IoT-based intelligent warehouse material precision positioning management system

Abstract

This invention discloses a smart warehouse board material precise positioning management system based on the Internet of Things. The system comprises: a board material data acquisition module that collects raw radio frequency signals and inertial data of the board material and performs preprocessing operations; a spatiotemporal coordinate acquisition module that constructs an adaptive weighted federated fusion network to fuse the preprocessed raw radio frequency signals and inertial data to obtain high-precision spatiotemporal coordinate data; a board material positioning management module that constructs a motion pattern classification model using an optimized and trained Gradient Boosting Machine (GBM) strategy, inputs the high-precision spatiotemporal coordinate data into the motion pattern classification model, and outputs the motion pattern prediction results; and a board material positioning management module that, based on the motion pattern prediction results and combined with dynamic geofencing, obtains the optimal board material motion pattern, maps the optimal board material motion pattern in real time to a three-dimensional geometric model to generate a real-time digital twin of the entire warehouse, achieving visualized and precise warehouse board material positioning management.

Description

Technical Field

[0001] This invention relates to the field of board management technology, and in particular to a smart warehouse board precision positioning management system based on the Internet of Things. Background Technology

[0002] With the maturity of technologies such as the Internet of Things (IoT), Radio Frequency Identification (RFID), and Inertial Measurement Unit (IMU), warehouse management is gradually evolving towards digitalization and intelligence. For example, the introduction of multi-sensor fusion technology, as well as the use of Kalman filtering or extended Kalman filtering to perform simple weighted fusion of RFID and IMU data, attempts to solve the problems of signal attenuation, noise interference, and drift of single sensors in complex environments. At the same time, the rise of the concept of digital twins has made the visualization and monitoring of warehouse dynamics through three-dimensional models a new development trend.

[0003] Currently, most existing fusion methods are simple algorithms with static weights (such as weighted average with fixed coefficients), which cannot adaptively cope with complex signal changes in the warehousing environment (such as the multipath effect of metal shelves on radio frequency and the interference of handling equipment vibration on IMU). This results in low accuracy and poor stability of the fused spatiotemporal coordinate data, which cannot provide a reliable data foundation for accurate positioning.

[0004] Secondly, the movement patterns of sheet metal in warehouses (such as linear transport, rotational repositioning, and cross-level lifting) are complex and varied, while existing systems lack the ability to intelligently identify and classify these movement patterns. They typically only provide raw coordinates and cannot understand the intended movement, leading to an inability to predict trajectories or promptly detect abnormal behavior (such as deviations from the path or unauthorized movement). Therefore, this paper proposes a smart warehouse sheet metal precision positioning management system based on the Internet of Things. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution:

[0006] The IoT-based smart warehouse board material precision positioning management system includes:

[0007] Sheet material data acquisition module: Collects raw radio frequency signals and inertial data of the sheet material and performs preprocessing operations;

[0008] Spatiotemporal coordinate acquisition module: Construct an adaptive weighted federated fusion network, and use the adaptive weighted federated fusion network to fuse the preprocessed original radio frequency signal and inertial data to obtain high-precision spatiotemporal coordinate data;

[0009] The adaptive weighted federated fusion network consists of a feature fusion branch and a particle filter solution branch. The feature fusion branch includes a radio frequency signal sub-branch, an inertial data sub-branch, and an attention weight calculation mechanism.

[0010] Motion classification and prediction module: Constructs a motion pattern classification model through an optimized and trained gradient boosting machine (GBM) strategy, inputs high-precision spatiotemporal coordinate data into the motion pattern classification model, and outputs motion pattern prediction results;

[0011] The optimization training is achieved by optimizing and adjusting the basic gradient booster strategy through a hierarchical progressive hyperparameter optimization strategy.

[0012] The motion pattern classification model is implemented by introducing a learning rate optimization mechanism through the optimized gradient boosting machine (GBM) strategy after training.

[0013] Board positioning management module: Based on the motion pattern prediction results, combined with dynamic geofencing, the optimal board motion pattern is obtained. The optimal board motion pattern is mapped in real time to a three-dimensional geometric model to generate a real-time digital twin of the entire warehouse, realizing visualized and accurate warehouse board positioning management.

[0014] The feature fusion branch extracts features from radio frequency signals and inertial data respectively through radio frequency signal sub-branch and inertial data sub-branch, and then calculates the fusion weight of the two through attention mechanism to perform weighted fusion of the extracted features;

[0015] The particle filter solution branch is used to solve the fused features to obtain high-precision spatiotemporal coordinate data.

[0016] The specific implementation process of the feature fusion branch is as follows:

[0017] Preprocessed radio frequency signal and inertial data Input the radio frequency signal sub-branch and the inertial data sub-branch respectively;

[0018] The radio frequency (RF) signal sub-branch utilizes a convolutional neural network (CNN) to extract the spatial features of the signal. Let the feature extraction function for the RF signal be... ,in The input radio frequency (RF) signal data is processed through convolutional and pooling layers in a CNN to obtain the feature vector of the RF signal. ;

[0019] The inertial data sub-branch utilizes a Long Short-Term Memory (LSTM) network to extract temporal features from the data. Let the feature extraction function for the inertial data be... ,in The input inertial data is processed by the memory cells of an LSTM to obtain the feature vector of the inertial data. ;

[0020] An attention mechanism is introduced to determine the weights of radio frequency (RF) signals and inertial data during fusion, and the RF signal weights are determined based on historical data. Influence weight and inertial data Influence weight

[0021] Then, the feature vectors of the two branches are fused: , This is the output of the feature fusion branch, i.e., the fused features.

[0022] The specific implementation process of the particle filter solution branch is as follows:

[0023] First, particle initialization is performed, defining the particle set as follows: , where i is the index The state of the particle. For the weight of the particles, each particle Includes spacetime And generate based on the warehouse space and the initial position of the boards. Spatial coordinates and timestamps of individual particles Set as the initial time ;

[0024] Constructing state transition functions combined with fusion features To predict the particle from time t. arrive The state changes are represented by the state transition function as follows: The predicted particle state is then represented as: ,in, For process noise, ;

[0025] After the prediction is completed, an update step is performed, and the likelihood function is calculated. This function represents the state of the particle. The current fusion features observed below The probability is expressed as a product of the total likelihood function. The particle weights are updated based on the likelihood function, and are represented as follows: ,in, It is the likelihood function;

[0026] After the update step, resampling is performed. Based on the resampled particle set, the optimal spatiotemporal coordinates are calculated using a weighted average method. .

[0027] The process of obtaining the resampled particle set is as follows:

[0028] First, by determining the effective number of particles... To determine whether a particle has degenerated, when Less than the preset threshold When resampling, a roulette wheel resampling method is used. During resampling, the cumulative weight array is calculated first. Then generate N uniformly distributed random numbers. For each Find the smallest i such that Select the i-th particle to obtain the new particle set after resampling. At this point, the state of the particle is represented as Complete resampling.

[0029] The process of optimizing and adjusting the basic gradient boosting machine strategy using a hierarchical progressive hyperparameter optimization strategy is as follows:

[0030] The search range of the decision tree depth D in the basic gradient boosting machine (GBM) policy is defined as two sub-intervals: and ;

[0031] At the same time, the number of leaf nodes L is divided into and Then, hyperparameter optimization is performed, first in the sub-interval. , Within, traverse all [items] with a step size of 1. and Combining these methods, cross-validation is used to evaluate model accuracy, and the optimal model within the given sub-interval is selected. and ( ), then, with and Centered on, in the sub-interval , Within this step, the step size is reduced to 0.5, and the combinations are iterated and cross-validated again to determine the optimal D and L that can adapt to both simple and complex motion scenarios.

[0032] After optimization training, the Gradient Boosting Machine (GBM) strategy determines the optimal decision tree depth D and the maximum number of leaf nodes L, serving as the basic framework for the motion pattern classification model.

[0033] The learning rate optimization mechanism is implemented as follows:

[0034] Let the initial learning rate be The learning rate for the t-th iteration is and high-precision spatiotemporal coordinate data Feature construction preprocessing is performed on the high-precision spatiotemporal coordinate data after preprocessing. Input into the motion pattern classification model;

[0035] The motion pattern classification model is based on preprocessed high-precision spatiotemporal coordinate data. Make a judgment and output a predicted value related to the movement of the plate;

[0036] ;

[0037] in, This represents the initial model prediction result. Let Y represent the logarithmic loss function, and Y represent the true label. It is a constant;

[0038] For the i-th iteration, calculate the negative gradient residuals related to the plate motion. ,in, This represents the model prediction result for the i-th iteration;

[0039] Then fit the decision tree. Minimize the negative gradient residual. Determine the optimal decision tree With this structure, the depth of the decision tree does not exceed the optimal D, and the number of leaf nodes does not exceed the optimal L, thus obtaining the optimal decision tree. ;

[0040] Based on the optimal decision tree Calculate the optimal learning rate Achieving the optimal learning rate Then, the predicted probability distribution of the motion pattern classification model is updated by the learning rate.

[0041] The process of obtaining the optimal motion mode of the sheet metal is as follows:

[0042] Define a dynamic geofence, assuming the warehouse's three-dimensional space is... The boundary function of dynamic geofencing is When inside the fence When outside the fence >0;

[0043] Then, the high-precision spatiotemporal coordinate point sequence The input is fed into the motion pattern classification model, which retains the motion patterns output from the coordinate points inside the fence and represents them as valid motion patterns. ;

[0044] For effective movement patterns Perform time series continuity analysis to identify continuous effective motion patterns. Perform sliding window smoothing, and let the prediction result within the window be... q represents the effective motion mode in the window. The number of occurrences is counted, and the most frequently occurring movement pattern is identified. Let the number of occurrences be denoted as . ,like Percentage within the window Exceeding the preset threshold If the prediction is correct, then the motion pattern of all points within the window is corrected to "linear motion"; otherwise, the original prediction result is retained to obtain the optimal board motion pattern. k represents the number of optimal board motion patterns.

[0045] The present invention has the following beneficial effects:

[0046] In this invention, firstly, by constructing a CNN (extracting spatial features), LSTM (extracting temporal features), and attention mechanism (dynamically calculating fusion weights) combined with a particle filter solution branch, the uncertainty and noise of sensor data can be effectively processed, and finally high-precision spatiotemporal coordinate data after denoising is output, providing an extremely reliable and accurate data foundation for the entire system;

[0047] Secondly, a hierarchical progressive hyperparameter optimization strategy was adopted to train the gradient booster machine (GBM) strategy, and a learning rate optimization mechanism was introduced to construct a high-performance motion pattern classification model. This model can adaptively adjust its complexity and accurately distinguish complex scenarios such as linear motion, rotation, and stillness, thereby achieving a deep understanding of the board's motion intention and providing core semantic information for subsequent abnormal behavior judgment and digital twin dynamic mapping.

[0048] Finally, by combining optimal motion patterns with dynamic geofencing technology, outliers in the location data were effectively filtered out, ensuring the reliability of the motion patterns. Then, precise data with semantic information (motion patterns) was mapped in real time onto a 3D geometric model, generating a real-time digital twin of the entire warehouse that not only displays location but also reflects status (such as rotation, movement) and trajectory. This allows managers to intuitively and comprehensively grasp the real-time dynamics of each piece of material, achieving precise and intelligent location management of the warehouse. Attached Figure Description

[0049] Figure 1 This is a system block diagram of the IoT-based smart warehouse material precision positioning management system proposed in this invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example

[0052] like Figure 1 As shown, the IoT-based smart warehouse material precision positioning management system proposed in this invention includes:

[0053] Sheet material data acquisition module: Collects raw radio frequency signals and inertial data of the sheet material and performs preprocessing operations;

[0054] For the raw radio frequency signal, UHF-RFID equipment is used. Multiple RFID readers and antennas are deployed in appropriate locations in the warehouse. When the RFID tag on the board enters the reader's reading range, the reader will capture the radio frequency signal emitted by the tag. These signals contain the tag's identification information as well as the raw data of the signal strength and phase.

[0055] For the raw inertial data, the acceleration and angular velocity inertial data of the plate during its motion are acquired in real time by a high-precision inertial measurement unit (IMU) installed on the plate.

[0056] The acquired raw radio frequency signals and inertial data are preprocessed separately:

[0057] The original radio frequency (RF) signal is subject to noise due to environmental electromagnetic interference and multipath effects during propagation. A wavelet threshold denoising algorithm is used to decompose the RF signal into different scale and frequency components, resulting in a denoised RF signal. ;

[0058] For inertial data, due to the measurement noise inherent in the IMU itself, a Kalman filter algorithm is used for smoothing. Through two steps of prediction and update, the estimation of inertial data is continuously optimized to obtain smoothed inertial data. .

[0059] Spatiotemporal coordinate acquisition module: Construct an adaptive weighted federated fusion network, and use the adaptive weighted federated fusion network to fuse the preprocessed original radio frequency signal and inertial data to obtain high-precision spatiotemporal coordinate data;

[0060] The adaptive weighted federated fusion network consists of a feature fusion branch and a particle filter solution branch;

[0061] The feature fusion branch extracts features from radio frequency signals and inertial data separately through two parallel feature extraction branches (radio frequency signal sub-branch and inertial data sub-branch), and then calculates the fusion weight of the two through an attention mechanism to perform weighted fusion of the extracted features;

[0062] The specific implementation process of the feature fusion branch is as follows:

[0063] Preprocessed radio frequency signal and inertial data Input the different branches of the network respectively;

[0064] The radio frequency (RF) signal sub-branch utilizes a convolutional neural network (CNN) to extract the spatial features of the signal. Let the feature extraction function for the RF signal be... ,in The input radio frequency (RF) signal data is processed through convolutional and pooling layers in a CNN to obtain the feature vector of the RF signal. ;

[0065] The inertial data sub-branch utilizes a Long Short-Term Memory (LSTM) network to extract temporal features from the data. Let the feature extraction function for the inertial data be... ,in The input inertial data is processed by the memory cells of an LSTM to obtain the feature vector of the inertial data. ;

[0066] An attention mechanism is introduced to determine the weights of radio frequency (RF) signals and inertial data during fusion, and the RF signal weights are determined based on historical data. Influence weight and inertial data Influence weight Then the feature vectors of the two branches are fused. , This is the output of the feature fusion branch;

[0067] Then, the spatiotemporal coordinates are obtained by solving the fused features using the particle filter solution branch. The specific process is as follows:

[0068] Using particle filtering algorithm to analyze the fused features To perform the calculation, particle filtering uses a large number of particles to represent the probability distribution of the state, defining the particle set as... , where i is the index The state of particles , For the weight of the particles, each particle Includes spacetime Then, based on the warehouse's spatial range and the initial location of the boards, random generation is performed within that range. Spatial coordinates and timestamps of individual particles Set as the initial time This completes the initialization of the particles;

[0069] After initialization, the prediction step begins, where a state transition function is constructed and fused features are combined. To predict the particle from time t. arrive The state changes are represented by the state transition function as follows: The predicted particle state is then represented as:

[0070] ,in, , To introduce process noise (due to uncertainties in actual motion (such as minute disturbances in transport equipment), Gaussian distributed process noise needs to be added to each particle). ), For time intervals;

[0071] Specifically, the predicted particle state includes three-dimensional coordinates and a timestamp, i.e. Among them, timestamp It is the current moment;

[0072] After the prediction is completed, an update step is performed, first calculating the likelihood function. This function represents the state of the particle. The current fusion features observed below The probability is expressed as a product of the total likelihood function. The particle weights are updated based on the likelihood function, and are represented as follows: ,in, It is the likelihood function;

[0073] After the update step, resampling is performed, first by checking the effective particle count. To determine whether a particle has degenerated, when Less than the preset threshold When resampling, a roulette wheel resampling method is used. During resampling, the cumulative weight array is calculated first. Then generate N uniformly distributed random numbers. For each Find the smallest i such that Select the i-th particle to obtain the new particle set after resampling. At this point, the state of the particle is represented as ;

[0074] After prediction, updating, and resampling steps, the optimal spatiotemporal coordinate point estimate is calculated based on the resampled particle set using a weighted average method.

[0075] ;

[0076] ;

[0077] From this, we obtain This is high-precision spatiotemporal coordinate data after noise reduction, which can accurately reflect the position and time information of the board at the current moment;

[0078] Specifically, through an adaptive weighted federated fusion network, on the one hand, a convolutional neural network is used to extract the spatial features of radio frequency signals, and a long short-term memory network is used to capture the temporal features of inertial data. On the other hand, an attention mechanism is used to dynamically adjust the fusion weights of the two types of data, making the fusion more consistent with the actual motion state of the board. Then, particle filtering is used for calculation. Particle filtering simulates the state probability distribution through a large number of particles, which can cope with the uncertainty of board motion in the storage environment (such as the slight disturbance of handling equipment). In actual engineering, when facing complex and uncertain system state estimation, particle filtering can gradually converge to the real spatiotemporal coordinates of the board through the process of initialization, prediction, update, and resampling, ensuring the practicality and authenticity of the design, and making the calculation results more consistent with the position changes of the board in actual storage.

[0079] Sheet positioning management module: A motion pattern classification model is constructed through an optimized and trained gradient boosting machine strategy (GBM), and high-precision spatiotemporal coordinate data is input into the motion pattern classification model to output motion pattern prediction results;

[0080] The motion pattern classification model is implemented by introducing a learning rate optimization mechanism through the Gradient Boosting Machine (GBM) strategy after optimization training.

[0081] The basic gradient booster strategy GBM is optimized and trained. The optimization training is achieved by optimizing and adjusting the basic gradient booster strategy through a hierarchical progressive hyperparameter optimization strategy.

[0082] The process of optimizing and adjusting the basic gradient boosting machine strategy using a hierarchical progressive hyperparameter optimization strategy is as follows:

[0083] The search range of the decision tree depth D in the basic gradient boosting machine (GBM) policy is defined as two sub-intervals: (Corresponding to simple motion scenarios, such as linear transport and short-distance displacement) and (For complex motion scenarios, such as multi-directional rotation and cross-layer transport).

[0084] Specifically, by leveraging domain knowledge in the board storage scenario, we clarify the typical feature complexity of board movement. For example, when boards are transported in a straight line, the movement features are relatively simple, and the decision tree does not need to be too deep. However, in complex movement scenarios such as rotation and vertical transport in the loading and unloading area, there are many feature interactions, requiring the decision tree to have a certain depth to capture them. Therefore, the search range of the decision tree depth D is defined as two sub-intervals.

[0085] At the same time, the number of leaf nodes L is also divided into (In a simple board storage scenario, with high feature differentiation, no need for excessive leaf node subdivision) and (In complex board storage scenarios, more leaf nodes are needed to finely divide the feature space.)

[0086] Then, hyperparameter optimization is performed, first in sub-intervals of a simple sheet metal storage scenario. Within, traverse all [items] with a step size of 1. and Combining these methods, cross-validation is used to evaluate model accuracy, and the optimal model within the given sub-interval is selected. and Then, with and Centered on this, in the complex board material storage scenario sub-area Within this step, the step size is reduced to 0.5, and the combinations are iterated and cross-validated again to determine the optimal D and L that can adapt to both simple and complex motion scenarios.

[0087] The optimized gradient booster strategy GBM (which determines the optimal decision tree depth D and the maximum number of leaf nodes L) is used as the basic framework for the motion pattern classification model.

[0088] The learning rate optimization mechanism is implemented as follows:

[0089] The learning rate determines the magnitude of model parameter updates in each iteration. Let the initial learning rate be... The learning rate for the t-th iteration is ;

[0090] Specifically, in the early stages of model training, a larger learning rate allows the model to quickly update parameters toward the optimal solution. As the number of iterations increases, the learning rate gradually decreases, allowing the model to make more refined adjustments when it is close to the optimal solution. This avoids the model from oscillating around the optimal solution due to an excessively large learning rate, thereby effectively preventing overfitting and improving the model's generalization ability.

[0091] High-precision spatiotemporal coordinate data of the plate Feature construction preprocessing is performed to expand a single spatiotemporal coordinate point into a feature vector containing multi-dimensional information such as position, displacement, velocity, and acceleration, thereby obtaining high-precision preprocessed spatiotemporal coordinate data. The preprocessed high-precision spatiotemporal coordinate data Input into the motion pattern classification model;

[0092] The motion pattern classification model makes predictions by ensembles of multiple decision trees trained using an optimized gradient booster. The structure of each decision tree is determined by the optimal D and L, and is used to learn the characteristic laws of plate motion. For the i-th decision tree... It will be based on preprocessed high-precision spatiotemporal coordinate data Make a judgment and output a predicted value related to the movement of the plate;

[0093] ;

[0094] in, This represents the initial model prediction results when the Basic Gradient Boosting Machine (GBM) strategy is used as the framework for a motion pattern classification model. Let Y represent the logarithmic loss function, and Y represent the true label. It is a constant. This means finding a constant. This makes all training samples loss function The sum is minimized, where m is the sum of all training samples. Quantity;

[0095] Then, for the i-th iteration, the negative gradient residuals related to the plate motion are calculated. ,in, This represents the model prediction result for the i-th iteration;

[0096] Then fit the decision tree. Minimize the negative gradient residual. Determine the optimal decision tree The structure is such that the depth of the decision tree does not exceed the optimal D, and the number of leaf nodes does not exceed the optimal L, ultimately yielding the optimal decision tree. To better capture the characteristics of sheet metal movement;

[0097] Based on the optimal decision tree Calculate the optimal learning rate m represents all training samples Quantity;

[0098] Achieving the optimal learning rate Then, the predicted probability distribution of the motion pattern classification model is updated using the learning rate, as follows: , To predict the probability distribution;

[0099] Specifically, when inputting preprocessed high-precision spatiotemporal coordinate data At that time, the motion pattern classification model will output the motion pattern prediction probability distribution corresponding to that coordinate point (prediction result). In this step, the optimal D and L ensure that each decision tree can effectively learn the motion characteristics of the board, without underfitting or overfitting, so that the integrated gradient booster model can accurately classify and predict the input feature vector.

[0100] Finally, the probability distribution is predicted based on the motion pattern output by the gradient booster. Select the prediction probability distribution with the highest probability. The corresponding motion pattern is used as the final motion pattern prediction result;

[0101] For example: if the probability distribution output by the model is {static})=0.1, {Linear motion} = 0.8 If {rotational motion} = 0.1, then the final motion pattern prediction result is: {Linear motion};

[0102] Specifically, this step transforms the probability distribution output by the model into specific motion pattern labels, completing the conversion from feature vectors to motion pattern prediction results, thus clearly representing the motion patterns of the board material corresponding to different high-precision spatiotemporal coordinate points.

[0103] Board positioning management module: Based on the motion pattern prediction results, combined with dynamic geofencing, the optimal board motion pattern is obtained, and the optimal board motion pattern is mapped in real time to a three-dimensional geometric model to generate a real-time digital twin of the entire warehouse, realizing visualized and accurate warehouse board positioning management.

[0104] Define a dynamic geofence, assuming the warehouse's three-dimensional space is... The boundary function of dynamic geofencing is Where t is time and (x,y,z) are spatial coordinates, satisfying the condition that when inside the fence... When outside the fence >0;

[0105] Then, the high-precision spatiotemporal coordinate point sequence The data is input into a motion pattern classification model to obtain the motion pattern prediction result for each coordinate point, and then the boundary function of the dynamic geofence is applied. Determine each high-precision spatiotemporal coordinate point Is it inside the fence (i.e.) For coordinates inside the fence, the output motion pattern prediction result is retained and represented as the valid motion pattern. For coordinate points outside the fence, since these points may be erroneous points caused by signal interference or other reasons, their motion pattern prediction results will be marked as invalid.

[0106] Finally, the effective movement patterns that were retained Perform time series continuity analysis to identify continuous effective motion patterns. Perform sliding window smoothing, and let the prediction result within the window be... q represents the effective motion mode in the window. The number of occurrences is counted, and the most frequently occurring motion pattern is determined. For example, if the most frequently occurring motion pattern in the window represents "linear motion", then the number of times "linear motion" occurs is counted. Let the number of occurrences be 1. ,like Percentage within the window Exceeding the preset threshold If the prediction is correct, then the motion pattern of all points within the window is corrected to "linear motion"; otherwise, the original prediction result is retained to obtain the optimal board motion pattern. k is the number of optimal board motion patterns;

[0107] Construct a basic 3D warehouse model. Based on static data such as actual warehouse building drawings, shelving layout, and aisle dimensions, use 3D modeling tools to generate a 3D geometric model containing static elements such as walls, shelves, and loading / unloading areas. Each static element in the model is associated with high-precision spatiotemporal coordinate data. Then, each high-precision spatiotemporal coordinate data is mapped to a 3D model. At the same spatial location, generate digital twin elements of the board;

[0108] At the same time, according to the optimal board movement mode By endowing digital twins with dynamic attributes corresponding to different motion modes, a full-database real-time digital twin can be generated. ;

[0109] For example, in the "linear motion" mode, the digital twin moves at a corresponding speed along the line connecting the coordinate points, and is marked with a blue trajectory line;

[0110] In "Rotation Motion" mode, the digital twin rotates around a specified axis, and the trajectory line is red;

[0111] In "Stationary Movement" mode, it remains in a fixed position and is highlighted in green;

[0112] Finally, through the real-time data interface MQTT protocol, for each new set of data received, the system immediately updates the position and status of the digital twin of the corresponding board according to the mapping rules, and synchronously updates the historical trajectory (retaining the movement path of the most recent hour).

[0113] Real-time digital twin of the entire database Using the coordinate information stored in the digital twin, the system automatically locates the position of the board in the 3D view and displays its movement trajectory and current movement mode. Managers can query the movement history of any time period (obtaining the best board movement mode) through the interface. When the digital twin detects an anomaly (such as the board movement mode conflicting with the preset path or the position exceeding the dynamic geofence), the system will remind the manager and display the reason for the anomaly on the interface (such as "Board 1 deviates from the aisle and is currently located in the blind spot of the shelf"), thus achieving precise warehouse board positioning management.

[0114] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0115] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart warehouse material precision positioning management system based on the Internet of Things, characterized in that, include: Sheet material data acquisition module: Collects raw radio frequency signals and inertial data of the sheet material and performs preprocessing operations; Spatiotemporal coordinate acquisition module: Constructs an adaptive weighted federated fusion network to fuse the preprocessed raw radio frequency signals and inertial data to obtain high-precision spatiotemporal coordinate data. Specifically, it includes: The adaptive weighted federated fusion network consists of a feature fusion branch and a particle filter solution branch. The feature fusion branch includes a radio frequency (RF) signal sub-branch, an inertial data sub-branch, and an attention weight calculation mechanism. Preprocessed RF signals are input into the RF signal sub-branch, and preprocessed inertial data is input into the inertial data sub-branch. The RF signal sub-branch uses a convolutional neural network (CNN) to extract the spatial features of the signal, obtaining the RF signal feature vector. The inertial data sub-branch uses a long short-term memory (LSTM) network to extract the temporal features of the data, obtaining the inertial data feature vector. The attention weight calculation mechanism determines the weights of the RF signal and inertial data during fusion, and the feature vectors obtained from the two sub-branches are fused. The motion classification prediction module constructs a motion pattern classification model using an optimized and trained gradient booster (GBM), inputs high-precision spatiotemporal coordinate data into the motion pattern classification model, and outputs the motion pattern prediction results. The optimization training is achieved by optimizing and adjusting the basic gradient booster machine through a hierarchical progressive hyperparameter optimization strategy, and the motion pattern classification model is achieved by introducing a learning rate optimization mechanism through the optimized gradient booster machine GBM. Board positioning management module: Based on the motion pattern prediction results, combined with dynamic geofencing, the optimal board motion pattern is obtained. The optimal board motion pattern is mapped in real time to a three-dimensional geometric model to generate a real-time digital twin of the entire warehouse, realizing visualized and accurate warehouse board positioning management.

2. The IoT-based intelligent warehouse material precision positioning management system according to claim 1, characterized in that, The feature fusion branch extracts features from radio frequency signals and inertial data through radio frequency signal sub-branch and inertial data sub-branch respectively, and then calculates fusion weights through attention weight calculation mechanism to perform weighted fusion of the extracted features; The particle filter solution branch calculates high-precision spatiotemporal coordinate data by solving the fused features.

3. The IoT-based intelligent warehouse material precision positioning management system according to claim 2, characterized in that, The specific implementation process of the feature fusion branch is as follows: Preprocessed radio frequency signal and inertial data Input the radio frequency signal sub-branch and the inertial data sub-branch respectively; The radio frequency (RF) signal sub-branch utilizes a convolutional neural network (CNN) to extract spatial features of the signal. The feature extraction function for the RF signal is: ,in The input radio frequency (RF) signal data is processed through the convolutional and pooling layers of a convolutional neural network (CNN) to obtain the feature vector of the RF signal. ; The inertial data sub-branch utilizes a Long Short-Term Memory (LSTM) network to extract temporal features from the data. The feature extraction function for inertial data is: ,in The input inertial data is processed by the memory units of the Long Short-Term Memory (LSTM) network to obtain the feature vector of the inertial data. ; An attention mechanism is introduced to determine the weights of radio frequency (RF) signals and inertial data during fusion, and the influence weight of RF signals is determined based on historical data. and the influence weight of inertial data The feature vectors of the two sub-branches are then fused: , This is the output of the feature fusion branch, i.e., the fused features.

4. The IoT-based intelligent warehouse material precision positioning management system according to claim 1, characterized in that, The specific implementation process of the particle filter solution branch is as follows: Perform particle initialization and define the particle set as follows: , where i is the index The state of a particle. For the weight of the particles, each particle's Includes spacetime coordinates Based on the spatial range of the warehouse and the initial position of the board, N particles' spatial coordinates are generated, and the timestamp is uniformly set to the initial time. Constructing state transition functions combined with fusion features To predict the state change of a particle from time t-1 to t, the state transition function is expressed as: The predicted particle state is then represented as: ,in, For process noise, ; After the prediction is completed, an update step is performed, and the likelihood function is calculated. This function represents the state of the particle. The probability of observing the current fused feature is used to update the particle weights based on the likelihood function, expressed as: ; After the update step, resampling is performed. Based on the resampled particle set, the optimal spatiotemporal coordinates are calculated using a weighted average method. .

5. The IoT-based intelligent warehouse material precision positioning management system according to claim 4, characterized in that, The process of obtaining the resampled particle set is as follows: First, by determining the effective number of particles... To determine whether a particle has degenerated, when Less than the preset threshold When resampling, a roulette wheel resampling method is used. During resampling, the cumulative weight array is calculated first. Then generate N uniformly distributed random numbers. Used to filter particles and obtain a new set of particles after resampling. ,in, Complete resampling.

6. The IoT-based intelligent warehouse material precision positioning management system according to claim 1, characterized in that, The process of optimizing and adjusting the base gradient booster using a hierarchical progressive hyperparameter optimization strategy is as follows: The search range of the decision tree depth D in the basic gradient boosting machine is defined as two sub-intervals: and ; At the same time, the number of leaf nodes L is divided into and Then, hyperparameter optimization is performed, first in the sub-interval. , Within, traverse all [items] with a step size of 1. and Combining these methods, we use cross-validation to evaluate model accuracy and select the optimal model for that sub-interval. and , Then, with and Centered on, in the sub-interval , Within the range, reduce the step size to 0.5, traverse and combine again and cross-validate to determine the optimal decision tree depth and number of leaf nodes that can simultaneously adapt to both simple and complex motion scenarios; After optimization and training, the Gradient Boosting Machine (GBM) determines the optimal decision tree depth and the maximum number of leaf nodes, serving as the basic framework for the motion pattern classification model.

7. The IoT-based intelligent warehouse material precision positioning management system according to claim 6, characterized in that, The learning rate optimization mechanism is implemented as follows: The initial learning rate is Feature construction preprocessing is performed on high-precision spatiotemporal coordinate data, and the preprocessed high-precision spatiotemporal coordinate data is then used. Input into the motion pattern classification model; The motion pattern classification model is based on preprocessed high-precision spatiotemporal coordinate data. Make a judgment and output a predicted value related to the movement of the plate; For the Mth iteration, calculate the negative gradient residuals related to the plate motion; Then, the decision tree is fitted to the negative gradient residual to obtain the optimal decision tree. The optimal learning rate is calculated based on the best decision tree. After obtaining the optimal learning rate, the predicted probability distribution of the motion pattern classification model is updated using the optimal learning rate.

8. The IoT-based intelligent warehouse material precision positioning management system according to claim 7, characterized in that, The process of obtaining the optimal motion mode of the sheet metal is as follows: Define a dynamic geofence; the warehouse's three-dimensional space is... The boundary function of dynamic geofencing is When inside the fence When outside the fence ; Then, the high-precision spatiotemporal coordinate data is input into the motion pattern classification model, retaining the motion patterns output from the coordinate points inside the fence and representing them as valid motion patterns. ; For effective movement patterns Perform time series continuity analysis, apply sliding window smoothing to continuous effective motion patterns, and count the most frequently occurring motion patterns. The proportion within the window exceeds a preset threshold. If the prediction is correct, the motion mode of all coordinate points within the window will be changed to "linear motion"; otherwise, the original prediction result will be retained to obtain the optimal motion mode of the board.

Images