Low-altitude logistics multi-modal intelligent scheduling system and method for high-noise environment

Abstract

This invention belongs to the field of intelligent logistics and industrial artificial intelligence technology, and more specifically, relates to a multimodal intelligent scheduling system and method for low-altitude logistics in high-noise environments. The invention includes the following steps: S1, logistics data acquisition and noise preprocessing; S2, multimodal network and dynamic gating construction; S3, model parameter training under low-altitude constraints; S4, multi-index robustness and accuracy evaluation. This invention also provides a multimodal intelligent scheduling system for low-altitude logistics in high-noise environments. This invention addresses the problems of insufficient single-modal feature resolution, OCR susceptibility to noise interference, lack of dynamic adaptation in multimodal fusion, and inability to couple airworthiness constraints by enhancing logistics data noise, multimodal dynamic gating fusion, coupling low-altitude physical constraints, lightweight model training, and multi-index evaluation. It achieves accurate product identification, automatic airworthiness determination, intelligent priority scheduling, and lightweight deployment of edge devices in high-noise environments.

Description

Technical Field

[0001] This invention belongs to the field of intelligent logistics and industrial artificial intelligence technology, and more specifically, relates to a multimodal intelligent scheduling system and method for low-altitude logistics in high-noise environments. Background Technology

[0002] With the rapid development of the low-altitude economy, drone logistics has become an important branch of smart logistics. Low-altitude logistics scheduling needs to take into account payload constraints, air traffic priority, and environmental adaptability. Existing logistics identification and scheduling systems mostly adopt single-modal perception. When faced with conditions such as rotor airflow vibration, sudden changes in lighting, homogenized packaging, and damaged waybills, feature extraction fails, fusion robustness is poor, and physical constraints cannot be coupled, making it difficult to meet the stable scheduling requirements in high-noise environments.

[0003] Chinese patent document CN121617001A discloses an intelligent low-altitude traffic UAV situational awareness method. It improves target recognition accuracy under adverse weather conditions by enhancing perception to generate boundary crossing samples, extracting boundary distance features, and dynamically fusing cross-specification air situation data. Although it solves the problems of target detail loss and feature aliasing in low-altitude noise environments, it still has shortcomings such as not being designed for multimodal features of logistics goods, being unable to couple physical constraints such as load capacity, and lacking intelligent scheduling and decision-making capabilities.

[0004] Chinese patent document CN121582833A discloses a radar low-altitude target recognition method based on a multi-dimensional feature nonlinear coupling network. It improves the robustness of target recognition through feature extraction enhancement, decoupling fusion, global correlation modeling and RCS-assisted decision-making. Although it solves the problems of insufficient nonlinear expression of radar echo features and susceptibility to interference in single decision-making, it still has the shortcomings of being only applicable to radar target recognition, not covering visual and textual multimodal recognition, and not being applicable to logistics commodity sorting and drone scheduling.

[0005] In summary, there is an urgent need to provide a multimodal intelligent scheduling method for low-altitude logistics in high-noise environments to solve the problems existing in the current technology. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the issues of insufficient single-modal feature resolution, OCR susceptibility to noise interference, lack of dynamic adaptation in multimodal fusion, and inability to couple airworthiness constraints by enhancing logistics data noise, multimodal dynamic gating fusion, low-altitude physical constraint coupling, lightweight model training, and multi-index evaluation. This will enable accurate identification of goods in high-noise environments, automatic airworthiness determination, intelligent priority scheduling, and lightweight deployment of edge devices.

[0007] The technical terms related to this invention are explained as follows: Multimodal intelligent scheduling model: a deep learning model that integrates visual and textual features and has dynamic noise resistance and airworthiness decision-making capabilities.

[0008] Dynamic gating: A network structure that adaptively adjusts the weights of visual and textual features and suppresses noise propagation.

[0009] Missing modality mask: A mechanism to block and prevent invalid gradients when text features are invalid.

[0010] Airworthiness indicator function: a function for determining the feasibility of UAV transportation by coupling weight, volume, and hazardous materials attributes.

[0011] C2f_Logi module: An improved YOLO feature extraction module adapted for low-altitude logistics goods.

[0012] Logi_Bottleneck: A bottleneck structure that separates semantic and detail tensors while preserving high-frequency information.

[0013] Linear mapping network: A network layer that unifies visual and text feature dimensions and achieves feature alignment.

[0014] Layer normalization: a standardized operation that stabilizes feature distribution and improves model noise resistance.

[0015] Dynamic learning rate: A parameter strategy that adaptively adjusts during training to accelerate model convergence.

[0016] F1-Score: The harmonic mean of precision and recall, a metric for comprehensively evaluating model performance.

[0017] The technical problem to be solved by this invention is achieved by the following technical solution: A multimodal intelligent scheduling method for low-altitude logistics in high-noise environments includes the following steps: S1. Logistics Data Acquisition and Noise Preprocessing: Using the visual sensors mounted on the UAV and the ground sorting sensors, RGB images of goods and OCR strings of express waybills are collected in the logistics scene. At the same time, the actual load, actual volume, and explosion-proof labels of hazardous chemicals corresponding to the goods are defined and labeled to form raw logistics data. The raw logistics data is processed using a sensor-specific DataLoader, and dual industrial disaster simulation noise is injected into the raw logistics data to enhance the preprocessing of the dataset. S2. Construction of Multimodal Networks and Dynamic Gating: This includes the following steps: S21. Hardware-level visual feature extraction and text encoding missing mask: The underlying video stream interface of the visual sensor is called to transfer the acquired RGB image array of the goods to the GPU memory of the edge node. The visual high-frequency tensor of the physical edge of the product packaging is extracted and preserved as visual features through the edge computing node. At the same time, the OCR string of the express waybill output by the ground sorting sensor and the signal strength indication of the RFID tag are read through the serial communication interface for text feature encoding. When the strength indication RSSI is detected to be lower than the preset hardware reception threshold, or the OCR string returns an empty byte, a hardware-level interrupt request is sent to the microprocessor to trigger the mask logic: a mask matrix with all zeros on the diagonal is generated by bitwise AND operation, and the text features are forced to be zeroed at the physical memory level to block the cross-layer propagation of invalid noise gradients. S22. Cross-modal feature affine mapping and tensor alignment: Visual high-frequency tensors in GPU memory and text features that have been masked and zeroed are input into a preset affine transformation matrix to perform linear projection operations. This aligns the tensor dimensions of the two heterogeneous feature sets to the same constant dimension through uniform filling or pooling. Then, the batch normalization function is called to perform floating-point standardization, forcibly constraining the values ​​of both to the same distribution range, eliminating the dimensional differences caused by sudden changes in lighting and character distortion. S23. Physical Perception Dynamic Gating and Multimodal Matrix Fusion: The microprocessor independently calculates the signal-to-noise ratio (SNR) of the current visual high-frequency tensor and text features, quantifying modal reliability using the physical SNR. The SNR is input into the Sigmoid activation function to generate a dynamic weight matrix in the range of 0 to 1, triggering the gating hardware logic: when the text SNR falls below the set safety threshold, the gating circuit automatically decays the value of the text weight matrix according to the function curve. Subsequently, the processor's underlying general matrix multiplication instruction is called, and the Hadamard product weighting operation is performed on the aligned visual features and text features using the dynamic weight matrix. Finally, the channel splicing instruction is executed to forcibly splice the dual-modal features in the channel dimension, fusing them into a multimodal feature tensor, and the predicted probability distribution Logits of the product category is calculated and output using a multilayer perceptron. S24. Safety constraint physical judgment and PLC joint scheduling instruction output: Retrieve the standard estimated physical attributes corresponding to the Logits prediction category from the local database, and compare them with the actual weight and actual volume fed back in real time by the gravity sensor and infrared sensor; Execute safety constraint judgment logic: Only when the actual weight does not trigger the threshold alarm of the UAV's maximum load, the actual volume does not exceed the cargo hold size, and the explosion-proof label identification bit is non-hazardous chemical, the airworthiness indicator function is judged to be 1, otherwise it is 0; When the judgment is 0, a hexadecimal encoded load rejection instruction is sent to the underlying PLC via the fieldbus, directly driving the pneumatic solenoid valve of the sorting line to pop out the mechanical blocking rod; When the judgment is 1, a departure message is sent to the UAV flight control system through the Mavlink communication protocol, and the probability distribution priority is extracted to generate 1-3 level scheduling takeoff sequence codes; S3. Model parameter training under low-altitude constraints: The multimodal intelligent scheduling model is trained and learned using the PyTorch deep learning tool. The model is configured according to the set total number of training rounds, the amount of input data each time, the uniform size of the images, the dynamic learning rate, and the weight decay coefficient. GPU graphics cards are used to accelerate the calculation. During the training process, the model parameters are continuously optimized and updated in combination with the actual payload, actual volume of the drone, and the explosion-proof label of hazardous chemicals. S4. Multi-metric robustness and accuracy evaluation: Using four metrics—accuracy, recall, precision, and F1-Score—the multimodal intelligent scheduling model is tested for its accuracy and anti-interference capability in high-noise environments. At the same time, the number of parameters of the multimodal intelligent scheduling model is counted to evaluate whether the multimodal intelligent scheduling model is lightweight and suitable for deployment on edge devices.

[0018] This technical solution simulates the harsh working conditions of a low-altitude hub through dual industrial disaster simulation noise, covering scenarios such as text failure, image mutation, and feature drift, thereby improving the model's generalization ability. It employs Logi_Bottleneck to achieve tensor hard separation, forcibly preserving high-frequency information of product edges and textures, addressing the problem of insufficient fine-grained resolution in homogeneous packaging. A missing modality masking mechanism is implemented to explicitly block invalid features when the waybill is damaged or missing, preventing noise-contaminated gradient propagation. Linear mapping projects visual and text features into a unified space, achieving dimensional alignment and normalization to ensure effective fusion. Dynamic gating is applied based on modality. Quality adaptive weighting reduces the weight of corresponding modes when noise is high, achieving modal complementarity and anti-interference fusion; layer normalization stabilizes feature distribution, avoiding internal network shifts and improving model convergence speed and robustness; coupling load, volume, and hazardous chemical explosion-proof tags to construct an airworthiness indicator function realizes a closed loop of perception and scheduling, automatically completing load rejection, priority allocation, and air rights control; a lightweight network structure and dynamic learning rate training are adopted to control the number of model parameters and adapt to the low-latency deployment requirements of edge computing devices; recognition accuracy, anti-interference ability, and lightweight level are comprehensively evaluated through four indicators to ensure that the solution meets the needs of practical engineering applications.

[0019] In addition, the low-altitude logistics multimodal intelligent scheduling method for high-noise environments proposed according to the present invention may also have the following additional technical features: According to a preferred embodiment of the present invention, the logistics data acquisition and noise preprocessing in S1, when no publicly available dedicated dataset is available, uses a homogeneous RPC commodity dataset as the visual basis, and supplements the actual load, actual volume, and hazardous chemical explosion-proof labels according to the low-altitude logistics scenario to complete the dataset reconstruction; all samples in the dataset are divided into training set, validation set and test set in a 7:2:1 ratio; and dual industrial disaster simulation noise is injected into the sensor-dedicated DataLoader: 30% of the samples are injected with incorrect category semantics and the text features are amplified by 5 times to simulate the failure of the OCR string of the express waybill and the mutation of the RGB image of the commodity; 70% of the samples retain the correct semantics and apply 3-5 times random feature amplitude drift to simulate the harsh working conditions of a real low-altitude logistics hub.

[0020] This technical solution constructs a dedicated visual foundation for low-altitude logistics using a homogenized commodity dataset based on RPC. It supplements this with load capacity, volumetric weight, and hazardous chemical explosion-proof labels to achieve scenario-based reconstruction of the dataset, ensuring that training samples closely match the actual transport attributes of UAV logistics. Dividing the training, validation, and test sets in a 7:2:1 ratio guarantees sufficient model training, reliable validation, and objective testing, improving the standardization of the training process and the credibility of the results. Simulating noise from dual industrial disasters highly replicates the harsh working conditions of real-world low-altitude logistics hubs. Injecting erroneous category semantics and amplifying textual features into 30% of the samples simulates extreme text anomalies such as damaged, torn, or failed recognition of express delivery waybills. Applying a 3-5 times random feature amplitude drift to 70% of the samples simulates visual feature distortion caused by rotor airflow vibration, sudden changes in illumination, and sensor drift. This preprocessing method allows the model to fully engage with various types of high-noise data during the training phase, learning noise distribution patterns and anti-interference characteristics. This significantly improves recognition stability, generalization ability, and robustness in real high-noise environments, preventing scheduling failures due to sudden changes in working conditions and ensuring the continuous and reliable operation of the UAV logistics system.

[0021] According to a preferred embodiment of the present invention, the hardware-level visual feature extraction and text encoding missing masking in S21, which extracts visual features from the RGB image of the product acquired by the visual sensor, includes the following specific steps: S211, Instruction-level edge feature extraction: Call the GPU of the edge computing node to load the pre-processed RGB image of the product into the video memory in the form of a block matrix; Deploy a custom frequency domain decoupling node in the deep feature flow of the visual backbone network, and the frequency domain decoupling node has a built-in high and low frequency separation bottleneck layer; Avoid direct modification of the visual backbone network through non-intrusive memory reuse. S212. Visual perception signal splitting and multi-band feature extraction based on Logi_Bottleneck: The microprocessor receives the initial visual feature matrix of the logistics scene obtained by the visual backbone network. To simultaneously capture the global physical contours and local high-frequency edges of the target object, the underlying tensor processing unit is scheduled, and the Logi_Bottleneck module is used to process the feature matrix. Along the channel dimension Performs signal segmentation and routing at the physical memory level; separates the hybrid visual signal into semantic tensors corresponding to the low-frequency global environment and detail tensors corresponding to high-frequency local mutations. The instruction set is expressed as follows:

[0022] Where: the separated low-frequency spatial signal Routed to a hardware-accelerated branch of dilated convolution with a large receptive field, used to perceive the overall size and positional topology of the cargo against a complex background; high-frequency spatial signals The features are synchronously distributed to the standard convolutional computation branch, specifically for accurately extracting barcode edges, packaging wrinkles, and textures from the surface of goods. After multi-band processing, the dual-band features are re-aligned and stitched together in the memory channel. Spatial resolution is compressed by calling global average pooling instructions, redundant background noise is filtered out, and finally, low-dimensional, high-purity visual features are output for subsequent cross-modal gating scheduling. .

[0023] This technical solution inputs product images into a YOLO model to construct a C2f_Logi module. Through a built-in Logi_Bottleneck structure, it achieves hard separation of visual feature channel dimensions, decomposing features into semantic tensors and detail tensors. The semantic tensor is fed into a dilated convolution branch to extract global semantic information, ensuring the accuracy of product category recognition. The detail tensor is fed into a standard convolution branch to extract high-frequency detail information such as edges and textures, solving the feature confusion problem caused by the homogenization of lightweight packaging. This dual-branch parallel structure can forcibly retain fine-grained features in deep networks, avoiding the loss of key discriminative information during dimensionality reduction in conventional convolutional networks, and improving the ability to distinguish between similar-looking but different-category logistics products. Finally, a one-dimensional visual feature vector is output through channel concatenation and global average pooling, achieving efficient compression of high-dimensional image information into low-dimensional discriminative features. This retains core discriminative information while reducing computational load, providing high-quality and robust visual input for subsequent cross-modal fusion, and improving the overall model's feature representation ability in noisy environments.

[0024] According to a preferred embodiment of the present invention, the hardware-level visual feature extraction and text encoding missing masking in S21, which extracts text features from the OCR string of the express waybill obtained by the ground sorting sensor, includes the following specific steps: S213. The preprocessed express waybill OCR string is introduced into the text modality encoder. The text modality encoder maps the express waybill OCR string into text features through OCR feature processing. ; S214. Security constraint determination and missing feature masking based on physical sensor status: The system continuously monitors the physical working status of the electronic tag reader or visual sensor OCR text encoding module; and analyzes the acquired text features... Perform safety constraint determination:

[0025] When RFID tags fall off, radio frequency signals are lost, or OCR lenses are obstructed in logistics scenarios, resulting in all valid information bits being empty, the sum of the absolute values ​​of the text features approaches zero. The underlying hardware triggers a text feature loss masking mechanism, which masks the current batch of text features missing. Forced zeroing prevents invalid or divergent gradient signals caused by sensor failure from propagating to subsequent physical control nodes. At the same time, it sends a low-level interrupt signal indicating missing text modality to the main control console to ensure the safe operation of physical devices.

[0026] This technical solution transforms express delivery waybill text into initial text feature vectors through text modal encoders and OCR feature processing, achieving digitized and standardized expression of text information. When the text is missing, blurred, damaged, or recognition fails, explicit masking is executed through conditional judgment to block it, making the sum of the absolute values ​​of the text feature vectors approach zero. This cuts off invalid features and noise signals from propagating to subsequent network layers at the source, preventing invalid gradients from interfering with model training and inference. This missing modality masking mechanism constitutes a pre-emptive physical circuit breaker protection, preventing model feature distribution shifts, abnormal outputs, and scheduling misjudgments due to waybill failure. Compared with traditional direct fusion methods, this solution significantly reduces the risk of noise propagation, improves the model's fault tolerance and stability in text modality failure scenarios, and ensures that even in the case of single-modality failure, the system can still rely on another modality to complete reliable recognition, providing a safety guarantee for multimodal scheduling in high-noise and strong-interference environments.

[0027] According to a preferred embodiment of the present invention, the cross-modal feature affine mapping and tensor alignment in S22 includes the following specific steps: S221. Hardware-level spatial projection alignment and physical noise suppression of heterogeneous sensor signals: Given the visual features captured by vision sensors Text features analyzed by radio frequency reading or image recognition modules Given the heterogeneity of the underlying data dimension, the microprocessor performs hardware-level spatial projection on the heterogeneous sensor signals by scheduling linear mapping network modules. This process maps to the underlying tensor operation logic, forcibly aligning the bimodal features to a unified register processing dimension. To match the hardware bandwidth requirements of subsequent fusion computing:

[0028]

[0029] in: and These are the projection weight matrices of visual and textual features in the linear mapping network, respectively. The visual feature projection bias vector; The text feature projection bias vector; For the text signal processing branch: Considering that the stacking and squeezing of logistics packages and the strong light reflection from the conveyor belt can easily cause garbled characters or radio frequency crosstalk noise in the OCR strings of express waybills, the system calls the BN(·) low-level instruction to force smoothing and standardize the distribution of physical feature data of the current batch, thereby isolating and suppressing OCR recognition noise at the hardware operation level; then, the ReLU(·) hardware activation instruction is executed in series, and its non-linear truncation characteristics are used to filter out invalid weak electrical signals or negative interference signals at the bottom level, ensuring that the text perception signal passed to the downstream node has extremely high purity; S222. Physical splicing and hybrid feature construction of multi-source sensing signals at the video memory level: After completing the dimensional alignment of heterogeneous data streams and text channel noise reduction filtering, the underlying main control chip uses memory block transfer instructions in the high-speed data bus of the on-chip shared video memory to transfer the processed pure visual enhancement features. Text Enhancement Features Physical-level data stitching is performed along the feature channel dimension to construct a hybrid feature matrix that integrates panoramic information of the logistics site. Its bus instruction logic is expressed as follows:

[0030] By stitching together the underlying layers, the signal conversion from an independent data stream from a single physical sensor to a unified multimodal hybrid bus is completed, enabling the gating network to dynamically quantize the signal-to-noise ratio of the physical sensor.

[0031] This technical solution projects visual and textual features into a unified feature space using the linear mapping network Text_Project, achieving dimensional alignment and numerical normalization. This eliminates differences in dimension, distribution, and numerical range among different modal features, laying the foundation for subsequent stable fusion. Batch normalization standardizes feature distribution, suppresses numerical fluctuations caused by OCR noise and image noise, and improves model training stability and convergence speed. Correcting the linear unit activation function introduces non-linear expressive power, filters out low-response invalid signals, and highlights effective discriminative features. Channel concatenation of the normalized visual and textual features forms a hybrid feature, preserving the original information of both modalities and achieving preliminary information integration, avoiding information loss. This feature alignment process ensures that image and textual features are fused at the same scale and with the same distribution, improving fusion quality and efficiency, reducing noise impact, and enabling subsequent dynamic gating to more accurately determine modal reliability, achieving more precise adaptive weighted fusion.

[0032] According to a preferred embodiment of the present invention, the rational perception dynamic gating and multimodal matrix fusion in S23, using the SmartFusion module to achieve noise-resistant feature weighting, includes the following specific steps: S231, Gating Weight Hardware-Level Calculation Instruction: The control system first calls the visual backbone network to complete the image feature extraction of the current scene and obtain visual features. ;Will Text features determined by security constraints Concatenate along the channel dimension to generate a hybrid feature sequence. ;Will The input is fed into a multilayer perceptron gating network deployed in an edge computing device; the gating network, acting as a modal reliability calculation module, dynamically calculates and outputs the visual channel reliability for the current environment by learning the physical consistency of cross-sensor data under different physical environments. Reliability of text channels :

[0033] Where: σ(·) is the Sigmoid activation function, which normalizes the weights to the (0,1) interval; This is the weight matrix of the gated network; For the gated network bias vector, the mapping process of the multilayer perceptron corresponds to the general matrix multiplication instruction of the underlying hardware, which is used to dynamically quantize the physical signal-to-noise ratio of visual features and text features. S232. Multimodal signal fusion and scheduling output control based on dynamic gating: The microprocessor uses an element-wise multiplication instruction sequence to apply the evaluated reliability weights to the corresponding sensor feature streams in real time, completing dynamic weighted multimodal fusion and obtaining the final gated fusion features used for the actuator. :

[0034] Scheduling output and physical response logic: Gated fusion features are then mapped to specific logistics equipment scheduling output commands; when excessive text noise is detected on-site, the control node automatically reduces... Weights are used to cut off the propagation of noise commands; when encountering harsh environments, the system automatically increases the data weight of the text sensor; through a dynamic gating mechanism, deep modal complementarity between visual and text signals is achieved, ultimately outputting highly robust sorting or path planning control commands; S233. Fusion signal standardization and physical disturbance rejection processing based on low-level registers: in acquiring multimodal gated fusion features Subsequently, to prevent hardware computational overflow caused by sudden loud noise in the industrial environment, the system invoked layer normalization of the edge processing chip. Hardware acceleration instructions force the signal distribution to stabilize along the feature channel dimension, outputting a smoothed fused feature for final terminal scheduling. :

[0035] in: This represents the statistical mean of the fused signals in the current batch. For characteristic variance, , These are the hardware affine compensation parameters obtained by training using historical logistics scenario data. It is a very small constant used to ensure the physical safety of the underlying division instructions; S234. Underlying hardware mapping from multimodal panoramic features to logistics attributes: standardizing smooth features... The decision output terminal, composed of a multi-layer sensing mechanism, is loaded. By scheduling the underlying computing core to execute multiple rounds of matrix-vector multiplication instructions, the multi-modal sensing signals in the high-dimensional space are reduced in dimensionality and mapped to the unnormalized log probabilities of each physical commodity category. This completes the physical fusion closed loop of the front-end cross-modal anti-interference features.

[0036] This technical solution employs the SmartFusion module to achieve dynamic noise-resistant feature fusion. The hybrid features are input into a multilayer perceptron gating network, and the gating weight vector is dynamically output through cross-modal consistency learning. This weight is then normalized to the (0,1) interval using the Sigmoid function, ensuring smooth and controllable weights. Element-wise multiplication achieves dynamic weighted fusion; text weights are automatically reduced when text noise is high and increased when visual blur is present, achieving dual-modal complementarity and active noise suppression, thus improving anti-interference capabilities at the fusion level. Layer normalization is introduced at the end of the fusion process to stabilize the feature distribution, avoiding feature drift and numerical oscillations, and improving model output consistency. Finally, the fused features are fed into the multilayer perceptron to output product category prediction Logits, forming a complete multimodal intelligent scheduling model. This dynamic gating mechanism can adaptively handle complex situations such as single-modal failure, strong noise pollution, and feature loss, significantly improving the model's robustness, recognition accuracy, and operational stability under extreme conditions, outperforming traditional fixed-weight fusion methods.

[0037] According to a preferred embodiment of the present invention, the physical determination of safety constraints and the joint scheduling instruction output of the PLC in step S24 include the following specific steps: S241. Decoding physical product categories based on multimodal probability distribution: Receive the predicted probability distribution of product categories output by the front-end perception module. , Given the pre-set total number of product categories in the logistics and warehousing system, the main control unit uses the underlying comparator circuit to execute the argmax instruction, retrieves the register index corresponding to the probability peak, and establishes the predicted product category of the physical package on the current sorting line. As a result of physical identification:

[0038] S242. Retrieve security inspection, explosion-proof, and scheduling attributes by calling the preset mapping table: The main control unit determines the physical category based on the identified attributes. The logistics attribute mapping table in local non-volatile memory is invoked via addressing instructions. This transforms category identifiers into underlying parameters for low-altitude logistics scheduling: basic scheduling priority instructions. Interruption signage for hazardous chemicals explosion prevention Among them, the logo Used to trigger the system's security control logic for fragile, contraband, or hazardous packages at the hardware level; S243. Joint airworthiness assessment of UAVs based on multi-source physical sensor closed-loop: Activate the physical sensor array at the ground sorting station to collect the real physical weight of the current product package in real time. Physical actual volume Read the hardware load limit of the currently assigned low-altitude logistics drone, i.e., the maximum load threshold. With maximum available warehouse volume In conjunction with the previously identified hazardous chemical symbols The airworthiness indication function is calculated jointly using hardware logic gates. :

[0039] in, These are indicator functions, mapped to underlying conditional comparison instructions; The control logic ensures that: only when the logistics package has a weight margin The volume does not exceed the limit, and When in the specified state, output the permission to mount command to the UAV flight control system. =1; otherwise, output blocking and rejection instructions. =0; then a physical control closed loop is achieved, from multimodal vision and text perception to the final scheduling of mechanical actuators; S244. Calculate the final hardware scheduling instruction and send it to the PLC: ;when When =0, the scheduling command collapses to level 0, sending a rejection command containing an anomaly code to the hardware, triggering the solenoid valves and mechanical blocking rods on the ground sorting line to intercept and unload the goods; when When =1, the basic scheduling priority is used. Generate specific low-altitude takeoff and landing control messages, allocate UAV flight rights and takeoff sequences, and output the final transport decision: send an emergency takeoff command with priority 1 or a normal takeoff sequence command with priority 3 to the UAV flight control system via the Mavlink communication protocol.

[0040] This technical solution determines the optimal commodity category based on the category probability distribution output by the model. It converts this into scheduling priorities and hazardous materials explosion-proof labels through a pre-set logistics attribute mapping relationship, achieving a transformation from visual recognition to logistics semantics. Real-time weight and volume data are read, and an airworthiness indicator function is calculated based on the UAV's maximum payload and maximum available volume. This indicator function enables a triple safety judgment: passing the airworthiness judgment requires meeting the following conditions simultaneously: weight not exceeding limits, volume not exceeding limits, and non-hazardous materials. When airworthiness is 0, scheduling instructions are reset to zero, and the system automatically executes load rejection and physical blocking to eliminate safety risks. When airworthiness is 1, flight rights and takeoff sequences are allocated according to priority, achieving priority scheduling for emergency items and orderly scheduling for general items. This mechanism deeply couples multimodal recognition results with low-altitude physical constraints, constructing a perception-decision-control closed loop. It strictly adheres to UAV operation safety regulations, achieving automated, safe, and orderly intelligent scheduling, reducing the cost of manual intervention and the probability of safety accidents.

[0041] According to a preferred embodiment of the present invention, the training of model parameters under low-altitude constraints in S3 includes the following specific steps: S31. Set training hyperparameters: Total training epochs = 50, Batch size = 128, Input image size uniformly scaled to Img-size = 224×224, Learning rate... Adopting a dynamic adjustment strategy 1e -4 / 3e -4 / 1e -5 Weight decay coefficient =1e -4 ; S32. Input the divided training and validation set data into the multimodal intelligent scheduling model, and perform forward and backward propagation to update the model parameters; S33. The entire training process uses GPU acceleration for computation, and the actual payload, actual volume, and hazardous chemical explosion-proof labels of the drone are coupled in the iterative optimization to continuously update the model weights to improve the recognition accuracy in high-noise scenarios. S34. After completing 50 rounds of training, a converged multimodal intelligent scheduling model is obtained, which is used for subsequent physical determination of safety constraints and joint scheduling instruction output of PLC.

[0042] This technical solution sets standardized training hyperparameters, with 50 training rounds to ensure full model convergence. A batch size of 128 balances computational efficiency and gradient stability, and a uniform 224×224 image size ensures standardized input. Dynamic learning rates of 1e-4 / 3e-4 / 1e-5 achieve rapid convergence in the early stages and fine-tuning in the later stages. A weight decay coefficient of 1e-4 suppresses overfitting and improves generalization ability. The training and validation sets are input into the model for forward and backward propagation, enabling adaptive parameter updates. GPU acceleration is used throughout to improve training speed and shorten the iteration cycle. The training process is coupled with low-altitude constraints such as load, volume, and hazardous chemical explosion-proof tags to ensure that the model's optimization objective is consistent with actual scheduling needs, improving the rationality and safety of output decisions. After training, a convergent and stable multimodal intelligent scheduling model is obtained, which can be directly deployed on edge devices to perform real-time recognition and scheduling, ensuring high accuracy, low latency, and high reliability even under high noise and strong interference conditions.

[0043] According to a preferred embodiment of the present invention, the multi-index robustness accuracy evaluation of S4 includes the following specific steps: S41. Accuracy Calculation: The proportion of correctly predicted samples to the total number of samples is used to characterize the overall classification accuracy of the model under extreme industrial noise.

[0044] S42. Recall rate calculation: The proportion of true positive examples that are successfully predicted represents the model's failure to detect the target product.

[0045] S43. Precision calculation: The proportion of true positives among predicted positives characterizes the reliability of the model's classification results;

[0046] S44. Comprehensive score calculation: The harmonic mean of precision and recall is used to comprehensively characterize the overall recognition performance of the model.

[0047] S45. Parameter Count Statistics: The lightweight level is assessed by the total number of learnable weights in the model. The parameter count of the multimodal intelligent scheduling model is adapted to the low-latency deployment of low-altitude logistics edge equipment.

[0048] This technical solution employs a comprehensive evaluation system based on four metrics: accuracy, recall, precision, and F1-Score. Accuracy reflects the overall correctness of classification, recall measures the number of missed items, precision reflects the reliability of classification results, and F1-Score comprehensively characterizes overall recognition performance. These four metrics complement each other, providing a comprehensive and objective evaluation of the model's recognition ability and robustness in high-noise environments. Simultaneously, the system statistically analyzes the number of learnable parameters of the model, assesses its lightweight nature, and determines whether it meets the low-power, low-latency, and real-time operation requirements of edge computing devices. The model of this invention has only 1.22M parameters, far lower than traditional multimodal models, significantly reducing computational and storage overhead while maintaining extremely high recognition accuracy, making it suitable for deployment on edge devices and embedded systems such as drones. This evaluation system effectively verifies the model's accuracy, efficiency, resource consumption, and engineering applicability, providing objective data support for model optimization, deployment selection, and practical applications, ensuring that the solution meets the needs of industrial implementation.

[0049] In another aspect, the present invention also provides a low-altitude logistics multimodal intelligent scheduling system for high-noise environments.

[0050] A low-altitude logistics multimodal intelligent scheduling system for high-noise environments includes: Multi-source sensing unit: including visual sensors carried by the drone and ground sorting sensors, used to collect RGB images of goods and OCR strings of express delivery labels; Intelligent computing unit: Deploys a multimodal intelligent scheduling model, including a visual feature extraction module, a text feature extraction module, a cross-modal feature alignment module, a SmartFusion module, and a joint decision-making module; Cloud-based scheduling unit: Used to receive the transportation decisions output by the intelligent computing unit and issue scheduling instructions to the drone; Drone execution unit: Receives scheduling instructions from the cloud-based scheduling unit and executes low-altitude logistics delivery tasks.

[0051] This technical solution constructs a four-layer collaborative low-altitude logistics multimodal intelligent scheduling system. The multi-source perception unit uses UAV visual sensors and ground sorting sensors to collect real-time data from multiple sources, including product images, waybills, load, and volume, providing comprehensive input for the system. The intelligent computing unit carries a complete multimodal intelligent scheduling model, achieving end-to-end intelligent processing through visual feature extraction, text feature extraction, cross-modal alignment, dynamic gating fusion, and joint decision-making modules, completing noise suppression, feature fusion, safety judgment, and schedule generation. The cloud-based scheduling unit handles command forwarding, task allocation, and air rights management, enabling collaborative scheduling of multiple UAVs. The UAV execution unit performs physical actions such as loading, takeoff, delivery, and rejection according to commands, achieving a closed-loop scheduling system. The entire system has a clear structure, distinct layers, and high collaborative efficiency, enabling stable operation in low-altitude logistics hubs with high noise, strong interference, and strong vibration. It achieves fully automated, intelligent, and safe scheduling, possessing high robustness, high real-time performance, and high engineering application value.

[0052] The one or more technical solutions provided by this invention have the following advantages compared with the prior art: (1) A dynamic gating multimodal fusion architecture is constructed. By combining missing mask and tensor separation technology, high-precision recognition is maintained in high noise, damaged waybill and homogeneous packaging scenarios. The robustness of the model is significantly improved and it is suitable for the harsh working conditions of low-altitude take-off and landing hubs. (2) Deeply coupled with physical constraints such as load, volume, and hazardous chemicals, the closed loop of perception and scheduling is realized through the airworthiness indicator function, and the safety judgment and priority allocation are automatically completed, which complies with the low-altitude logistics specifications and reduces human intervention and safety risks; (3) The model has a small number of parameters and high computational efficiency, supports low-latency deployment of edge devices, significantly reduces space complexity compared with similar models, balances accuracy and resource consumption, and has excellent engineering implementation value. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the principle of the method of the present invention.

[0054] Figure 2 This is a framework diagram for constructing multimodal networks and dynamic gating.

[0055] Figure 3 This is an architecture diagram of a multimodal intelligent scheduling model.

[0056] Figure 4 This is the architecture diagram of the SmartFusion module.

[0057] Figure 5 This is a schematic diagram of the system of the present invention.

[0058] Figure 6 This is a comparison curve between the model of this invention and existing models.

[0059] In the diagram: 1. Visual sensor; 2. Ground sorting sensor; 3. Multimodal intelligent scheduling model; 4. Cloud scheduling unit. Detailed Implementation

[0060] Please refer to Figure 1 , Figure 2 , Figure 3 This embodiment provides a basic method for multimodal intelligent scheduling of low-altitude logistics in high-noise environments, which is applicable to the scheduling scenarios of standard low-altitude logistics hubs.

[0061] The execution environment of this embodiment is a low-altitude unmanned aerial vehicle (UAV) logistics distribution center. The operating area has high noise interference factors such as rotor airflow vibration, sudden changes in lighting, easy wear and tear on express waybills, and highly homogenized product packaging. The system needs to maintain stable identification and scheduling under the above operating conditions.

[0062] The execution flow of this embodiment is carried out in four steps, S1 to S4, and the structure of each step is explained below with reference to the accompanying drawings: S1 is for logistics data acquisition and noise preprocessing. For example... Figure 1 As shown at the starting point of the process, the system synchronously collects data through a visual sensor 1 mounted on a drone and a ground sorting sensor 2. The visual sensor 1 acquires images of the product's outer packaging, while the ground sorting sensor 2 collects text information from the express delivery label. Simultaneously, it labels each product with three key attributes: load capacity, volume, and hazardous materials explosion-proof label, forming raw logistics data. The raw data suffers from problems such as high noise, missing modalities, and unstable features. Therefore, a dedicated DataLoader for the sensors is used to load the data, and dual industrial disaster simulation noise is injected into the dataset to augment it. This operation simulates extreme situations in real-world scenarios such as label failure, image blurring, and feature drift, allowing the model to adapt to a high-noise environment during the training phase and improving its generalization ability.

[0063] S2 is constructed using a multimodal network and dynamic gating, corresponding to... Figure 2 The overall framework and Figure 3 The model architecture is as follows. This step is divided into four sub-steps, S21 to S24, which constitute a complete multimodal processing chain.

[0064] S21 performs visual feature extraction and missing modality masking. Visual features are extracted from product images, and text features are extracted from express delivery labels. When text is missing, blurry, or invalid, the system automatically performs a masking operation on the text features, setting invalid features to zero to prevent noise from propagating along the network and interfering with the model output. This mechanism is one of the core guarantees for the stability of this invention in high-noise environments, preventing single-modal failures from causing overall system collapse.

[0065] S22 performs cross-modal feature linear mapping and alignment. Visual and textual features are fed into a linear mapping network and projected onto the same feature space, achieving dimensionality unification and numerical normalization. OCR noise is suppressed through batch normalization, and invalid signals are filtered out using the ReLU activation function. Finally, the two feature channels are concatenated to form a hybrid feature. This process eliminates modal differences, ensuring the effectiveness and stability of subsequent fusion.

[0066] S23 performs dynamic intelligent gating and feature fusion. (Corresponding to...) Figure 3 The central structure incorporates a dynamic gating unit for hybrid feature input. The system adaptively assesses the reliability of the two modalities, automatically reducing the weight of noisy data and increasing the weight of clean data to achieve anti-interference feature fusion and form a multimodal intelligent scheduling model. The dynamic gating adjusts the modal ratio based on real-time input quality; high text noise reduces text weight, while visual blurriness increases it, achieving adaptive complementarity.

[0067] The S24 performs intelligent joint decision-making for low-altitude logistics. Based on the commodity identification results output by the model, combined with load capacity, volume, and hazardous chemical explosion-proof tags, it calculates airworthiness and scheduling priority. The system automatically determines whether the commodity meets the requirements for drone transport. If it does, it allocates air rights and takeoff order; if it does not, it directly refuses to carry the commodity and executes physical blocking, ultimately outputting the drone transport decision.

[0068] S3 is for training model parameters under low-altitude constraints. A multimodal intelligent scheduling model 3 was built using the PyTorch framework, with hyperparameters such as training epochs, batch size, image size, dynamic learning rate, and weight decay coefficient set, and GPU acceleration used for computation. During training, physical constraints such as UAV payload, volume, and hazardous materials explosion-proof tags were continuously coupled, and the model weights were iteratively updated to make the model output more closely match the actual operating rules of low-altitude logistics.

[0069] S4 is a multi-metric robustness and accuracy evaluation. It comprehensively tests the model's recognition accuracy and anti-interference capability under high noise conditions using four metrics: accuracy, recall, precision, and F1-Score. Simultaneously, it counts the number of model parameters to determine whether it meets the requirements for lightweight deployment on edge devices. In this embodiment, the model can run stably on the edge computing module with a latency of less than 100ms, meeting the real-time scheduling requirements of low-altitude logistics.

[0070] This embodiment fully implements the basic technical solution of the present invention. Through multimodal perception, dynamic gating fusion, airworthiness decision-making, and lightweight training, it solves the problems of inaccurate identification, unstable scheduling, and uncoupling of constraints in high-noise environments. It can be directly applied to small and medium-sized low-altitude logistics stations and has advantages such as clear structure, stable operation, and simple deployment.

[0071] Example 2 Please refer to Figure 2 , Figure 3 , Figure 4 Based on Example 1, this embodiment focuses on improving anti-interference capability, feature extraction accuracy, dynamic fusion effect and scheduling security, and is suitable for large low-altitude logistics hubs with high noise, strong interference and high-density take-off and landing.

[0072] The core improvements in this embodiment include: high-precision dataset reconstruction, visual feature enhancement of C2f_Logi and Logi_Bottleneck, explicit masking of missing modalities, linear mapping alignment, SmartFusion dynamic fusion, accurate calculation of airworthiness indicator functions, normalized hyperparameter training, multi-index quantitative evaluation, and systematic hardware deployment.

[0073] In the S1 data preprocessing stage, this embodiment uses a homogenized RPC commodity dataset as the visual foundation, supplementing it with labels for load capacity, volume, and hazardous chemical explosion-proof tags for low-altitude logistics scenarios, thus completing the reconstruction of a dedicated dataset. All samples are divided into training, validation, and test sets in a 7:2:1 ratio to ensure sufficient training, reliable validation, and objective testing. The system is injected with dual industrial disaster simulation noise: 30% of the samples are injected with incorrect category semantics and their text features are amplified to simulate waybill failure; 70% of the samples are subjected to a 3–5 times random feature amplitude shift to simulate visual distortion caused by rotor vibration and sudden changes in illumination. After this processing, the model maintains high-precision output even under extreme noise conditions.

[0074] The logistics data acquisition and noise preprocessing in S1, when no publicly available dedicated dataset is available, uses a homogeneous RPC commodity dataset as the visual foundation. The dataset is reconstructed by supplementing it with labels indicating load capacity, volume, and hazardous chemical explosion-proof tags according to low-altitude logistics scenarios. All samples in the dataset are divided into training, validation, and test sets in a 7:2:1 ratio. Dual industrial disaster simulation noise is injected into the sensor-dedicated DataLoader: 30% of the samples are injected with incorrect category semantics and their text features are amplified by 5 times to simulate the failure of express waybill text and abrupt changes in commodity images; the remaining 70% of the samples retain correct semantics and are subjected to a 3-5 times random feature amplitude drift to simulate the harsh working conditions of real low-altitude logistics hubs. This technical solution constructs a dedicated visual foundation for low-altitude logistics using a homogeneous RPC commodity dataset, supplementing it with load capacity, volume, and hazardous chemical explosion-proof tags to achieve scenario-based reconstruction of the dataset, ensuring that the training samples closely match the actual transport attributes of UAV logistics. Dividing the training, validation, and test sets in a 7:2:1 ratio ensures sufficient model training, reliable validation, and objective testing, improving the standardization of the training process and the credibility of the results. The dual industrial disaster simulation noise can highly reproduce the harsh working conditions of low-altitude logistics hubs. Injecting incorrect category semantics and amplifying text features into 30% of the samples can simulate extreme text anomalies such as damaged, torn, and malfunctioning waybills. Applying a 3-5 times random feature amplitude drift to 70% of the samples can simulate visual feature distortion caused by rotor airflow vibration, sudden changes in illumination, and sensor drift. This preprocessing method allows the model to fully engage with various types of high-noise data during the training phase, learning noise distribution patterns and anti-interference characteristics. This significantly improves recognition stability, generalization ability, and robustness in real high-noise environments, preventing scheduling failures due to sudden changes in operating conditions and ensuring the continuous and reliable operation of the UAV logistics system.

[0075] The visual feature extraction and missing modality masking in step S21, which extracts visual features from the product image acquired by the visual sensor, includes the following specific steps: S211, Instruction-level Edge Feature Extraction: The GPU of the edge computing node is invoked to load the pre-processed RGB tensor of the product image into the video memory as a block matrix; a custom frequency-decoupled operator is deployed in the deep feature stream of the general object detection network backbone (based on a standard convolutional network or Transformer framework). This operator has a built-in high-low frequency separation bottleneck layer; through non-intrusive memory reuse, direct modification of the standard visual backbone is avoided. S212. Visual perception signal splitting and multi-band feature extraction based on Logi_Bottleneck: The control unit receives the initial visual feature matrix of the logistics scene obtained by the visual backbone network. To simultaneously capture the global physical contours and local high-frequency edges (such as damaged textures and tape edges) of a target object (e.g., packages on a sorting conveyor belt), the system schedules the underlying tensor processing unit and utilizes the Logi_Bottleneck module to process the feature matrix. Along the channel dimension Performs signal segmentation and routing at the physical memory level. This operation separates the mixed visual signal into semantic tensors corresponding to the low-frequency global environment and detail tensors corresponding to high-frequency local mutations. The instruction set is expressed as follows:

[0076] Where: the separated low-frequency spatial signal Routed to a hardware-accelerated branch of dilated convolution with a large receptive field, used to perceive the global physical semantics of the target in a complex context (such as the overall size and location topology of the cargo); high-frequency spatial signals The features are synchronously distributed to the standard convolutional computation branch, specifically for accurately extracting barcode edges, packaging wrinkles, and textures from the cargo surface. After multi-band processing, the system re-aligns and concatenates the aforementioned dual-band features in the video memory channel. It then compresses the spatial resolution by calling the Global Average Pooling instruction, filters redundant background noise, and finally outputs a low-dimensional, high-purity visual feature vector v_feat (i.e., ...) for subsequent cross-modal gating scheduling. ).

[0077] This technical solution inputs product images into the backbone of a general object detection network (e.g., using a lightweight backbone network) to construct a C2f_Logi module. Through a built-in Logi_Bottleneck structure, it achieves hard separation of visual feature channel dimensions, decomposing features into semantic tensors and detail tensors. The semantic tensor is fed into a dilated convolutional branch to extract global semantic information, ensuring the accuracy of product category recognition; the detail tensor is fed into a standard convolutional branch to extract high-frequency detail information such as edges and textures, solving the feature confusion problem caused by the homogenization of lightweight packaging. This dual-branch parallel structure can forcibly retain fine-grained features in deep networks, avoiding the loss of key discriminative information during dimensionality reduction in conventional convolutional networks, and improving the ability to distinguish between similar-looking but different-category logistics products. Finally, a one-dimensional visual feature vector is output through channel concatenation and global average pooling, achieving efficient compression of high-dimensional image information into low-dimensional discriminative features. This retains core discriminative information while reducing computational load, providing high-quality and robust visual input for subsequent cross-modal fusion, and improving the overall model's feature representation ability in noisy environments.

[0078] The visual feature extraction and missing modality masking in S21, which extracts text features from the express waybill text obtained by the ground sorting sensor, includes the following specific steps: S213. The preprocessed express waybill text is introduced into the text modality encoder. The text modality encoder maps the express waybill text into an initial text feature vector through OCR feature processing. ; S214. Security Constraint Determination and Missing Mask Blocking Based on Physical Sensor Status: The system continuously monitors the physical working status of the electronic tag reader or industrial camera OCR text encoding module. The initial text feature vector t_feat (i.e., in the formula...) is used... Perform safety constraint determination:

[0079] When RFID tags fall off, radio frequency signals are lost, or OCR lenses are blocked, resulting in all valid information bits being empty (i.e., the sum of the absolute values ​​of the text feature vectors approaches zero) in a logistics scenario. The system triggers a hardware-level explicit masking mechanism at the system level, forcibly setting t_feat of the current batch to zero. This physical-level truncation not only prevents invalid or divergent gradient signals caused by sensor failure from propagating to subsequent physical control nodes (such as motor drive modules), but also sends a low-level interrupt signal of "text modality missing" to the main control console, ensuring the safe operation of physical devices.

[0080] This technical solution transforms express delivery waybill text into initial text feature vectors through text modal encoders and OCR feature processing, achieving digitized and standardized expression of text information. When the text is missing, blurred, damaged, or recognition fails, explicit masking is executed through conditional judgment to block it, making the sum of the absolute values ​​of the text feature vectors approach zero. This cuts off invalid features and noise signals from propagating to subsequent network layers at the source, preventing invalid gradients from interfering with model training and inference. This missing modality masking mechanism constitutes a pre-emptive physical circuit breaker protection, preventing model feature distribution shifts, abnormal outputs, and scheduling misjudgments due to waybill failure. Compared with traditional direct fusion methods, this solution significantly reduces the risk of noise propagation, improves the model's fault tolerance and stability in text modality failure scenarios, and ensures that even in the case of single-modality failure, the system can still rely on another modality to complete reliable recognition, providing a safety guarantee for multimodal scheduling in high-noise and strong-interference environments.

[0081] The cross-modal feature linear mapping and alignment of S22 includes the following specific steps: S221. Hardware-level spatial projection alignment and physical noise suppression of heterogeneous sensor signals: Given the visual characteristics v_feat captured by industrial cameras (i.e., in the formula) ) and the text features t_feat (i.e., in the formula) parsed by the radio frequency reading or image recognition module Due to the heterogeneity of the underlying data dimensions, the system control unit performs hardware-level spatial projection on the two heterogeneous sensor signals by scheduling a linear mapping network module named Text_Project. This process maps to the underlying tensor operation logic, forcibly aligning the bimodal features to a unified register processing dimension. To match the hardware bandwidth requirements of subsequent fusion computing:

[0082]

[0083] in: and These are the projection weight matrices of visual and textual features in the linear mapping network, respectively. This is the projection bias vector for visual features. Specifically for the text feature projection bias vector, and specifically for the text signal processing branch: considering that the complex environment of the logistics site (such as package stacking and squeezing, strong light reflection from the conveyor belt) can easily cause garbled characters or radio frequency crosstalk noise in the optical character recognition (OCR) module, the system calls the BN(·) (batch normalization) underlying instruction to force smoothing and standardize the distribution of physical feature data of the current batch, thereby isolating and suppressing OCR recognition noise at the hardware operation level; then, the ReLU(·) (corrected linear unit) hardware activation instruction is executed in series, and its nonlinear truncation characteristics are used to filter out invalid weak electrical signals or negative interference signals at the underlying level, ensuring that the text perception signal passed to the downstream node has extremely high purity; S222. Physical splicing and hybrid feature construction of multi-source sensing signals at the video memory level: After completing the dimensional alignment of heterogeneous data streams and text channel noise reduction filtering, the underlying main control chip uses memory block transfer instructions in the high-speed data bus of on-chip shared memory to transfer the processed pure visual enhancement features. Text Enhancement Features Physical-level data stitching is performed along the feature channel dimension to construct a hybrid feature matrix that integrates panoramic information from the logistics site. Its internal bus instruction logic is expressed as follows:

[0084] The aforementioned underlying splicing process completes the signal conversion from independent data streams from a single physical sensor to a unified multimodal hybrid bus, providing a structurally complete data foundation for the subsequent dynamic quantization of the signal-to-noise ratio of each physical sensor in the gating network.

[0085] This technical solution projects visual and textual features into a unified feature space using the linear mapping network Text_Project, achieving dimensional alignment and numerical normalization. This eliminates differences in dimension, distribution, and numerical range among different modal features, laying the foundation for subsequent stable fusion. Batch normalization standardizes feature distribution, suppresses numerical fluctuations caused by OCR noise and image noise, and improves model training stability and convergence speed. Correcting the linear unit activation function introduces non-linear expressive power, filters out low-response invalid signals, and highlights effective discriminative features. Channel concatenation of the normalized visual and textual features forms a hybrid feature, preserving the original information of both modalities and achieving preliminary information integration, avoiding information loss. This feature alignment process ensures that image and textual features are fused at the same scale and with the same distribution, improving fusion quality and efficiency, reducing noise impact, and enabling subsequent dynamic gating to more accurately determine modal reliability, achieving more precise adaptive weighted fusion.

[0086] The dynamic intelligent gating and feature fusion of S23, using the SmartFusion module to achieve noise-resistant feature weighting, includes the following specific steps: S231, Gating Weight Hardware-Level Calculation Instructions: The control system first calls the Visual Backbone network to extract image features of the current scene, obtaining a visual feature vector v_feat. v_feat is then concatenated with the text feature vector t_feat (which has undergone safety constraint checks) along the channel dimension to generate a mixed feature sequence. . Will The input is fed into a multilayer perceptron (MLP) gating network deployed in an edge computing device. This gating network, acting as a modal reliability calculation module, learns the physical consistency of cross-sensor data under different physical environments (e.g., comparing visual feature quality with RF signal strength under current lighting conditions), dynamically calculates and outputs a gating control weight vector specific to the current environment. (Visual channel reliability) and (Text channel reliability):

[0087] Where: σ(·) is the Sigmoid activation function, which normalizes the weights to the (0,1) interval; This is the weight matrix of the gated network; For the gated network bias vector, the mapping process of the multilayer perceptron corresponds to the general matrix multiplication (GEMM) instruction of the underlying hardware, which is used to dynamically quantize the physical signal-to-noise ratio of visual features (v_feat) and text features (t_feat). S232. Multimodal signal fusion and scheduling output control based on dynamic gating: The system's main control chip uses an element-wise multiplication instruction sequence to apply the evaluated reliability weights to the corresponding sensor feature streams in real time, completing dynamically weighted multimodal fusion and obtaining the final gated fusion features used for the actuator. :

[0088] Scheduling Output and Physical Response Logic: This gating fusion feature is then mapped to specific logistics equipment scheduling output commands. When excessive text noise is detected on-site (such as RFID cross-reading or OCR misidentification in dense environments), the control node automatically reduces... Weights are used to cut off the propagation of noise commands; when encountering harsh environments (such as heavy fog or strong light causing blurred vision in industrial cameras), the system automatically increases the data weights of the text sensor. Through the above-mentioned dynamic gating mechanism based on physical environment feedback, deep modal complementarity between visual and text signals is achieved, ultimately outputting highly robust sorting or path planning control commands; S233. Fusion signal standardization and physical disturbance rejection processing based on low-level registers: In acquiring multimodal gated fusion features Subsequently, to prevent hardware computation overflow caused by sudden strong noise in industrial environments (such as severe vibration of conveyor belts or sudden changes in lighting), the system calls the layer normalization function of the edge processing chip. Hardware acceleration instructions force the signal distribution to stabilize along the feature channel dimension, outputting a smoothed fused feature for final terminal scheduling. :

[0089] in: This represents the statistical mean of the fused signals in the current batch. For characteristic variance, , These are the hardware affine compensation parameters obtained by training using historical logistics scenario data. It is a very small constant used to ensure the physical safety of the underlying division instruction (to prevent division by zero errors). S234. Underlying hardware mapping from multimodal panoramic features to logistics attributes: This involves mapping standardized smooth features... The decision output is loaded from a multilayer perceptron (MLP). This module executes multiple rounds of matrix-vector multiplication instructions by scheduling the underlying computing core to reduce the dimensionality of the high-dimensional multimodal sensing signals and map them to the unnormalized log probabilities (Logits) of each physical commodity category. This completes the physical fusion closed loop of the front-end cross-modal anti-interference features.

[0090] This technical solution employs the SmartFusion module to achieve dynamic noise-resistant feature fusion. The hybrid features are input into a multilayer perceptron gating network, and the gating weight vector is dynamically output through cross-modal consistency learning. This weight is then normalized to the (0,1) interval using the Sigmoid function, ensuring smooth and controllable weights. Element-wise multiplication achieves dynamic weighted fusion; text weights are automatically reduced when text noise is high and increased when visual blur is present, achieving dual-modal complementarity and active noise suppression, thus improving anti-interference capabilities at the fusion level. Layer normalization is introduced at the end of the fusion process to stabilize the feature distribution, avoiding feature drift and numerical oscillations, and improving model output consistency. Finally, the fused features are fed into the multilayer perceptron to output product category prediction Logits, forming a complete multimodal intelligent scheduling model. This dynamic gating mechanism can adaptively handle complex situations such as single-modal failure, strong noise pollution, and feature loss, significantly improving the model's robustness, recognition accuracy, and operational stability under extreme conditions, outperforming traditional fixed-weight fusion methods.

[0091] The intelligent joint decision-making process for low-altitude logistics in S24 includes the following specific steps: S241. Decoding physical product categories based on multimodal probability distribution: Receive the predicted probability distribution of product categories output by the front-end perception module. ( (Given the total number of physical commodity categories pre-set for the logistics and warehousing system), the main control unit uses the underlying comparator circuit to execute the arg max instruction, retrieves the register index corresponding to the probability peak, and establishes the predicted commodity category of the physical package on the current sorting line. As a result of physical identification:

[0092] S242. Retrieve security inspection, explosion-proof, and scheduling attributes by calling the preset mapping table: The main control unit determines the attributes based on the identified physical category. The logistics attribute mapping table in local non-volatile memory (such as ROM or Flash) is invoked through addressing instructions. This transforms category identifiers into underlying parameters for low-altitude logistics scheduling: basic scheduling priority instructions. Interruption signage for hazardous chemicals explosion prevention Among them, the logo Used to trigger the system's security control logic for fragile, contraband, or hazardous packages at the hardware level; S243. Joint airworthiness assessment of UAVs based on multi-source physical sensor closed-loop: Activate the physical sensor array (including industrial weighing sensors and 3D structured light volume scanners) at the ground sorting station to collect the real physical weight of the current product package in real time. Physical actual volume The system reads the hardware load limit, i.e., the maximum load threshold, of the currently assigned low-altitude logistics drone (UAV). With maximum available warehouse volume In conjunction with the previously identified hazardous chemical symbols The airworthiness indication function is calculated jointly using hardware logic gates. :

[0093] in, As an indicator function, it maps to underlying conditional comparison instructions. This physical control logic ensures that: only if the weight of the logistics package is not overloaded (load margin)... ), volume does not exceed the limit, and explosion-proof markings are not triggered (i.e. When in a certain state, output a permission to mount command to the UAV flight control system. =1; otherwise, output blocking and rejection instructions. =0. This step completely realizes the physical control closed loop from multimodal vision / text perception to the final scheduling of mechanical actuators; S244. Calculate the final hardware scheduling instruction and send it to the programmable logic controller (PLC). ;when When =0, the scheduling command collapses to level 0, issuing a "Reject Command" containing an exception code to the hardware, triggering the solenoid valves and mechanical blocking rods on the ground sorting line to intercept and unload the goods; when When =1, the basic scheduling priority is used. Specific low-altitude takeoff and landing control messages (including UAV_ID, Time_Slot, Path_Node) are generated to allocate UAV flight rights and takeoff sequences, and the final transport decision is output (specifically, by sending a priority-1 "Emergency Item Priority Takeoff Command (Priority_Takeoff)" or a priority-3 "Regular Item Waiting Sequence Command (Queue_Wait)" to the UAV flight control system via the Mavlink communication protocol). This technical solution determines the optimal commodity category based on the category probability distribution output by the model, and converts it into scheduling priority and hazardous chemical explosion-proof labels through a pre-set logistics attribute mapping relationship, realizing the conversion from visual recognition to logistics semantics. Real-time weight and volume data are read, and airworthiness indicator functions are calculated by combining the UAV's maximum payload and maximum available volume. The indicator function achieves a triple safety judgment: weight not exceeding limits, volume not exceeding limits, and non-hazardous materials simultaneously must meet these conditions to pass the airworthiness judgment. When airworthiness is 0, scheduling instructions are reset to zero, and the system automatically executes load rejection and physical blocking to eliminate safety risks. When airworthiness is 1, flight rights and takeoff sequences are allocated according to priority, achieving priority scheduling for emergency missions and orderly scheduling for general missions. This mechanism deeply couples multimodal recognition results with low-altitude physical constraints to construct a perception-decision-control closed loop, strictly adhering to UAV operation safety standards, achieving automated, safe, and orderly intelligent scheduling, and reducing the cost of manual intervention and the probability of safety accidents.

[0094] The training of model parameters under the low-altitude constraint of S3 includes the following specific steps: S31. Set training hyperparameters: Total training epochs = 50, Batch size = 128, Input image size uniformly scaled to Img-size = 224×224, Learning rate... Adopting a dynamic adjustment strategy 1e -4 / 3e -4 / 1e -5 Weight decay coefficient =1e -4 ; S32. Input the divided training and validation set data into the multimodal intelligent scheduling model, and perform forward and backward propagation to update the model parameters; S33. The entire training process uses GPU acceleration for computation, and in the iterative optimization, the drone's payload, volume, and hazardous chemical explosion-proof tags are coupled to continuously update the model weights to improve the recognition accuracy in high-noise scenarios. S34. After completing 50 rounds of training, a converged multimodal intelligent scheduling model is obtained, which can be used for subsequent intelligent joint decision-making in low-altitude logistics.

[0095] The model training process of this invention adopts conventional training methods, and some hyperparameter settings are shown in Table 1: Table 1 Hyperparameter Settings Table

[0096] The S4 evaluation phase uses four indicators for comprehensive assessment, and the statistical model has only 1.22M parameters, which is far lower than traditional multimodal models. It can be easily deployed on edge computing devices to meet the low latency requirements of low-altitude logistics.

[0097] S41. Accuracy Calculation: The proportion of correctly predicted samples to the total number of samples is used to characterize the overall classification accuracy of the model under extreme industrial noise.

[0098] S42. Recall rate calculation: The proportion of true positive examples that are successfully predicted represents the model's failure to detect the target product.

[0099] S43. Precision calculation: The proportion of true positives among predicted positives characterizes the reliability of the model's classification results;

[0100] S44. Comprehensive score calculation: The harmonic mean of precision and recall is used to comprehensively characterize the overall recognition performance of the model.

[0101] S45. Parameter Count Statistics: The lightweight level is assessed by the total number of learnable weights in the model. The parameter count of the multimodal intelligent scheduling model is adapted to the low-latency deployment of low-altitude logistics edge equipment.

[0102] Meanwhile, this embodiment also fully implements the system structure corresponding to the present invention, including a multi-source sensing unit, an intelligent computing unit, a cloud scheduling unit 4, and a drone execution unit. The multi-source sensing unit is responsible for data acquisition, the intelligent computing unit runs the model, the cloud scheduling unit 4 issues instructions, and the drone execution unit completes the delivery. The four units work together to form a complete scheduling closed loop.

[0103] This embodiment uses a multi-level optimization technology to enable the system to maintain high recognition accuracy, stable scheduling, and strong security even under complex working conditions such as high noise, strong vibration, damaged waybills, and homogeneous packaging. It is suitable for use in high-requirement scenarios such as large logistics hubs and high-density drone take-off and landing sites.

[0104] Example 3 Please refer to Figure 1 , Figure 5 , Figure 6 This embodiment quantifies and demonstrates the technical effects of the present invention from three levels: experimental verification, model comparison, and engineering deployment, providing data support and deployment guidance for practical applications.

[0105] This embodiment uses a public dataset and a self-built low-altitude logistics dataset for testing. The hardware environment consists of an Intel i9 processor, an NVIDIA RTX 4090 GPU, and 16GB of memory. The edge deployment platform is an NVIDIA Jetson AGX Orin, and the software environment is PyTorch 2.0 and CUDA 12.2.

[0106] Experiment 1 compares the performance of the model of this invention with existing mainstream multimodal models. Figure 6 The comparison curves are shown. The test set contains 10,000 samples of typical working conditions such as high noise, damaged shipping labels, sudden changes in lighting, and homogenized packaging. The comparison models cover representative multimodal fusion schemes such as FiLM, Cross-At, Concat, and MulT. Experimental results show that the multimodal intelligent scheduling model of this invention achieves an accuracy of 0.982, a precision of 0.981, a recall of 0.982, and an F1-Score of 0.982, significantly better than Concat (early fusion) (0.940), DCN (0.942), and CrossViT (0.962). In terms of parameter count, this invention has only 1.22M, far lower than CrossViT's 3.06M and Bi-CoAttn's 2.74M, achieving a dual breakthrough in recognition accuracy and model lightweighting.

[0107] Experiment 2 is the system engineering deployment test, corresponding to Figure 5 The system architecture is described. In this embodiment, the model is deployed to a Jetson AGX Orin edge device, connected to a real low-altitude logistics station, and subjected to a continuous 72-hour uninterrupted operation test. During the test, a total of 120,000 goods were processed, with a waybill damage rate of approximately 22%, and frequent environmental vibrations and sudden changes in lighting. The overall system recognition accuracy was 99.6%, the rejection accuracy was greater than 98%, and the average scheduling latency was 78ms. No safety issues such as missed scheduling, incorrect scheduling, or overloaded takeoff occurred, fully meeting the requirements for stable industrial-grade operation.

[0108] Experiment 3 was a modal failure tolerance test. Even with the text modality completely failing and relying solely on the visual modality, the system maintained an accuracy of 92.3%. When visual blur was severe and relying solely on text features, the accuracy remained at 89.1%, significantly higher than traditional fusion models. This demonstrates that dynamic gating and the missing mask mechanism can effectively achieve single-modal failure tolerance, greatly improving adaptability to complex environments.

[0109] From an engineering deployment perspective, the system of this invention can be quickly integrated into existing drone dispatching platforms, with simple hardware integration: the visual sensor 1 uses a USB high-definition camera, the ground sorting sensor 2 integrates barcode scanning and weighing modules, the intelligent computing unit uses an edge host, the cloud dispatching unit 4 uses a cloud server or a local server, and the drone execution unit is compatible with mainstream models such as multi-rotor and VTOL drones. The entire system supports hot-swapping, dynamic configuration, and remote upgrades, and can adapt to low-altitude logistics needs of different scales and scenarios.

[0110] In actual operation, the system starts automatically after power-on, and the multi-source sensing unit collects images, shipping labels, weight, and volume data in real time. After the data is transmitted to the intelligent computing unit, noise preprocessing, feature extraction, modality alignment, dynamic fusion, category prediction, and airworthiness determination are performed in sequence. The cloud scheduling unit 4 receives the decision results in real time and sends take-off, waiting, and rejection commands to the UAV. The UAV execution unit completes the physical actions, forming a complete automated closed loop.

[0111] This embodiment demonstrates through three sets of experiments that the present invention possesses five major advantages in high-noise environments: high precision, strong robustness, lightweight, low latency, and high safety. It can effectively solve the pain points of existing technologies, such as single-modal fragility, poor fusion robustness, inability to couple airworthiness constraints, and bloated models that are difficult to deploy. It fully meets the development needs of intelligent, safe, and efficient low-altitude logistics.

[0112] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for multi-modal intelligent scheduling of low-altitude logistics in a high-noise environment, characterized in that, Includes the following steps: S1. Logistics data acquisition and noise preprocessing: The visual sensor (1) and ground sorting sensor (2) on the UAV are used to collect RGB images of goods and OCR strings of express waybills in the logistics scene. At the same time, the actual load, actual volume and explosion-proof label of hazardous chemicals corresponding to the goods are defined and labeled to form raw logistics data. The raw logistics data is processed by the sensor-specific DataLoader and double industrial disaster simulation noise is injected into the raw logistics data to realize the preprocessing enhancement of the dataset. S2. Construction of Multimodal Networks and Dynamic Gating: This includes the following steps: S21, Hardware-level visual feature extraction and text encoding missing mask: Call the underlying video stream interface of the visual sensor (1), transfer the acquired RGB image array of the product to the GPU memory of the edge node, and extract the visual high-frequency tensor that retains the physical edge of the product packaging as a visual feature through the edge computing node. The OCR string of the express waybill output by the ground sorting sensor (2) and the signal strength indication of the RFID tag are read synchronously through the serial communication interface for text feature encoding; When the RSSI intensity indicator is detected to be lower than the preset hardware reception threshold, or when the OCR string returns an empty byte, a hardware-level interrupt request is sent to the microprocessor to trigger the masking logic: a mask matrix with all zeros on the diagonal is generated by bitwise AND operation, and the text features are forced to be zeroed at the physical memory level to block the cross-layer propagation of invalid noise gradients. S22, cross-modal feature affine mapping and tensor alignment: input the visual high-frequency tensor in the GPU memory and the text feature after mask zero processing into the preset affine transformation matrix respectively to perform linear projection operation, unify the tensor dimensions of the two groups of heterogeneous features or pool and align to the same dimension constant ; then call the batch normalization function to perform floating point standardization operation, forcibly constrain the values of the two to the same distribution interval, and eliminate the dimensional difference caused by light sudden change and character distortion; S23. Physical Perception Dynamic Gating and Multimodal Matrix Fusion: The microprocessor independently calculates the signal-to-noise ratio (SNR) of the current visual high-frequency tensor and text features, quantifying modal reliability using the physical SNR. The SNR is input into the Sigmoid activation function to generate a dynamic weight matrix in the range of 0 to 1, triggering the gating hardware logic: when the text SNR falls below the set safety threshold, the gating circuit automatically decays the value of the text weight matrix according to the function curve. Subsequently, the processor's underlying general matrix multiplication instruction is called, and the Hadamard product weighting operation is performed on the aligned visual features and text features using the dynamic weight matrix. Finally, the channel splicing instruction is executed to forcibly splice the dual-modal features in the channel dimension, fusing them into a multimodal feature tensor, and the predicted probability distribution Logits of the product category is calculated and output using a multilayer perceptron. S24. Safety constraint physical judgment and PLC joint scheduling instruction output: Retrieve the standard estimated physical attributes corresponding to the Logits prediction category from the local database, and compare them with the actual weight and actual volume fed back in real time by the gravity sensor and infrared sensor; Execute safety constraint judgment logic: Only when the actual weight does not trigger the threshold alarm of the UAV's maximum load, the actual volume does not exceed the cargo hold size, and the explosion-proof label identification bit is non-hazardous chemical, the airworthiness indicator function is judged to be 1, otherwise it is 0; When the judgment is 0, a hexadecimal encoded load rejection instruction is sent to the underlying PLC via the fieldbus, directly driving the pneumatic solenoid valve of the sorting line to pop out the mechanical blocking rod; When the judgment is 1, a departure message is sent to the UAV flight control system through the Mavlink communication protocol, and the probability distribution priority is extracted to generate 1-3 level scheduling takeoff sequence codes; S3. Model parameter training under low-altitude constraints: The multimodal intelligent scheduling model (3) was trained using the PyTorch deep learning tool. The model was configured according to the set total training rounds, the amount of input data each time, the uniform size of the images, the dynamic learning rate, and the weight decay coefficient. The GPU graphics card was used to accelerate the calculation. During the training process, the model parameters were continuously optimized and updated in combination with the actual payload, actual volume, and explosion-proof label of the UAV. S4. Multi-index robustness and accuracy evaluation: Using four indicators, accuracy, recall, precision and F1-Score, the accuracy and anti-interference ability of the multimodal intelligent scheduling model (3) in a high-noise environment are tested; at the same time, the number of parameters of the multimodal intelligent scheduling model (3) is counted to evaluate whether the multimodal intelligent scheduling model (3) is lightweight and suitable for edge device deployment.

2. The low-altitude logistics multimodal intelligent scheduling method for high-noise environments as described in claim 1, characterized in that: The logistics data acquisition and noise preprocessing in S1, when no publicly available dedicated dataset is available, uses a homogeneous RPC commodity dataset as the visual basis. Actual load, actual volume, and hazardous chemical explosion-proof labels are added according to the low-altitude logistics scenario to complete dataset reconstruction. All samples in the dataset are divided into training, validation, and test sets in a 7:2:1 ratio. Dual industrial disaster simulation noise is injected into the sensor-dedicated DataLoader: 30% of the samples are injected with incorrect category semantics and the text features are amplified by 5 times to simulate the failure of the OCR string on the express waybill and the abrupt change in the RGB image of the commodity; 70% of the samples retain correct semantics and are subjected to a 3-5 times random feature amplitude drift to simulate the harsh working conditions of a real low-altitude logistics hub.

3. The low-altitude logistics multimodal intelligent scheduling method for high-noise environments as described in claim 1, characterized in that: The hardware-level visual feature extraction and text encoding missing masking in S21 extracts visual features from the RGB image of the product acquired by the visual sensor, including the following specific steps: S211, Instruction-level edge feature extraction: Call the GPU of the edge computing node to load the pre-processed RGB image of the product into the video memory in the form of a block matrix; Deploy a custom frequency domain decoupling node in the deep feature flow of the visual backbone network, and the frequency domain decoupling node has a built-in high and low frequency separation bottleneck layer; Avoid direct modification of the visual backbone network through non-intrusive memory reuse. S212. Visual perception signal splitting and multi-band feature extraction based on Logi_Bottleneck: The microprocessor receives the initial visual feature matrix of the logistics scene obtained by the visual backbone network. To simultaneously capture the global physical contours and local high-frequency edges of the target object, the underlying tensor processing unit is scheduled, and the Logi_Bottleneck module is used to process the feature matrix. Along the channel dimension Performs signal segmentation and routing at the physical memory level; separates the hybrid visual signal into semantic tensors corresponding to the low-frequency global environment and detail tensors corresponding to high-frequency local mutations. The instruction set is expressed as follows: Where: the separated low-frequency spatial signal Routed to a hardware-accelerated branch of dilated convolution with a large receptive field, used to perceive the overall size and positional topology of the cargo against a complex background; high-frequency spatial signals The features are synchronously distributed to the standard convolutional computation branch, specifically for accurately extracting barcode edges, packaging wrinkles, and textures from the surface of goods. After multi-band processing, the dual-band features are re-aligned and stitched together in the memory channel. Spatial resolution is compressed by calling global average pooling instructions, redundant background noise is filtered out, and finally, low-dimensional, high-purity visual features are output for subsequent cross-modal gating scheduling. .

4. The low-altitude logistics multimodal intelligent scheduling method for high-noise environments as described in claim 3, characterized in that: The hardware-level visual feature extraction and text encoding missing masking in S21 extracts text features from the OCR string of the express waybill obtained by the ground sorting sensor, including the following specific steps: S213. The preprocessed express waybill OCR string is introduced into the text modality encoder. The text modality encoder maps the express waybill OCR string into text features through OCR feature processing. ; S214. Security constraint determination and missing feature masking based on physical sensor status: The system continuously monitors the physical working status of the electronic tag reader or visual sensor OCR text encoding module; and analyzes the acquired text features... Perform safety constraint determination: When RFID tags fall off, radio frequency signals are lost, or OCR lenses are obstructed in logistics scenarios, resulting in all valid information bits being empty, the sum of the absolute values ​​of the text features approaches zero. The underlying hardware triggers a text feature loss masking mechanism, which masks the current batch of text features missing. Forced zeroing prevents invalid or divergent gradient signals caused by sensor failure from propagating to subsequent physical control nodes. At the same time, it sends a low-level interrupt signal indicating missing text modality to the main control console to ensure the safe operation of physical devices.

5. The low-altitude logistics multimodal intelligent scheduling method for high-noise environments as described in claim 1, characterized in that: The cross-modal feature affine mapping and tensor alignment of S22 includes the following specific steps: S221. Hardware-level spatial projection alignment and physical noise suppression of heterogeneous sensor signals: Given the visual features captured by vision sensors Text features analyzed by radio frequency reading or image recognition modules Given the heterogeneity of the underlying data dimension, the microprocessor performs hardware-level spatial projection on the heterogeneous sensor signals by scheduling linear mapping network modules. This process maps to the underlying tensor operation logic, forcibly aligning the bimodal features to a unified register processing dimension. To match the hardware bandwidth requirements of subsequent fusion computing: in: and These are the projection weight matrices of visual and textual features in the linear mapping network, respectively. The visual feature projection bias vector; The text feature projection bias vector; For the text signal processing branch: Considering that the stacking and squeezing of logistics packages and the strong light reflection from the conveyor belt can easily cause garbled characters or radio frequency crosstalk noise in the OCR strings of express waybills, the system calls the BN(·) low-level instruction to force smoothing and standardize the distribution of physical feature data of the current batch, thereby isolating and suppressing OCR recognition noise at the hardware operation level; then, the ReLU(·) hardware activation instruction is executed in series, and its non-linear truncation characteristics are used to filter out invalid weak electrical signals or negative interference signals at the bottom level, ensuring that the text perception signal passed to the downstream node has extremely high purity; S222. Physical splicing and hybrid feature construction of multi-source sensing signals at the video memory level: After completing the dimensional alignment of heterogeneous data streams and text channel noise reduction filtering, the underlying main control chip uses memory block transfer instructions in the high-speed data bus of the on-chip shared video memory to transfer the processed pure visual enhancement features. Text Enhancement Features Physical-level data stitching is performed along the feature channel dimension to construct a hybrid feature matrix that integrates panoramic information of the logistics site. Its bus instruction logic is expressed as follows: By stitching together the underlying layers, the signal conversion from an independent data stream from a single physical sensor to a unified multimodal hybrid bus is completed, enabling the gating network to dynamically quantize the signal-to-noise ratio of the physical sensor.

6. The low-altitude logistics multimodal intelligent scheduling method for high-noise environments as described in claim 1, characterized in that: The S23 rational perception dynamic gating and multimodal matrix fusion, using the SmartFusion module to implement noise-resistant feature weighting, includes the following specific steps: S231, Gating Weight Hardware-Level Calculation Instruction: The control system first calls the visual backbone network to complete the image feature extraction of the current scene and obtain visual features. ;Will Text features determined by security constraints Concatenate along the channel dimension to generate a hybrid feature sequence. ;Will The input is fed into a multilayer perceptron gating network deployed in an edge computing device; the gating network, acting as a modal reliability calculation module, dynamically calculates and outputs the visual channel reliability for the current environment by learning the physical consistency of cross-sensor data under different physical environments. Reliability of text channels : Where: σ(·) is the Sigmoid activation function, which normalizes the weights to the (0,1) interval; This is the weight matrix of the gated network; For the gated network bias vector, the mapping process of the multilayer perceptron corresponds to the general matrix multiplication instruction of the underlying hardware, which is used to dynamically quantize the physical signal-to-noise ratio of visual features and text features. S232. Multimodal signal fusion and scheduling output control based on dynamic gating: The microprocessor uses an element-wise multiplication instruction sequence to apply the evaluated reliability weights to the corresponding sensor feature streams in real time, completing dynamic weighted multimodal fusion and obtaining the final gated fusion features used for the actuator. : Scheduling output and physical response logic: Gated fusion features are then mapped to specific logistics equipment scheduling output commands; when excessive text noise is detected on-site, the control node automatically reduces... Weights are used to cut off the propagation of noise commands; when encountering harsh environments, the system automatically increases the data weight of the text sensor; through a dynamic gating mechanism, deep modal complementarity between visual and text signals is achieved, ultimately outputting highly robust sorting or path planning control commands; S233. Fusion signal standardization and physical disturbance rejection processing based on low-level registers: in acquiring multimodal gated fusion features Subsequently, to prevent hardware computational overflow caused by sudden loud noise in the industrial environment, the system invoked layer normalization of the edge processing chip. Hardware acceleration instructions force the signal distribution to stabilize along the feature channel dimension, outputting a smoothed fused feature for final terminal scheduling. : in: This represents the statistical mean of the fused signals in the current batch. For characteristic variance, , These are the hardware affine compensation parameters obtained by training using historical logistics scenario data. It is a very small constant used to ensure the physical safety of the underlying division instructions; S234. Underlying hardware mapping from multimodal panoramic features to logistics attributes: standardizing smooth features... The decision output terminal, composed of a multi-layer sensing mechanism, is loaded. By scheduling the underlying computing core to execute multiple rounds of matrix-vector multiplication instructions, the multi-modal sensing signals in the high-dimensional space are reduced in dimensionality and mapped to the unnormalized log probabilities of each physical commodity category. This completes the physical fusion closed loop of the front-end cross-modal anti-interference features.

7. The low-altitude logistics multimodal intelligent scheduling method for high-noise environments as described in claim 1, characterized in that: The physical determination of safety constraints and the joint scheduling instruction output of PLC in S24 include the following specific steps: S241. Decoding physical product categories based on multimodal probability distribution: Receive the predicted probability distribution of product categories output by the front-end perception module. , Given the pre-set total number of product categories in the logistics and warehousing system, the main control unit uses the underlying comparator circuit to execute the argmax instruction, retrieves the register index corresponding to the probability peak, and establishes the predicted product category of the physical package on the current sorting line. As a result of physical identification: S242. Retrieve security inspection, explosion-proof, and scheduling attributes by calling the preset mapping table: The main control unit determines the physical category based on the identified attributes. The logistics attribute mapping table in local non-volatile memory is invoked via addressing instructions. This transforms category identifiers into underlying parameters for low-altitude logistics scheduling: basic scheduling priority instructions. Interruption signage for hazardous chemicals explosion prevention Among them, the logo Used to trigger the system's security control logic for fragile, contraband, or hazardous packages at the hardware level; S243. Joint airworthiness assessment of UAVs based on multi-source physical sensor closed-loop: Activate the physical sensor array at the ground sorting station to collect the real physical weight of the current product package in real time. Physical actual volume Read the hardware load limit of the currently assigned low-altitude logistics drone, i.e., the maximum load threshold. With maximum available warehouse volume In conjunction with the previously identified hazardous chemical symbols The airworthiness indication function is calculated jointly using hardware logic gates. : in, These are indicator functions, mapped to underlying conditional comparison instructions; The control logic ensures that: only when the logistics package has a weight margin The volume does not exceed the limit, and When in the specified state, output the permission to mount command to the UAV flight control system. =1; otherwise, output blocking and rejection instructions. =0; then a physical control closed loop is achieved, from multimodal vision and text perception to the final scheduling of mechanical actuators; S244. Calculate the final hardware scheduling instruction and send it to the PLC: ;when When =0, the scheduling command collapses to level 0, sending a rejection command containing an anomaly code to the hardware, triggering the solenoid valves and mechanical blocking rods on the ground sorting line to intercept and unload the goods; when When =1, the basic scheduling priority is used. Generate specific low-altitude takeoff and landing control messages, allocate UAV flight rights and takeoff sequences, and output the final transport decision: send an emergency takeoff command with priority 1 or a normal takeoff sequence command with priority 3 to the UAV flight control system via the Mavlink communication protocol.

8. The low-altitude logistics multimodal intelligent scheduling method for high-noise environments as described in claim 1, characterized in that: The training of model parameters under the low-altitude constraint of S3 includes the following specific steps: S31. Set training hyperparameters: Total training epochs = 50, Batch size = 128, Input image size uniformly scaled to Img-size = 224×224, Learning rate... Adopting a dynamic adjustment strategy 1e -4 / 3e -4 / 1e -5 Weight decay coefficient =1e -4 ; S32. Input the divided training and validation set data into the multimodal intelligent scheduling model, and perform forward and backward propagation to update the model parameters; S33. The entire training process uses GPU acceleration for computation, and the actual payload, actual volume, and hazardous chemical explosion-proof labels of the drone are coupled in the iterative optimization to continuously update the model weights to improve the recognition accuracy in high-noise scenarios. S34. After completing 50 rounds of training, a converged multimodal intelligent scheduling model is obtained, which is used for subsequent physical determination of safety constraints and joint scheduling instruction output of PLC.

9. The low-altitude logistics multimodal intelligent scheduling method for high-noise environments as described in claim 1, characterized in that: The multi-index robustness accuracy evaluation of S4 includes the following specific steps: S41. Accuracy Calculation: The proportion of correctly predicted samples to the total number of samples is used to characterize the overall classification accuracy of the model under extreme industrial noise. S42. Recall rate calculation: The proportion of true positive examples that are successfully predicted represents the model's failure to detect the target product. S43. Precision calculation: The proportion of true positives among predicted positives characterizes the reliability of the model's classification results; S44. Comprehensive score calculation: The harmonic mean of precision and recall is used to comprehensively characterize the overall recognition performance of the model. S45. Parameter Count Statistics: The lightweight level is assessed by the total number of learnable weights in the model. The parameter count of the multimodal intelligent scheduling model is adapted to the low-latency deployment of low-altitude logistics edge equipment.

10. A multimodal intelligent scheduling system for low-altitude logistics in high-noise environments, employing the multimodal intelligent scheduling method for low-altitude logistics in high-noise environments as described in any one of claims 1-9, characterized in that: include: Multi-source sensing unit: including a visual sensor (1) carried by the drone and a ground sorting sensor (2) for collecting RGB images of goods and OCR strings of express waybills; Intelligent computing unit: Deploys a multimodal intelligent scheduling model (3), including a visual feature extraction module, a text feature extraction module, a cross-modal feature alignment module, a SmartFusion module and a joint decision module; Cloud scheduling unit (4): used to receive the transportation decision output by the intelligent computing unit and issue scheduling instructions to the UAV; Unmanned aerial vehicle (UAV) execution unit: receives scheduling instructions from cloud scheduling unit (4) and performs low-altitude logistics delivery tasks.

Images