A state diagnosis method, system, device and medium based on unmanned aerial vehicle operation data

Abstract

This application discloses a state diagnosis method, system, device, and medium based on UAV operational data, mainly relating to the field of state diagnosis technology. It addresses the problems of single-channel statistical features failing to capture cross-channel gradient correlations and dynamic coupling relationships between sensor signals, and unidirectional LSTMs only utilizing historical information and neglecting the inverse constraints of future temporal context on the current state. The method includes: inputting training samples into a forward LSTM and a backward LSTM respectively to obtain forward and backward outputs; concatenating the forward and backward outputs; compressing the concatenated data into a fixed-length semantic vector; obtaining the labeled state of the semantic vector; and obtaining a trained classification function based on the semantic vector and the labeled state; when new multi-channel UAV operational data is obtained, obtaining the corresponding new semantic vector; and inputting the new semantic vector into the trained classification function to obtain the predicted state.

Description

Technical Field

[0001] This application relates to the field of condition diagnosis technology, and in particular to a condition diagnosis method, system, device and medium based on UAV operation data. Background Technology

[0002] Current UAV condition diagnosis generally relies on rule-based judgments based on preset thresholds and traditional machine learning models. Mainstream methods typically collect data from multiple sensors such as inertial measurement units (IMUs), GPS, barometers, and motor current, construct feature sets by calculating single-channel statistical features (such as mean, variance, and peak value), and then use support vector machines (SVMs) or random forests for state classification. Some studies have attempted to introduce unidirectional long short-term memory networks (LSTMs) to process time series, but these are limited to forward time-dependent modeling, and feature extraction is mostly based on raw signals or simple sliding window statistics, lacking a deep representation of the nonlinear gradient evolution patterns between multiple channels. Data processing often relies on cloud transmission and centralized analysis, resulting in significant response delays and making it difficult to meet the real-time diagnostic needs in complex low-altitude environments.

[0003] Existing technologies have significant structural defects in time series data modeling: on the one hand, single-channel statistical features cannot capture the cross-channel gradient correlation and dynamic coupling relationship between sensor signals, resulting in insufficient sensitivity to minor anomalies (such as progressive bearing wear and motor torque fluctuations); on the other hand, unidirectional LSTM only uses historical information and ignores the reverse constraint of future time series context on the current state, making it difficult to accurately identify composite fault modes with precursors and non-stationarity. Summary of the Invention

[0004] This application provides a state diagnosis method, system, device, and medium based on UAV operation data to solve the problems of existing solutions, such as the inability of single-channel statistical features to capture cross-channel gradient correlation and dynamic coupling between sensor signals, and the difficulty of accurately identifying complex fault modes with precursors and non-stationarity by using only historical information and ignoring the reverse constraints of future time context on the current state in unidirectional LSTM.

[0005] Firstly, this application provides a state diagnosis method based on UAV operational data, the method comprising: Acquire multi-channel operation data of the UAV, divide the multi-channel operation data according to time windows, and calculate the multi-channel gradient distribution histogram for each time window; construct feature vectors based on the specific distribution of a preset number of multi-channel gradient distribution histograms; use principal component analysis to reduce the dimensionality of the feature vectors to obtain the dimensionality-reduced vectors. Based on the dimensionality-reduced vectors and the preset number of time steps, training samples are constructed; The training samples are input into the forward LSTM and the backward LSTM respectively to obtain the forward output and the backward output. The forward output and the backward output are concatenated according to the preset time steps to obtain the concatenated data. After global average pooling, the concatenated data is compressed into a semantic vector of fixed length. The annotation state of the semantic vector is obtained. Based on the semantic vector and the annotation state, the trained classification function is obtained. When new multi-channel operation data of a drone is obtained, a new semantic vector corresponding to the new drone multi-channel operation data is obtained; the new semantic vector is input into a trained classification function to obtain the predicted state.

[0006] In one implementation of this application, the multi-channel operating data of the UAV includes at least three types of sensor data: flight attitude angle, motor speed, battery voltage, accelerometer reading, and gyroscope angular velocity.

[0007] In one implementation of this application, multi-channel UAV operational data is acquired, the data is divided into time windows, and a multi-channel gradient distribution histogram for each time window is calculated. Specifically, this includes: The time window of the continuously collected UAV multi-channel operation data of a preset length is divided by sliding segmentation. The first difference of the data sequence of each sensor channel within each time window is calculated to obtain the gradient sequence. The gradient sequence of each channel is divided into equal-width bins, and the frequency of occurrence of gradient values ​​in each bin is counted to form a multi-channel gradient distribution histogram.

[0008] In one implementation of this application, a feature vector is constructed by the specific distribution of a preset number of multi-channel gradient distribution histograms, specifically including: The multi-channel gradient distribution histograms corresponding to N consecutive time windows are concatenated according to the channel dimension to form a high-dimensional original feature matrix; The histogram of each channel is normalized so that the sum of the frequencies of all histograms is 1. All normalized histograms are concatenated in chronological order to form a one-dimensional feature vector.

[0009] In one implementation of this application, principal component analysis is used to reduce the dimensionality of the eigenvectors to obtain the dimensionality-reduced vectors, specifically including: Center all constructed eigenvectors and calculate the covariance matrix; Perform eigenvalue decomposition on the covariance matrix and select the eigenvectors corresponding to the top K principal components with a cumulative contribution rate of not less than 95%. Each feature vector is projected onto a predefined low-dimensional subspace consisting of K principal components to obtain a dimension-K reduced vector.

[0010] In one implementation of this application, forward and backward outputs are concatenated according to a preset number of time steps to obtain concatenated data, specifically including: The dimensionality-reduced vector is divided into continuous subsequences according to a preset time step T, and then input into the forward LSTM network and the backward LSTM network respectively. The forward LSTM processes the subsequences in ascending time order and outputs the hidden state at each time step. The backward LSTM processes the same subsequence in reverse time order and outputs the hidden state at each time step. The forward and backward hidden states at the corresponding time steps are concatenated along the feature dimension to form bidirectional context-aware concatenated data.

[0011] In one implementation of this application, when new UAV multi-channel operation data is obtained, a new semantic vector corresponding to the new UAV multi-channel operation data is obtained; the new semantic vector is input into a trained classification function to obtain a predicted state, specifically including: For the newly added UAV multi-channel operation data, the time window division, gradient histogram calculation, feature vector construction and principal component dimensionality reduction process are completely consistent with the training phase to generate the corresponding newly added dimensionality-reduced vector; The newly added dimensionality-reduced vector is input into the trained bidirectional LSTM and global average pooling network, and the output is a new semantic vector of fixed length; the new semantic vector is input into the classification function obtained by training with labeled states; the class with the highest probability is selected as the final predicted state.

[0012] Secondly, this application provides a status diagnosis system based on UAV operational data, the system comprising: The dimensionality reduction module is used to acquire multi-channel UAV operational data, divide the multi-channel operational data according to time windows, and calculate the multi-channel gradient distribution histogram for each time window; construct feature vectors based on the specific distribution of a preset number of multi-channel gradient distribution histograms; and perform dimensionality reduction processing on the feature vectors using principal component analysis to obtain the dimensionality-reduced vectors. The training module is used to construct training samples based on the dimensionality-reduced vectors and a preset number of time steps; input the training samples into the forward LSTM and backward LSTM respectively to obtain the forward output and backward output; concatenate the forward output and backward output according to the preset number of time steps to obtain concatenated data; compress the concatenated data into a fixed-length semantic vector through global average pooling; obtain the annotation state of the semantic vector; and obtain the trained classification function based on the semantic vector and the annotation state. The prediction module is used to obtain the new semantic vector corresponding to the new multi-channel operation data of the new UAV when new multi-channel operation data is obtained; and input the new semantic vector into the trained classification function to obtain the predicted state.

[0013] In one implementation of this application, the dimensionality reduction module includes a graphing unit. This method is used to slide and divide the continuously collected multi-channel UAV operation data into a time window of a preset length. Within each time window, the first-order difference of the data sequence of each sensor channel is calculated to obtain the gradient sequence. The gradient sequence of each channel is divided into equal-width bins, and the frequency of the gradient value in each bin is counted to form a multi-channel gradient distribution histogram.

[0014] Thirdly, this application provides a status diagnostic device based on UAV operational data, the device comprising: processor; And a memory that stores executable code, which, when executed, causes the processor to perform a state diagnosis method based on UAV operational data, as described above.

[0015] Fourthly, this application provides a non-volatile computer storage medium storing computer instructions, which, when executed, implement a state diagnosis method based on UAV operation data as described above.

[0016] As can be seen from the above technical solutions, this application has the following advantages: By constructing a multi-channel gradient distribution histogram, this approach overcomes the information silo limitations of traditional single-channel statistical features in time-series modeling. Traditional methods only extract scalar statistics such as mean and variance independently for each sensor channel, failing to reflect the cooperative gradient relationships of instantaneous changes between channels. This results in a lack of responsiveness to cross-channel coupled micro-anomalies (such as the synchronous evolution of rotor motor torque fluctuations and gyroscope angular velocity offsets). This application jointly models the gradient changes of multi-channel signals within each time window, forming a high-dimensional gradient distribution histogram, thereby explicitly encoding the dynamic coupling structure between sensor signals. This structured feature representation enables the model to perceive the cooperative perturbation patterns of multi-dimensional physical quantities in the time dimension, improving the structural sensitivity to complex early faults. It eliminates the need for manual feature engineering or pre-assumptions about inter-channel correlations, achieving a fundamental leap from "single-point observation" to "multi-dimensional linkage."

[0017] This paper employs a bidirectional concatenation architecture of forward and backward LSTM to overcome the one-sided temporal perception caused by unidirectional LSTM relying solely on historical information. In UAV operational data, many complex fault modes exhibit precursory and non-stationarity characteristics, and their current state is often constrained by the inverse of future temporal trends (e.g., a slow current rise before a voltage drop, or periodic modulation of vibration amplitude before a fault eruption). Traditional unidirectional models, unable to perceive future context, easily misjudge such gradual signs as normal fluctuations. This application captures historical dependencies using forward LSTM and models future constraints using backward LSTM, concatenating their outputs along the time step dimension to form a bidirectional temporal semantic fusion representation. This design allows the model to utilize complete past and future contextual information at every moment, enhancing its ability to discriminate non-stationary, delayed-trigger fault modes. Furthermore, combining principal component analysis for dimensionality reduction and global average pooling ensures that the semantic vector retains key temporal correlations while possessing high robustness and fixed-dimensional consistency, providing a stable and generalizable input foundation for subsequent classification functions. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a state diagnosis method based on UAV operation data provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the internal structure of a status diagnosis system based on UAV operation data provided in an embodiment of this application.

[0021] Figure 3 This is a schematic diagram of the internal structure of a status diagnostic device based on UAV operation data provided in an embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0024] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0025] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0026] The embodiment provides a status diagnosis method based on UAV operational data, such as Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps: Step 110: Acquire multi-channel UAV operation data, divide the multi-channel operation data according to time windows, and calculate the multi-channel gradient distribution histogram for each time window; construct feature vectors by using the specific distribution of a preset number of multi-channel gradient distribution histograms; use principal component analysis to reduce the dimensionality of the feature vectors to obtain the dimensionality-reduced vectors.

[0027] In some embodiments, the multi-channel operating data of the UAV includes at least three types of sensor data, including flight attitude angle, motor speed, battery voltage, accelerometer reading, and gyroscope angular velocity.

[0028] This includes acquiring multi-channel UAV operational data, dividing the data into time windows, and calculating the multi-channel gradient distribution histogram for each time window. Specifically, this includes: The continuously collected multi-channel UAV operation data is divided into sliding segments according to a preset time window. Within each time window, the first-order difference of the data sequence of each sensor channel is calculated to obtain the gradient sequence. The gradient sequence of each channel is divided into equal-width bins, and the frequency of gradient values ​​in each bin is counted to form a multi-channel gradient distribution histogram.

[0029] It should be noted that the preset length of the time window can be determined based on the sampling frequency of the UAV sensor data and the typical fault evolution rate. In practical engineering applications, when the sensor sampling frequency is 100 Hz, the time window length is usually taken as 100 ms to 500 ms, corresponding to 10 to 50 sampling points.

[0030] Furthermore, when binning the gradient sequences of each channel into equal-width bins, the number of bins can be set to 16 to 32. This range can balance the fineness of the representation with the computational cost. The bin boundaries are evenly divided according to the global minimum and maximum values ​​of the gradient values ​​to ensure that the width of each bin is consistent.

[0031] Specifically, feature vectors are constructed by utilizing the specific distribution of a predetermined number of multi-channel gradient distribution histograms, including: The multi-channel gradient distribution histograms corresponding to N consecutive time windows are concatenated according to the channel dimension to form a high-dimensional original feature matrix; The histogram of each channel is normalized so that the sum of the frequencies of all histograms is 1. All normalized histograms are concatenated in chronological order to form a one-dimensional feature vector.

[0032] It should be noted that when constructing the feature vector, the selection of N consecutive time windows covers the typical length of at least one complete fault evolution cycle. In UAV condition monitoring scenarios, N is typically between 5 and 15, corresponding to observation durations of 0.5 to 7.5 seconds.

[0033] Principal component analysis is used to reduce the dimensionality of the eigenvectors, resulting in dimensionality-reduced vectors. Specifically, this includes: Center all constructed eigenvectors and calculate the covariance matrix; Perform eigenvalue decomposition on the covariance matrix and select the eigenvectors corresponding to the top K principal components with a cumulative contribution rate of not less than 95%. Each feature vector is projected onto a predefined low-dimensional subspace consisting of K principal components to obtain a dimension-K reduced vector.

[0034] It should be noted that in principal component analysis, the K value, which has a cumulative contribution rate of no less than 95%, is determined by the eigenvalue decomposition results of the covariance matrix. In the high-dimensional feature space composed of multi-channel sensor data (such as attitude angle, rotational speed, voltage, acceleration, and angular velocity) involved in this scheme, the K value typically falls between 5 and 15. Furthermore, the K-dimensional vector after dimensionality reduction serves as the input to the bidirectional LSTM, and its structure is consistent with the time step. Each time step inputs a K-dimensional vector, corresponding to the dimensionality reduction features of a time window. The forward LSTM processes the historical sequence, and the backward LSTM processes the future sequence. The outputs of both are concatenated along the time dimension and then subjected to global average pooling to generate a fixed-length semantic vector for use by the subsequent classification function.

[0035] Step 120: Construct training samples based on the dimensionality-reduced vectors and a preset number of time steps; input the training samples into the forward LSTM and backward LSTM respectively to obtain the forward output and backward output; concatenate the forward output and backward output according to the preset number of time steps to obtain concatenated data; compress the concatenated data into a fixed-length semantic vector through global average pooling; obtain the annotation state of the semantic vector; and obtain the trained classification function based on the semantic vector and the annotation state.

[0036] Specifically, based on a preset number of time steps, the forward and backward outputs are concatenated to obtain concatenated data, including: The dimensionality-reduced vector is divided into continuous subsequences according to a preset time step T, and then input into the forward LSTM network and the backward LSTM network respectively. The forward LSTM processes the subsequences in ascending time order and outputs the hidden state at each time step. The backward LSTM processes the same subsequence in reverse time order and outputs the hidden state at each time step. The forward and backward hidden states at the corresponding time steps are concatenated along the feature dimension to form bidirectional context-aware concatenated data.

[0037] It should be noted that when dividing the dimensionality-reduced vector into continuous subsequences according to a preset number of time steps, the value of the preset time step should match the time scale of the fault evolution. In UAV operation data, the preset time step is usually set to 8 to 20, corresponding to an observation window of 0.8 seconds to 2.0 seconds (based on a 100Hz sampling frequency).

[0038] In addition, the hidden layer dimensions of both forward and backward LSTMs should ideally be set to 64 or 128, with 1 to 2 layers. Setting the hidden layer dimension to a power of 2 facilitates memory alignment for hardware accelerators (such as GPUs). The LSTM unit uses the hyperbolic tangent activation function, with the bias term enabled by default. The input dimension is equal to the dimension K of the reduced vector, and the output dimension is the hidden layer dimension.

[0039] Furthermore, the hidden states of the forward and backward LSTMs are concatenated along the feature dimension at each time step to form a bidirectional output sequence with a dimension of 2×hidden_size. The global average pooling operation performs uniform compression on the concatenated sequence in the time dimension, that is, it calculates the arithmetic mean of each feature channel over T time steps, and outputs a fixed-length semantic vector with a dimension of 2×hidden_size.

[0040] In addition, the semantic vector labels are obtained manually, with labels representing discrete fault categories, such as "normal," "abnormal motor speed," "battery voltage decay," and "gyroscope drift." The labels are encoded as integers (0, 1, 2, ...). The labeling is based on flight test records, maintenance logs, and expert diagnostic conclusions, with each sample corresponding to the final state label of a time window sequence.

[0041] Furthermore, the classification function consists of a fully connected layer and a softmax function. The input is the semantic vector output by global average pooling, and the output is the probability distribution of each fault category. Training uses the cross-entropy loss function, with Adam as the optimizer, an initial learning rate of 0.001, and a batch size of 32. An early stopping mechanism (patience=10) is used during training; training terminates when the validation set loss does not decrease for 10 consecutive rounds to prevent overfitting.

[0042] Step 130: When new UAV multi-channel operation data is obtained, obtain the new semantic vector corresponding to the new UAV multi-channel operation data; input the new semantic vector into the trained classification function to obtain the predicted state.

[0043] In some embodiments, when new UAV multi-channel operation data is obtained, a new semantic vector corresponding to the new UAV multi-channel operation data is obtained; the new semantic vector is input into a trained classification function to obtain a predicted state, specifically including: For the newly added UAV multi-channel operation data, the time window division, gradient histogram calculation, feature vector construction and principal component dimensionality reduction process are completely consistent with the training phase to generate the corresponding newly added dimensionality-reduced vector; The newly added dimensionality-reduced vector is input into the trained bidirectional LSTM and global average pooling network, and the output is a new semantic vector of fixed length; the new semantic vector is input into the classification function obtained by training with labeled states; the class with the highest probability is selected as the final predicted state.

[0044] It should be noted that the preprocessing procedure for newly added UAV multi-channel operational data must be completely consistent with that of the training phase: the time window length, the number of gradient bins, and the eigenvectors used for PCA dimensionality reduction all adopt parameters fixed during the training phase. The time window length is fixed at 100 ms to 500 ms, the number of gradient bins is fixed at 16 to 32, and the projection matrix used for principal component analysis is determined by the eigenvectors of the covariance matrix during the training phase and loaded as a constant at the inference end to ensure consistency of the feature space.

[0045] The newly added semantic vector is processed by a classification function to output the probability distribution of each fault category. The maximum probability value is used to determine the predicted state. When the maximum probability is less than 0.7, the system outputs "unknown state" instead of forcibly assigning a category to avoid misjudgment due to low confidence.

[0046] When a sensor channel fails in the new data (e.g., the standard deviation of the gradient sequence is 0 or exceeds the historical 99th percentile within three consecutive time windows), the system sets the frequency of all bins in the histogram corresponding to that channel to zero during the feature vector construction stage, while the other channels are processed normally.

[0047] In addition, this application Figure 2 This application provides a status diagnosis system based on UAV operational data. For example... Figure 2 As shown in the embodiments of this application, the system mainly includes: The dimensionality reduction module 210 is used to acquire multi-channel operation data of UAVs, divide the multi-channel operation data according to time windows, calculate the multi-channel gradient distribution histogram for each time window, construct feature vectors through the specific distribution of a preset number of multi-channel gradient distribution histograms, and use principal component analysis to perform dimensionality reduction processing on the feature vectors to obtain the dimensionality-reduced vectors.

[0048] Dimensionality reduction module 210 includes graphing units. This method is used to slide and divide continuously collected multi-channel UAV operation data according to a preset time window. Within each time window, the first-order difference of the data sequence of each sensor channel is calculated to obtain the gradient sequence. The gradient sequence of each channel is divided into equal-width bins, and the frequency of the gradient value in each bin is counted to form a multi-channel gradient distribution histogram.

[0049] The training module 220 is used to construct training samples based on the dimensionality-reduced vectors and a preset number of time steps; input the training samples into the forward LSTM and the backward LSTM respectively to obtain the forward output and the backward output; concatenate the forward output and the backward output according to the preset number of time steps to obtain concatenated data; compress the concatenated data into a fixed-length semantic vector through global average pooling; obtain the annotation state of the semantic vector; and obtain the trained classification function based on the semantic vector and the annotation state.

[0050] The prediction module 230 is used to obtain the new semantic vector corresponding to the new UAV multi-channel operation data when new UAV multi-channel operation data is obtained; and input the new semantic vector into the trained classification function to obtain the predicted state.

[0051] The above are method embodiments of this application. Based on the same inventive concept, this application also provides a status diagnostic device based on UAV operational data. Figure 3 As shown, the device includes: a processor; and a memory storing executable code thereon, which, when executed, causes the processor to perform a state diagnosis method based on UAV operating data as described in the above embodiment.

[0052] Specifically, the server acquires multi-channel UAV operation data, divides the data into time windows, and calculates the multi-channel gradient distribution histogram for each time window. Feature vectors are constructed based on the specific distribution of a preset number of multi-channel gradient distribution histograms. Principal component analysis is used to reduce the dimensionality of the feature vectors, resulting in dimensionality-reduced vectors. Training samples are constructed based on the dimensionality-reduced vectors and a preset number of time steps. The training samples are input into a forward LSTM and a backward LSTM respectively to obtain forward and backward outputs. The forward and backward outputs are concatenated according to the preset number of time steps to obtain concatenated data. Global average pooling is used to compress the concatenated data into a fixed-length semantic vector. The annotation state of the semantic vector is acquired, and a trained classification function is obtained based on the semantic vector and the annotation state. When new multi-channel UAV operation data is acquired, a new semantic vector corresponding to the new data is obtained. The new semantic vector is input into the trained classification function to obtain the predicted state.

[0053] In addition, this application embodiment also provides a non-volatile computer storage medium storing executable instructions, which, when executed, implement the above-described state diagnosis method based on UAV operation data.

[0054] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A state diagnosis method based on UAV operational data, characterized in that, The method includes: Acquire multi-channel operation data of the UAV, divide the multi-channel operation data according to time windows, and calculate the multi-channel gradient distribution histogram for each time window; construct feature vectors based on the specific distribution of a preset number of multi-channel gradient distribution histograms; use principal component analysis to reduce the dimensionality of the feature vectors to obtain the dimensionality-reduced vectors; Based on the dimensionality-reduced vectors and a preset number of time steps, training samples are constructed; the training samples are input into a forward LSTM and a backward LSTM respectively to obtain forward and backward outputs; the forward and backward outputs are concatenated according to the preset number of time steps to obtain concatenated data; the concatenated data is compressed into a fixed-length semantic vector through global average pooling; the annotation state of the semantic vector is obtained, and the trained classification function is obtained based on the semantic vector and the annotation state. When new multi-channel operation data of a drone is obtained, a new semantic vector corresponding to the new drone multi-channel operation data is obtained; the new semantic vector is input into a trained classification function to obtain the predicted state.

2. The condition diagnosis method based on UAV operational data according to claim 1, characterized in that, The multi-channel operation data of the UAV includes data from at least three of the following sensors: flight attitude angle, motor speed, battery voltage, accelerometer reading, and gyroscope angular velocity.

3. The condition diagnosis method based on UAV operational data according to claim 1, characterized in that, Acquire multi-channel UAV operational data, divide the data into time windows, and calculate the multi-channel gradient distribution histogram for each time window. Specifically, this includes: The continuously collected multi-channel UAV operation data is divided into sliding segments according to a preset time window. Within each time window, the first-order difference of the data sequence of each sensor channel is calculated to obtain the gradient sequence. The gradient sequence of each channel is divided into equal-width bins, and the frequency of gradient values ​​in each bin is counted to form a multi-channel gradient distribution histogram.

4. The status diagnosis method based on UAV operational data according to claim 1, characterized in that, Feature vectors are constructed by analyzing the specific distribution of a predetermined number of multi-channel gradient distribution histograms, specifically including: The histograms of multi-channel gradient distributions corresponding to N consecutive time windows are concatenated according to the channel dimension to form a high-dimensional original feature matrix; The histogram of each channel is normalized so that the sum of the frequencies of all histograms is 1. All normalized histograms are concatenated in chronological order to form a one-dimensional feature vector.

5. The condition diagnosis method based on UAV operational data according to claim 1, characterized in that, Principal component analysis is used to reduce the dimensionality of the eigenvectors, resulting in dimensionality-reduced vectors, specifically including: Center all constructed eigenvectors and calculate the covariance matrix; Perform eigenvalue decomposition on the covariance matrix and select the eigenvectors corresponding to the top K principal components with a cumulative contribution rate of not less than 95%. Each feature vector is projected onto a predefined low-dimensional subspace consisting of K principal components to obtain a dimension-K reduced vector.

6. The condition diagnosis method based on UAV operational data according to claim 1, characterized in that, Based on a preset number of time steps, the forward and backward outputs are concatenated to obtain the concatenated data, specifically including: The dimensionality-reduced vector is divided into continuous subsequences according to a preset time step T, and then input into the forward LSTM network and the backward LSTM network respectively. The forward LSTM processes the subsequences in ascending time order and outputs the hidden state at each time step. The backward LSTM processes the same subsequence in reverse time order and outputs the hidden state at each time step. The forward and backward hidden states at the corresponding time steps are concatenated along the feature dimension to form bidirectional context-aware concatenated data.

7. The condition diagnosis method based on UAV operational data according to claim 1, characterized in that, When new multi-channel operation data of a drone is obtained, a new semantic vector corresponding to the new drone multi-channel operation data is obtained; the new semantic vector is input into a trained classification function to obtain the predicted state, specifically including: For the newly added UAV multi-channel operation data, the time window division, gradient histogram calculation, feature vector construction and principal component dimensionality reduction process are completely consistent with the training phase to generate the corresponding newly added dimensionality-reduced vector; The newly added dimensionality-reduced vector is input into the trained bidirectional LSTM and global average pooling network, and the output is a new semantic vector of fixed length; the new semantic vector is input into the classification function obtained by training with labeled states; the class with the highest probability is selected as the final predicted state.

8. A condition diagnosis system based on UAV operational data, characterized in that, The system includes: The dimensionality reduction module is used to acquire multi-channel UAV operational data, divide the multi-channel operational data according to time windows, and calculate the multi-channel gradient distribution histogram for each time window; construct feature vectors based on the specific distribution of a preset number of multi-channel gradient distribution histograms; and perform dimensionality reduction processing on the feature vectors using principal component analysis to obtain the dimensionality-reduced vectors. The training module is used to construct training samples based on the dimensionality-reduced vectors and a preset number of time steps; input the training samples into the forward LSTM and backward LSTM respectively to obtain the forward output and backward output; concatenate the forward output and backward output according to the preset number of time steps to obtain concatenated data; compress the concatenated data into a fixed-length semantic vector through global average pooling; obtain the annotation state of the semantic vector; and obtain the trained classification function based on the semantic vector and the annotation state. The prediction module is used to obtain the new semantic vector corresponding to the new multi-channel operation data of the new UAV when new multi-channel operation data is obtained; and input the new semantic vector into the trained classification function to obtain the predicted state.

9. A condition diagnostic device based on UAV operational data, characterized in that, The device includes: processor; And a memory having executable code stored thereon, which, when executed, causes the processor to perform a state diagnosis method based on UAV operating data as described in any one of claims 1-7.

10. A non-volatile computer storage medium, characterized in that, It stores computer instructions, which, when executed, implement a state diagnosis method based on UAV operational data as described in any one of claims 1-7.

Images