Unmanned aerial vehicle anomaly analysis method and system

Abstract

The invention provides an unmanned aerial vehicle anomaly analysis method and system, and belongs to the technical field of data processing, and the method comprises the steps: receiving a monitoring data stream of an unmanned aerial vehicle sensor, and carrying out the data distribution processing of the monitoring data stream, and generating at least one sub-data stream; according to the data features of any sub-data stream, determining an anomaly diagnosis result of any sub-data stream, and calling a corresponding machine learning model from a model library according to the anomaly diagnosis result; inputting the data features of any sub-data stream into a machine learning model to obtain a model analysis result of any sub-data stream; and obtaining a comprehensive analysis result of the unmanned aerial vehicle according to the model analysis results of all the sub-data streams. According to the method, through a collaborative mechanism of data distribution processing, multi-model dynamic calling and intelligent result fusion, the accuracy and reliability of analysis results of large-batch real-time monitoring data are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to an abnormal analysis method and system for unmanned aerial vehicles (UAVs). Background Art

[0002] Under the background of the rapid development of the low-altitude economy, UAVs are increasingly widely used in fields such as logistics distribution, agricultural and forestry inspection, and emergency rescue. With the improvement of the complexity of UAV systems, their anomaly monitoring and prognostics and health management (PHM) face unprecedented challenges.

[0003] Currently, UAV PHM systems mainly use a single model to perform anomaly monitoring and diagnostic analysis on specific components or sensors. However, when dealing with large quantities of multi-source heterogeneous data, a single model cannot adapt to complex working conditions (such as compound faults and environmental mutation interference), and it is difficult to ensure the accuracy and reliability of the analysis results when facing multi-dimensional coupled abnormal patterns.

[0004] Therefore, how to improve the accuracy and reliability of the analysis results has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The present invention provides an abnormal analysis method, system, electronic device, and storage medium for UAVs to solve the defect of low accuracy and reliability of analysis results in the prior art and achieve the improvement of the accuracy and reliability of analysis results.

[0006] The present invention provides an abnormal analysis method for UAVs, including the following steps: Receiving the monitoring data stream of UAV sensors and performing data shunting processing on the monitoring data stream to generate at least one sub-data stream; Determining the abnormal diagnosis result of any one of the sub-data streams according to the data characteristics of the any one of the sub-data streams, and calling a corresponding machine learning model from a model library according to the abnormal diagnosis result; the model library includes a fault diagnosis model and a life prediction model, the fault diagnosis model is built based on the Ghost network, or the fault diagnosis model is built based on a first autoencoder and a long short-term memory network, the first autoencoder is composed of a Fourier analysis network, and the Fourier analysis network includes at least one FAN layer; the life prediction model is built based on a second autoencoder, a long short-term memory network, and a fully connected layer; Inputting the data characteristics of the any one of the sub-data streams into the machine learning model to obtain the model analysis result of the any one of the sub-data streams; Obtaining the comprehensive analysis result of the UAV according to the model analysis results of all the sub-data streams.

[0007] According to an unmanned aerial vehicle (UAV) anomaly analysis method provided by the present invention, the model analysis result includes a first diagnosis result; the step of inputting the data features of any one of the sub-data streams into the machine learning model to obtain the model analysis result of any one of the sub-data streams specifically includes: When the anomaly diagnosis result is of the fault diagnosis type, call the fault diagnosis model from the model library; Extract features from the data features through the feature extraction network in the fault diagnosis model to obtain fault diagnosis features; Classify the fault diagnosis features through the feature classification layer in the fault diagnosis model to obtain the first diagnosis result; Among them, the feature extraction network includes a feature extraction module, a convolutional layer, and a fully connected layer connected in sequence; the feature extraction module includes a plurality of Ghost Bottleneck modules connected in sequence; the Ghost Bottleneck module includes an expansion layer and a compression layer connected in sequence; the output of the Ghost Bottleneck module is the sum of the input of the expansion layer and the output of the compression layer.

[0008] According to an unmanned aerial vehicle (UAV) anomaly analysis method provided by the present invention, the model analysis result includes a second diagnosis result; the step of inputting the data features of any one of the sub-data streams into the machine learning model to obtain the model analysis result of any one of the sub-data streams specifically includes: When the anomaly diagnosis result is of the fault diagnosis type, call the fault diagnosis model from the model library; Extract periodic features from the data features through any one of the FAN layers in the first autoencoder; Stack all the FAN layers to obtain the periodic features; Perform sequence modeling on the periodic features through the gating mechanism of the long short-term memory network to obtain the first external state at the current moment; Input the first external state at the current moment into the fully connected layer to obtain the second diagnosis result.

[0009] According to an unmanned aerial vehicle (UAV) anomaly analysis method provided by the present invention, the model analysis result includes a life prediction result; the step of inputting the data features of any one of the sub-data streams into the machine learning model to obtain the model analysis result of any one of the sub-data streams specifically includes: When the anomaly diagnosis result is of the life prediction type, call the life prediction model from the model library; Perform feature extraction operations on the data features through the encoder and decoder in the second autoencoder to obtain effective features; Sequence modeling is performed on the effective features through the gating mechanism of the long short-term memory network to obtain the second external state at the current moment; The second external state at the current moment is input into the fully connected layer to obtain the life prediction result.

[0010] According to an unmanned aerial vehicle (UAV) anomaly analysis method provided by the present invention, obtaining the comprehensive analysis result of the UAV according to the model analysis results of all the sub-data streams specifically includes: When the anomaly diagnosis result is of the fault diagnosis type, weighted fusion processing is performed on all the model analysis results to obtain the comprehensive analysis result; When the anomaly diagnosis result is of the life prediction type, all the model analysis results are directly output.

[0011] According to an unmanned aerial vehicle (UAV) anomaly analysis method provided by the present invention, receiving the monitoring data stream of the UAV sensor and performing data shunt processing on the monitoring data stream to generate at least one sub-data stream specifically includes: A stream processing engine configured based on a preset communication protocol is connected to the sensor data source of the UAV to receive the sensor data stream in real time; Performing data cleaning and format standardization processing on the sensor data stream to generate a standardized monitoring data stream; Inputting the monitoring data stream into a distributed message queue for buffered storage, and segmenting the monitoring data stream to generate at least one of the sub-data streams; Wherein, the distributed message queue is used for persistent storage of the received data.

[0012] According to an unmanned aerial vehicle (UAV) anomaly analysis method provided by the present invention, the method further includes: Associating the model analysis result with the corresponding sub-data stream, and respectively pushing the model analysis result to at least two types of servers through different message topic channels.

[0013] The present invention also provides an unmanned aerial vehicle (UAV) anomaly analysis system, including the following modules: A first processing module, configured to receive the monitoring data stream of the UAV sensor, and perform data shunt processing on the monitoring data stream to generate at least one sub-data stream; A second processing module, configured to determine the anomaly diagnosis result of any one of the sub-data streams according to the data characteristics of any one of the sub-data streams, and call a corresponding machine learning model from a model library according to the anomaly diagnosis result; A third processing module, configured to input the data characteristics of any one of the sub-data streams into the machine learning model to obtain the model analysis result of any one of the sub-data streams; The fourth processing module is configured to obtain a comprehensive analysis result of the drone according to the model analysis results of all the sub-data streams.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the method for analyzing drone anomalies as described in any one of the above is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method for analyzing drone anomalies as described in any one of the above is implemented.

[0016] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, the method for analyzing drone anomalies as described in any one of the above is implemented.

[0017] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: By receiving the monitoring data stream of the drone sensor and performing data shunt processing on the monitoring data stream, the effective management and efficient processing of multi-source heterogeneous data are realized. By determining the anomaly diagnosis result of any sub-data stream according to the data characteristics of any sub-data stream, calling the corresponding machine learning model from the model library according to the anomaly diagnosis result, inputting the data characteristics of any sub-data stream into the machine learning model, and obtaining the model analysis result of any sub-data stream, the dimensionality reduction of high-dimensional data and the extraction of key features are realized, and the most suitable model is determined according to the anomaly diagnosis result to realize the accurate analysis of data characteristics. By obtaining the comprehensive analysis result of the drone according to the model analysis results of all sub-data streams, and processing all model analysis results by selecting the method of weighted fusion or direct output, the fusion and efficient output of multi-model results are realized, thereby improving the accuracy and reliability of the analysis result. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 is one of the flow diagrams of the method for analyzing drone anomalies provided by the present invention.

[0020] Figure 2 is the second flow diagram of the method for analyzing drone anomalies provided by the present invention.

[0021] Figure 3 It is the third schematic flow chart of the UAV anomaly analysis method provided by the present invention.

[0022] Figure 4 It is the fourth schematic flow chart of the UAV anomaly analysis method provided by the present invention.

[0023] Figure 5 It is the fifth schematic flow chart of the UAV anomaly analysis method provided by the present invention.

[0024] Figure 6 It is the sixth schematic flow chart of the UAV anomaly analysis method provided by the present invention.

[0025] Figure 7 It is the schematic structural diagram of the UAV anomaly analysis system provided by the present invention.

[0026] Figure 8 It is the schematic structural diagram of the electronic device provided by the present invention. Detailed implementation manners

[0027] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] It should be noted that in the description of the present invention, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or device. Without further limitations, the element defined by the phrase "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element. The orientation or positional relationship indicated by the terms "upper", "lower", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the system or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be construed as a limitation to the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0029] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, rather than to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same category, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" means at least one of the connected objects, and the character " / " generally indicates an "or" relationship between the associated objects before and after.

[0030] The following combines Figures 1-8 to describe the unmanned aerial vehicle anomaly analysis method, system, electronic device and storage medium provided by the present invention.

[0031] Figure 1 is one of the flow diagrams of the unmanned aerial vehicle anomaly analysis method provided by the present invention. As Figure 1 shown, it includes but is not limited to the following steps: Step 101: Receive the monitoring data stream of the unmanned aerial vehicle sensor, and perform data shunting processing on the monitoring data stream to generate at least one sub-data stream.

[0032] During the operation of the unmanned aerial vehicle anomaly analysis system, step 101 is the core link of data access and preprocessing. In this embodiment, the reception and shunting processing of the monitoring data stream of the unmanned aerial vehicle sensor are realized, and the core lies in solving the technical problem that the traditional PHM system cannot effectively process a large number of real-time data streams. The specific implementation process is as follows: In a possible implementation manner, Figure 2 is the second flow diagram of the unmanned aerial vehicle anomaly analysis method provided by the present invention. As Figure 2 shown, step 101 specifically includes steps 201-203: Step 201: Connect the stream processing engine configured based on the preset communication protocol to the sensor data source of the unmanned aerial vehicle, and receive the sensor data stream in real time.

[0033] Step 202: Perform data cleaning and format standardization processing on the sensor data stream to generate a standardized monitoring data stream; Step 203: Input the monitoring data stream into the distributed message queue for buffered storage, and segment the monitoring data stream to generate at least one sub-data stream; wherein, the distributed message queue is used for persistent storage of the received data.

[0034] Specifically, first, configure a stream processing engine (such as Flink) based on a preset communication protocol to establish a real-time connection with the UAV sensor data source. This design solves the problem that traditional batch processing systems cannot meet real-time requirements. Through the real-time data pulling mechanism of the stream processing engine, the timeliness of monitoring data is ensured. In actual deployment, the stream processing engine pulls the original data stream from the UAV sensor data source at fixed time intervals (such as 100 ms), effectively avoiding data backlog and delay.

[0035] Next, perform data cleaning and format standardization processing on the received original data stream. Since UAV sensor data has multi-source heterogeneous characteristics (including controller status data, environmental monitoring data, equipment operation data, etc.), in this embodiment, unified data format conversion rules are adopted to convert various types of sensor data into a standardized JSON format data stream. This processing process not only solves the problem of inconsistent multi-source data formats but also ensures data quality through data verification mechanisms (such as range checking, null value processing, etc.), providing a reliable data basis for subsequent processing.

[0036] Then, input the standardized monitoring data stream into a distributed message queue system (such as Pulsar) for buffered storage. This design solves the problem of data loss in traditional systems during network anomalies. Through the persistent storage mechanism of the message queue, even if a network interruption occurs, the data can be completely saved and continue to be processed after the network resumes. At the same time, the message queue system adopts a multi-partition storage strategy, evenly distributing the data stream to different partitions according to timestamps, significantly improving the data throughput capacity.

[0037] Finally, split the monitoring data stream to generate at least one sub-data stream, and then input each sub-data stream into the machine learning model of a subsystem for further processing, so as to meet real-time requirements and ensure the accuracy of data analysis.

[0038] Step 102: Determine the anomaly diagnosis result of any sub-data stream according to the data characteristics of any sub-data stream, so as to call the corresponding machine learning model from the model library; the model library includes a fault diagnosis model and a life prediction model. The fault diagnosis model is built based on the Ghost network, or the fault diagnosis model is built based on the first autoencoder and the long short-term memory network. The first autoencoder consists of a Fourier analysis network, and the Fourier analysis network includes at least one FAN layer; the life prediction model is built based on the second autoencoder, the long short-term memory network, and the fully connected layer.

[0039] As a key link in data processing and anomaly recognition, Step 102 performs multi-dimensional analysis on the UAV sensor sub-data stream based on the feature extraction method and the anomaly mode switching mechanism.

[0040] In specific implementation, first, data preprocessing is performed on the sub-data stream generated in step 101: a data segment of a fixed length is intercepted, and the data within the window is normalized to eliminate the dimension difference. Subsequently, multi-scale feature extraction is performed, including time-domain statistical features (maximum value, minimum value, mean value, standard deviation) and frequency-domain features (the first 5-order main frequency amplitudes are extracted based on the fast Fourier transform).

[0041] In the abnormal mode recognition stage, the system inputs the extracted feature vector into a preset threshold determination module: if the peak factor in three consecutive windows of the time-domain features exceeds the safety threshold (for example, the threshold of the acceleration sensor is set to 4.2), the abnormal diagnosis result is set to the fault diagnosis type; in other cases, the normal monitoring state is maintained. For the remaining useful life prediction task, when the time-domain mean value of the sensor data shows a monotonically decreasing trend and the decreasing rate exceeds the historical baseline (for example, the daily voltage drop rate of the lithium battery > 0.5%), the abnormal diagnosis result is set to the remaining useful life prediction type.

[0042] Furthermore, for the identified abnormal diagnosis result, the fault diagnosis or remaining useful life prediction model in the model library is dynamically called to solve the problem of insufficient adaptability of a single model in the prior art.

[0043] When the fault diagnosis type is detected, the system provides two optional model architecture implementation methods: The first is based on the Ghost network structure, and efficient feature extraction is achieved by cascading Ghost Bottleneck modules. The Ghost network (also known as the Phantom network) can generate more feature maps through simple linear transformations. While reducing the computational complexity, it can generate sufficiently flexible feature representations to meet the requirements of different tasks. In each module, the expansion layer uses a 1×1 convolution to expand the input channels by 3 times, then a 3×3 depthwise separable convolution is used to extract spatial features, and finally, the channel dimension is restored through the compression layer and connected with the input residual. This design achieves an accuracy of 94.2% in the motor bearing fault classification task, and the inference speed is 2.3 times faster than that of ResNet-50.

[0044] The second implementation method adopts a hybrid architecture of the Fourier analysis network (FAN layer) and the long short-term memory network (LSTM). The FAN layer extracts frequency-domain features through a configurable Fourier fundamental frequency filter bank, then models the temporal dependence relationship through the gating mechanism of the LSTM, and finally outputs the diagnosis result through a fully connected layer.

[0045] For the remaining useful life prediction type, the system calls a dedicated model composed of a stacked autoencoder and LSTM. The autoencoder extracts the low-dimensional degradation features of the sensor data through the encoder, and after the decoder reconstructs, it is concatenated with the temporal state vector of the LSTM, and the remaining useful life (RUL) is calculated through regression by the fully connected layer.

[0046] Step 103: Input the data features of any sub-data stream into the machine learning model to obtain the model analysis result of any sub-data stream.

[0047] Specifically, in this application, since the machine learning model can be a fault diagnosis model and a life prediction model, and among them, the fault diagnosis model includes two model architectures. Therefore, the model analysis results output by these three different architecture models are the first diagnosis result, the second diagnosis result, and the life prediction result respectively. The following embodiments will elaborate on the specific implementation manners of Step 103.

[0048] In a possible implementation manner, Figure 3 is the third flowchart of the unmanned aerial vehicle anomaly analysis method provided by the present invention. As Figure 3 shown, Step 103 includes Steps 301 - 303: Step 301: When the anomaly diagnosis result is of the fault diagnosis type, call the fault diagnosis model from the model library.

[0049] Step 302: Through the feature extraction network in the fault diagnosis model, perform feature extraction on the data features to obtain fault diagnosis features; wherein, the feature extraction network includes a feature extraction module, a convolutional layer, and a fully connected layer connected in sequence; the feature extraction module includes multiple Ghost Bottleneck modules connected in sequence; the Ghost Bottleneck module includes an expansion layer and a compression layer connected in sequence; the input of the expansion layer is added to the output of the compression layer and used as the output of the Ghost Bottleneck module.

[0050] Step 303: Through the feature classification layer in the fault diagnosis model, perform feature classification on the fault diagnosis features to obtain the first diagnosis result.

[0051] Specifically, when the anomaly diagnosis result identifier is of the fault diagnosis type, the system calls the fault diagnosis type model from the model library. In view of the strict requirements of the unmanned aerial vehicle PHM system for real-time performance and computing resources, this application adopts a lightweight model architecture based on the Ghost network. To solve the problems of large number of parameters and high inference latency existing in the deployment of traditional convolutional neural networks (CNNs) on edge computing devices, Steps 301 - 303 achieve efficient computing through a feature extraction network including multiple Ghost Bottleneck modules, and the Ghost Bottleneck module includes an expansion layer and a compression layer connected in sequence.

[0052] Both the expansion layer and the compression layer can be built through Ghost modules. The Ghost module can generate phantom features and improve the network computing efficiency. For the input data features , c, k, and q are the dimension, length, and width of the input data respectively. The Ghost module first uses the convolution kernel to perform a normal convolution operation on the input data , and the output feature map is , and are the length and width of the output features, is the dimension of the output features, and the width and height of the convolution kernel are both . Then, a linear transformation is performed on the output feature map according to the following formula to generate a series of phantom features , as follows:

[0053]

[0054] where, is the data in the output phantom features , is the -th dimension in the output feature map , represents the linear transformation function, indicates the type of linear transformation, that is, the existing feature maps, and each will be sequentially combined with s linear transformation operations to obtain s types of phantom features. The final feature dimension is . Finally, and are concatenated to obtain the output of the Ghost module.

[0055] In the Ghost Bottleneck module, the first Ghost module that the input data goes through is the expansion layer, which is used to increase the number of channels of the features. The ratio of the number of input channels to the number of output channels is the expansion ratio. The second Ghost module is used to reduce the number of channels of the features so that the number of output features is the same as the input, which is called the compression layer. Finally, the input and output are added in the way of a shortcut connection. Batch Norm and ReLU are batch normalization processing and activation function respectively. That is, the output of the expansion layer and the output of the compression layer are added together as the output of the Ghost Bottleneck module.

[0056] The fault diagnosis model includes a feature extraction network and a feature classification layer. The feature extraction network includes a feature extraction module, a convolution layer, and a fully connected layer connected in sequence. The feature extraction module includes multiple Ghost Bottleneck modules connected in sequence. For example, the feature extraction module can include 6 Ghost Bottleneck modules. The size of the convolution layer can be (input channel, output channel, convolutional window length, convolutional window width). The feature classification layer can be a softmax classifier.

[0057] Input the data features into the feature extraction network in the fault diagnosis model to obtain the fault diagnosis features output by the feature extraction network; the fault diagnosis features are deep features. Then input the fault diagnosis features into the feature classification layer in the fault diagnosis model to obtain the first diagnosis result output by the feature classification layer.

[0058] In another possible implementation, Figure 4 is the fourth schematic diagram of the process of the UAV anomaly analysis method provided by the present invention, as Figure 4 shown, step 103 includes steps 401-405: Step 401: When the anomaly diagnosis result is of the fault diagnosis type, call the fault diagnosis model from the model library.

[0059] Step 402: Perform periodic feature extraction on the data features through any one of the FAN layers in the first autoencoder.

[0060] Step 403: Stack all the FAN layers to obtain periodic features.

[0061] Step 404: Perform sequence modeling on the periodic features through the gating mechanism of the long short-term memory network to obtain the first external state at the current moment.

[0062] Step 405: Input the first external state at the current moment into the fully connected layer to obtain the second diagnosis result.

[0063] Steps 401-405 are alternative technical solutions for fault diagnosis. When the anomaly diagnosis result identifier is of the fault diagnosis type, the system calls the fault diagnosis type model from the model library. This design solves the problem for scenarios where the periodic features in UAV sensor data are significant but the time-domain noise interference is severe (such as propeller imbalance faults), and thus proposes a hybrid architecture based on the Fourier analysis network (FAN layer) and the long short-term memory network (LSTM).

[0064] Each layer of the Fourier analysis network (FAN) can effectively capture the periodic fault patterns in the data features by performing feature extraction on the data features.

[0065] For any one FAN layer, it can be expressed by the following formula: ; In the formula, L is the number of layers of the FAN layer, is the l-th FAN layer, x is the data feature, is the first learnable weight matrix of the l-th FAN layer, is the bias term of the l-th FAN layer, is the second learnable weight matrix of the l-th FAN layer, is the activation function, and || denotes concatenation along the first feature dimension; is the learnable weight matrix of the last FAN layer, is the bias vector of the last FAN layer.

[0066] Through the cosine term and the sine term , the FAN layer simulates the basis functions of the Fourier series and directly captures the periodic features in the data features; through the non-linear term , the FAN layer enhances the non-linear expression ability of the model. For the last FAN layer, its output features are calculated through the linear transformation , which simplifies the output calculation process. By stacking multiple FAN layers, multi-level periodic features in the data features can be gradually extracted, providing high-quality feature inputs for subsequent autoencoders and long short-term memory networks.

[0067] For each FAN layer, it is calculated through the cosine term, sine term and non-linear term, simulating the basis functions of the Fourier series and enhancing the non-linear expression ability respectively. Stacking all FAN layers can synthesize the multi-level periodic features extracted by each layer to generate periodic features containing rich information. This stacking operation not only retains the local periodic patterns extracted by each FAN layer, but also enhances the expression ability of global periodic features.

[0068] By stacking all FAN layers, high-quality feature inputs can be provided for subsequent autoencoders, thereby improving the accuracy and robustness of periodic features. This method significantly improves the effect of periodic feature extraction through multi-level feature extraction.

[0069] Furthermore, the autoencoder further performs dimensionality reduction through the stacking of all FAN layers, and can generate periodic feature representations, thereby removing redundant information and retaining key features.

[0070] Furthermore, the first external state at the current moment is calculated through the following formula: ; ; ; ; ; ; wherein, is the output of the forget gate at the current moment, is the output of the input gate at the current moment, is the output of the output gate at the current moment, is the candidate state at the current moment, is the internal state at the current moment, is the first external state at the current moment, is the periodic feature at the current moment, is the first external state at the previous moment, is the internal state at the previous moment, is the first weight matrix of the forget gate, is the first weight matrix of the input gate, is the first weight matrix of the output gate, is the first weight matrix of the candidate state, is the second weight matrix of the forget gate, is the second weight matrix of the input gate, is the second weight matrix of the output gate, is the second weight matrix of the candidate state, is the bias vector of the forget gate, is the bias vector of the input gate, is the bias vector of the output gate, is the bias vector of the candidate state, is the Logistic function, represents element-wise multiplication.

[0071] The gating mechanism of the long short-term memory network includes a forget gate, an input gate, and an output gate. First, the forget gate calculates based on the periodic feature of the current input and the first external state at the previous moment . Its value range is [0, 1], and it is used to quantify the retention ratio of information in the historical internal state .

[0072] Next, the input gate calculates and generates the candidate state at the same time. Subsequently, the internal state is updated through .

[0073] Finally, the output gate calculates and generates the first external state based on .

[0074] Next, a second diagnostic result is obtained through non-linear transformation and probability mapping of by a fully connected layer. The specific implementation is as follows: First, the first external state at the current moment output by the LSTM (with a dimension of 64) is input into the fully connected layer and linearly projected through the weight matrix and the bias vector to calculate the logical value .

[0075] Subsequently, the Softmax function is applied to the logical value z to generate the probability distribution of the fault categories , where , i ∈ {0, 1, …, 9} corresponds to 10 fault types, that is, the second diagnosis result is obtained. This step converts the non-linear combination of high-dimensional features into probability output through the normalized exponential function, enabling the model to quantify the confidence of different fault types.

[0076] In another possible implementation manner, Figure 5 is the fifth flow schematic diagram of the UAV anomaly analysis method provided by the present invention. As Figure 5 shown, step 103 includes steps 501 - 504: Step 501: When the anomaly diagnosis result is of the life prediction type, call the life prediction model from the model library.

[0077] Step 502: Perform feature extraction operations on the data features through the encoder and decoder in the second autoencoder to obtain effective features.

[0078] Step 503: Perform sequence modeling on the effective features through the gating mechanism of the long short-term memory network to obtain the second external state at the current moment.

[0079] Step 504: Input the second external state at the current moment into the fully connected layer to obtain the life prediction result.

[0080] When the anomaly diagnosis result is marked as the life prediction type, the system calls the life prediction model from the model library. This design solves the problem that traditional single models cannot adapt to complex working conditions. Through the dynamic model calling mechanism, it ensures that the life prediction task can select the optimal model according to actual needs. In actual deployment, the life prediction model adopts a three-level architecture of "second autoencoder + long short-term memory network + fully connected layer", where the second autoencoder is responsible for feature dimensionality reduction, the long short-term memory network is responsible for time series modeling, and the fully connected layer is responsible for regression prediction. This hierarchical design significantly improves the computational efficiency and prediction accuracy of the model.

[0081] In the feature extraction stage, the encoder in the second autoencoder is used to compress and represent the input data features. In a specific implementation, the encoder adopts a three-layer fully connected network structure. The first layer compresses the original feature dimension from 256 dimensions to 128 dimensions, the second layer further compresses it to 64 dimensions, and the third layer outputs a feature representation of 32 dimensions. This dimensionality reduction process solves the computational complexity problem brought by high-dimensional data, and at the same time retains the key feature information through non-linear transformation. The training of the encoder adopts an unsupervised method, and by minimizing the mean square error between the input data and the reconstructed data of the decoder, the reliability of the feature representation is ensured.

[0082] Next, the decoder in the second autoencoder is used to reconstruct the compressed features. The decoder adopts a network structure symmetric to the encoder to gradually restore the 32-dimensional features to the original dimension. This reconstruction process not only verifies the effectiveness of feature extraction, but also provides additional regularization constraints for the model to avoid overfitting problems.

[0083] Furthermore, the specific implementation methods of step 503 and step 504 are the same as those of step 404 and step 405, so they will not be elaborated here.

[0084] Step 104: Obtain the comprehensive analysis result of the UAV according to the model analysis results of all sub-data streams.

[0085] This embodiment realizes the fusion processing and multi-terminal distribution of the model analysis results of the UAV PHM system. The core lies in solving the technical problem that the traditional PHM system cannot dynamically adjust the result processing method according to the abnormal diagnosis results. The specific implementation process is as follows: In a possible implementation manner, Figure 6 is the sixth flowchart of the UAV abnormal analysis method provided by the present invention. As Figure 6 shown, step 104 specifically includes steps 601-602: Step 601: When the abnormal diagnosis result is of the fault diagnosis type, perform weighted fusion processing on all model analysis results to obtain the comprehensive analysis result.

[0086] Step 602: When the abnormal diagnosis result is of the life prediction type, directly output all model analysis results.

[0087] Specifically, for the multi-model calculation results of the fault diagnosis type, this application adopts a fusion strategy of dynamically adjusting weights based on confidence to solve the problem of insufficient adaptability of the traditional fixed-weight fusion mechanism to the real-time performance of the model.

[0088] In specific implementation, when the abnormal diagnosis result is marked as the fault diagnosis type, the system first obtains the initial diagnosis results output by the fault diagnosis models of each subsystem and their corresponding confidence indicators: for the fault diagnosis model based on the Ghost network, the confidence is calculated through the entropy value of the probability distribution output by Softmax; while for the fault diagnosis model of the hybrid model based on FAN-LSTM, the prediction variance is calculated through Monte Carlo Dropout sampling as the confidence. Further, the system dynamically generates weight coefficients according to the confidence.

[0089] For example, in the fault diagnosis scenario of the power system of a logistics UAV, when the confidence levels of the fault diagnosis models of four subsystems are 0.92 (fault diagnosis model based on the Ghost network), 0.90 (fault diagnosis model based on the Ghost network), 0.86 (fault diagnosis model of the hybrid model based on FAN-LSTM), and 0.85 (fault diagnosis model of the hybrid model based on FAN-LSTM) respectively, the weight allocation can be 0.3, 0.28, 0.22, 0.2, and the finally obtained comprehensive analysis result is y =0.3× y 1 + 0.28× y 2 + 0.22× y 3 + 0.2× y 4, where, y 1, y 2, y 3, y 4 are the initial diagnosis results of the fault diagnosis models of the four subsystems respectively. For the case where the confidence difference is less than the threshold, the system can switch to the average weighting mode to ensure decision robustness.

[0090] Through this dynamic weight allocation strategy, both the reliability of the model and the importance of the subsystems are considered, and finally a comprehensive analysis result is generated.

[0091] When the abnormal diagnosis result is marked as the life prediction type, the system directly outputs the prediction result of the life prediction model. This design simplifies the processing flow of the life prediction task, avoids unnecessary fusion calculations, and improves the system efficiency. In actual deployment, the life prediction result is output in the form of the remaining useful life (RUL), and is accompanied by a confidence interval (such as "remaining life: 120 hours, confidence: 95%"), providing a more intuitive decision-making basis for users.

[0092] In a possible implementation manner, after step 104, the method further includes the following steps: Associate the model analysis result with the corresponding sub-data stream, and push it to at least two types of servers through different message topic channels respectively.

[0093] This embodiment realizes the association and multi - terminal distribution of the result data of the UAV PHM system. The core lies in solving the technical problems that the traditional PHM system cannot effectively manage the correlation of multi - source data and the reliability of distribution. The specific implementation process is as follows: Associate the model analysis results with the corresponding sub - data streams, that is, associate the first diagnostic and prediction results, the second diagnostic and prediction results, and the life prediction results with the corresponding sub - data streams, and push them to at least two types of servers through different message topic channels. The types of servers at least include the ground user server type and the UAV server type. The system first performs spatio - temporal alignment on the first diagnostic and prediction results, the second diagnostic and prediction results, and the life prediction results with the corresponding sub - data streams to ensure that each calculation result can be traced back to its corresponding original monitoring data. In actual implementation, the system adds a unified timestamp (accurate to milliseconds) and subsystem identifier to each calculation result and sub - data stream to ensure the consistency of data in time and space.

[0094] Next, the system distributes the associated data to at least two types of servers through different message topic channels. This design solves the reliability problem of the traditional single - point distribution system. Through the multi - channel distribution mechanism, it ensures that even if a certain server fails, other servers can still receive data normally. In specific implementation, the system establishes multiple independent message topic channels, including the first message topic channel for real - time monitoring, the second message topic channel for decision - making analysis, and the third message topic channel as a backup. The real - time monitoring server receives the original monitoring data stream and calculation results through the first message topic channel to support real - time visual display; the decision - making analysis server receives the calculation results and diagnostic reports through the second message topic channel to support offline data analysis; the backup channel temporarily takes over the data distribution task in case of network anomalies to ensure the continuity and reliability of data distribution.

[0095] Refer to Figure 7 , Figure 7 which is the structural schematic diagram of the UAV anomaly analysis system provided by the present invention. The system includes: A first processing module, configured to receive the monitoring data stream of the UAV sensor and perform data shunting processing on the monitoring data stream to generate at least one sub - data stream; A second processing module, configured to determine the anomaly diagnosis result of any sub - data stream according to the data characteristics of any sub - data stream, and call the corresponding machine learning model from the model library according to the anomaly diagnosis result. The model library includes a fault diagnosis model and a life prediction model. The fault diagnosis model is built based on the Ghost network, or the fault diagnosis model is built based on the first auto - encoder and the long - short - term memory network. The first auto - encoder consists of a Fourier analysis network, and the Fourier analysis network includes at least one FAN layer. The life prediction model is built based on the second auto - encoder, the long - short - term memory network, and the fully - connected layer; The third processing module is configured to input the data features of any sub-data stream into a machine learning model to obtain the model analysis result of any sub-data stream; The fourth processing module is configured to obtain the comprehensive analysis result of the drone according to the model analysis results of all sub-data streams.

[0096] In a possible implementation manner, the third processing module is further configured to: When the anomaly diagnosis result is of the fault diagnosis type, call a fault diagnosis model from the model library; When the anomaly diagnosis result is of the fault diagnosis type, call a fault diagnosis model from the model library; Extract features from the data features through the feature extraction network in the fault diagnosis model to obtain fault diagnosis features; Classify the fault diagnosis features through the feature classification layer in the fault diagnosis model to obtain the first diagnosis result; Wherein, the feature extraction network includes a feature extraction module, a convolutional layer, and a fully connected layer connected in sequence; the feature extraction module includes a plurality of Ghost Bottleneck modules connected in sequence; the Ghost Bottleneck module includes an expansion layer and a compression layer connected in sequence; the output of the expansion layer and the input of the compression layer are added together as the output of the Ghost Bottleneck module.

[0097] In a possible implementation manner, the third processing module is further configured to: When the anomaly diagnosis result is of the fault diagnosis type, call a fault diagnosis model from the model library; Extract periodic features from the data features through any one of the FAN layers in the first autoencoder; Stack all the FAN layers to obtain periodic features; Perform sequence modeling on the periodic features through the gating mechanism of the long short-term memory network to obtain the first external state at the current moment; Input the first external state at the current moment into the fully connected layer to obtain the second diagnosis result.

[0098] In a possible implementation manner, the third processing module is further configured to: When the anomaly diagnosis result is of the life prediction type, call a life prediction model from the model library; Perform feature extraction operations on the data features through the encoder and decoder in the second autoencoder to obtain effective features; Perform sequence modeling on the effective features through the gating mechanism of the long short-term memory network to obtain the second external state at the current moment; Input the second external state at the current moment into the fully connected layer to obtain the life prediction result.

[0099] In a possible implementation manner, the fourth processing module is further configured to: When the abnormal diagnosis result is of the fault diagnosis type, perform weighted fusion processing on all model analysis results to obtain a comprehensive analysis result; When the abnormal diagnosis result is of the life prediction type, directly output all model analysis results.

[0100] In a possible implementation manner, the first processing module is further configured to: Connect the flow processing engine configured based on the preset communication protocol to the sensor data source of the drone, and receive the sensor data stream in real time; Perform data cleaning and format standardization processing on the sensor data stream to generate a standardized monitoring data stream; Input the monitoring data stream into the distributed message queue for buffered storage, and segment the monitoring data stream to generate at least one sub-data stream; Wherein, the distributed message queue is used for persistent storage of the received data.

[0101] In a possible implementation manner, the system further includes a fifth processing module, configured to associate the model analysis result with the corresponding sub-data stream, and push them to at least two types of servers through different message topic channels.

[0102] It should be noted that the drone abnormal analysis system provided by the present invention can execute the drone abnormal analysis method of any of the above embodiments during specific operation, and this embodiment will not be elaborated herein.

[0103] Figure 8 is a schematic structural diagram of the electronic device provided by the present invention, as Figure 8As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840. Among them, the processor 810, the communications interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call logic instructions in the memory 830 to execute an unmanned aerial vehicle (UAV) anomaly analysis method, which includes: receiving a monitoring data stream from a UAV sensor, and performing data shunting processing on the monitoring data stream to generate at least one sub-data stream; determining an anomaly diagnosis result of any sub-data stream according to the data characteristics of any sub-data stream, so as to call a corresponding machine learning model from a model library according to the anomaly diagnosis result; the model library includes a fault diagnosis model and a life prediction model, the fault diagnosis model is built based on a Ghost network, or the fault diagnosis model is built based on a first autoencoder and a long short-term memory network, the first autoencoder consists of a Fourier analysis network, and the Fourier analysis network includes at least one FAN layer; the life prediction model is built based on a second autoencoder, a long short-term memory network, and a fully connected layer; inputting the data characteristics of any sub-data stream into the machine learning model to obtain a model analysis result of any sub-data stream; obtaining a comprehensive analysis result of the UAV according to the model analysis results of all sub-data streams.

[0104] In addition, when the logic instructions in the above-mentioned memory 830 are implemented in the form of software functional units and sold or used as an independent product, they may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, may be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0105] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the drone anomaly analysis method provided by each of the above embodiments. The method includes: receiving a monitoring data stream of a drone sensor, and performing data splitting processing on the monitoring data stream to generate at least one sub-data stream; determining an anomaly diagnosis result of any one of the sub-data streams according to the data characteristics of any one of the sub-data streams, so as to call a corresponding machine learning model from a model library according to the anomaly diagnosis result; the model library includes a fault diagnosis model and a life prediction model. The fault diagnosis model is built based on the Ghost network, or the fault diagnosis model is built based on a first autoencoder and a long short-term memory network. The first autoencoder is composed of a Fourier analysis network, and the Fourier analysis network includes at least one FAN layer; the life prediction model is built based on a second autoencoder, a long short-term memory network, and a fully connected layer; inputting the data characteristics of any one of the sub-data streams into the machine learning model to obtain a model analysis result of any one of the sub-data streams; obtaining a comprehensive analysis result of the drone according to the model analysis results of all the sub-data streams.

[0106] On another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is configured to execute the drone anomaly analysis method provided by each of the above embodiments. The method includes: receiving a monitoring data stream of a drone sensor, and performing data splitting processing on the monitoring data stream to generate at least one sub-data stream; determining an anomaly diagnosis result of any one of the sub-data streams according to the data characteristics of any one of the sub-data streams, so as to call a corresponding machine learning model from a model library according to the anomaly diagnosis result; the model library includes a fault diagnosis model and a life prediction model. The fault diagnosis model is built based on the Ghost network, or the fault diagnosis model is built based on a first autoencoder and a long short-term memory network. The first autoencoder is composed of a Fourier analysis network, and the Fourier analysis network includes at least one FAN layer; the life prediction model is built based on a second autoencoder, a long short-term memory network, and a fully connected layer; inputting the data characteristics of any one of the sub-data streams into the machine learning model to obtain a model analysis result of any one of the sub-data streams; obtaining a comprehensive analysis result of the drone according to the model analysis results of all the sub-data streams.

[0107] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative work.

[0108] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiments.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments or equivalently replace some of the technical features. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. A method for analyzing abnormal conditions of an unmanned aerial vehicle, characterized in that, Including: Receiving the monitoring data stream of the drone sensor, and performing data shunting processing on the monitoring data stream to generate at least one sub-data stream; Determining the abnormal diagnosis result of any one of the sub-data streams according to the data characteristics of any one of the sub-data streams, so as to call the corresponding machine learning model from the model library according to the abnormal diagnosis result; the model library includes a fault diagnosis model and a life prediction model, the fault diagnosis model is built based on the Ghost network, or the fault diagnosis model is built based on the first autoencoder and the long short-term memory network, the first autoencoder consists of a Fourier analysis network, and the Fourier analysis network includes at least one FAN layer; the life prediction model is built based on the second autoencoder, the long short-term memory network and the fully connected layer; Inputting the data characteristics of any one of the sub-data streams into the machine learning model to obtain the model analysis result of any one of the sub-data streams; Obtaining the comprehensive analysis result of the drone according to the model analysis results of all the sub-data streams.

2. The method for analyzing drone anomalies according to claim 1, wherein The model analysis result includes a first diagnosis result; the step of inputting the data characteristics of any one of the sub-data streams into the machine learning model to obtain the model analysis result of any one of the sub-data streams specifically includes: When the abnormal diagnosis result is of the fault diagnosis type, calling the fault diagnosis model from the model library; Extracting features from the data characteristics through the feature extraction network in the fault diagnosis model to obtain fault diagnosis features; Classifying the fault diagnosis features through the feature classification layer in the fault diagnosis model to obtain the first diagnosis result; Wherein, the feature extraction network includes a feature extraction module, a convolutional layer and a fully connected layer connected in sequence; the feature extraction module includes a plurality of Ghost Bottleneck modules connected in sequence; the Ghost Bottleneck module includes an expansion layer and a compression layer connected in sequence; the output of the expansion layer and the input of the compression layer are added to be used as the output of the Ghost Bottleneck module.

3. The method for analyzing drone anomalies according to claim 1, wherein The model analysis result includes a second diagnosis result; the step of inputting the data characteristics of any one of the sub-data streams into the machine learning model to obtain the model analysis result of any one of the sub-data streams specifically includes: When the abnormal diagnosis result is of the fault diagnosis type, calling the fault diagnosis model from the model library; Performing periodic feature extraction on the data characteristics through any one of the FAN layers in the first autoencoder; Stacking all the FAN layers to obtain the periodic features; Performing sequence modeling on the periodic features through the gating mechanism of the long short-term memory network to obtain the first external state at the current moment; Inputting the first external state at the current moment into the fully connected layer to obtain the second diagnosis result.

4. The method for analyzing drone anomalies according to claim 1, wherein, The model analysis result includes a life prediction result; the step of inputting the data characteristics of any one of the sub-data streams into the machine learning model to obtain the model analysis result of any one of the sub-data streams specifically includes: When the abnormal diagnosis result is of the life prediction type, call the life prediction model from the model library; Perform feature extraction operations on the data features through the encoder and decoder in the second autoencoder to obtain effective features; Perform sequence modeling on the effective features through the gating mechanism of the long short-term memory network to obtain the second external state at the current moment; Input the second external state at the current moment into the fully connected layer to obtain the life prediction result.

5. The method for analyzing drone anomalies according to claim 1, wherein The obtaining of the comprehensive analysis result of the UAV according to the model analysis results of all the sub-data streams specifically includes: When the abnormal diagnosis result is of the fault diagnosis type, perform weighted fusion processing on all the model analysis results to obtain the comprehensive analysis result; When the abnormal diagnosis result is of the life prediction type, directly output all the model analysis results.

6. The method for analyzing drone anomalies according to claim 1, characterized in that, The receiving of the monitoring data stream of the UAV sensor and the data splitting process of the monitoring data stream to generate at least one sub-data stream specifically includes: The stream processing engine configured based on the preset communication protocol is connected to the sensor data source of the UAV to receive the sensor data stream in real time; Perform data cleaning and format standardization processing on the sensor data stream to generate a standardized monitoring data stream; Input the monitoring data stream into the distributed message queue for buffered storage, and split the monitoring data stream to generate at least one of the sub-data streams; Wherein, the distributed message queue is used for persistent storage of the received data.

7. The method for analyzing drone anomalies according to claim 1, characterized in that, The method further includes: Associate the model analysis result with the corresponding sub-data stream, and push them to at least two types of servers through different message topic channels.

8. An unmanned aerial vehicle anomaly analysis system, characterized in that, Including: A first processing module, configured to receive the monitoring data stream of the UAV sensor and perform data splitting processing on the monitoring data stream to generate at least one sub-data stream; A second processing module, configured to determine the abnormal diagnosis result of any one of the sub-data streams according to the data features of any one of the sub-data streams, and call the corresponding machine learning model from the model library according to the abnormal diagnosis result; the model library includes a fault diagnosis model and a life prediction model, the fault diagnosis model is built based on the Ghost network, or the fault diagnosis model is built based on the first autoencoder and the long short-term memory network, the first autoencoder consists of a Fourier analysis network, and the Fourier analysis network includes at least one FAN layer; the life prediction model is built based on the second autoencoder, the long short-term memory network, and the fully connected layer; A third processing module, configured to input the data features of any one of the sub-data streams into the machine learning model to obtain the model analysis result of any one of the sub-data streams; A fourth processing module, configured to obtain the comprehensive analysis result of the UAV according to the model analysis results of all the sub-data streams.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the UAV abnormal analysis method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the UAV abnormal analysis method according to any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the unmanned aerial vehicle anomaly analysis method according to any one of claims 1 to 7.

Images