​on the x, y, and z axes respectively. The mapping process is as follows: multiply the components of F on the corresponding axes by the spatial calibration coefficient, then multiply by the correction term composed of the environmental correction weight γ and the environmental disturbance vector components, and finally add the origin reference value of the coordinate axes. The mapping formulas are as follows: ; ; ; Where Fx, Fy, Fz and Dx, Dy, Dz are the components of vectors F and D in the three coordinate axes, respectively, Sx, Sy, Sz are the pre-calibrated spatial calibration coefficients, and (x0, y0, z0) is the reference value of the origin of the coordinate system. By defining a four-step calculation process for real-time and accurate location coordinates, the system first extracts four core feature vectors—position, movement trend, environmental interference, and visual matching—from standardized fused data. Then, a trainable weighted learning model dynamically generates attention weight coefficients adapted to the current scene, achieving adaptive weighted fusion of multi-dimensional features. Finally, a pre-trained 3D spatial calibration model and an environmental correction term are combined to complete coordinate mapping. This approach takes into account the differentiated contributions of different features to positioning while using the environmental correction term to offset the impact of environmental interference on coordinate calculation. It avoids positioning deviations caused by reliance on a single feature or fixed weights, significantly improving the accuracy, anti-interference capabilities, and scene adaptability of location coordinate calculation. This provides a reliable real-time location basis for subsequent path optimization algorithms, further ensuring the stability of target tracking within the warehouse.

[0045] In an AGV warehouse tracking scenario, the AGV moves within a 10m×10m×6m warehouse area. Standardized fused data has undergone adaptive wavelet threshold denoising preprocessing. At time t, the real-time 3D position coordinates of the AGV need to be calculated. The preset parameters are as follows: spatial calibration coefficient. The reference value of the origin of the coordinate system Attention weight coefficients dynamically generated by the weighted learning model (satisfy ).

[0046] Feature vector extraction: Four types of 3D feature vectors (corresponding to x, y, and z axis components) are extracted from the standardized fused data: Location feature vector (Data extraction based on ultra-wideband positioning module); Motion trend vector (Calculated based on motion data from inertial measurement units); Environmental disturbance vector D = (0.05, 0.03, 0.02) (based on data collected by a distributed environmental monitoring module, dimensionless); Visual matching vector (Based on machine vision semantic segmentation results extraction).

[0047] Feature fusion calculation: according to the formula (Since the sum of the weights is 1, it is simplified to) ), calculate the x, y, and z-axis components of the comprehensive eigenvector F respectively: ; ; ; That is, the comprehensive feature vector F = (1.4575, 1.7295, 0.428).

[0048] Coordinate mapping calculation: Calculate the real-time position coordinates by substituting the parameters according to the x, y, z axis coordinate mapping formula: x-axis coordinate: ; y-axis coordinate: ; z-axis coordinate: ; The final real-time accurate position coordinates of the AGV are (2.1716, 2.5633, 0.9293)m.

[0049] If dynamic attention weights and environment correction terms are not used, fixed weights are employed. Without environmental correction, the calculation yields: ; The deviation from the actual position is approximately 0.45m; The deviation between the above calculation results and the actual position of the AGV (2.2, 2.6, 0.9)m is only within 0.03m, highlighting the role of dynamic weight fusion and environmental correction in improving positioning accuracy and effectively offsetting the impact of environmental interference and dependence on single features.

[0050] The path optimization algorithm includes path evaluation and dynamic triggering steps: The path evaluation uses an improved A* algorithm to calculate the evaluation function value f(n) from the starting node to any node n to be evaluated in the search graph consisting of path nodes within the warehouse. The evaluation function f(n) consists of three parts: The first part, g(n), represents the actual path cost from the starting node to the current node n, which is obtained by accumulating the weights of the edges along the path. The second part, h(n), is the heuristic estimated cost from the current node n to the target node, calculated using Euclidean distance or Manhattan distance. The third part is the congestion cost term, which is obtained by multiplying the real-time congestion coefficient C(n) of the path segment where node n is located by a preset congestion impact weight ω; the real-time congestion coefficient C(n) is calculated based on the current occupancy status data of the associated cargo spaces of the path segment, and the higher the cargo space occupancy rate, the larger the value of C(n); the calculation formula of the evaluation function f(n) of the improved A* algorithm is as follows: ; The dynamic triggering step uses the rate of change in cargo space occupancy, the rate of change in movement speed, and the rate of change in environmental disturbance as dynamic adjustment factors. Among them, the rate of change in storage space occupancy The calculation process is as follows: Take the number of occupied storage spaces within the path planning area at the current time t. The number of occupied storage spaces compared to the previous time t-1 The absolute value of the difference is then divided by the total number of storage locations in that area. The calculation process is as follows: ; Rate of change of velocity The calculation process is as follows: take the absolute value of the difference between the target object's velocity V(t) at the current time t and the velocity V(t-1) at the previous time t-1, and then divide it by the historical average velocity Vavg. Environmental disturbance change rate The calculation process is as follows: Take the absolute value of the difference between the environmental disturbance factor evaluation value D(t) at the current time t and the evaluation value D(t-1) at the previous time t-1, and then divide it by the historical average environmental disturbance factor evaluation value Davg. When any of the dynamic adjustment factors exceeds its preset threshold, the replanning of the tracking path is triggered; This path optimization algorithm combines path evaluation and dynamic triggering. The path evaluation uses an improved A* algorithm that incorporates a congestion cost term, transforming the real-time occupancy status of storage locations into a congestion coefficient that is integrated into the evaluation function. This ensures that path planning considers both actual costs and estimated distances, while also adapting to the dynamic occupancy of storage locations, thus avoiding planned paths that pass through congested areas. The dynamic triggering step uses the rates of change of three factors—storage location occupancy, movement speed, and environmental interference—as adjustment factors. When any factor exceeds a preset threshold, path replanning is automatically triggered, achieving real-time dynamic updates of the path. This effectively solves the problems of inefficiency or congestion caused by the static nature of traditional path planning and its failure to consider dynamic changes in the warehouse. It improves the optimality, real-time performance, and adaptability of tracking paths, ensuring the smooth and efficient movement of target objects (such as AGVs) within the warehouse.

[0051] In scenarios such as AGV warehouse tracking, the AGV needs to start from the initial node. Move to target node In the search graph formed by the warehouse path nodes, the node n to be evaluated is The preset parameters are as follows: Congestion impact weight Threshold for change in storage space occupancy rate Threshold for rate of change of motion speed Environmental disturbance change rate threshold Total number of storage spaces within the route planning area Historical average speed of AGV Historical average environmental disturbance factor evaluation value .

[0052] Improved A* algorithm path evaluation: Basic parameter settings: Actual path cost g(n): From the starting node S(1,1) to the node to be evaluated n(5,5), passing through nodes (1,3) and (3,5), the weights (path lengths) of each edge are as follows. , cumulatively ; Heuristic cost estimation h(n): Calculated using Euclidean distance, the formula is as follows: Substitute ,have to ; Congestion coefficient C(n): The number of currently occupied storage spaces associated with the path segment containing node n. Cargo space occupancy rate The congestion coefficient C(n) is set to be linearly related to the occupancy rate (with a value range of 0-1), so C(n) = 0.7.

[0053] Evaluation function calculation: Based on the improved A* algorithm evaluation function formula Substituting the data, we get: .

[0054] If the traditional A* algorithm (without congestion cost) is used, the evaluation function is: While seemingly superior, the traditional algorithm fails to account for congestion at 70% occupancy on that route segment. Consequently, the AGV's actual travel speed drops to 0.2 m / s due to congestion, and the travel time is approximately [missing information]. The optimal path planned by the improved A* algorithm has a storage space occupancy rate of only 30% (C(n) = 0.3), corresponding to an evaluation function value of... The AGV maintains a travel speed of 0.5 m / s, and the travel time is approximately... This significantly improves traffic efficiency and highlights the core role of congestion cost in route optimization.

[0055] Dynamic trigger path replanning: Basic data collection: Storage space occupancy data: Number of storage spaces occupied in the route planning area at time t-1 time t ; Motion speed data: AGV speed at time t-1 velocity at time t ; Environmental disturbance data: Environmental disturbance evaluation value D(t-1) = 0.08 at time t-1, and D(t) = 0.13 at time t.

[0056] Calculation of dynamic adjustment factor: Cargo space occupancy change rate According to the formula Substituting the data yields ; Rate of change of velocity According to the formula Substituting the data yields ; Environmental disturbance change rate According to the formula Substituting the data yields .

[0057] Trigger judgment and optimization effect: Threshold comparison: , , If all three factors exceed the preset threshold, path replanning is triggered. Optimization Results: The original planned route passed through the congested section where node n(5,5) was located. After replanning, the route was adjusted to S(1,1)→(1,5)→(5,7)→T(9,9). The new route has a cargo space occupancy rate of only 28%, with no significant environmental interference. The AGV's travel speed remains stable at 0.5m / s, avoiding congestion and delays caused by a sudden increase in cargo space occupancy, abrupt speed changes, and increased environmental interference, thus ensuring mobility efficiency and stability.

[0058] In step S5, the anomaly identification employs a two-layer mechanism: The first layer of identification is based on the dynamic 3σ criterion. Its calculation process is as follows: Within a sliding time window, the mean μ and standard deviation σ of the data are calculated, and the adaptive coefficient k(t) that changes with time is obtained. The mean, standard deviation, and adaptive coefficient are combined to calculate the dynamic anomaly detection threshold θ. The calculation process is as follows: ; The second layer of recognition is based on a graph neural network model for semantic anomaly detection. Its calculation process includes constructing a feature graph and calculating the correlation between nodes: using data samples as graph nodes, the semantic correlation Sij between any two samples i and j is calculated; this calculation first yields the cosine similarity between the feature vectors of the two samples. Then multiply by an exponential decay term with time distance dij as the parameter. Finally, this result is normalized by dividing by the sum of the products of the similarity between sample i and all other samples k and the attenuation term. The formula for calculating semantic relevance is: ; By employing a two-layer anomaly detection mechanism combining the dynamic 3σ criterion and graph neural network semantic recognition, the first-layer dynamic 3σ criterion generates precise thresholds based on the mean, standard deviation, and adaptive coefficients dynamically adjusted over time within a sliding time window, quickly capturing sudden anomalies in numerical values ​​that deviate from the normal range. The second-layer graph neural network accurately identifies semantic anomalies in temporal logic and feature association by constructing feature maps and calculating the semantic correlation between samples. The two mechanisms work synergistically and complementarily, avoiding the omission of semantic logic anomalies by single numerical recognition and solving the problem of delayed response to sudden numerical anomalies by pure semantic recognition. This significantly improves the comprehensiveness, accuracy, and real-time performance of anomaly detection, effectively filtering various anomalies during the tracking process, providing accurate basis for subsequent adaptive correction, and ensuring the reliability and stability of target tracking results within the warehouse.

[0059] In the AGV warehouse tracking scenario, the AGV collects x-axis position coordinate data at a sampling frequency of 10Hz. The sliding time window length is set to 5 sampling points. The adaptive coefficient k(t) is dynamically adjusted according to environmental interference (k(t) = 3.0 when the environmental interference factor D = 0.1, and k(t) is larger when the interference is greater). The sample nodes in the graph neural network are the position feature vectors of the AGV continuously sampled (including x-axis position, movement speed, and environmental interference components). The time distance benchmark value d0 = 2s, and a total of N = 5 data samples (times t1-t5) are collected for semantic anomaly judgment.

[0060] First-level dynamic 3σ criterion (identifying numerical anomalies): Basic data acquisition: The x-axis position data of 5 sampling points within the sliding time window is x=[2.17,2.19,2.18,3.50,2.20]m (the data at time t4 is abnormal due to temporary sensor interference).

[0061] Calculation of mean and standard deviation: mean ; Standard deviation: .

[0062] Dynamic threshold calculation: Current environmental interference factor D = 0.1, adaptive coefficient k(t) = 3.0, according to the formula... ,have to The lower threshold is .

[0063] Anomaly detection: The data at time t4, 3.50m, is within the range of [0.708, 4.188], and no numerical anomaly has been triggered; If environmental disturbances intensify (D=0.3, k(t)=3.5), then If the data at time t4 is 4.60m, it exceeds the threshold and is judged as a numerical anomaly, thus realizing the identification of numerical anomalies that dynamically adapt to environmental interference.

[0064] Second-layer graph neural network semantic anomaly recognition (identifying logical anomalies): Sample feature vectors: The feature vectors at 5 time points are , , , , (The position, velocity, and interference components of X4 all deviate from those of other samples).

[0065] Time distance calculation: The time interval between adjacent moments is Δt = 0.1s, and the time distance is... .

[0066] Semantic relevance calculation (taking sample i=4 as an example): Cosine similarity calculation: Similarly, we can calculate... ; Calculation of the exponential decay term: Similarly ; Molecular calculations (taking j=1 as an example): ; Denominator calculation: ; Semantic relevance Similarly, we can calculate... All of them are below the normal correlation threshold of 0.4.

[0067] Anomaly detection: Based on the first layer of numerical recognition results, the semantic relevance of X4 is below the threshold, which is determined to be a semantic logic anomaly (abnormal position accompanied by speed and interference, which does not conform to the normal movement logic of AGV). The abnormal data is confirmed by double verification.

[0068] It also includes step S6: constructing a trajectory knowledge graph of the target object; Graph embedding algorithms are used to transform graphs into low-dimensional vector representations; Based on the low-dimensional vector, motion patterns are analyzed through association rule mining. The calculation process for the support of the association rule X→Y is as follows: The formula for calculating the number of samples that contain both events X and Y is: count(X∪Y), divided by the total number of samples count(U). ; Output an operational optimization plan based on the analysis results; By adding steps for constructing a trajectory knowledge graph, graph embedding transformation, and association rule mining, the tracking data of target objects is extended from simple location records to mineable operational data assets. It leverages the knowledge graph to structurally integrate the temporal and spatial correlation information of trajectories, and uses graph embedding algorithms to transform high-dimensional graphs into low-dimensional vectors to simplify feature processing. Furthermore, association rules support the measurement of the correlation strength of motion patterns, accurately identifying potential patterns such as high-frequency paths and storage location correlation needs of target objects (such as AGVs). This avoids the limitations of passive scheduling based solely on tracking results, providing data-driven decision-making basis for storage location layout optimization, transportation route planning, and resource allocation adjustment. It achieves an upgrade from "precise tracking" to "intelligent operation," further improving the overall efficiency, rationality, and intelligence level of warehouse management.

[0069] For example, in an AGV warehouse tracking scenario, the warehouse includes an inbound area (A), an outbound area (B), frequently used storage locations (C), less frequently used storage locations (D), and a transit area (E). The AGV has cumulatively completed a total of [number] transport tasks. The trajectory knowledge graph nodes are "pick up goods in warehouse area A" (event X), "pick up goods in storage location C" (event M), "park in transit area E" (event N), "deliver goods in warehouse area B" (event Y), and "pick up goods in storage location D" (event P), with edges representing the temporal relationships between events. The graph embedding uses the Node2Vec algorithm to transform the graph nodes into 2D low-dimensional vectors. The association rule mining focuses on the association relationships between "pick up goods - delivery goods" and "storage location - area".

[0070] Constructing a trajectory knowledge graph: Based on the complete trajectory of each AGV transportation task, the nodes and edges of the graph are constructed: nodes are the above 5 types of events (including area and location-related actions), and edges represent the temporal connection relationship between events (such as the graph path X→M→N→Y corresponding to the trajectory "A picks up goods → C picks up goods → E stops → B delivers goods"). A total of 1,000 trajectory data are integrated to form a structured knowledge graph, which clearly presents the frequency of association and temporal logic between events.

[0071] Graph embedding transforms into low-dimensional vectors: The Node2Vec algorithm is used, with a walk length l=4, window size w=2, return parameter p=1, and input / output parameters q=2. After traversing the graph to generate a node sequence, the high-dimensional graph nodes are transformed into 2D low-dimensional vectors through Skip-gram model training. Partial node vector results are as follows: X=(1.23,0.81) (Pick up goods in warehouse area A), Y=(1.19,0.88) (Deliver goods in warehouse area B), M=(0.92,0.73) (Pick up goods at location C), N=(0.51,0.62) (Dock at transit area E), P=(0.32,0.21) (Pick up goods at location D); Vector cosine similarity verification: This confirms the strong correlation between X and Y.

[0072] Association rule mining (support calculation): Focusing on core operational association rules, we statistically analyze the number of co-occurrences of events in the sample and calculate support. Rule 1: X→Y (goods are picked up in warehouse area A and delivered to warehouse area B), count the number of samples that simultaneously contain both X and Y. According to the formula ,have to ; Rule 2: M→Y (Goods are picked up at location C and delivered to outbound area B). Calculated ; Rule 3: P→Y (After picking up goods at location D, deliver the goods to the outbound area B). Calculated ; Rule 4: X→M (After picking up goods in warehouse area A, proceed to warehouse location C first). Calculated .

[0073] Analysis results: The support of rules 1, 2, and 4 is ≥0.58, indicating that the correlation between the AGV picking up goods from the inbound area A, prioritizing the replenishment of goods at storage location C, and ultimately delivering goods to the outbound area B is extremely strong; the support of rule 3 is only 0.08, indicating that the transportation correlation between storage location D and outbound area B is extremely weak. Optimization plan: Adjust the location C (highly associated location) from its original coordinates (8,6,2)m to (5,5,2)m between the inbound area A (1,2,2)m and the outbound area B (9,8,2)m, shortening the transportation distance of high-frequency paths; add a dedicated channel between the inbound area A and location C to reduce congestion; reduce the proportion of frequently used goods inventory in location D, and transfer 80% of frequently used goods to location C to improve AGV picking efficiency.

[0074] Before optimization, the average transportation time for the AGV to execute the X→M→Y path was 80 seconds. After optimization, the average transportation time was reduced to 62 seconds due to the adjustment of the location of storage location C and the addition of a dedicated channel, thus improving efficiency. After frequently used goods were concentrated in storage location C, the number of invalid detours by the AGV decreased from 32 times per day to 8 times, further verifying the practical value of association rule mining for operational optimization.

[0075] In step S5, for abnormal data in strong interference scenarios, a weighted voting mechanism is used to perform multimodal data cross-validation correction. This correction process includes: First, calculating the voting weight of each modality: Based on the historical tracking error rate (error(m)) of the m-th modality, its credibility (credibility(m)) is calculated. The calculation process involves subtracting the historical error rate of the modality from 1, then dividing by the sum of (1 - historical error rates) of all modalities to achieve weight normalization. The formula for calculating credibility is: ; Then, weighted fusion is performed: the independent correction result Rm of the m-th modality data is multiplied by its credibility(m), and the product results of all M modalities are summed to obtain the final correction result R, which is calculated using the following formula: ; For abnormal data in scenarios with strong interference, a weighted voting mechanism is used to achieve cross-validation correction of multimodal data. First, a normalized confidence weight is calculated based on the historical tracking error rate of each modality. Then, the independent correction results of each modality are weighted and fused with their corresponding weights to obtain the final result. This not only gives higher decision weights to modalities with low error rates and high confidence, but also offsets the interference bias of a single modality through multimodal cross-validation. This avoids the inaccuracy problems caused by traditional single-modal correction or equal-weighted fusion, significantly improving the accuracy and stability of abnormal data correction in scenarios with strong interference, ensuring the reliability of the final tracking results, and enhancing the adaptability of intelligent tracking methods in warehouses to complex interference environments.

[0076] In an AGV warehouse tracking scenario, when the AGV moves in a strong electromagnetic interference area (environmental interference factor D=0.4), the x-axis position data becomes abnormal, requiring correction through cross-validation of multimodal data. This involves four core modal data types: ultra-wideband positioning mode (modulus 1, m=1), inertial measurement mode (modulus 2, m=2), machine vision mode (modulus 3, m=3), and environmental correction auxiliary mode (modulus 4, m=4). The true x-axis position of the AGV is... .

[0077] Basic parameter settings: Historical tracking error rates for each modality: error(1)=0.08 (ultra-wideband positioning, higher error due to electromagnetic interference), error(2)=0.03 (inertial measurement, low error due to rigid connection), error(3)=0.05 (machine vision, medium error due to multi-camera collaboration), error(4)=0.10 (environmental correction, higher error due to auxiliary modality). Independent correction results for each modality (correction values ​​for abnormal original data): R1=2.35m (ultra-wideband positioning correction result), R2=2.18m (inertial measurement correction result), R3=2.22m (machine vision correction result), R4=2.40m (environmental correction auxiliary result).

[0078] Calculate the credibility (voting weight) of each modality: According to the credibility formula First, calculate the sum of the denominators: ; Then calculate the confidence level of each mode separately: ; ; ; ; The weights are verified to be 0.246 + 0.259 + 0.254 + 0.241 = 1, which satisfies the normalization requirement.

[0079] Final corrected result of weighted fusion calculation: According to the weighted fusion formula Substitute the data to calculate: R=0.246×2.35+0.259×2.18+0.254×2.22+0.241×2.40≈0.578+0.565+0.564+0.578≈2.285m.

[0080] Single-mode correction deviation: If only the ultra-wideband positioning mode (mode 1) with a higher error rate is used for correction, the result R1 = 2.35m, and the deviation from the true value is |2.35-2.2| = 0.15m; if only the environmental correction mode (mode 4) is used, the deviation is |2.40-2.2| = 0.20m. Correcting bias through equal-weighted fusion: If confidence levels are not weighted, and equal-weighted fusion (each weight 0.25) is used, the result... The deviation is |2.2875 - 2.2| = 0.0875m; Correction bias: |2.285-2.2|=0.085m, the bias is reduced by a maximum of 63.3% compared with single mode and by 2.9% compared with equal weight fusion; Strong interference reinforcement verification: If the electromagnetic interference intensifies and the error rate of mode 1 rises to error(1)=0.15, the confidence level is recalculated. With the weights reduced, the final correction result R≈2.27m, and the deviation was still controlled within 0.07m, highlighting the adaptive correction capability of the weighted voting mechanism for strong interference scenarios and effectively avoiding the impact of single-mode error expansion.

[0081] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0082] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0083] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A data analysis-based intelligent tracking method for warehouses, characterized in that, Includes the following steps: S1: Deploy a multi-source heterogeneous sensing network to collect full-dimensional raw data of target objects in the warehouse. The full-dimensional raw data includes location sensing data, motion feature data, environmental interference data, and visual semantic data. S2: Perform multimodal fusion preprocessing on the raw data across all dimensions to generate standardized fused data; S3: Based on standardized fused data, the dynamic tracking model calculates the real-time accurate location coordinates of the target object. The dynamic tracking model uses an attention mechanism to fuse multi-dimensional features. S4: Based on real-time accurate location coordinates and the dynamic status of goods in the warehouse, a path optimization algorithm is used to generate and update the tracking path; S5: Performs anomaly identification and adaptive correction on the data during the tracking process, and outputs the final tracking result.

2. The intelligent tracking method for warehouses based on data analysis according to claim 1, characterized in that: In step S1, the multi-source heterogeneous sensing network includes: The ultra-wideband positioning module has dynamically adjustable anchor points, and the tag has an adaptive transmission power adjustment function. An inertial measurement unit, rigidly connected to the target object, is used to collect motion data and calculate motion characteristics; The machine vision module employs multi-camera collaboration and extracts visual features through image semantic segmentation; The distributed environmental monitoring module is deployed according to a spatial grid and is used to collect environmental physical quantity data of each grid.

3. The intelligent tracking method for warehouses based on data analysis according to claim 2, characterized in that: In step S2, the multimodal fusion preprocessing includes a noise suppression process, which employs an adaptive wavelet threshold denoising algorithm. The process of calculating the dynamic threshold is as follows: First, the data is decomposed into wavelet coefficients at each scale; then, the noise variance σ of these coefficients is estimated; and finally, the dynamic wavelet threshold τ is calculated based on the noise variance σ and the data length N.

4. The intelligent tracking method for warehouses based on data analysis according to claim 3, characterized in that: In step S3, the features used by the dynamic tracking model are extracted from the standardized fused data through a feature extraction network; the feature extraction network is a hybrid architecture of convolutional neural network and long short-term memory network. The attention module is introduced into the convolutional neural network to enhance key spatial features. The calculation process of the attention module is as follows: the input basic feature map F is subjected to two linear transformations and a non-linear activation function is applied. Finally, the attention weight matrix A is generated by the sigmoid function. The forgetting gate in Long Short-Term Memory (LSTM) networks introduces a regularization term based on historical states. The calculation process involves outputting the hidden state sequence from the first k time steps. The summation is multiplied by the regularization coefficient λ, then normalized using the softmax function, and finally multiplied by the original forget gate output ft to obtain the regularized forget gate output. .

5. The intelligent tracking method for warehouses based on data analysis according to claim 4, characterized in that: Step S3, the process of calculating the real-time precise location coordinates includes: The first step is to extract the location feature vector X, motion trend vector V, environmental interference vector D, and visual matching vector M from the standardized fused data; The second step is to dynamically generate corresponding attention weight coefficients based on the current features through a trainable weight learning model, including position feature weight α, motion feature weight β, environment correction weight γ, and visual assistance weight δ. The third step is feature fusion: the feature vector is multiplied by its corresponding attention weight coefficient, then the sum is weighted and normalized to obtain the comprehensive feature vector F. The fourth step is coordinate mapping: the comprehensive feature vector F is input into the pre-trained 3D spatial calibration model, and its coordinate values ​​on the x, y, and z axes are calculated respectively. The mapping process is as follows: multiply the component of F on the corresponding axis by the spatial calibration coefficient, then multiply by the correction term consisting of the environmental correction weight γ and the environmental disturbance vector component, and finally add the reference value of the coordinate axis origin.

6. The intelligent tracking method for warehouses based on data analysis according to claim 5, characterized in that: Path optimization algorithms include path evaluation and dynamic triggering steps: The path evaluation uses an improved A* algorithm to calculate the evaluation function value f(n) from the starting node to any node n to be evaluated in the search graph consisting of path nodes within the warehouse. The evaluation function f(n) consists of three parts: The first part, g(n), represents the actual path cost from the starting node to the current node n, which is obtained by accumulating the weights of the edges along the path. The second part, h(n), is the heuristic estimated cost from the current node n to the target node, calculated using Euclidean distance or Manhattan distance. The third part is the congestion cost item, which is obtained by multiplying the real-time congestion coefficient C(n) of the path segment where node n is located by the preset congestion impact weight ω. The real-time congestion coefficient C(n) is calculated based on the current occupancy status data of the associated cargo space of the path segment. The higher the cargo space occupancy rate, the larger the value of C(n). The dynamic triggering step uses the rate of change in cargo space occupancy, the rate of change in movement speed, and the rate of change in environmental disturbance as dynamic adjustment factors. Among them, the rate of change in storage space occupancy The calculation process is as follows: Take the number of occupied storage spaces within the path planning area at the current time t. The number of occupied storage spaces compared to the previous time t-1 The absolute value of the difference is then divided by the total number of storage locations in that area. ; Rate of change of velocity The calculation process is as follows: take the absolute value of the difference between the target object's velocity V(t) at the current time t and the velocity V(t-1) at the previous time t-1, and then divide it by the historical average velocity Vavg. Environmental disturbance change rate The calculation process is as follows: Take the absolute value of the difference between the environmental disturbance factor evaluation value D(t) at the current time t and the evaluation value D(t-1) at the previous time t-1, and then divide it by the historical average environmental disturbance factor evaluation value Davg. When any dynamic adjustment factor exceeds its preset threshold, the replanning of the tracking path is triggered.

7. The intelligent tracking method for warehouses based on data analysis according to claim 6, characterized in that: In step S5, anomaly detection employs a two-layer mechanism: The first layer of identification is based on the dynamic 3σ criterion. Its calculation process is as follows: calculate the mean μ and standard deviation σ of the data within the sliding time window, and obtain the adaptive coefficient k(t) that changes with time. Combine the mean, standard deviation and adaptive coefficient to calculate the dynamic anomaly judgment threshold θ. The second layer of recognition is based on a graph neural network model for semantic anomaly detection. Its calculation process includes constructing a feature graph and calculating the correlation between nodes: using data samples as graph nodes, the semantic correlation Sij between any two samples i and j is calculated; this calculation first yields the cosine similarity between the feature vectors of the two samples. Then multiply by an exponential decay term with time distance dij as the parameter. Finally, this result is normalized by dividing it by the sum of the products of the similarity between sample i and all other samples k and the decay term.

8. The intelligent tracking method for warehouses based on data analysis according to claim 7, characterized in that: It also includes step S6: constructing a trajectory knowledge graph of the target object; Graph embedding algorithms are used to transform graphs into low-dimensional vector representations; Based on low-dimensional vectors, motion patterns are analyzed through association rule mining. The calculation process for the support of the association rule X→Y is as follows: Count the number of samples that contain both event X and event Y, count(X∪Y), and then divide by the total number of samples, count(U). Based on the analysis results, output an operational optimization plan.

9. The intelligent tracking method for warehouses based on data analysis according to claim 8, characterized in that: In step S5, for abnormal data under strong interference scenarios, a weighted voting mechanism is used to perform multimodal data cross-validation correction. The correction process includes: First, calculating the voting weight of each modality data: Based on the historical tracking error rate error(m) of the m-th modality data, calculating its credibility(m) is performed by subtracting the historical error rate of the modality from 1 and then dividing by the sum of (1-historical error rates) of all modalities to achieve weight normalization; Then, a weighted fusion is performed: the independent correction result Rm of the m-th modality data is multiplied by its credibility(m), and the product results of all M modalities are summed to obtain the final correction result R.

Images