Unmanned aerial vehicle performance intelligent analysis method and system based on cloud platform

Abstract

The invention discloses an unmanned aerial vehicle performance intelligent analysis method and system based on a cloud platform, and relates to the field of unmanned aerial vehicle fault analysis. Small-scale abrupt change features (such as instantaneous change caused by out-of-step of a single motor) and large-scale slow change features (such as global performance reduction caused by battery aging) in flight control log time sequence data are synchronously captured. And then, a gating fusion mechanism is creatively introduced, and dynamic weighting is performed on the extracted multi-scale features, so that the model can adaptively judge which scale feature is more decisive for the current fault classification according to the input data. In this way, the model can mine the deep time sequence mode with the identification degree from the similar macroscopic representation, faults with different roots but similar representations are effectively distinguished, and therefore the accuracy of fault diagnosis is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle fault analysis, and more specifically, to an unmanned aerial vehicle performance intelligent analysis method and system based on a cloud platform. BACKGROUND

[0002] With the wide application of unmanned aerial vehicles in aerial photography, logistics, inspection and other fields, its flight safety and performance reliability become crucial. Unmanned aerial vehicles generate a large amount of flight control log data during flight, which contains rich information about the state of the aircraft. In order to achieve efficient and centralized management and health assessment of large-scale unmanned aerial vehicle fleets, it is inevitable for the industry to develop an intelligent analysis scheme based on a cloud platform. Through the cloud platform, flight data from different unmanned aerial vehicles can be conveniently collected, stored and processed, providing a data basis for in-depth analysis and fault diagnosis.

[0003] However, existing unmanned aerial vehicle data analysis methods mostly rely on human experience or simple threshold judgments, and the level of intelligence is insufficient. Even some researches begin to try to apply intelligent algorithms such as machine learning, but they often focus on general anomaly detection, i.e. judging whether there is a fault or not, and cannot accurately distinguish what kind of fault it is. The root of this limitation lies in the fact that an unmanned aerial vehicle is a highly coupled complex system, and different root faults (such as single motor ESC abnormality or battery voltage drop) may trigger a chain reaction, eventually showing very similar macroscopic characteristics in the flight control log, such as increasing attitude error and abnormal current. Traditional models only focus on these macroscopic features, and it is difficult to distinguish the subtle differences behind them, resulting in confusion and low diagnostic accuracy.

[0004] Therefore, how to effectively capture and distinguish the subtle features with unique evolution patterns in time series caused by different root faults from highly similar and confused fault features has become a core technical bottleneck to improve the practicality of unmanned aerial vehicle intelligent diagnosis. SUMMARY

[0005] To solve the problems in the prior art, according to an aspect of the present application, an unmanned aerial vehicle performance intelligent analysis method based on a cloud platform is provided, which comprises: obtaining an original flight control log file from the cloud platform; performing data preprocessing and feature window generation on the original flight control log file to obtain a flight state feature window sequence; performing multi-scale feature extraction and channel refinement on the flight state feature window sequence to obtain flight state refined small-scale features, flight state refined medium-scale features and flight state refined large-scale features; The flight state refined small-scale features, the flight state refined medium-scale features and the flight state refined large-scale features are adaptively fused based on a gating mechanism to obtain a flight state fused diagnosis vector; Fault classification and diagnosis output are performed based on the flight state fused diagnosis vector to obtain a fault probability.

[0006] According to another aspect of the present application, a cloud platform-based unmanned aerial vehicle performance intelligent analysis system is provided, which comprises: An original flight control log acquisition module is configured to acquire an original flight control log file from a cloud platform; A flight control data preprocessing module is configured to perform data preprocessing and feature window generation on the original flight control log file to obtain a flight state feature window sequence; A flight state multi-scale feature extraction module is configured to perform multi-scale feature extraction and channel refinement on the flight state feature window sequence to obtain flight state refined small-scale features, flight state refined medium-scale features and flight state refined large-scale features; A gating adaptive feature fusion module is configured to adaptively fuse the flight state refined small-scale features, the flight state refined medium-scale features and the flight state refined large-scale features based on a gating mechanism to obtain a flight state fused diagnosis vector; A fault classification and diagnosis output module is configured to perform fault classification and diagnosis output based on the flight state fused diagnosis vector to obtain a fault probability.

[0007] Compared with the prior art, the cloud platform-based unmanned aerial vehicle performance intelligent analysis method and system provided by the present application solves the technical problem that different unmanned aerial vehicle faults cause similar macro representations and are easily confused. A multi-scale feature extraction and adaptive fusion diagnosis method is proposed. The method synchronously captures small-scale mutation features (such as instantaneous changes caused by single motor step loss) and large-scale slowly changing features (such as global performance decline caused by battery aging) in flight control log time series data through parallel multi-scale convolution branches. Subsequently, a gating fusion mechanism is creatively introduced to dynamically weight the extracted multi-scale features, so that the model can adaptively determine which scale of features is more decisive for the current fault classification. In this way, the model can excavate deep time series patterns with discriminability from similar macro representations, effectively distinguish faults with different roots but similar representations, and thus significantly improve the accuracy of fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0008] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings.

[0009] Figure 1A flow chart of the cloud platform-based unmanned aerial vehicle performance intelligent analysis method according to the embodiment of the present application.

[0010] Figure 2 A data flow schematic diagram of the cloud platform-based unmanned aerial vehicle performance intelligent analysis method according to the embodiment of the present application.

[0011] Figure 3 A flow chart of step S2 in the cloud platform-based unmanned aerial vehicle performance intelligent analysis method according to the embodiment of the present application.

[0012] Figure 4 A flow chart of step S4 in the cloud platform-based unmanned aerial vehicle performance intelligent analysis method according to the embodiment of the present application.

[0013] Figure 5 A block diagram of the cloud platform-based unmanned aerial vehicle performance intelligent analysis system according to the embodiment of the present application. DETAILED DESCRIPTION

[0014] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. It should be understood that the accompanying drawings and embodiments of the present disclosure are only for exemplary purposes, and are not intended to limit the protection scope of the present disclosure.

[0015] Based on the deficiencies in the above technical field, the present application proposes a cloud platform-based unmanned aerial vehicle performance intelligent analysis method. Figure 1 A flow chart of the cloud platform-based unmanned aerial vehicle performance intelligent analysis method according to the embodiment of the present application. Figure 2 A data flow schematic diagram of the cloud platform-based unmanned aerial vehicle performance intelligent analysis method according to the embodiment of the present application. As shown in Figure 1 and Figure 2 As shown in the above, the cloud platform-based unmanned aerial vehicle performance intelligent analysis method according to the embodiment of the present application comprises: S1, obtaining an original flight control log file from a cloud platform; S2, performing data preprocessing and feature window generation on the original flight control log file to obtain a flight state feature window sequence; S3, performing multi-scale feature extraction and channel refining on the flight state feature window sequence to obtain flight state refined small-scale features, flight state refined medium-scale features, and flight state refined large-scale features; S4, performing adaptive fusion based on a gating mechanism on the flight state refined small-scale features, the flight state refined medium-scale features, and the flight state refined large-scale features to obtain a flight state fusion diagnosis vector; and S5, performing fault classification and diagnosis output based on the flight state fusion diagnosis vector to obtain a fault probability.

[0016] In step S1, the original flight control log file is obtained from the cloud platform. It should be understood that as the scale of the unmanned aerial vehicle fleet continues to expand, the massive, multi-source flight control logs generated during its flight are stored centrally in the cloud, forming a huge data asset. These data are the fundamental basis for evaluating the performance of unmanned aerial vehicles and diagnosing faults. However, the data itself cannot directly reveal the performance status or fault mode and needs to be analyzed and processed in a series of intelligent ways. Therefore, in order to start the entire intelligent analysis and diagnosis process, the first and essential step is to accurately retrieve and obtain the original record of the specified unmanned aerial vehicle in a specific flight task from this centralized data center. This step is the data entry and basis for all subsequent data processing, feature extraction, and fault diagnosis links.

[0017] Exemplarily, a feasible implementation of step S1 of the present application is as follows: The original flight control log file is a binary data file that records various sensor data, controller status, instruction output, and environmental parameters of the unmanned aerial vehicle during the entire flight cycle. It is directly generated by the unmanned aerial vehicle flight control firmware without any analysis or conversion, and reproduces all details in the flight process with the highest fidelity. For example, a log file with the suffix.ulg generated by the PX4 open source flight control firmware records the original readings of the inertial measurement unit (IMU), the position information of the global positioning module (GPS), the voltage and current data of the battery management module (BMS), and the pulse width modulation (PWM) control signal values of each motor electronic speed controller (ESC), etc. with high-frequency timestamps. These log files are automatically uploaded and archived to the designated cloud platform object storage service through the on-board wireless communication module after the unmanned aerial vehicle completes the task, so as to facilitate centralized management and subsequent calling.

[0018] The execution process begins with the analysis server initiating a data request to the data service interface of the cloud platform. The request is not blind acquisition, but with explicit query parameters, such as the unique device identifier of the specified unmanned aerial vehicle, the flight number that needs to be analyzed, or a precise start and end time range. After receiving the request, the back-end service of the cloud platform will first perform access authorization verification to ensure the legality of the request. After verification, the data service will retrieve the matching log file in its metadata database according to the query parameters in the request. For example, a request aims to analyze the log of the P03 inspection task performed by the unmanned aerial vehicle numbered SN20251020A in the morning of a certain day. The cloud platform will locate to the specific file named SN20251020A-20251020-P03.ulg in the storage bucket.

[0019] After locating the target file, the cloud platform transmits the binary data stream of the original flight control log file to the analysis server initiating the request through a secure data transmission protocol such as HTTPS. After the analysis server receives the complete data stream, it temporarily stores it as an independent binary file in the local working directory, ready for the next step of processing.

[0020] In step S2, data preprocessing and feature window generation are performed on the original flight control log file to obtain a flight state feature window sequence. Accordingly, the original flight control log file obtained from the cloud platform is essentially a highly compressed binary data archive that encapsulates multi-source heterogeneous data streams from different sensors and control modules recorded at their own independent frequencies and timestamps. The data structure of this original format is complex and the time reference is not unified, and it cannot be directly read or processed by subsequent deep learning models. Without preprocessing, the fine timing correlation contained in the data cannot be utilized at all. Therefore, the data preprocessing and feature window generation performed by the present application can convert the original, chaotic log data into a structured, time-aligned, and fixed-length serialized sample format, i.e., a flight state feature window sequence. This provides a regular and high-quality input for the subsequent multi-scale feature extraction module, and is the fundamental prerequisite for ensuring that the model can effectively learn and distinguish different fault modes.

[0021] Figure 3 The flowchart of step S2 in the cloud platform-based unmanned aerial vehicle performance intelligent analysis method according to the embodiments of the present application is shown in FIG. 2. As shown in FIG. 2, the data preprocessing and feature window generation process includes the following steps: Figure 3As shown, one feasible manner of step S2 of the present application is to perform data preprocessing and feature window generation on the original flight control log file to obtain a flight state feature window sequence, including: S21, performing multi-source heterogeneous log analysis on the original flight control log file to obtain a flight state synchronization data table; S22, performing data cleaning and standardization on the flight state synchronization data table to obtain a flight state standardized data table; S23, performing time window segmentation and sequence construction on the flight state standardized data table to obtain the flight state feature window sequence. In particular, each flight state feature window in the flight state feature window sequence includes roll angle, pitch angle, target roll angle, target pitch angle, X-axis angular velocity, Y-axis angular velocity, Z-axis angular velocity, battery voltage, PWM control signal value sent by the flight control to the 1st motor ESC, PWM control signal value sent by the flight control to the 2nd motor ESC, PWM control signal value sent by the flight control to the 3rd motor ESC, and PWM control signal value sent by the flight control to the 4th motor ESC. That is, in the complex flight control mechanism of the unmanned aerial vehicle, the roll angle and the pitch angle are the core parameters for describing the current attitude of the machine body, and the target roll angle and the target pitch angle represent the ideal attitude expected by the flight control, and the difference between the two groups of parameters directly reflects the error of attitude control. The angular velocities of the X, Y and Z axes describe the speed of the attitude change of the machine body and are key dynamic indicators for measuring flight stability. The battery voltage, as a core state parameter of the power system, is directly related to the power output and endurance of the whole machine. The pulse width modulation (PWM) control signal values sent by the flight control to the four motor ESCs are the most direct instructions issued by the flight control to the executing mechanism to maintain or change the attitude, and reflect the load condition of each power unit. These twelve parameters constitute flight state features from multiple dimensions such as instructions, states, responses and energy, and together form a minimal and complete feature set that can completely describe the flight dynamics behavior of the unmanned aerial vehicle. By jointly analyzing this highly coupled group of parameters, the unique fingerprints caused by different faults in the flight control closed loop can be captured, thereby providing the most fundamental data basis for solving the technical problems mentioned in the background art, i.e., the difficulty in distinguishing specific fault sources due to similar macroscopic representations.

[0022] Exemplarily, the implementation of step S2 of the present application is as follows: first, in S21, a parsing library component specially designed for parsing this type of log file is called to perform deep decoding on the binary file. The decoding operation will traverse the data structure inside the file, identifying and extracting each independent data topic recorded one by one. Each topic corresponds to a specific aspect of data recording of the UAV, forming an independent time series data stream. For example, the parser will extract multiple topic data streams related to flight attitude, motor control and power state. For example, it will extract the aircraft attitude topic data stream from the log, which contains the time series data of roll angle and pitch angle; at the same time, it will extract the target attitude topic data stream, which contains the target roll angle and target pitch angle; it will also extract the angular velocity topic data stream, which contains the angular velocity of X, Y and Z axes; and the battery state topic data stream, which contains the battery voltage information; finally, it will extract the actuator output topic data stream, which records the PWM control signal value sent by the flight control to the four motor ESCs. At this time, although the extracted data streams have clear content, their respective timestamps are completely independent, for example, the recording frequency of the aircraft attitude data stream may be 250 Hz, while the recording frequency of the battery state may be only 50 Hz, and their time points cannot be directly corresponded. In order to solve this heterogeneity problem, a unified reference time grid needs to be constructed next. This process will first scan all the extracted data streams completely to determine the earliest start timestamp and the latest end timestamp of the entire flight task. Then, a target resampling frequency is set, which needs to take into account both computational efficiency and information fidelity, and is determined according to the variation characteristics of the physical quantities analyzed, for example, set to 100 Hz, which is sufficient to capture the dynamic response of the UAV attitude and motor. Based on the start and end time and the 100 Hz sampling rate, the program will generate a new, completely equidistant reference timestamp sequence, with each time point in the sequence separated by 0.01 seconds. This sequence will serve as a unified time scale for aligning all subsequent data. Next, enter the resampling and alignment section. This section will traverse each of the aforementioned extracted independent topic data streams. For each feature in each data stream (such as the battery voltage feature in the battery state data stream), a linear interpolation algorithm will be applied to map its value to the newly generated reference timestamp sequence. Specifically, for any time point in the reference timestamp sequence, the interpolation algorithm will find the two nearest original time points and their corresponding values in the original data stream of the feature, and then perform a weighted average according to the time distance to calculate an accurate interpolation of the feature at the reference time point.For example, if there is a time point of 15.255 seconds in the reference time series, and the raw battery voltage data stream records a value of 14.8 volts at 15.250 seconds and a value of 14.7 volts at 15.270 seconds, then the battery voltage at 15.255 seconds should be 14.775 volts by linear interpolation. This interpolation process is performed for all twelve features (roll angle, pitch angle, target roll angle, target pitch angle, three-axis angular velocity, battery voltage, four motor PWM control signal values) that need to be analyzed. Finally, all the interpolated and precisely aligned feature data to the same reference time stamp sequence are merged to construct a regular two-dimensional data structure, i.e., a flight state synchronization data table. The data table is similar in structure to an electronic spreadsheet, with each row representing a unique time point in the reference time stamp sequence and each column representing a specific flight state feature. For example, the first column is the time stamp, the second column is the roll angle, the third column is the pitch angle, and so on until the last column is the PWM control signal value of the fourth motor. Each cell in the table is filled with the interpolated result of the corresponding time and corresponding feature, and the output is a flight state synchronization data table with dimensions of T rows by F columns, where T is the total number of reference time stamps, such as 50000 consecutive time points, and F is the total number of selected features, which is 12 in this example. The data in this data table is completely synchronized.

[0023] Afterwards, the missing value processing step is firstly initiated in S22. The analysis program scans through each data point in the flight state synchronous data table column by column to identify the null values (represented as non-numbers, i.e. NaN) that failed to be assigned successfully in the previous interpolation process due to boundary conditions or missing original data segments. Once such missing values are detected, a forward filling strategy is employed to fill them in. The specific operation of this strategy is to use the last valid observation in the feature column closest to the missing position to fill in the current missing value. For example, in the battery voltage column, if the value at time point t = 30.5 seconds is 15.1 volts, and the value after t = 30.6 seconds is null, the program will automatically copy the value 15.1 volts to the position of t = 30.6 seconds. This processing method is based on a reasonable prior that under high-frequency sampling, the physical quantity at adjacent time points will not jump drastically, so the value at the previous time point is the best approximation of the current missing value. After completing the missing value filling, the data table enters the outlier processing step. This step is also carried out independently for each feature column. It uses the statistical-based 3-Sigma principle to identify and correct extreme outliers in the data stream caused by sensor transient interference or data transmission errors. For any feature column, for example, the X-axis angular velocity, the program will first calculate the arithmetic mean and standard deviation of all values in the column. In a specific example, the average value of the X-axis angular velocity is calculated to be 0.05 rad / s, and the standard deviation is 0.8 rad / s. According to the 3-Sigma principle, the normal value range of the feature is defined as the range of the average value plus or minus three times the standard deviation, i.e. [0.05-3*0.8, 0.05+3*0.8], which calculates to [-2.35, 2.45] rad / s. Subsequently, the program will traverse all data points in the column, and if a data point value, for example -2.8 rad / s, is found to be outside the calculated normal range, the data point is determined to be an outlier. For this outlier, a clipping process is taken, i.e. the nearest boundary value is used to replace it. In this example, -2.8 is less than the lower boundary -2.35, so the value of the data point will be corrected to -2.35. This can effectively suppress extreme noise interference on subsequent model training without deleting data points and maintaining the integrity of the time series. Finally, the data table after the above two steps of cleaning is subjected to data standardization processing to eliminate the dimensional influence of different features due to differences in physical units and numerical ranges, so that all features are on a unified and comparable scale. The Z-Score standardization method is used in this embodiment. This method calculates the updated mean and standard deviation for each feature column of the cleaned data table. Taking the X-axis angular velocity as an example, after outlier clipping, the mean and standard deviation may be adjusted to 0.04 rad / s and 0.75 rad / s.At this point, for any one raw value in the column, for example 1.54 rad / sec, the new value after normalization will be calculated by a specific formula: new value = (raw value - mean value) / standard deviation. Substituting the values, i.e. (1.54 - 0.04) / 0.75 = 1.5 / 0.75 = 2.0. This means that the original state of 1.54 rad / sec is represented in the normalized feature space as a level that is two standard deviations away from the mean value. This normalization process will be applied to each data point for all twelve feature columns. After the normalization conversion is completed for all data points, the final flight state normalized data table is obtained, which has exactly the same dimension as the input, i.e. T rows by 12 columns, but all the values in the table have been converted to dimensionless standard normal distribution values, with each column having a mean value of approximately 0 and a standard deviation of approximately 1.

[0024] Finally, two key hyperparameters need to be pre-set before starting the segmentation: window length W and step size S. Window length W defines the number of time points each individual analysis sample contains, and its selection is based on the ability to completely cover the feature evolution process of a typical fault event. In this embodiment, the window length W is set to 200 time points. Considering that the data sampling frequency is 100 Hz, this means that each window contains 2 seconds of flight data, which is sufficient to capture a series of fast dynamic processes such as single motor step loss, attitude controller response, etc. Step size S defines the time interval between the starting points of adjacent two windows, determining the degree of overlap between windows. In order to ensure the continuity of the timing information and avoid missing key information at the window boundary, a step size smaller than the window length is set to produce overlap. Here, the step size S is set to 50 time points, which means that there are 150 time points overlapping between adjacent two windows, i.e. 75% overlap rate, which helps the model capture the changes in timing features more smoothly and finely. After setting the parameters, the timing window segmentation process starts. The program starts from the first time point of the flight state standardized data table, i.e. row 0, and extracts all 12 columns of data from row 0 to row 199, a total of 200 rows. This 200-row, 12-column data segment forms the first flight state feature window. It is a two-dimensional matrix that completely records the standardized state evolution of the aircraft in the first 2 seconds. Subsequently, the starting point of the segmentation slides forward according to the set step size S. The starting position of the second extraction becomes row 50, and the program starts from row 50 and extracts data of the same length W (200 rows) backward, i.e. from row 50 to row 249. This new 200x12 data segment forms the second flight state feature window. Then, the process is repeated: the third window is extracted from row 100 to row 299, the fourth window is extracted from row 150 to row 349, and so on. This sliding extraction process will continue until the end position of the window exceeds the end of the entire data table (i.e. row 49999). During the sliding segmentation process, all the extracted window samples with a size of 200x12 are collected and stacked in the order they are generated. This stacking is done in a new dimension, thereby constructing a three-dimensional data structure, i.e. a sequence of flight state feature windows. The first dimension M of this three-dimensional tensor represents the total number of windows generated from the entire flight log. According to the formula M = ⌊(T-W) / S⌋+1, and substituting the values in this embodiment, where T is the total number of time points, W is the window length, S is the step size, and ⌊·⌋ represents the floor function, the total number of windows M = ⌊(50000-200) / 50⌋+1 = ⌊49800 / 50⌋+1 = 996+1 = 997 can be calculated. Therefore, the final generated sequence of flight state feature windows is a three-dimensional tensor with a size of 997x200x12.

[0025] In step S3, multi-scale feature extraction and channel refinement are performed on the flight state feature window sequence to obtain flight state refined small-scale features, flight state refined medium-scale features, and flight state refined large-scale features. It can be understood that the flight state feature window sequence generated in the previous step meets the model input requirements in structure, but the fault information contained therein is multi-level and intertwined. The core problem disclosed in the background art is that the time evolution patterns of faults of different origins have significant differences in scale: for example, the loss of step of a single motor electronic governor may exhibit a millisecond-level, high-frequency mutation signal, while a voltage drop caused by battery internal resistance aging may be a relatively flat downward trend that lasts for hundreds of milliseconds. If a single-scale feature extractor is used, it is easy to lose one and gain the other, and it is difficult to capture both of these two completely different dynamic patterns. Therefore, the present application constructs a parallel analysis structure with different time receptive fields through multi-scale feature extraction and channel refinement, so that deep features representing short-time mutations, medium-range fluctuations, and long-term trends can be simultaneously and unbiasedly extracted from the same flight data window, providing rich and identifiable information for subsequent accurate differentiation of these macroscopically similar faults.

[0026] One feasible way of step S3 of the present application is to perform multi-scale feature extraction and channel refinement on the flight state feature window sequence to obtain flight state refined small-scale features, flight state refined medium-scale features, and flight state refined large-scale features, including: S31, inputting the flight state feature window sequence into a first 1D convolution branch, a second 1D convolution branch, and a third 1D convolution branch to obtain flight state small-scale features, flight state medium-scale features, and flight state large-scale features, wherein the first 1D convolution branch, the second 1D convolution branch, and the third 1D convolution branch use convolution kernels with different scales; S32, inputting the flight state small-scale features, the flight state medium-scale features, and the flight state large-scale features into a channel attention module to obtain the flight state refined small-scale features, the flight state refined medium-scale features, and the flight state refined large-scale features.

[0027] Exemplarily, the implementation of step S3 of the present application is as follows: firstly in S31, a parallel computing architecture containing three independent sub-networks is constructed. The input flight state feature window sequence tensor is simultaneously copied and fed into the entrances of the three convolution branches. The three branches are similar in structure, each consisting of a one-dimensional convolution layer and a rectified linear unit (ReLU) activation function in series, but the main difference lies in the size of the convolution kernel in the one-dimensional convolution layer, which determines the field of view or the time receptive field of each branch. In particular, the convolution kernel size of the first 1D convolution branch is 3x1, the convolution kernel size of the second 1D convolution branch is 7x1, and the convolution kernel size of the third 1D convolution branch is 15x1. The first branch is a small-scale feature extraction branch. Its one-dimensional convolution layer is configured with a convolution kernel of size 3x1. This means that the convolution kernel will only cover 3 consecutive time points at a time when it slides along the time dimension (dimension of length 200). This small size of the convolution kernel is most sensitive to high-frequency changes and transient mutations in the data stream. The convolution layer is set with D equal to 64 convolution kernels, meaning it will learn and extract 64 different types of short-term local patterns. The weights and bias parameters within these convolution kernels are learned iteratively through the backpropagation algorithm and optimizer (such as the Adam optimizer) during the end-to-end training of the entire diagnostic model on a large amount of labeled flight data. Initially, they are random values. When an input tensor of 997x200x12 passes through this layer, a feature map of dimensions 997x198x64 is generated. The formula for the length of the output after one-dimensional convolution is: output length = input length - convolution kernel size + 1, i.e. 200-3+1=198. Subsequently, the feature map is fed into the ReLU activation function, and all values less than zero are set to zero to introduce non-linear representation capability. The time dimension of the feature map after activation function processing is retained and directly used as the output of the branch. Therefore, the output of this operation is the flight state small-scale feature, which is a three-dimensional feature map of dimensions 997x198x64. The second branch is a medium-scale feature extraction branch, which is similar in design to the small-scale branch, but its one-dimensional convolution layer uses a convolution kernel of size 7x1. This medium-sized convolution kernel allows it to observe 7 consecutive time points at a time, effectively capturing oscillation or fluctuation patterns with slightly longer duration. The layer is also configured with 64 convolution kernels, and their parameters are obtained through model training. Similarly, after the input tensor passes through the convolution and activation operations of this branch, a three-dimensional feature map of dimensions 997x194x64 is finally obtained, which is the flight state medium-scale feature. The third branch is a large-scale feature extraction branch, designed to capture long-term dependencies and slow trends in the data. To this end, its one-dimensional convolution layer is configured with a large convolution kernel of size 15x1. This wider receptive field allows it to integrate information across a longer time span, identifying patterns such as slow battery voltage decline or low-frequency drift of the aircraft's attitude.Consistent with the first two branches, this branch also contains 64 learnable convolution kernels and goes through the same processing procedure. The final output of the large-scale flight state feature is also a three-dimensional feature map with a dimension of 997x186x64, which retains the time sequence information.

[0028] Then in S32, a feasible way of step S32 of the present application, the flight state small-scale feature, the flight state medium-scale feature and the flight state large-scale feature are input into a channel attention module to obtain a flight state refined small-scale feature, a flight state refined medium-scale feature and a flight state refined large-scale feature, including: S321, performing global average pooling on the flight state small-scale feature to obtain a flight state small-scale channel descriptor; S322, performing feature excitation based on full connection coding on the flight state small-scale channel descriptor to obtain a flight state small-scale channel weight vector; S323, performing feature scaling on the flight state small-scale feature based on the flight state small-scale channel weight vector to obtain the flight state refined small-scale feature.

[0029] For clarity, the following will be elaborated with the process of handling the flight state small-scale features as an example, which is also applied to the medium-scale and large-scale features in parallel and independently. First, step S321 is performed, which is the step of global average pooling for the flight state small-scale features. The input of this step is a three-dimensional feature map with dimensions of 997x198x64. The global average pooling operation will be performed along the time dimension of each feature channel, i.e., the dimension with length of 198. Specifically, for each of the 997 samples and each of the 64 feature channels, the program will calculate the arithmetic mean of the values at all 198 time points on this channel. This operation will compress a sequence with length of 198 into a single scalar value, which can be regarded as the global response intensity of this feature channel within the entire time window. After performing this operation on all 64 channels, the original feature map with dimensions of 997x198x64 is converted into a two-dimensional tensor with dimensions of 997x64. This tensor is the flight state small-scale channel descriptor, each row of which represents a sample and each column represents a global information summary of a channel. Next, step S322 is performed, which is the feature excitation based on fully connected coding for the channel descriptor to generate channel weights. This process is a small bottleneck structure containing two fully connected layers, which is used to learn the nonlinear correlation between channels. The first layer is a fully connected layer that maps the input 64-dimensional channel descriptor to a lower-dimensional hidden space. To determine the dimension of this hidden space, a reduction rate needs to be set, which is a key parameter for balancing the complexity and expressiveness of the model, and is set to 16 in this embodiment. Therefore, the output dimension of this fully connected layer is 64 divided by 16, i.e., 4. This layer is followed by a rectified linear unit (ReLU) activation function to increase nonlinearity. The second layer is another fully connected layer that maps the 4-dimensional hidden vector back to the original 64 dimensions. The weights and bias parameters in these two fully connected layers are learned through the backpropagation algorithm during the training of the entire diagnostic model. After the second fully connected layer, a Sigmoid activation function is applied. The function of the Sigmoid function is to normalize the output 64 values to the range of 0 to 1, so that each value obtained can be directly interpreted as the importance weight of the corresponding channel. The final output of this step is a flight state small-scale channel weight vector with dimensions of 997x64. Finally, step S323 is performed, which is the feature scaling for the original flight state small-scale features based on the generated channel weight vector. The importance information learned in the previous two steps is applied back to the original feature map. The input has two: one is the flight state small-scale features output by step S31 without processing, and the other is the flight state small-scale channel weight vector generated in the previous step. The scaling operation is a channel-by-channel multiplication. To achieve this, the dimension of the weight vector needs to be expanded to match the dimension of the feature map.Specifically, for each sample, the i-th weight value in its 64-dimensional weight vector is broadcasted and multiplied to all 198 time points of the i-th channel of the original feature map of the sample. In this way, the original features of the channels with importance weight values close to 1 are basically retained, and the features of the channels with importance weight values close to 0 are significantly weakened, and the output obtained is the flight state refined small-scale feature, which has the same dimension as the input feature map, still 997x198x64, but the internal feature information has been reweighted according to the importance of the channels.

[0030] For the flight state medium-scale feature and the flight state large-scale feature, they are also processed according to the above three sub-steps through respective independent channel attention modules, and finally the flight state refined medium-scale feature with a dimension of 997x194x64 and the flight state refined large-scale feature with a dimension of 997x186x64 are obtained.

[0031] In particular, after multi-scale convolution processing, the small, medium and large scale features obtained have essential differences in physical meaning and data statistical distribution. For example, the small-scale convolution kernel captures high-frequency transient signals such as fine jitter of motor control or instantaneous spikes of sensor readings; while the large-scale convolution kernel focuses on low-frequency trend signals such as slow decline of battery voltage or long-term drift of body posture. If the same, one-size-fits-all channel attention mechanism is used for the three types of heterogeneous features, the depth dependence of the importance relationship between channels on the scale will be ignored. This one-size-fits-all method cannot fully exploit the most recognizable information at each scale. Therefore, in order to maximize the use of prior distribution information of scale features, the present application introduces a channel attention module based on scale conditioning. That is, during the calculation of the excitation process of the channel weight, the model is actively provided with the context information of the scale (small, medium or large) to which the feature being processed belongs. In this way, the model can dynamically adjust its excitation behavior and learn different, more appropriate inter-channel dependencies for features of different scales, thereby achieving fine and customized purification of the features.

[0032] Based on this, a preferred way of step S32 of the present application is S32, inputting the flight state small-scale feature, the flight state medium-scale feature and the flight state large-scale feature into a channel attention module to obtain the flight state refined small-scale feature, the flight state refined medium-scale feature and the flight state refined large-scale feature, comprising: inputting the flight state small-scale feature, the flight state medium-scale feature and the flight state large-scale feature into a channel attention module based on scale conditioning to obtain the flight state refined small-scale feature, the flight state refined medium-scale feature and the flight state refined large-scale feature.

[0033] In detail, first, scale identity embedding and channel information compression are performed, a unique, learnable identity is created for each scale, and at the same time, the input time sequence feature map is compressed into a channel descriptor that can represent global information, preparing for subsequent conditional excitation. Specifically, first, a low-dimensional, learnable embedding vector is created for small, medium, and large scales, denoted as Ek, where k takes values s (small), m (medium), and l (large). The dimension of this embedding vector is a preset hyperparameter, for example, set to 8. During model training, Es, Em, and El will learn to represent the intrinsic semantics of their respective scales through the backpropagation algorithm. Here, the flight state small scale feature is taken as an example to illustrate, that is, k is s. The input flight state small scale feature with a dimension of 997x198x64 is compressed through global average pooling along the time dimension with a length of 198 to obtain a channel descriptor zk with a dimension of 997x64. In particular, this process is the same as the S321 process in the above method. Subsequently, the channel descriptor zk is sent to the first scale-specific fully connected layer (FC1), which has an independent weight matrix W1k. The fully connected layer (FC1) is specifically composed of a weight matrix W1k and a bias vector b1k. The dimension of the weight matrix W1k is 16x64, and the dimension of the bias vector b1k is 16x1. After linear transformation and ReLU activation function processing, an intermediate hidden representation hk with a compressed dimension is generated, hk = ReLU(W1k·zk), which describes the transformation process from the channel descriptor to the intermediate representation. The input zk is a 64-dimensional vector representing global channel information. W1k is the scale-specific weight matrix learned during model training, which maps the 64-dimensional zk to a lower-dimensional space, such as 16-dimensional. The ReLU function introduces nonlinearity, enhancing the model's expression ability. The output hk is a 16-dimensional intermediate vector that condenses the core information.

[0034] Next, scale-conditioned modulation is performed, which is the core step to achieve scale conditioning. The purpose is to inject the scale identity information into the feature processing flow in a differentiable way, dynamically adjusting the intermediate features to carry the scale context information. Specifically, the embedding vector Es corresponding to the scale, such as the small scale S, is input into an independent small FiLM parameter generator network. The generator network can be implemented as a single fully connected linear layer in architecture, with the input dimension being the dimension of the embedding vector Ek (i.e. 8) and the output dimension being twice the dimension of the intermediate representation hk (i.e. 32). The linear layer contains a 32x8 weight matrix and a 32x1 bias vector inside, and the parameters are also learned through model training. The 32-dimensional vector output is then split into two 16-dimensional vectors, which are used as the scaling parameter vector γk and the bias parameter vector βk, respectively. Then, the feature-level linear modulation is performed on hk using the two parameters, i.e. hk' = γk·hk+ βk, which is the core of the feature-level linear modulation (FiLM). It performs independent scaling and translation on each element in the hk vector. γk is responsible for feature scaling of each channel of hk, while βk is responsible for feature translation. The multiplication and addition here are element-wise operations of vectors. For example, if an element of hk is 0.8, the corresponding element of γk is 1.2, and the element of βk is -0.1, then the new element hk' after modulation is 1.2*0.8-0.1 = 0.86. Through this operation, the original hk is modulated by the scale information Es to generate a new hk' that contains scale context information.

[0035] Then, the most weight generation and feature refinement are performed. The intermediate representation modulated by the scale information is mapped back to the original channel dimension, and the final channel attention weight with scale perception ability is generated, thereby completing the refinement of the original feature map. Specifically, the modulated intermediate representation hk' is sent to a second fully connected layer (FC2). The fully connected layer (FC2) is also composed of a weight matrix W2k and a bias vector b2k in architecture, the dimension of the weight matrix W2k is 64x16, and the dimension of the bias vector b2k is 64x1, which is responsible for mapping the 16-dimensional modulated feature back to the 64-dimensional channel space. Then, through a Sigmoid activation function, the output of 64 values is normalized to the range of 0 to 1 to obtain the final channel attention weight sk. This sk is a 64-dimensional weight vector, and each value represents the importance of the corresponding channel at the current scale. Finally, the weight vector sk and the original flight state small-scale feature (dimension 997x198x64) output in step S31 without pooling are multiplied (broadcast operation) by channel to obtain the final flight state refined small-scale feature with the dimension of 997x198x64. In particular, the flight state medium-scale feature and the flight state large-scale feature are also obtained according to the above three processes to obtain the flight state refined medium-scale feature and the flight state refined large-scale feature.

[0036] In particular, in the training process of the entire model, in order to ensure that the learning of the scale embedding vector Ek has actual physical meaning rather than being formal, an additional constraint term is introduced, that is: The formula is a regularization constraint, which is part of the loss function during training. It aims to ensure that the internal structure of the scale embedding vector Ek is consistent with the structure of the fully connected layer weight W1k using it. Where V1Ek is an eigenvector (or related projection) of W1k. This projection V1Ek is dynamically calculated according to the internal structure of the weight matrix W1k. In specific implementation, by analyzing the weight matrix W1k, the most important structural direction (e.g. by singular value decomposition or other techniques to find its main components) can be extracted, thereby defining a feature subspace most relevant to the scale transformation. V1Ek is the result of mathematically projecting the current scale embedding vector Ek into this feature subspace. By minimizing the difference between Ek and its projection V1Ek, the model can learn a more meaningful and more structured scale representation.

[0037] In this way, the model can learn channel relationships that are highly dependent on scale. For example, for small-scale features, the model may learn to assign high weights to vibration-related sensor channels; while for large-scale features, it may assign high weights to energy-related channels such as voltage and current, thereby greatly improving the model's ability to distinguish similar fault representations caused by different physical causes (such as high-frequency vibration or low-frequency voltage drop).

[0038] In step S4, the refined small-scale features, refined medium-scale features, and refined large-scale features of the flight state are adaptively fused using a gating mechanism to obtain the flight state fused diagnostic vector. It should be understood that after multi-scale feature extraction and channel refinement, although optimized feature sets capable of representing dynamics at different time scales are obtained, these three sets of features remain independent. However, for different types of UAV faults, the most identifiable information may explicitly exist at a specific scale, or require a specific combination of features at multiple scales for accurate identification. For example, an instantaneous fault is most significant in small-scale features, while a gradual performance degradation is more reflected in large-scale features. If a simple static fusion strategy such as addition or concatenation is used, it will be impossible to dynamically adjust the contribution of each scale feature in the final diagnostic decision based on the actual situation of the input data. Therefore, the core purpose of implementing adaptive fusion based on gating mechanism is to introduce an intelligent arbitration module. This module can learn autonomously and determine which scale of feature information is more critical for the current flight status window to be diagnosed, thereby generating dynamic fusion weights, realizing the optimized integration of multi-scale features, and providing a diagnostic vector with higher information entropy and stronger identification ability for the final accurate fault classification.

[0039] Figure 4 This is a flowchart of step S4 in the cloud-based intelligent performance analysis method for unmanned aerial vehicles according to an embodiment of this application. Figure 4 As shown, one feasible method of step S4 in this application involves adaptively fusing the refined small-scale features, refined medium-scale features, and refined large-scale features of the flight state based on a gating mechanism to obtain a flight state fusion diagnostic vector. This includes: S41, integrating the information from the refined small-scale features, refined medium-scale features, and refined large-scale features of the flight state to obtain a multi-scale spliced ​​vector of the flight state; S42, inputting the multi-scale spliced ​​vector of the flight state into a small fully connected network to obtain a first gating weight, a second gating weight, and a third gating weight; and S43, performing a weighted fusion of the refined small-scale features, refined medium-scale features, and refined large-scale features of the flight state based on the first gating weight, the second gating weight, and the third gating weight to obtain the flight state fusion diagnostic vector.

[0040] Exemplarily, the implementation of step S4 of the present application is as follows: Firstly, information integration is performed in S41. Since the time dimensions (198, 194, 186 respectively) of the three feature tensors are inconsistent, direct concatenation is not possible. Therefore, before concatenation, information summary of the time dimension needs to be performed on each feature tensor. This embodiment uses global max-pooling operation to achieve this purpose. Specifically, for the flight state refined small-scale feature with a dimension of 997x198x64, global max-pooling will be performed along the time dimension with a length of 198, and for each of the 64 feature channels, the maximum value at the 198 time points will be found. This operation will compress the small-scale feature of each sample from a 198x64 feature map into a 64-dimensional vector, which captures the strongest response of each small-scale pattern within the time window. After this operation, the feature tensor becomes 997x64. Similarly, the same global max-pooling operation is performed on the flight state refined medium-scale feature and the flight state refined large-scale feature, and two feature tensors with a dimension of 997x64 are obtained. After obtaining three feature tensors with a dimension of 997x64, which represent the global information of each scale, information integration can be performed. This process is performed in the feature dimension. That is, the three tensors are connected along the second dimension (feature dimension) to form a longer feature vector. The dimension of the concatenated tensor is 997x(64+64+64), i.e. 997x192. The tensor with a dimension of 997x192 is the flight state multi-scale concatenated vector, and each row of the vector contains the condensed feature information of small, medium and large scales from the same flight data window.

[0041] Next, in step S42, the gating weights for weighted fusion are generated. This small fully connected network is a multi-layer perceptron (MLP) whose architecture is designed to learn the complex nonlinear relationships among the three scale features with low computational cost. The network consists of an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is 192, which exactly matches the dimension of the input vector. The role of the hidden layer is to extract more deep-level combined features, and the number of neurons in the hidden layer is a hyperparameter that needs to be preset, which is set to 32 in this embodiment to reflect its small size. After the hidden layer, a rectified linear unit (ReLU) is applied as the activation function to introduce nonlinear expression capability to the network. The output layer of the network contains 3 neurons, which correspond to the gating weights to be generated for the small, medium, and large scale features, respectively. In particular, after the output layer, a Softmax activation function is applied. The Softmax function can convert the 3 original output values of the output layer into a probability distribution, i.e., the 3 converted values are all positive numbers, and their sum is strictly equal to 1. For example, if the original values output by the 3 neurons of the output layer are [2.5, -0.5, 0.1], they may become [0.88, 0.04, 0.08] after the Softmax function processing. This property makes its output perfectly interpreted as the relative importance weights of the three feature scales. In particular, all the weights and bias parameters in this small fully connected network are learned uniformly through the backpropagation algorithm and the optimizer during the end-to-end training of the entire diagnostic model. The final output is a tensor with a dimension of 997x3, each row of which contains three weight values corresponding to the first gating weight, the second gating weight, and the third gating weight.

[0042] Then, in S43, weighted fusion is performed. One feasible way of step S43 of the present application is to perform weighted fusion on the flight state refined small-scale feature, the flight state refined medium-scale feature, and the flight state refined large-scale feature based on the first gating weight, the second gating weight, and the third gating weight to obtain the flight state fusion diagnostic vector, including: performing weighted fusion on the flight state refined small-scale feature, the flight state refined medium-scale feature, and the flight state refined large-scale feature based on the first gating weight, the second gating weight, and the third gating weight according to the following formula: ; wherein, and is the first gating weight, the second gating weight, and the third gating weight, and is the feature obtained after pooling the flight state refined small-scale feature, the flight state refined medium-scale feature, and the flight state refined large-scale feature, The flight state fusion diagnosis vector is fused. This process is consistent with the information integration operation for learning weights in step S41, and is implemented using global max pooling. Specifically, for the flight state refined small-scale feature of 997x198x64, global max pooling is performed along the time dimension of length 198, obtaining a vectorized feature of 997x64, denoted as . Similarly, the same operation is performed on the flight state refined medium-scale feature and the flight state refined large-scale feature, respectively, obtaining vectorized features of 997x64, denoted as and After obtaining three vectorized refined features of 997x64, the weighted fusion operation can be performed. This calculation process is performed in parallel on all 997 samples.

[0043] For a clearer illustration, take the processing of the ith sample as an example. This sample corresponds to three 64-dimensional feature vectors: and . At the same time, it also corresponds to the ith row in the gating weight tensor, that is, three scalar weight values: and . For this sample, for example, the gating network determines that the fault feature is mainly reflected in the high-frequency mutation, so the learned weight may be . The fusion calculation first performs element-wise scalar and vector multiplication: multiplied by the 64-dimensional , multiplied by the 64-dimensional , multiplied by the 64-dimensional . This operation significantly amplifies the values in , while greatly attenuating the values in and . Subsequently, the three scaled 64-dimensional vectors are added element by element, and finally a single new 64-dimensional vector is obtained, which is the fusion diagnosis vector of the ith sample. Since this fusion vector is mainly contributed by the small-scale feature that is assigned the highest weight, it maximally retains the most valuable information for diagnosing this particular fault. This calculation process is efficiently completed on the entire tensor through the broadcast mechanism, and finally outputs a two-dimensional tensor of 997x64, that is, the flight state fusion diagnosis vector. Each row of this tensor is a 64-dimensional diagnosis vector obtained after adaptive weighted fusion, with highly concentrated information and high discriminability.

[0044] In step S5, fault classification and diagnosis output is performed based on the flight state fusion diagnosis vector to obtain a fault probability. That is, after the complex processing of all previous steps, a highly condensed, large amount of information and fixed dimension flight state fusion diagnosis vector has been successfully extracted and generated from the original, heterogeneous flight control log. This vector is the essence of the entire data processing flow, and it contains the fingerprints of possible fault patterns in the current flight state window in a numerical form that can be understood by a deep learning model. However, this vector itself is completely opaque to the user, it is just a string of abstract floating-point numbers, and cannot directly tell whether the UAV has a fault, let alone indicate what specific fault. Therefore, the final fault classification and diagnosis output step is to build a decoder or translator to convert this machine-readable, highly abstract feature vector into a human-readable, clearly business-meaning, decision-making diagnosis conclusion, i.e. the occurrence probability of each pre-set fault type.

[0045] Exemplarily, a feasible implementation of step S5 of the present application is as follows: the core of this process is a fully connected neural network for classification, which is a multi-layer perceptron. Before deploying the classifier, a complete set of fault categories is first defined. This set is determined in advance based on statistical analysis of historical UAV fault data, expert experience and failure mode and effects analysis (FMEA). In a specific application scenario, the set may contain C equal to 7 categories, for example: category 0 - normal flight, category 1 - No. 1 motor electronic speed controller step loss, category 2 - No. 2 motor electronic speed controller step loss, category 3 - No. 3 motor electronic speed controller step loss, category 4 - No. 4 motor electronic speed controller step loss, category 5 - battery voltage drop, and category 6 - severe body vibration. The definition of this category directly determines the output structure of the classifier.

[0046] The architecture of the classifier consists of an input layer, a hidden layer and an output layer. The number of neurons in the input layer is set to 64, which completely matches the dimension of the input flight state fusion diagnosis vector, to ensure that the vector can be received as input. Then there is a hidden layer, which learns more complex nonlinear combinations between the internal features of the diagnosis vector to enhance the accuracy of classification. The number of neurons in the hidden layer is an adjustable hyperparameter, which is set to 32 in this embodiment. After the hidden layer, a rectified linear unit (ReLU) is applied as the activation function. Finally, the output layer has a number of neurons equal to the total number of pre-defined fault categories C, so in this embodiment there are 7. Each neuron in the output layer uniquely corresponds to a fault category. For example, the first neuron corresponds to normal flight, the second neuron corresponds to No. 1 motor electronic speed controller step loss, and so on.

[0047] All the weight and bias parameters in the classifier are learned uniformly as an integral part of the whole diagnostic model (from the convolutional branch in S3 to the classifier in S5) through an end-to-end training process. This process uses a labeled dataset containing a large number of flight log samples and their corresponding true fault labels, adopts a loss function such as cross-entropy, and iteratively adjusts all parameters through a backpropagation algorithm and an optimizer such as Adam until the difference between the predicted results output by the model and the true labels is minimized.

[0048] In a specific implementation, a flight state fusion diagnostic vector with a dimension of 997 x 64 is sent to the input layer of the classifier. For any 64-dimensional vector in the 997 samples, it is first linearly transformed and nonlinearly activated by the hidden layer to be mapped to a 32-dimensional hidden feature. Then, the 32-dimensional hidden feature is linearly transformed by the output layer to obtain a 7-dimensional original output vector, also known as a logical value. Each value in the 7-dimensional vector represents the original confidence score of the model that the sample belongs to the corresponding fault category.

[0049] Finally, in order to convert these original scores into intuitive probabilities, the 7-dimensional logical value vector is sent to a Softmax activation function. The Softmax function performs exponential operation and normalization on it, outputting a 7-dimensional vector, but all element values in the vector are between 0 and 1, and their sum is strictly equal to 1. This vector processed by the Softmax function is the fault probability. For example, for a certain input sample, the output fault probability vector may be [0.01, 0.88, 0.02, 0.01, 0.03, 0.04, 0.01]. This set of values clearly shows that the model diagnosis believes that the probability of the flight data window appearing No. 1 motor governor step loss is 88%, the probability of battery voltage drop is 4%, and the probability of belonging to other categories is very low. The entire process is parallelly processed for all 997 samples, and finally outputs a fault probability matrix with a dimension of 997 x 7, completing the conversion from features to diagnostic conclusions. The fault probability matrix not only gives the final diagnostic conclusion, but also presents the dynamic evolution process of the probability of each type of fault in a time-sequential manner, providing fine and quantitative decision-making basis for fault location, health assessment and predictive maintenance of unmanned aerial vehicles.

[0050] In summary, the cloud platform-based unmanned aerial vehicle performance intelligent analysis method based on the embodiments of the present application is illustrated. For the technical problems of similar macroscopic manifestations and easy confusion caused by different unmanned aerial vehicle faults, the present application proposes a multi-scale feature extraction and adaptive fusion diagnosis method. Through parallel multi-scale convolution branches, the method synchronously captures small-scale mutation features (such as instantaneous changes caused by single motor step loss) and large-scale slowly changing features (such as global performance decline caused by battery aging) in the flight control log time series data. Subsequently, a gating fusion mechanism is creatively introduced to dynamically weight the extracted multi-scale features, enabling the model to adaptively determine which scale of features is more decisive for the current fault classification. In this way, the model can excavate the deep temporal patterns with discriminability from similar macroscopic manifestations, effectively distinguish faults with different roots but similar manifestations, and significantly improve the accuracy of fault diagnosis.

[0051] Figure 5 The block diagram of the cloud platform-based unmanned aerial vehicle performance intelligent analysis system according to the embodiments of the present application is shown. As shown in Figure 5 The cloud platform-based unmanned aerial vehicle performance intelligent analysis system 100 according to the embodiments of the present application includes: an original flight control log acquisition module 110 for acquiring original flight control log files from a cloud platform; a flight control data preprocessing module 120 for data preprocessing and feature window generation on the original flight control log files to obtain a flight state feature window sequence; a flight state multi-scale feature extraction module 130 for multi-scale feature extraction and channel refining on the flight state feature window sequence to obtain flight state refined small-scale features, flight state refined medium-scale features, and flight state refined large-scale features; a gating adaptive feature fusion module 140 for adaptive fusion based on a gating mechanism on the flight state refined small-scale features, the flight state refined medium-scale features, and the flight state refined large-scale features to obtain a flight state fusion diagnosis vector; and a fault classification and diagnosis output module 150 for fault classification and diagnosis output based on the flight state fusion diagnosis vector to obtain a fault probability.

[0052] Here, those skilled in the art can understand that the specific operations of each step in the above cloud platform-based unmanned aerial vehicle performance intelligent analysis system have been described in detail above with reference to the description of the cloud platform-based unmanned aerial vehicle performance intelligent analysis method of Figures 1 to 4 Therefore, the repeated description will be omitted.

Claims

1.A cloud platform-based intelligent analysis method for unmanned aerial vehicle performance, characterized in that, The method comprises the following steps: obtaining an original flight control log file from a cloud platform; performing data preprocessing and feature window generation on the original flight control log file to obtain a flight state feature window sequence; performing multi-scale feature extraction and channel refinement on the flight state feature window sequence to obtain flight state refined small-scale features, flight state refined medium-scale features and flight state refined large-scale features; performing adaptive fusion based on a gating mechanism on the flight state refined small-scale features, the flight state refined medium-scale features and the flight state refined large-scale features to obtain a flight state fusion diagnosis vector; performing fault classification and diagnosis output based on the flight state fusion diagnosis vector to obtain a fault probability. 2.The cloud platform-based unmanned aerial vehicle performance intelligent analysis method according to claim 1, characterized in that, The method for performing data preprocessing and feature window generation on the original flight control log file to obtain a flight state feature window sequence comprises the following steps: performing multi-source heterogeneous log analysis on the original flight control log file to obtain a flight state synchronization data table; performing data cleaning and standardization on the flight state synchronization data table to obtain a flight state standardized data table; performing time window segmentation and sequence construction on the flight state standardized data table to obtain the flight state feature window sequence. 3.The cloud platform-based UAV performance intelligent analysis method of claim 2, wherein, Each flight state feature window in the flight state feature window sequence comprises a roll angle, a pitch angle, a target roll angle, a target pitch angle, an X-axis angular velocity, a Y-axis angular velocity, a Z-axis angular velocity, a battery voltage, a PWM control signal value sent by the flight control to a No. 1 motor electronic speed controller, a PWM control signal value sent by the flight control to a No. 2 motor electronic speed controller, a PWM control signal value sent by the flight control to a No. 3 motor electronic speed controller and a PWM control signal value sent by the flight control to a No. 4 motor electronic speed controller. 4.The cloud platform based UAV performance intelligent analysis method of claim 1, wherein, The method for performing multi-scale feature extraction and channel refinement on the flight state feature window sequence to obtain flight state refined small-scale features, flight state refined medium-scale features and flight state refined large-scale features comprises the following steps: inputting the flight state feature window sequence into a first 1D convolution branch, a second 1D convolution branch and a third 1D convolution branch respectively to obtain flight state small-scale features, flight state medium-scale features and flight state large-scale features, wherein the first 1D convolution branch, the second 1D convolution branch and the third 1D convolution branch respectively use convolution kernels with different scales; inputting the flight state small-scale features, the flight state medium-scale features and the flight state large-scale features into a channel attention module to obtain flight state refined small-scale features, flight state refined medium-scale features and flight state refined large-scale features. 5.The cloud platform based UAV performance intelligent analysis method of claim 4, wherein, The size of the convolution kernel of the first 1D convolution branch is 3×1, the size of the convolution kernel of the second 1D convolution branch is 7×1 and the size of the convolution kernel of the third 1D convolution branch is 15×1. 6.The cloud platform based UAV performance intelligent analysis method according to claim 4, wherein, The method for inputting the flight state small-scale features, the flight state medium-scale features and the flight state large-scale features into the channel attention module to obtain the flight state refined small-scale features, the flight state refined medium-scale features and the flight state refined large-scale features comprises the following steps: performing global average pooling on the flight state small-scale features to obtain flight state small-scale channel descriptors; performing feature excitation based on fully connected coding on the flight state small-scale channel descriptors to obtain a flight state small-scale channel weight vector; The flight state small-scale features are scaled based on a flight state small-scale channel weight vector to obtain flight state refined small-scale features. 7.The cloud platform based UAV performance intelligent analysis method according to claim 1, wherein, The flight state refined small-scale features, the flight state refined medium-scale features, and the flight state refined large-scale features are adaptively fused based on a gating mechanism to obtain a flight state fusion diagnosis vector, including: The flight state refined small-scale features, the flight state refined medium-scale features, and the flight state refined large-scale features are information-integrated to obtain a flight state multi-scale splicing vector; The flight state multi-scale splicing vector is input into a small full connection network to obtain a first gating weight, a second gating weight, and a third gating weight; The flight state refined small-scale features, the flight state refined medium-scale features, and the flight state refined large-scale features are weighted fused based on the first gating weight, the second gating weight, and the third gating weight to obtain the flight state fusion diagnosis vector. 8.The cloud platform based UAV performance intelligent analysis method of claim 7, wherein, The flight state refined small-scale feature, the flight state refined medium-scale feature and the flight state refined large-scale feature are weighted and fused based on the first gating weight, the second gating weight and the third gating weight to obtain the flight state fusion diagnosis vector, comprising: the flight state refined small-scale feature, the flight state refined medium-scale feature and the flight state refined large-scale feature are weighted and fused based on the first gating weight, the second gating weight and the third gating weight according to the following formula: ; wherein, and the first gating weight, the second gating weight and the third gating weight, and the feature obtained after the flight state refined small-scale feature, the flight state refined medium-scale feature and the flight state refined large-scale feature are pooled, the flight state fusion diagnosis vector. 9.A cloud platform-based unmanned aerial vehicle (UAV) performance intelligent analysis system, characterized in that, including: An original flight control log acquisition module configured to acquire an original flight control log file from a cloud platform; A flight control data preprocessing module configured to perform data preprocessing and feature window generation on the original flight control log file to obtain a flight state feature window sequence; A flight state multi-scale feature extraction module configured to perform multi-scale feature extraction and channel refinement on the flight state feature window sequence to obtain flight state refined small-scale features, flight state refined medium-scale features, and flight state refined large-scale features; A gating adaptive feature fusion module configured to adaptively fuse the flight state refined small-scale features, the flight state refined medium-scale features, and the flight state refined large-scale features based on a gating mechanism to obtain a flight state fusion diagnosis vector; A fault classification and diagnosis output module configured to perform fault classification and diagnosis output based on the flight state fusion diagnosis vector to obtain a fault probability.

Images