Mountainous area low-altitude logistics unmanned aerial vehicle anti-interference navigation method based on Beidou and visual SLAM fusion

Abstract

The invention provides a mountainous area low-altitude logistics unmanned aerial vehicle anti-interference navigation method based on Beidou and visual SLAM fusion, and relates to the technical field of low-altitude logistics. Comprising the steps that S1, flight control data and navigation data are collected and preprocessed; s2, training and deploying an LSTM model, and outputting an emergency response probability and a bandwidth preemption time window in the future 50 ms; s3, triggering navigation data pre-caching according to a prediction result; s4, constructing a linear extrapolation prediction model, and generating a navigation data prediction value when the flow is cut off; and S5, performing fusion navigation switching decision by using the actual measurement data when no flow is cut off, and performing fusion navigation switching decision by using a continuous sequence formed by the actual measurement data and the prediction data when the flow is cut off. According to the invention, flight control emergency response and bandwidth preemption time periods can be accurately pre-judged, cut-off data can be efficiently complemented, the anti-interference capability and positioning stability of a navigation system are improved, and the unmanned aerial vehicle is ensured to safely and efficiently complete logistics transportation tasks in a mountainous complex environment.

Description

Technical Field

[0001] This invention relates to the field of low-altitude logistics technology, and in particular to an anti-interference navigation method for low-altitude logistics drones in mountainous areas based on the fusion of BeiDou and visual SLAM. Background Technology

[0002] In the field of low-altitude logistics transportation in mountainous areas, drones need to rely on high-precision navigation systems to achieve autonomous flight and accurate delivery of goods. The reliability of their navigation directly determines the safety and efficiency of transportation. At present, the mainstream navigation solutions for low-altitude logistics drones in mountainous areas mostly adopt the fusion technology of Beidou satellite navigation and visual SLAM (simultaneous localization and mapping). The two complement each other: the Beidou system can provide global position information for long-distance navigation in open areas; visual SLAM can collect environmental feature points through cameras and maintain local positioning accuracy in scenarios where satellite signals are blocked (such as valleys and dense forests). The fusion of the two can cover the navigation needs of complex mountainous terrain.

[0003] However, the low-altitude environment in mountainous areas is complex, and drones are prone to encountering obstacles such as falling rocks, birds, and sudden turbulence, requiring the flight control system to respond quickly to avoid risks. Simultaneously, navigation relies on the fusion of BeiDou and visual SLAM. Signal blockage and interference in mountainous areas can easily trigger navigation system switching, requiring both systems to work together via bus data transmission. Existing flight control systems, to ensure safety during risk avoidance, have an emergency attitude adjustment priority mechanism. When triggered, this mechanism preempts the main bus bandwidth, causing temporary interruptions or delays in navigation data transmission. Current anti-interference navigation only addresses scenarios such as sensor failure and signal loss, failing to consider conflict scenarios where the flight control system's emergency response preempts the bus. This makes it unsuitable for dynamic interference caused by flight control bandwidth preemption. If the navigation system is in a critical state of switching between BeiDou and visual SLAM, it needs to continuously receive sensor data to complete decision-making and transition. When the flight control system preempts bandwidth, causing data interruption, it disrupts the switching process. When navigation data interruption lasts for more than 30ms, the fused navigation switching is prone to logic interruption due to the lack of continuous data input, leading to a sharp drop in positioning accuracy and, in severe cases, the risk of drone yaw and collision. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an anti-interference navigation method for low-altitude logistics drones in mountainous areas based on the fusion of BeiDou and visual SLAM.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] An anti-jamming navigation method for low-altitude logistics UAVs in mountainous areas based on the fusion of BeiDou and visual SLAM includes the following steps:

[0007] S1: Data acquisition and preprocessing;

[0008] S11: Data Acquisition;

[0009] Real-time acquisition of flight control and navigation data; the flight control data includes obstacle distance, relative speed, and attitude adjustment range; the navigation data includes: BeiDou CN0 value and visual SLAM feature point matching rate.

[0010] S12: Data preprocessing; outlier removal and time synchronization of flight control and navigation data; normalization of flight control data to obtain a normalized sequence as input to the model;

[0011] S2: Flight control emergency response prediction model training and deployment;

[0012] S21: LSTM model training; Collect over 100,000 historical flight control data points to train the LSTM model, and obtain the trained model after training is completed;

[0013] S22: Real-time predictive deployment;

[0014] Deploy the trained model into the navigation processing algorithm, and set a sliding window input:

[0015] Every 10ms, input a sequence of normalized values ​​of obstacle distance, relative speed, and attitude adjustment amplitude from the flight control data of the most recent 50ms, and output the emergency response probability and bandwidth preemption time window for the next 50ms.

[0016] S3: Determine whether to trigger pre-caching based on the prediction results;

[0017] If the probability of an emergency response in the next 50ms exceeds 60% and the bandwidth preemption window exceeds 15s, pre-caching is triggered, and the collected navigation data is cached starting from 100ms before the current time; otherwise, caching is not performed.

[0018] S4: Navigation data prediction;

[0019] A linear extrapolation prediction model is constructed for navigation data. When a disconnection is detected, the prediction value for the next 50ms is obtained based on the three sets of navigation data before the disconnection.

[0020] S5: Integrated navigation switching data input;

[0021] If there is no interruption in the data flow, the collected measured navigation data is used to make decisions on fusion navigation switching;

[0022] If the data flow is interrupted, the last three sets of measured data and linear extrapolation prediction data before the interruption are obtained to form a continuous data sequence, which replaces the interrupted data input and is used to make decisions on fusion navigation switching.

[0023] Furthermore, step S12 specifically includes:

[0024] S121: Outlier removal;

[0025] Collect flight control and navigation data of the UAV during normal flight, and calculate the mean μ and standard deviation σ for each data dimension; remove data that exceed the range of [μ-3σ, μ+3σ].

[0026] S122: Time synchronization;

[0027] Deploy timers, record the relative time of the flight control data / navigation data according to the time of receipt, and align the timestamps of the flight control data and navigation data;

[0028] S123: Flight control data normalization;

[0029] Map the flight control data to the interval [0,1].

[0030] Furthermore, step S21 includes:

[0031] S211: Construct the dataset;

[0032] More than 100,000 historical flight control data points were collected, each including input features and labels;

[0033] The input features are a sequence of normalized values ​​of obstacle distance, relative speed, and attitude adjustment range from the flight control data of the first 50ms.

[0034] The tag indicates whether an emergency response will be triggered within the next 50ms. 真 ;y 真 It is 0 or 1; and in y 真 When the value is 1, the start time t1 and end time t2 of the emergency trigger response are marked.

[0035] The flight control historical data was divided into training set, verification set, and test set in a 7:2:1 ratio.

[0036] S212: Building the model;

[0037] Select an LSTM model for construction; including:

[0038] Input layer: Configure the input layer as 3D features with a time step of 50;

[0039] LSTM layer: Set to 2 layers, with 64 hidden units;

[0040] Fully connected layer: Outputs three values: the emergency response probability P and the bandwidth preemption time window T. window : [t 始 ,t 终 ];

[0041] Output layer: P is activated with sigmoid, T window Use linear activation;

[0042] S213: Design the loss function;

[0043] The joint loss function is adopted: Loss = α·Loss emer +β·Loss window Training is performed; where α and β are the weight parameters in the joint loss function; Loss emer It is a binary classification cross-entropy loss; Loss window For time window MSE loss;

[0044] S214: Model Evaluation and Optimization;

[0045] After training, the test set is used for metric evaluation and optimization until emergency response prediction is achieved:

[0046] Accuracy > 95%, recall > 90%, confidence > 90%; mean absolute error of prediction within the time window < 5ms.

[0047] Furthermore, S213 specifically includes:

[0048] 1) Initialize the weight parameters in the joint loss function. Set the initial values ​​of the weight parameters as follows: α = 0.7, β = 0.3;

[0049] 2) Calculate the binary cross-entropy loss to obtain the emergency response prediction value y1 of the model's original output; and the label y 真 First, the emergency response prediction value y1 is converted into a probability P using the sigmoid function.

[0050] Next, calculate the single-sample loss, Loss i =-y 真 ×logP-(1-y 真 )×log(1-P); where y 真 The label indicates whether an emergency response has been triggered, including 1 or 0; P is the probability P obtained by the sigmoid function transformation; log represents the natural logarithm.

[0051] Then, the mean of the sample loss within the batch is calculated to obtain the binary classification cross-loss: Where N is the number of batches configured for the optimizer; Loss i This is the single-sample loss for calculation;

[0052] 3) Calculate the time window MSE loss to obtain the time window [t] for model prediction. 始 ,t 终 ]; Real-time window [t1,t2]; Label y 真 ;

[0053] First, the samples within the batch are screened, retaining those labeled y.真 For a sample with a value of 1, the number of recorded samples is M;

[0054] Next, calculate the single-sample MSE loss: Where t 始 , t 终 t1 and t2 are the time windows for model prediction, and t1 and t2 are the time windows for label annotation.

[0055] Then, the mean of the M filtered samples is calculated to obtain the time window MSE loss;

[0056] Where M is the number of samples selected; Loss k To calculate the single-sample MSE loss, if M is 0, then the loss is... window =0;

[0057] 4) Calculate the total loss Loss = α·Loss emer +β·Loss window ;

[0058] 5) Minimize the total loss using the optimizer and update the model weights;

[0059] Configure the Adam optimizer parameters with an initial learning rate of 0.001 and a decay rate of 1e. -5 The batch size is 32, the number of iterations is 50, and the iteration optimization is performed. If the validation set loss does not decrease for 5 consecutive rounds, training is stopped.

[0060] Furthermore, step S4 includes:

[0061] S41: Configure prediction parameters;

[0062] A linear extrapolation prediction model is constructed for the BeiDou CN0 value;

[0063] Configuration parameters:

[0064] State vector: Where C k The value of BeiDou CN0 at time k; The rate of change of the BeiDou CN0 value;

[0065] State transition matrix: Where Δt is the time interval for collecting BeiDou CN0 values, the sampling frequency is 50Hz, and the time interval Δt is 20ms;

[0066] Process noise covariance: Where σ q The value is 0.01, and Δt is 20ms;

[0067] Observation matrix: H = [1 0];

[0068] The observation noise variance R = 0.25;

[0069] To improve the visual SLAM feature point matching rate, we constructed prediction models for the visual SLAM feature point matching rate in the same way and adjusted the state vectors accordingly. Where M k Let k be the visual SLAM feature point matching rate at time k; For visual SLAM feature point matching rate; adjust Δt to 33.3ms, σ q The variance of the observation noise is 0.005, and the variance of the observation noise is R = 0.0004.

[0070] S42: Disconnection detection, to determine whether linear extrapolation prediction is triggered;

[0071] If no new data is received for three consecutive BeiDou CN0 value sampling cycles or visual SLAM feature point matching rate sampling cycles, it is determined to be a data interruption, and linear extrapolation prediction is triggered; otherwise, linear extrapolation prediction is not triggered.

[0072] S43: Perform linear extrapolation prediction calculations to obtain predicted values.

[0073] Furthermore, S43 includes:

[0074] S431: After triggering linear extrapolation prediction, read the last 3 sets of navigation data before the disconnection from the cached navigation data and use them as the initial input of the linear extrapolation predictor;

[0075] S432: Calculate the initial rate of change:

[0076] Regarding the BeiDou CN0 value, based on the three sets of navigation data before the disconnection, it is set as C. k-2 C k-1 C k The initial rate of change was determined by linear fitting:

[0077] S433: State prediction, generates CN0 prediction value within the next 50ms;

[0078] According to the prediction formula: Obtain the predicted value of BeiDou CN0 within a future time step, where Δt' is the prediction time step, which is preset to 10ms;

[0079] According to X k+1|k =F×X k Update the state vector at time k+1, where F is the state transition matrix and X k The state vector at time k; obtained from the state vector at time k+1. And the calculated BeiDou CN0 prediction value C k+1|kBased on the prediction formula, we continue to obtain the predicted value of BeiDou CN0 within the future time step, and so on, to obtain the prediction sequence within 50ms.

[0080] S434: Using the same prediction logic, obtain the predicted value of the visual SLAM feature point matching rate within the next 50ms.

[0081] Furthermore, obstacle distance, relative velocity, and attitude adjustment amplitude are collected synchronously at a frequency of 1 kHz; the BeiDou CN0 value sampling frequency is 50 Hz; and the visual SLAM feature point matching rate is the proportion of successful inter-frame matching output by the SLAM algorithm, with a sampling frequency of 30 Hz.

[0082] Furthermore, in step S22, the trained LSTM model is quantized into TensorRT INT8 format.

[0083] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention uses the LSTM model to perform in-depth analysis of flight control data, which can accurately predict the probability of emergency response triggering and the bandwidth preemption time window within the next 50ms, so that the navigation system can identify the risk of flight control bandwidth preemption in advance, avoid the navigation data interruption problem caused by the sudden preemption of the bus due to emergency response in the traditional solution, reserve sufficient time for subsequent data caching and completion, and reduce the risk of navigation data transmission interruption from the source. (2) In response to the navigation data interruption problem caused by flight control bandwidth preemption, the present invention combines pre-cached data and linear extrapolation prediction model to achieve accurate data completion. When no new data is detected for three consecutive sampling periods, the system extracts the last three sets of measured data before the interruption from the cache as input, and generates the prediction sequence of Beidou CN0 value and visual SLAM feature point matching rate for the next 50ms through customized linear extrapolation prediction, forming a complete and continuous data sequence to replace the interrupted data. It effectively solves the problem of switching logic interruption when the data interruption exceeds 30ms in traditional fusion navigation, ensures stable navigation and positioning accuracy, significantly reduces the risk of UAV deviation and collision, and ensures the safety of low-altitude logistics transportation in mountainous areas. (3) This invention breaks through the limitations of traditional anti-interference navigation which is only for sensor failure and signal failure. It can effectively adapt to dynamic interference scenarios where flight control emergency response preempts the bus in mountainous areas, greatly improve the adaptability of the fusion navigation system to complex mountainous environments, and ensure that UAVs can continuously and stably complete autonomous flight and precise material delivery tasks in complex terrains such as valleys and dense forests. Attached Figure Description

[0084] Figure 1 This is a flowchart illustrating the steps of the anti-interference navigation method for low-altitude logistics drones in mountainous areas based on the fusion of BeiDou and visual SLAM, as described in this invention. Detailed Implementation

[0085] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0086] An anti-jamming navigation method for low-altitude logistics drones in mountainous areas based on the fusion of BeiDou and visual SLAM includes the following steps:

[0087] S1: Data acquisition and preprocessing;

[0088] S11: Data Acquisition;

[0089] Real-time acquisition of flight control and navigation data;

[0090] Flight control data is read and acquired from the flight control bus (CAN bus); the flight control data includes obstacle distance, relative speed, and attitude adjustment range;

[0091] Obstacle distance, relative velocity, and attitude adjustment amplitude are collected synchronously at a frequency of 1kHz.

[0092] S112: Connects to navigation sensors to collect navigation data, including: BeiDou CN0 value and visual SLAM feature point matching rate;

[0093] The BeiDou CN0 value is the carrier-to-noise ratio of the BeiDou module output, with a sampling frequency of 50Hz;

[0094] Visual SLAM feature point matching rate is the proportion of successful inter-frame matching output by a SLAM algorithm (such as ORB-SLAM3), with a sampling frequency of 30Hz.

[0095] S12: Data preprocessing;

[0096] S121: Outlier removal;

[0097] Collect flight control and navigation data of the UAV during normal flight, and calculate the mean μ and standard deviation σ for each data dimension; remove data that are outside the range of [μ-3σ, μ+3σ].

[0098] Obtain flight control and navigation data for normal flight with a period of 10 minutes, and calculate the mean μ1 and standard deviation σ1 of obstacle distances respectively; obtain real-time acquired obstacle distances, compare whether the obstacle distances are within the range of [μ1-3σ1, μ1+3σ1], and remove obstacle distances that are outside the range; similarly, remove outliers in relative speed, attitude adjustment amplitude, BeiDou CN0 value, and visual SLAM feature point matching rate that are outside the range respectively.

[0099] S122: Time synchronization;

[0100] Deploy a timer, record the relative time of the flight control data / navigation data according to the time of receipt, align the timestamps of the flight control data and navigation data to ensure that the time deviation is less than 1ms.

[0101] S123: Flight control data normalization;

[0102] Map the flight control data to the interval [0,1].

[0103] Through formula Where x represents the distance to the obstacle, relative velocity, and attitude adjustment range; x min With x max These are the physical limits corresponding to the flight control data;

[0104] For the distance to the obstacle, x min For 0m, x max The maximum detection range of the radar is set to 500m;

[0105] For relative velocity, set x min =-50m / s, x max It is 50 m / s.

[0106] For the attitude adjustment range, set x min -30°, x max It is 30°.

[0107] This invention employs the 3σ rule to eliminate outliers, ensures that the data time synchronization deviation is less than 1ms, and normalizes the flight control data to the [0,1] interval, providing high-quality input for model prediction.

[0108] S2: Flight control emergency response prediction model training and deployment;

[0109] S21: LSTM model training;

[0110] S211: Construct the dataset;

[0111] More than 100,000 historical flight control data points were collected (including samples of emergency response trigger scenarios and samples of normal flight states), with each data point including input features and labels;

[0112] The input features are a sequence of normalized values ​​of obstacle distance, relative speed, and attitude adjustment amplitude from the flight control data of the first 50ms; the flight control data is mapped to the interval [0,1] using the same method as S123.

[0113] The tag indicates whether an emergency response will be triggered within the next 50ms. 真 Emergency response commands from the flight control system are marked as 1 or 0, where 1 indicates activation and 0 indicates non-activation. And in y 真When the value is 1, the start time t1 and end time t2 of the emergency trigger response are marked.

[0114] The flight control historical data was divided into training set, verification set, and test set in a 7:2:1 ratio.

[0115] S212: Building the model;

[0116] Select an LSTM model for construction; including:

[0117] Input layer: Configure the input layer as 3D features with a time step of 50;

[0118] LSTM layer: Set to 2 layers, with 64 hidden units;

[0119] Fully connected layer: Outputs three values: the emergency response probability P and the bandwidth preemption time window T. window : [t 始 ,t 终 ];

[0120] Output layer: P is activated with sigmoid (range [0,1]), T window Use linear activation (unit: ms).

[0121] S213: Design the loss function;

[0122] The joint loss function is adopted: Loss = α·Loss emer +β·Loss window Conduct training;

[0123] Where α and β are the weight parameters of the joint loss function; Loss emer It is a binary classification cross-entropy loss; Loss window For time window MSE loss;

[0124] Specifically, it includes:

[0125] 1) Initialize the weight parameters in the joint loss function. Set the initial values ​​of the weight parameters as follows: α = 0.7, β = 0.3;

[0126] 2) Calculate the binary cross-entropy loss to obtain the emergency response prediction value y1 (LSTM fully connected layer output, no activation, arbitrary range) of the original model output; and the label y 真 First, the emergency response prediction value y1 is converted into a probability P using the sigmoid function.

[0127] Where y1 is the predicted value of the emergency response, and e is the base of the natural logarithm function;

[0128] Next, calculate the single-sample loss, Loss i =-y真 ×logP-(1-y 真 )×log(1-P); where y 真 The label indicates whether an emergency response has been triggered, including 1 or 0; P is the probability P obtained by the sigmoid function transformation; log represents the natural logarithm.

[0129] Then, the mean of the sample loss within the batch is calculated to obtain the binary classification cross-loss: Where N is the number of batches configured for the optimizer; Loss i This is the single-sample loss calculated.

[0130] 3) Calculate the time window MSE loss to obtain the time window [t] for model prediction. 始 ,t 终 ]; Real-time window [t1,t2]; Label y 真 ;

[0131] First, the samples within the batch are screened, retaining those labeled y. 真 For a sample with a value of 1, the number of recorded samples is M;

[0132] Next, calculate the single-sample MSE loss: Where t 始 , t 终 t1 and t2 represent the time windows predicted by the model, while t1 and t2 represent the actual time windows for label annotation.

[0133] Then, the mean of the M filtered samples is calculated to obtain the time window MSE loss;

[0134] Where M is the number of samples selected; Loss k To calculate the single-sample MSE loss, if M is 0, then the loss is... window It is 0.

[0135] 4) Calculate the total loss Loss = α·Loss emer +β·Loss window ;

[0136] 5) Minimize the total loss using the optimizer and update the model weights;

[0137] Configure the Adam optimizer parameters with an initial learning rate of 0.001 and a decay rate of 1e. -5 The batch size is 32, the number of iterations is 50, and the iteration optimization is performed. If the validation set loss does not decrease for 5 consecutive rounds, training is stopped.

[0138] S214: Model Evaluation and Optimization;

[0139] After training, the test set is used for metric evaluation and optimization until emergency response predictions are achieved with the following results: accuracy > 95%, recall > 90%, confidence > 90%; and mean absolute error of prediction within the time window < 5ms.

[0140] The mean absolute error is: 1 / 2(|t 始 -t1|+|t 终 -t2|).

[0141] S22: Real-time predictive deployment;

[0142] The trained LSTM model was quantized into TensorRT INT8 format and deployed into the navigation processing algorithm. A sliding window input was set: every 10ms, a sequence of normalized values ​​of obstacle distance, relative velocity, and attitude adjustment amplitude from the flight control data of the most recent 50ms was input, and the output was the emergency response probability and bandwidth preemption time window for the next 50ms. Quantizing the LSTM model into TensorRT INT8 format reduces computational resource consumption by approximately 55% compared to floating-point models. Under the constraints of limited hardware computing power and storage space of UAVs, this significantly improves the operational efficiency of the navigation system, extends the single-charge endurance of the UAV, and indirectly improves the efficiency of logistics transportation in mountainous areas.

[0143] S3: Determine whether to trigger pre-caching based on the prediction results;

[0144] If the probability of an emergency response within the next 50ms exceeds 60% and the bandwidth preemption window exceeds 15s, pre-caching is triggered, caching navigation data collected starting 100ms prior to the current time. Otherwise, no caching occurs. Pre-caching provides data support for data processing during subsequent network outages. Navigation data pre-caching is only initiated when the probability of an emergency response exceeds 60% and the bandwidth preemption window exceeds 15s, rather than indiscriminately caching data throughout the entire process. Optimizing the caching strategy by combining differentiated sampling frequencies—1kHz for flight control data, 50Hz for BeiDou data, and 30Hz for visual SLAM—can reduce storage resource usage by approximately 40%.

[0145] S4: Navigation data prediction;

[0146] S41: Configure prediction parameters;

[0147] Linear extrapolation prediction models are constructed for the BeiDou CN0 value and the visual SLAM feature point matching rate, respectively; BeiDou CN0 value is used as an example.

[0148] Configuration parameters:

[0149] State vector: Where C k The value of BeiDou CN0 at time k; The rate of change of the BeiDou CN0 value;

[0150] State transition matrix: Where Δt is the time interval for collecting BeiDou CN0 values, the sampling frequency is 50Hz, and the time interval Δt is 20ms;

[0151] Process noise covariance: Where σ q The value is 0.01, and Δt is 20ms;

[0152] Observation matrix: H = [1 0];

[0153] The observation noise variance R = 0.25.

[0154] Similarly, construct the visual SLAM feature point matching rate prediction model and adjust the state vector: Where M k Let k be the visual SLAM feature point matching rate at time k; For visual SLAM feature point matching rate; adjust Δt to 33.3ms, σ q The variance of the observation noise is 0.005, and the variance of the observation noise is R = 0.0004.

[0155] S42: Disconnection detection, to determine whether linear extrapolation prediction is triggered;

[0156] If no new data is received for three consecutive BeiDou CN0 value sampling cycles or visual SLAM feature point matching rate sampling cycles, it is determined as a data interruption, and linear extrapolation prediction is triggered; otherwise, linear extrapolation prediction is not triggered.

[0157] S43: Perform linear extrapolation prediction calculations to obtain the predicted values;

[0158] S431: After triggering linear extrapolation prediction, read the last 3 sets of navigation data before the disconnection from the cached navigation data and use them as the initial input of the linear extrapolation prediction model.

[0159] S432: Calculate the initial rate of change:

[0160] Based on the three sets of navigation data before the disconnection, taking the BeiDou CN0 value as an example, let the three sets of BeiDou CN0 values ​​be C. k-2 C k-1 C k ;

[0161] The initial rate of change was determined by linear fitting: Where Δt is the time interval for collecting BeiDou CN0 values.

[0162] S433: State prediction, generates CN0 prediction value within the next 50ms;

[0163] According to the prediction formula: Obtain the predicted value of BeiDou CN0 within a future time step, where Δt' is the prediction time step, which is preset to 10ms;

[0164] According to X k+1|k =F×X k Update the state vector at time k+1, where F is the state transition matrix and X k The state vector at time k; obtained from the state vector at time k+1. And the calculated BeiDou CN0 prediction value C k+1|k Based on the prediction formula, we continue to obtain the predicted value of BeiDou CN0 within the future time step, and so on, to obtain the prediction sequence within 50ms.

[0165] S434: Using the same linear extrapolation prediction logic, obtain the predicted value of the visual SLAM feature point matching rate within the next 50ms.

[0166] Based on three sets of measured data before the data interruption, the initial rate of change was calculated through linear fitting, making the completed data more closely match the actual data change trend. Tests showed that in a 50ms data interruption scenario, the BeiDou CN0 value completion error was ≤0.8dB-Hz, effectively ensuring positioning continuity during fusion navigation switching and avoiding a sharp drop in positioning accuracy due to completion errors.

[0167] This invention's linear extrapolation prediction model features independent parameter configuration modules for both BeiDou and visual SLAM. If a higher-precision BeiDou module (such as a BeiDou-3 high-precision board) is subsequently replaced, or the visual SLAM algorithm is upgraded (such as switching from ORB-SLAM3 to DSO-SLAM), only the state transition matrix and process noise covariance parameters of the corresponding modules need to be adjusted. This avoids overall system reconstruction and significantly simplifies later maintenance, reducing maintenance costs and improving system iteration flexibility.

[0168] S5: Integrated navigation switching data input;

[0169] If there is no interruption in the data flow, the collected measured navigation data is used to make decisions on fusion navigation switching;

[0170] If the data flow is interrupted, the last three sets of measured data and linear extrapolation prediction data before the interruption are obtained to form a continuous data sequence, which replaces the interrupted data input and is used to make decisions on fusion navigation switching.

[0171] This invention can accurately predict emergency flight control responses and bandwidth preemption periods, efficiently replenish interrupted data, improve the anti-interference capability and positioning stability of the navigation system, and ensure that UAVs can safely and efficiently complete logistics transportation tasks in complex mountainous environments.

[0172] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method for anti-interference navigation of low-altitude logistics UAVs in mountainous areas based on the fusion of BeiDou and visual SLAM, characterized in that: Includes the following steps: S1: Data acquisition and preprocessing; S11: Data Acquisition; Real-time acquisition of flight control and navigation data; the flight control data includes obstacle distance, relative speed, and attitude adjustment range. Navigation data includes: BeiDou CN0 value and visual SLAM feature point matching rate; S12: Data preprocessing; outlier removal and time synchronization of flight control and navigation data; normalization of flight control data to obtain a normalized sequence as input to the model; S2: Flight control emergency response prediction model training and deployment; S21: LSTM model training; Collect over 100,000 historical flight control data points to train the LSTM model, and obtain the trained model after training is completed; S22: Real-time predictive deployment; The trained model is deployed to the navigation processing algorithm. The sliding window input is set to input a sequence of normalized values ​​of obstacle distance, relative speed and attitude adjustment amplitude of the flight control data of the most recent 50ms every 10ms, and output the emergency response probability and bandwidth preemption time window for the next 50ms. S3: Determine whether to trigger pre-caching based on the prediction results; If the probability of an emergency response in the next 50ms exceeds 60% and the bandwidth preemption time window exceeds 15s, pre-caching is triggered, and the collected navigation data is cached starting from 100ms before the current time; otherwise, caching is not performed. S4: Navigation data prediction; A linear extrapolation prediction model is constructed for navigation data. When a disconnection is detected, the prediction value for the next 50ms is obtained based on the three sets of navigation data before the disconnection. S5: Integrated navigation switching data input; If there is no interruption in the data flow, the collected measured navigation data is used to make decisions on fusion navigation switching; If the data flow is interrupted, the last three sets of measured data and linear extrapolation prediction data before the interruption are obtained to form a continuous data sequence, which replaces the interrupted data input and is used to make decisions on fusion navigation switching.

2. The anti-interference navigation method for low-altitude logistics UAVs in mountainous areas based on the fusion of BeiDou and visual SLAM as described in claim 1, characterized in that: Step S12 specifically includes: S121: Outlier removal; Collect flight control and navigation data of the UAV during normal flight, and calculate the mean μ and standard deviation σ for each data dimension; remove data that exceed the range of [μ-3σ, μ+3σ]. S122: Time synchronization; Deploy timers, record the relative time of the flight control data / navigation data according to the time of receipt, and align the timestamps of the flight control data and navigation data; S123: Flight control data normalization; Map the flight control data to the interval [0,1].

3. The anti-interference navigation method for low-altitude logistics UAVs in mountainous areas based on the fusion of BeiDou and visual SLAM as described in claim 1, characterized in that: Step S21 includes: S211: Construct the dataset; More than 100,000 historical flight control data points were collected, each including input features and labels; The input features are a sequence of normalized values ​​of obstacle distance, relative speed, and attitude adjustment range from the flight control data of the first 50ms. The tag indicates whether an emergency response will be triggered within the next 50ms. 真 ;y 真 It is 0 or 1; and in y 真 When the value is 1, the start time t1 and end time t2 of the emergency trigger response are marked. The flight control historical data was divided into training set, verification set, and test set in a 7:2:1 ratio. S212: Building the model; Choose an LSTM model for construction; including: Input layer: Configure the input layer as 3D features with a time step of 50; LSTM layers: set to 2 layers, with 64 hidden units; Fully connected layer: Outputs three values: the emergency response probability P and the bandwidth preemption time window T. window : [t 始 , t 终 ]; Output layer: P is activated with sigmoid, T window Use linear activation; S213: Design the loss function; Using joint loss function Loss = α ⋅ Loss emer + β ⋅ Loss window Training is performed; where α and β are the weight parameters in the joint loss function. Loss emer It is a binary cross-entropy loss; Loss window For time window MSE loss; S214: Model Evaluation and Optimization; After training, the test set is used for metric evaluation and optimization until emergency response predictions are achieved with the following results: accuracy > 95%, recall > 90%, confidence > 90%; and mean absolute error of time window prediction < 5ms.

4. The anti-interference navigation method for low-altitude logistics UAVs in mountainous areas based on the fusion of BeiDou and visual SLAM as described in claim 3, characterized in that: S213 specifically includes: 1) Initialize the weight parameters in the joint loss function. Set the initial values ​​of the weight parameters as follows: α=0.7, β=0.3; 2) Calculate the binary cross-entropy loss to obtain the emergency response prediction value y1 of the model's original output; and the label y 真 First, the emergency response prediction value y1 is converted into a probability P using the sigmoid function. Next, the single-sample loss is calculated. ; where y 真 The label indicates whether an emergency response has been triggered, including 1 or 0; P is the probability P obtained by the sigmoid function transformation; log represents the natural logarithm. Then, the mean of the sample loss within the batch is calculated to obtain the binary classification cross-loss: Where N is the number of batches configured for the optimizer; Loss i This is the single-sample loss for calculation; 3) Calculate the time window MSE loss to obtain the time window [t] for model prediction. 始 , t 终 ]; Real-time window [t1, t2]; Label y 真 ; First, the samples within the batch are screened, retaining those labeled y. 真 For samples with a value of 1, the number of recorded samples is M; Next, calculate the single-sample MSE loss: ; where t 始 , t 终 t1 and t2 are the time windows for model prediction and the actual time windows for label annotation. Then, the mean of the M filtered samples is calculated to obtain the time window MSE loss; Where M is the number of samples selected; Loss k To calculate the single-sample MSE loss, if M is 0, then Loss window =0; 4) Calculate the total loss Loss = α ⋅ Loss emer + β ⋅ Loss window ; 5) Minimize the total loss using the optimizer and update the model weights; Configure the Adam optimizer parameters with an initial learning rate of 0.001 and a decay rate of 1e. -5 The batch size is 32, the number of iterations is 50, and the iteration optimization is performed. If the validation set loss does not decrease for 5 consecutive rounds, training is stopped.

5. The anti-interference navigation method for low-altitude logistics UAVs in mountainous areas based on the fusion of BeiDou and visual SLAM as described in claim 1, characterized in that: Step S4 includes: S41: Configure prediction parameters; A linear extrapolation prediction model is constructed for the BeiDou CN0 value; Configuration parameters: State vector: X k = ,in Let be the BeiDou CN0 value at time k; be the rate of change of the BeiDou CN0 value; State transition matrix: F= ;in The time interval for collecting BeiDou CN0 values, with a sampling frequency of 50Hz. It takes 20ms; Process noise covariance: Q= ;in σ q It is 0.

01. It takes 20ms; Observation matrix: H= ; The observation noise variance R = 0.25; To improve the visual SLAM feature point matching rate, a prediction model for the visual SLAM feature point matching rate is constructed using the same method, and the state vector is adjusted: X k = ,in Let k be the visual SLAM feature point matching rate at time k; Adjust the feature point matching rate for visual SLAM. It is 33.3ms. σ q The variance of the observation noise is 0.005, and the variance of the observation noise is R=0.0004; S42: Disconnection detection, to determine whether linear extrapolation prediction is triggered; If no new data is received for three consecutive BeiDou CN0 value sampling cycles or visual SLAM feature point matching rate sampling cycles, it is determined to be a data interruption, and linear extrapolation prediction is triggered; otherwise, linear extrapolation prediction is not triggered. S43: Perform linear extrapolation prediction calculations to obtain predicted values.

6. The anti-interference navigation method for low-altitude logistics UAVs in mountainous areas based on the fusion of BeiDou and visual SLAM as described in claim 5, characterized in that: S43 includes: S431: After triggering linear extrapolation prediction, read the last 3 sets of navigation data before the disconnection from the cached navigation data as the initial input of the prediction model; S432: Calculate the initial rate of change: Regarding the BeiDou CN0 value, based on the three sets of navigation data before the disconnection, it is set as C. k-2 C k-1 C k The initial rate of change was determined by linear fitting: ; S433: State prediction, generates CN0 prediction value within the next 50ms; According to the prediction formula: Obtain the predicted value of BeiDou CN0 within a future time step, where is the prediction time step, which is preset to 10ms; according to Update the state vector at time k+1, where F is the state transition matrix. The state vector at time k; obtained from the state vector at time k+1. And the calculated BeiDou CN0 prediction value Based on the prediction formula, we continue to obtain the predicted value of BeiDou CN0 within the future time step, and so on, to obtain the prediction sequence within 50ms. S434: Using the same linear extrapolation prediction logic, obtain the predicted value of the visual SLAM feature point matching rate within the next 50ms.

7. The anti-interference navigation method for low-altitude logistics UAVs in mountainous areas based on the fusion of BeiDou and visual SLAM as described in claim 1, characterized in that: Obstacle distance, relative velocity, and attitude adjustment amplitude are collected synchronously at a frequency of 1 kHz; the BeiDou CN0 value sampling frequency is 50 Hz; the visual SLAM feature point matching rate is the proportion of successful inter-frame matching output by the SLAM algorithm, with a sampling frequency of 30 Hz.

8. The anti-interference navigation method for low-altitude logistics UAVs in mountainous areas based on the fusion of BeiDou and visual SLAM as described in claim 1, characterized in that: In step S22, the trained LSTM model is quantized into TensorRT INT8 format.

Images