A flight data-based delivery unmanned aerial vehicle anomaly detection and processing method

Abstract

The application discloses a kind of based on flight data's distribution unmanned plane abnormal detection and processing method and system, detect unmanned plane dispatch package abnormal situation, improve delivery rate and accuracy.It is technical solution for: to flight data from multiple angles parameter index selection;To the parameter standardization processing of unmanned plane obtained;Unmanned plane data is input into adaptive bidirectional GRU model network, model carries out forward propagation and adaptive back propagation to input data, and respectively extracts the eigenvalue of forward and reverse state information by attention mechanism, then the eigenvalue of forward and reverse state information is combined, and the final state information is output;State information is input into multi-head attention layer to extract features, and obtain attention information;Determine abnormal data points by fuzzy isolated forest algorithm based on feature correlation degree;Abnormal type is analyzed based on abnormal data points and corresponding processing method is carried out.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and more specifically to a method for detecting and handling anomalies in delivery UAVs based on flight data. Background Technology

[0002] The diverse applications of drones have brought revolutionary changes to the logistics field, offering advantages such as speed, flexibility, and cost-effectiveness, making the logistics process more efficient, reliable, and convenient. First, drones enable rapid delivery. In traditional logistics, traffic congestion and road restrictions often lead to extended transportation times. Drones, however, can fly in straight lines, avoiding ground traffic conditions, and deliver packages, documents, or urgently needed supplies directly to their destinations, providing rapid delivery services and significantly shortening transportation time. Second, drones can fly in diverse geographical environments and complex terrains, adapting to the needs of different scenarios. They can reach remote areas that are difficult to access by traditional transportation methods or densely populated urban areas that are hard to penetrate, opening up new possibilities for logistics operations. Furthermore, the operating costs of drones are relatively low. Compared to traditional transportation methods, drones do not require significant human resources and infrastructure investment. They have autonomous flight capabilities, enabling them to complete tasks automatically, reducing labor costs and expenses related to fuel and maintenance.

[0003] During the use of drones, we need to monitor their environmental information and flight parameters to ensure their safety and reliability. Common drone anomaly detection methods involve collecting environmental information and flight parameters using sensor technology, and then using machine learning and data analysis techniques to model and analyze the collected data. By training the model, normal and abnormal behavior patterns can be identified, thereby detecting whether the drone is malfunctioning.

[0004] However, current common methods for detecting anomalies in drones are still insufficient in terms of efficiency and accuracy. Summary of the Invention

[0005] The following provides a brief overview of one or more aspects to offer a basic understanding of them. This overview is not an exhaustive summary of all conceived aspects, nor is it intended to identify key or decisive elements of all aspects, nor to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form to prepare for the more detailed descriptions that follow.

[0006] The purpose of this invention is to solve the above problems and provide a method for anomaly detection and handling of delivery drones based on flight data, so as to detect abnormal situations that drones may encounter when delivering packages and improve the delivery rate and accuracy of packages.

[0007] The technical solution of this invention is as follows: This invention discloses a method for anomaly detection and processing of delivery drones based on flight data, the method comprising:

[0008] Step S1: Select parameter indicators from multiple perspectives for the UAV's flight data as subsequent input;

[0009] Step S2: The parameters of the UAV obtained in Step S1 are standardized.

[0010] Step S3: Input the standardized UAV data into the adaptive bidirectional GRU model network. The bidirectional GRU model performs forward propagation and adaptive backward propagation on the UAV input data, and extracts the feature values ​​of the forward and backward state information through the attention mechanism. Then, the feature values ​​of the extracted forward and backward state information are combined to output the final state information.

[0011] Step S4: Input the state information obtained in step S3 into the multi-head attention layer for feature extraction to obtain attention information;

[0012] Step S5: Identify outlier data points using a fuzzy isolated forest algorithm based on feature correlation.

[0013] Step S6: Analyze the anomaly types based on the abnormal data points, and the UAV will take corresponding processing methods according to different anomaly types.

[0014] According to an embodiment of the method for anomaly detection and processing of delivery drones based on flight data of the present invention, in step S1, the flight data of the drone includes: positioning data, moving speed, and operation data, wherein the positioning data includes longitude, latitude, and altitude, the moving speed includes x, y, and z axes, and pitch angular velocity, and the operation data includes battery power and weight;

[0015] The parameters of a drone include: climb difference coefficient, turn difference coefficient, speed difference coefficient, battery power difference coefficient, and weight difference coefficient.

[0016] According to an embodiment of the method for anomaly detection and handling of delivery drones based on flight data according to the present invention, the standardization process in step S2 further includes:

[0017] Step S2-1: Standardize the UAV data using the z-score method;

[0018] Step S2-2: Calculate an unbiased estimate for the standardized UAV data to obtain the autocorrelation coefficient;

[0019] Step S2-3: Obtain multiple time points according to the sorting of autocorrelation coefficients, and obtain the reconstructed phase space based on these multiple time points.

[0020] According to an embodiment of the method for anomaly detection and handling of delivery drones based on flight data according to the present invention, step S3 further includes:

[0021] Step S3-1: Input the preprocessed data into the bidirectional GRU model network in the form of inputting data at two time points each time;

[0022] Step S3-2: The input data is propagated forward in the forward GRU network to obtain forward state information;

[0023] Step S3-3: The input data is backpropagated in the reverse GRU network to obtain the reverse state information;

[0024] Step S3-4: Extract features from the positive and negative state information respectively using an attention mechanism;

[0025] Steps 3-5: Combine the features extracted from the positive and negative state information through the attention mechanism to obtain the final output.

[0026] According to an embodiment of the anomaly detection and processing method for delivery drones based on flight data of the present invention, in step S3, an adaptive mechanism adjusts the input during backpropagation to improve the accuracy and speed of overall feature extraction, including:

[0027] In the early stages of training, backpropagation is used to improve the accuracy of forward propagation; in the later stages of training, an adaptive mechanism is introduced to accelerate the training process, as shown in the following formula:

[0028]

[0029] Where L and L′ represent the information lengths during forward and backward propagation, respectively, and t and T represent the current training round number and the total training round number, respectively.

[0030] According to an embodiment of the method for anomaly detection and handling of delivery drones based on flight data according to the present invention, step S4 further includes:

[0031] Step S4-1: Each attention head undergoes dimensionality transformation based on the final output from the bidirectional GRU network to generate corresponding Q, K, and V, where Q, K, and V are obtained by multiplying the same data by three weight matrices.

[0032] Step 4-2: After generating Q, K, and V, calculate the value of Multi-Attention;

[0033] Step 4-3: Process multiple attention heads in parallel, and then pass them through a neural network layer to condense the information of multiple attention heads so that the output word vectors have the same length.

[0034] Step 4-4: The output generated by the multi-head attention mechanism layer is added to the output from the bidirectional GRU network. At this point, feature extraction is complete, including state information and attention information.

[0035] According to an embodiment of the method for anomaly detection and processing of delivery drones based on flight data of the present invention, step S5 is to construct a binary tree using a four-part search method, including:

[0036] Step S5-1: Select a portion of the total data and calculate the degree of feature correlation for each portion;

[0037] Step S5-2: Sort the selected data according to the degree of feature association to obtain a sequence in which the feature association process is enhanced sequentially;

[0038] Step S5-3: Construct a binary tree using the four-part method, where a discriminant coefficient is used as the criterion for the binary tree to stop growing;

[0039] Step S5-4: Calculate the average path length for each isolated tree;

[0040] Step S5-5: Determine the set of evaluation factors;

[0041] Step S5-6: Determine the set of comments for the evaluation factors;

[0042] Step S5-7: Conduct single-factor evaluation: Use the expert estimation method to score the evaluation objects, finally count the scoring results, and form a fuzzy relation matrix to represent the degree of membership of the evaluation factors in the evaluation factor set to the various possible evaluation results in the comment set;

[0043] Step S5-8: Determine the fuzzy membership set of the evaluation factors;

[0044] Step S5-9: Calculate the multi-factor fuzzy evaluation vector using fuzzy sets and fuzzy relation matrices;

[0045] Step S5-10: Fuzzy comprehensive evaluation result analysis: The relative evaluation result is obtained by summing the corresponding components and their corresponding ranks in the multi-factor fuzzy evaluation vector.

[0046] According to an embodiment of the method for anomaly detection and handling of delivery drones based on flight data according to the present invention, step S6 includes the following steps:

[0047] Step S6-1: Determine if the drone is experiencing an abnormal climb and provide a solution;

[0048] Step S6-2: Determine if the drone is experiencing a turning anomaly and provide a solution;

[0049] Step S6-3: Determine if the drone is experiencing abnormal speed and provide a solution;

[0050] Step S6-4: Determine if the drone has an abnormal battery level and provide a solution;

[0051] Step S6-5: Determine if the drone has an abnormal weight and provide a solution.

[0052] Compared with the prior art, the present invention has the following beneficial effects: The present invention makes important innovations in the following three aspects.

[0053] 1. This invention selects the parameters of the drone from multiple perspectives, which can better determine the actual situation of the drone;

[0054] 2. This invention achieves better prediction results by fusing bidirectional GRU networks and multi-head attention mechanisms;

[0055] 3. The present invention employs fuzzy rules in the isolated forest algorithm, which can better handle uncertainties during the flight of UAVs. Attached Figure Description

[0056] The above-described features and advantages of the present invention will be better understood after reading the following detailed description of embodiments of the present disclosure in conjunction with the accompanying drawings. In the drawings, components are not necessarily drawn to scale, and components having similar related properties or features may have the same or similar reference numerals.

[0057] Figure 1A and 1B A flowchart of an embodiment of the delivery drone anomaly detection and handling method based on flight data of the present invention is shown.

[0058] Figure 2 It shows Figure 1A and 1B A schematic diagram of a bidirectional GRU network based on an attention mechanism in the method embodiment shown.

[0059] Figure 3 It shows Figure 1A and 1B The flowchart of the fuzzy isolated forest algorithm in the method embodiment shown. Detailed Implementation

[0060] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be noted that the aspects described below with reference to the accompanying drawings and specific embodiments are merely exemplary and should not be construed as limiting the scope of protection of the present invention in any way.

[0061] Figure 1A and 1BThe flowchart illustrates an embodiment of the anomaly detection and handling method for delivery drones based on flight data according to the present invention. Please refer to... Figure 1A and 1B The implementation steps of the method in this embodiment are described in detail below.

[0062] Step S1: Select parameter indicators from multiple perspectives for the UAV's flight data as subsequent input.

[0063] In step S1, the UAV's flight data includes: positioning data (longitude, latitude, altitude), movement speed (x, y, z axes, pitch angular velocity), and operational data (battery power, weight).

[0064] The parameters of a drone include: climb difference coefficient δ, turn difference coefficient E(ΔR), speed difference coefficient E(Δv), battery difference coefficient E(Δe), and weight difference coefficient E(Δw).

[0065] (1) Climb difference coefficient δ is the degree of difference between the actual angle and the predicted angle of the UAV during climb or descent. The calculation formula is as follows:

[0066] δ=|cos(E(θ t ))-cosθ AB |

[0067] Where, θ AB θ represents the climb angle of the drone between points A and B. t E(θ) represents the climb angle of the UAV at time t. t () represents the average climb angle of the UAV at time t. The calculation formulas are as follows:

[0068]

[0069]

[0070] Here, (x,y) represents the coordinates of the UAV on the x and y axes.

[0071] (2) The turning difference coefficient E(ΔR) represents the degree of fluctuation in the turning radius deviation of the UAV during the turning process. The calculation formula is as follows:

[0072]

[0073] Where N represents the number of data points during the drone's movement from point A to point B, and the formula represents the turning deviation ΔR during the entire process after removing the start and end points. I The formula for calculating the turning radius deviation is as follows:

[0074] ΔR I =|Rsi -R AB |

[0075] Among them, R si This indicates the drone's real-time turning radius, R. AB The turning radius of the drone on route AB is calculated using the following formula:

[0076]

[0077] Where, ax 2 +ay 2 +bx+by+c=0 is the equation of the circle formed by the arc traversed by the drone at times t, t+1, and t+2.

[0078] (3) The speed difference coefficient E(Δv) represents the degree of difference in speed change of the UAV during flight. The calculation formula is as follows:

[0079]

[0080] Where, Δv i The deviation in flight speed is represented by the following formula:

[0081] Δv i =|v si -v AB |

[0082] Among them, v si This indicates the drone's real-time flight speed, v AB The formula for calculating the drone's flight speed along route AB is as follows:

[0083]

[0084] Where (x,y,z) represents the coordinates of the UAV on the x, y, and z axes.

[0085] (4) The power difference coefficient E(Δe) represents the degree of difference in power change before and after unmanned operation. The calculation formula is as follows:

[0086]

[0087] Where, Δe i The formula for calculating the deviation in the drone's battery level is as follows:

[0088]

[0089] Among them, e si and e AB These represent the real-time battery level change of the drone and the battery level change before and after AB, respectively.

[0090] (5) Weight difference coefficient E(Δw), which represents the degree of difference in weight change before and after unmanned operation, is calculated as follows:

[0091]

[0092] Where, Δw i The formula for calculating the deviation in the weight of the drone is as follows:

[0093]

[0094] Among them, w si and w AB These represent the real-time weight change of the drone and the weight change before and after AB, respectively.

[0095] Step S2: Standardize the parameters of the UAV obtained in Step S1.

[0096] In step S2, the specific steps of the standardization process include:

[0097] Step S2-1: Use the z-score method to process the UAV data X = (x1, x2, ..., x...). n The standardization is performed using the following formula:

[0098]

[0099] in, Let represent the arithmetic mean of the parameters in the drone data, and s represent the standard deviation of the parameters in the drone data. The calculation formulas are as follows:

[0100]

[0101] Step 2-2: Process the standardized UAV data Y = (y1, y2, ..., y n The unbiased estimate R(m) is calculated using the following formula:

[0102]

[0103] In the above formula, n represents the number of data points, and m is a value in the range [1, n].

[0104] Taking m=3 as an example,

[0105] By performing nonlinear normalization on R(m), the autocorrelation coefficient r(m) in the range of [-1, 1] is obtained, as shown in the following formula:

[0106]

[0107] Steps 2-3: Sort the data according to the autocorrelation coefficient r(m) from largest to smallest to obtain the k time points with the largest autocorrelation coefficients, and use these to determine m1, m2, ..., m k The values ​​of are used to reconstruct the phase space as follows:

[0108]

[0109] Among them, X c X(t) is the input sequence of the bidirectional GRU network, and X(t) is the corresponding true value.

[0110] Step S3: Input the standardized UAV data into the adaptive bidirectional GRU (Gated Recurrent Unit) model network. The bidirectional GRU model performs forward propagation and adaptive backward propagation on the UAV input data, and extracts the feature values ​​of the forward and backward state information through the attention mechanism. Then, the extracted feature values ​​of the forward and backward state information are combined to output the final state information.

[0111] like Figure 2 As shown, the further processing of step S3 includes the following steps.

[0112] Step 3-1: Divide the preprocessed data according to X i ′=[X c (2i)+X c The input layer is in the form of (2i+1)] (i.e., inputting data at two time points each time), which reduces the computational scale.

[0113] Step 3-2: Input X′ into the forward GRU network for forward propagation. The calculation process is shown in steps a to d below:

[0114] Step a. Input the new information x t Memory of the previous moment h t-1 Combining these factors, the calculation formula is as follows:

[0115] r t =σ(W r ·[h t-1 ,x t ])

[0116] Among them, W r Indicates resetting the connection weight of the gate, h t-1 x represents the memory state information from the previous moment. t For the new input information at the current moment, r t This represents the memory state information after the input information enters the reset gate in the forward GRU network. σ represents the sigmoid function, which compresses the data to the range [0, 1]. Its general form is:

[0117] Step b. Record the memory of the previous moment H t-1 The calculation formula for saving the current state information is as follows:

[0118] z t =σ(W z ·[h t-1 ,x t ])

[0119] Among them, W z This indicates updating the connection weights of the gate, h. t-1 x represents the memory state information from the previous moment. t For the new input information at the current moment, z t The update gate controls the extent to which information from the previous state is incorporated into the current state, i.e., h. t-1 The information that is memorized or retained is a weight, as shown in the formula in step d below.

[0120] Step c. Calculate the state information of the forward candidate hidden layer The calculation formula is as follows:

[0121]

[0122] Where W represents the weight of the forward candidate hidden layer.

[0123] Step d. Calculate the positive state information. The calculation formula is as follows:

[0124]

[0125] Among them, z t This represents the weight of information retained by the forward GRU, with a value ranging from [0,1]. A value closer to 1 indicates more past data has been "remembered," while a value closer to 0 indicates more past data has been "forgotten." t-1 This represents the state information propagated from the previous moment; This represents the state information of the forward GRU candidate hidden layer.

[0126] Step 3-3: Input X′ into the reverse GRU network for backpropagation. The calculation process is shown in steps a to d below:

[0127] Step a. Input the new information x t Memory h′ from the previous moment t-1 Combining these factors, the calculation formula is as follows:

[0128] r t ′=σ(W r ′·[h′t-1 ,x t ])

[0129] Among them, W r h′ represents resetting the connection weight of the gate. t-1 For the memory and state information of the previous moment, x t For the new input information at the current moment, σ represents the sigmoid function, which compresses the data to the range [0, 1]. Its general form is:

[0130] Step b. Record the memory h′ from the previous moment. t-1 The calculation formula for saving the current state information is as follows:

[0131] z t ′=σ(W z ′·[h′ t-1 ,x t ])

[0132] Among them, W z h′ represents the re-updated connection weights of the gate. t-1 x represents the memory state information from the previous moment. t For the new input information at the current moment, z t This represents the memory state information after the input information enters the update gate in the reverse GRU network.

[0133] Step c. Calculate the state information of the reverse candidate hidden layer The calculation formula is as follows:

[0134]

[0135] Where W′ represents the weights of the reverse candidate hidden layer, This represents the state information of the candidate hidden layer in the reverse GRU.

[0136] Step d. Calculate the reverse state information. The calculation formula is as follows:

[0137]

[0138] Among them, z t This represents the information remembered by the reverse GRU, with values ​​ranging from [0,1]. Values ​​closer to 1 indicate more past data "remembered," while values ​​closer to 0 indicate more past data "forgotten." t-1 This represents the state information propagated from the previous moment; This represents the state information of the candidate hidden layer in the reverse GRU.

[0139] Steps 3-4: Apply attention mechanisms to the positive and negative state information h respectively.t and h′ t Feature extraction is performed, and the calculation process is shown in steps a to c below:

[0140] Step a. Calculate the forward and reverse state information h t and h′ t The internal similarity is calculated using the following formula:

[0141] s i =F(h) t ,h ti );s i ′=F(h′ t ,ht ti )

[0142]

[0143] Among them, s i and s i ′ represent the forward and reverse similarity scores, respectively; F(Q,K) i () represents the attention scoring function, and the calculation formula is as follows:

[0144]

[0145] Where d represents the length of the input data, and Q represents s. i =F(h) t ,h ti ) and s i ′=F(h′ t ,h′ ti h in ) t and h′ t , representing overall state information; K i Refers to s i =F(h) t ,h ti ) and s i ′=F(h′ t ,h′ ti h in ) ti and h′ ti , representing the i-th state information.

[0146] Step b. Perform numerical transformation on the forward and reverse similarity scores respectively to obtain the weights of each element within the forward and reverse state information. The calculation formula is as follows:

[0147]

[0148] Where, α i and α i ′ represents the weight of each element within the forward and reverse state information, respectively.

[0149] Step c. Based on the weight coefficients obtained in step b, perform a weighted summation of the forward and reverse final state information to obtain the features extracted from the forward and reverse state information through the attention mechanism. The calculation formula is as follows:

[0150]

[0151] Among them, A t and A′ t These represent the features extracted from the positive and negative state information through the attention mechanism, respectively.

[0152] Steps 3-5: Extract feature A from the positive and negative state information using an attention mechanism. t and A′ t Combining these, we obtain the final output, calculated using the following formula:

[0153] H t =(1-W) a -W a ′)H t-1 +w a A t +W a 'A' t

[0154] Among them, W a and W a ′ represents the weights of the forward and reverse state information, respectively.

[0155] In step S3, the adaptive mechanism can effectively adjust the input during backpropagation to improve the overall accuracy and speed of feature extraction. This mainly includes the following:

[0156] In the early stages of training, the accuracy of forward propagation is relatively low. Backpropagation with a larger amount of information can effectively improve its accuracy. In the later stages of training, the overall feature extraction accuracy tends to stabilize, so there is no longer a need for complex backpropagation with large amounts of information. Introducing an adaptive mechanism can effectively speed up the training process, as shown in the following formula:

[0157]

[0158] Where L and L′ represent the information lengths during forward and backward propagation, respectively, and t and T represent the current training round number and the total training round number, respectively.

[0159] Step S4: Input the state information into the multi-head attention layer for feature extraction to obtain attention information.

[0160] Step S4 further includes the following processing.

[0161] Step S4-1: Each attention head undergoes dimensionality transformation based on the final output from the bidirectional GRU network to generate corresponding Q, K, and V. Q, K, and V are obtained by multiplying the same data by three weight matrices.

[0162] Step 4-2: After generating Q, K, and V, calculate the Multi-Attention value:

[0163]

[0164] d k Indicates the length of the input data.

[0165] Step 4-3: Process multiple attention heads in parallel, and then pass them through a neural network layer to condense the information from multiple attention heads, thereby making the output word vectors of the same length.

[0166] Step 4-4: The output generated by the multi-head attention mechanism layer is added to the output from the bidirectional GRU network. At this point, feature extraction is complete, including state information and attention information.

[0167] Step S5: Identify outlier data points using a fuzzy isolated forest algorithm based on the degree of feature correlation.

[0168] like Figure 3 As shown, step S5 is to construct a binary tree using the four-part method, which mainly includes the following steps.

[0169] Step S5-1: Select m data points X = {x1, x2, ..., xm} from the total data. n} and calculate the feature correlation degree p(i) for each, as shown in the following formula:

[0170]

[0171] Where, x min x mean and x max These represent the minimum, average, and maximum values ​​in the fuzzy set X, respectively.

[0172] Step S5-2: Sort the m data according to the degree of feature association p(i) to obtain a sequence in which the feature association process is enhanced in turn.

[0173] Step S5-3: Construct a binary tree using the quartic method, and assign the three quartiles as (X... l X m X r In addition, a discriminant coefficient σ is designed as the criterion for the binary tree to stop growing, and the formula is as follows:

[0174]

[0175] Where L≠1 and R≠1. If σ∈[0.8, 1.28], then the number of elements in the sets on the left and right sides of the cut point P is almost equal, and it can be determined that this growth node is a poor node, so the binary tree stops growing.

[0176] Step S5-4: Calculate the average path length c(n) for each isolated tree, using the following formula:

[0177]

[0178] Where H(n-1) represents the harmonic number, which can be estimated as ln(i) + ζ (ζ represents Euler's constant) and the duration E(h(x)) of the path length h(x). Finally, the outlier score s(x,n) of the sample is obtained from E(h(x)), as shown in the following formula:

[0179]

[0180] Step S5-5: Determine the set of evaluation factors U, u i Let m represent the number of evaluation factors.

[0181] U = {u1, u2, ..., u} m}={δ, E(ΔR), E(Δv), E(Δe), E(Δw)}

[0182] Where δ represents the climbing difference coefficient, E(ΔR) represents the turning difference coefficient, E(Δv) represents the speed difference coefficient, E(Δe) represents the battery power difference coefficient, and E(Δw) represents the weight difference coefficient.

[0183] Steps S5-6: Determine the set of comments V, v for the evaluation factors. i Let n represent the possible evaluation results, and n represent the number of evaluation results.

[0184] V = {v1, v2, ..., v} n} = {extremely abnormal, abnormal, somewhat abnormal, normal}

[0185] Step S5-7: Conduct single-factor evaluation. Using expert estimation, score the evaluation objects, and finally statistically analyze the scores to form a fuzzy relation matrix R representing the evaluation factors u in the set of evaluation factors U. i For each possible evaluation result v in the set of comments V j The degree of subordination.

[0186] Step S5-8: Determine the fuzzy membership set S, s′ of the evaluation factors. i This represents the fuzzy membership degree, where m represents the number of members.

[0187] S=[s′1,s′2,…,s′ m ]

[0188]

[0189] Among them, s i This represents the abnormal scores of each evaluation factor calculated in step S5.

[0190] Step S5-9: Multi-factor fuzzy evaluation B, which can be obtained by calculating the fuzzy set S and the fuzzy relation matrix R:

[0191] B = S × R = {b1, b2, ..., b} n}

[0192] Among them, b i This represents the fuzzy comprehensive evaluation result of each object to be evaluated.

[0193] Step S5-10: Fuzzy Comprehensive Evaluation Result Analysis. The relative evaluation result is obtained by summing the corresponding components and their corresponding ranks in vector B. The calculation formula is as follows:

[0194]

[0195] Where k is an undetermined coefficient (k = 1, 2), the purpose of which is to prevent b j A larger value will affect the result; j represents the rank of each level corresponding to the component in vector B.

[0196] Step S6: Analyze the anomaly types based on the abnormal data points, and the UAV will take corresponding processing methods according to different anomaly types.

[0197] Step S6 includes the following specific processing steps.

[0198] Step S6-1: Determine if the drone is experiencing an abnormal climb and provide a solution;

[0199]

[0200] When δ>0.5, the drone's climb difference coefficient is too large, indicating abnormal climb. This can result in the drone being too high or too low. In this case, it is necessary to adjust the drone's climb angle to return to the normal trajectory.

[0201] Step S6-2: Determine if the drone is experiencing a turning anomaly and provide a solution;

[0202]

[0203] When E(ΔR)>0.5, the drone will deviate from the original route and have abnormal turning, such as spinning in place or turning too large. At this time, it is necessary to adjust the turning angle of the drone to bring it back to the correct route and continue to operate.

[0204] Step S6-3: Determine if the drone is experiencing abnormal speed and provide a solution;

[0205]

[0206] When E(Δv)>0.5, the drone's speed will be too fast or too slow, resulting in the inability to avoid unknown obstacles or reach the designated location on time. In this case, it is necessary to adjust the drone's speed to return it to the correct flight path and continue operating.

[0207] Step S6-4: Determine if the drone has an abnormal battery level and provide a solution;

[0208]

[0209] When E(Δe)>0.5, the drone's battery will be rapidly depleted, resulting in the inability to reach the designated location or even falling into the air. In this case, it is necessary to command the drone to land immediately and send a recovery signal.

[0210] Step S6-5: Determine if the drone has an abnormal weight and provide a solution;

[0211]

[0212] When E(Δw)>0.5, there is a possibility that the package carried by the drone may be lost. In this case, it is necessary to search the area near the drone to find the lost package before completing the subsequent delivery task.

[0213] Although the methods described above are illustrated and depicted as a series of actions for the sake of simplicity, it should be understood and appreciated that these methods are not limited by the order of the actions, as some actions may occur in a different order and / or concurrently with other actions from the illustrations and descriptions herein or not illustrated and described herein but which may be understood by those skilled in the art, according to one or more embodiments.

[0214] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in a generalized manner in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described functionality in different ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the invention.

[0215] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein can be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0216] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0217] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or through a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0218] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for anomaly detection and handling of delivery drones based on flight data, characterized in that, The methods include: Step S1: Select parameter indicators from multiple perspectives for the UAV's flight data as subsequent input; Step S2: The parameters of the UAV obtained in Step S1 are standardized. Step S3: Input the standardized UAV data into the adaptive bidirectional GRU model network. The bidirectional GRU model performs forward propagation and adaptive backward propagation on the UAV input data, and extracts the feature values ​​of the forward and backward state information through the attention mechanism. Then, the feature values ​​of the extracted forward and backward state information are combined to output the final state information. Step S4: Input the state information obtained in step S3 into the multi-head attention layer for feature extraction to obtain attention information; Step S5: Identify outlier data points using a fuzzy isolated forest algorithm based on feature correlation. Step S6: Analyze the anomaly types based on the abnormal data points, and the UAV will take corresponding processing methods according to different anomaly types; Step S3 further includes: Step S3-1: Input the preprocessed data into the bidirectional GRU model network in the form of inputting data at two time points each time; Step S3-2: The input data is propagated forward in the forward GRU network to obtain forward state information; Step S3-3: The input data is backpropagated in the reverse GRU network to obtain the reverse state information; Step S3-4: Extract features from the positive and negative state information respectively using an attention mechanism; Steps 3-5: Combine the features extracted from the positive and negative state information through the attention mechanism to obtain the final output; In step S3, the adaptive mechanism adjusts the input during backpropagation to improve the accuracy and speed of overall feature extraction, including: In the early stages of training, backpropagation is used to improve the accuracy of forward propagation; in the later stages of training, an adaptive mechanism is introduced to accelerate the training process, as shown in the following formula: in, and These represent the information lengths during forward and backward propagation, respectively. and These represent the current training round number and the total number of training rounds, respectively. Step S3-2 further includes the following steps a to d: Step a: Input new information Memory of the previous moment Combining these factors, the calculation formula is as follows: in, This indicates resetting the connection weight of the door. This refers to the memory state information from the previous moment. For the new input information at the current moment, This represents the memory state information after the input information enters the reset gate in the forward GRU network, and σ represents the sigmoid function; Step b: Record the memory of the previous moment. The calculation formula for saving the current state information is as follows: in, This indicates that the connection weights of the updated gates are being updated. The update gate controls the extent to which the state information from the previous moment is incorporated into the current state; Step c: Calculate the state information of the forward candidate hidden layer The calculation formula is as follows: in, Indicates the weights of the positive candidate hidden layers; Step d: Calculate the positive state information The calculation formula is as follows: 。 2. The method for anomaly detection and processing of delivery drones based on flight data according to claim 1, characterized in that, In step S1, the drone's flight data includes: positioning data, moving speed, and operational data. The positioning data includes longitude, latitude, and altitude; the moving speed includes x, y, and z axes, and pitch angular velocity; and the operational data includes battery power and weight. The parameters of a drone include: climb difference coefficient, turn difference coefficient, speed difference coefficient, battery power difference coefficient, and weight difference coefficient.

3. The method for anomaly detection and processing of delivery drones based on flight data according to claim 1, characterized in that, The standardization process in step S2 further includes: Step S2-1: Standardize the UAV data using the z-score method; Step S2-2: Calculate an unbiased estimate for the standardized UAV data to obtain the autocorrelation coefficient; Step S2-3: Obtain multiple time points according to the sorting of autocorrelation coefficients, and obtain the reconstructed phase space based on these multiple time points.

4. The method for anomaly detection and processing of delivery drones based on flight data according to claim 1, characterized in that, Step S4 further includes: Step S4-1: Each attention head undergoes dimensionality transformation based on the final output from the bidirectional GRU network to generate corresponding Q, K, and V, where Q, K, and V are obtained by multiplying the same data by three weight matrices. Step 4-2: After generating Q, K, and V, calculate the value of Multi-Attention; Step 4-3: Process multiple attention heads in parallel, and then pass them through a neural network layer to condense the information of multiple attention heads so that the output word vectors have the same length. Step 4-4: The output generated by the multi-head attention mechanism layer is added to the output from the bidirectional GRU network. At this point, feature extraction is complete, including state information and attention information.

5. The method for anomaly detection and processing of delivery drones based on flight data according to claim 1, characterized in that, Step S5 involves constructing a binary tree using the four-part search method, including: Step S5-1: Select a portion of the total data and calculate the degree of feature correlation for each portion; Step S5-2: Sort the selected data according to the degree of feature association to obtain a sequence in which the feature association process is enhanced sequentially; Step S5-3: Construct a binary tree using the four-part method, where a discriminant coefficient is used as the criterion for the binary tree to stop growing; Step S5-4: Calculate the average path length for each isolated tree; Step S5-5: Determine the set of evaluation factors; Step S5-6: Determine the set of comments for the evaluation factors; Step S5-7: Conduct single-factor evaluation: Use the expert estimation method to score the evaluation objects, finally count the scoring results, and form a fuzzy relation matrix to represent the degree of membership of the evaluation factors in the evaluation factor set to the various possible evaluation results in the comment set; Step S5-8: Determine the fuzzy membership set of the evaluation factors; Step S5-9: Calculate the multi-factor fuzzy evaluation vector using fuzzy sets and fuzzy relation matrices. ； Step S5-10: Fuzzy comprehensive evaluation result analysis: The relative evaluation result is obtained by summing the corresponding components and their corresponding ranks in the multi-factor fuzzy evaluation vector.

6. The method for anomaly detection and processing of delivery drones based on flight data according to claim 1, characterized in that, Step S6 includes the following steps: Step S6-1: Determine if the drone is experiencing an abnormal climb and provide a solution; Step S6-2: Determine if the drone is experiencing a turning anomaly and provide a solution; Step S6-3: Determine if the drone is experiencing abnormal speed and provide a solution; Step S6-4: Determine if the drone has an abnormal battery level and provide a solution; Step S6-5: Determine if the drone has an abnormal weight and provide a solution.

Images