Data-driven unmanned aerial vehicle wing dynamic load anomaly identification method and system

Abstract

The invention discloses a data-driven unmanned aerial vehicle wing dynamic load anomaly identification method and system, and relates to the technical field of data processing. The method comprises the following steps: acquiring reference load interval distribution, and monitoring flight load distribution in real time; and when the monitored flight load distribution exceeds the reference load interval distribution, a historical environment data sequence is called. And on the basis of the historical environment data sequence, predicting the abnormal state of the target cargo to obtain abnormal state probability distribution. And in combination with the historical environment data sequence and the abnormal state probability distribution, generating influence load distribution, calculating the similarity between the distribution and actual flight load distribution, and calculating a wing load anomaly rate as a load anomaly identification result. Through a data driving mode, load abnormity caused by environment interference, cargo state change and unmanned aerial vehicle body faults is effectively distinguished, identification accuracy and diagnosis pertinence are improved, and reliable technical support is provided for unmanned aerial vehicle freight safety.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a data-driven method and system for identifying dynamic load anomalies on the wings of unmanned aerial vehicles (UAVs). Background Technology

[0002] When drones carry cargo to perform missions, the wings, as a key load-bearing structure, directly affect the stability and safety of flight due to their dynamic load status. The wings may experience abnormal loads due to factors such as gusts of wind, structural damage, or changes in the cargo's condition. If these abnormalities are not identified in time, they can easily lead to loss of flight attitude control or even serious accidents.

[0003] Currently, most technologies for identifying abnormal payloads in UAVs rely on pre-set fixed thresholds or judgment methods based on simple physical models. Load data from the wings is collected by sensors and compared to safety thresholds; if the thresholds are exceeded, an anomaly is identified. However, in actual flight, payload signals are subject to various complex interferences, such as sudden gusts of wind, unstable airflow, and sensor noise. These interference signals are easily misinterpreted by existing methods as payload anomalies caused by structural damage or control malfunctions, leading to frequent false alarms and reducing the accuracy and reliability of the identification. Furthermore, existing methods typically only provide a binary judgment of abnormality versus normality, lacking the ability to analyze and quantify the causes of anomalies. They cannot accurately distinguish whether an anomaly stems from external environmental interference or structural problems within the aircraft itself, making it difficult to provide effective subsequent decision support. Summary of the Invention

[0004] This invention addresses the technical problems of existing technologies where load anomaly identification is susceptible to environmental interference, has low accuracy, and cannot effectively distinguish the root cause of anomalies. It provides a data-driven method and system for identifying dynamic load anomalies on UAV wings.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] In a first aspect, the present invention provides a data-driven method for identifying dynamic load anomalies on the wings of unmanned aerial vehicles (UAVs), comprising:

[0007] Obtain the baseline load range distribution of the wing of the UAV carrying the target cargo, and monitor the flight load distribution through sensors during flight;

[0008] When the flight load distribution exceeds the reference load range distribution, historical environmental data sequences are retrieved;

[0009] Based on the historical environmental data sequence, the abnormal state of the target cargo is predicted to obtain the probability distribution of the abnormal state.

[0010] Based on the historical environmental data sequence and the probability distribution of abnormal states, the distribution of influencing loads is analyzed, the similarity with the flight load distribution is calculated, and the wing load anomaly rate is calculated as the load anomaly identification result.

[0011] Secondly, the present invention provides a data-driven unmanned aerial vehicle (UAV) wing dynamic load anomaly identification system, comprising:

[0012] The load monitoring module is used to obtain the baseline load range distribution of the wing of the UAV carrying the target cargo. During flight, the load distribution is monitored by sensors.

[0013] The data retrieval module is used to retrieve historical environmental data sequences when the flight load distribution exceeds the reference load range distribution.

[0014] An anomaly prediction module is used to predict the abnormal state of the target cargo based on the historical environmental data sequence and obtain the probability distribution of the abnormal state.

[0015] The anomaly identification module is used to analyze and obtain the influence load distribution based on the historical environmental data sequence and the probability distribution of the abnormal state, calculate the similarity with the flight load distribution, and calculate the wing load anomaly rate as the load anomaly identification result.

[0016] The beneficial effects of this invention are:

[0017] Compared to existing technologies, this invention firstly establishes a dynamic baseline load range distribution, replacing the traditional fixed threshold method, thus enhancing adaptability to different flight conditions and reducing misjudgments caused by changes in conditions. Secondly, it introduces a cargo anomaly state prediction step based on historical environmental data, correlating load changes with the cargo's internal state to effectively distinguish between two different anomaly sources: external environmental interference and cargo's own abnormal state. Thirdly, by fusing environmental data and state prediction results to calculate the theoretical influence on load distribution and comparing similarity with measured data, it achieves a quantitative assessment of the degree of anomaly, making the identification results more accurate and objective. Finally, the entire solution is data-driven at its core, enabling continuous learning and optimization, providing a more reliable and interpretable decision-making basis for the UAV flight control system, and improving the flight safety and mission reliability of cargo UAVs in complex environments. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the data-driven method for identifying dynamic load anomalies on UAV wings provided by this invention;

[0019] Figure 2 A schematic diagram of the structure of the data-driven UAV wing dynamic load anomaly identification system provided by the present invention.

[0020] In the attached diagram, the components represented by each number are as follows:

[0021] Load monitoring module 11, data retrieval module 12, anomaly prediction module 13, anomaly identification module 14. Detailed Implementation

[0022] Example 1, as Figure 1 As shown, this embodiment of the invention provides a data-driven method for identifying dynamic load anomalies on the wings of unmanned aerial vehicles (UAVs), including:

[0023] It should be noted that, in the following embodiments of the present invention, unless otherwise explicitly defined, the term "load" specifically refers to the pressure acting on a unit area of ​​the UAV wing, i.e., pressure. The term "load value" refers to the numerical value of the pressure. The SI unit of pressure is Pascal (Pa), but kilopascal (kPa) can also be used in practical applications. Correspondingly, "load distribution" refers to the spatial distribution of pressure on the wing surface.

[0024] S10: Obtain the baseline load range distribution of the wing of the UAV carrying the target cargo, and monitor the flight load distribution through sensors during flight;

[0025] First, the baseline load range distribution of the drone's wings after it carries the target cargo is obtained. The target cargo refers to the specific transported item carried by the drone in this flight mission. The baseline load range distribution refers to the reasonable fluctuation range of the load borne by multiple key monitoring areas of the wing under normal flight conditions after the drone carries the specific target cargo. This baseline load range distribution is not a single fixed value, but rather a load upper and lower limit range determined for each monitoring area of ​​the wing based on a large amount of historical normal flight test data through statistical analysis methods.

[0026] Specifically, the main function of the baseline load range distribution is to serve as a dynamic reference for subsequent judgment of whether real-time flight loads are abnormal. This baseline load range distribution takes into account the inherent effects of specific cargo mass, center of gravity distribution, and standard flight attitude on wing loads, thus more accurately reflecting load characteristics under normal conditions and providing a scientific basis for comparison in subsequent anomaly monitoring.

[0027] Specifically, the baseline load range distribution of the drone's wings after it carries the target cargo is obtained, and the flight load distribution is monitored through sensors during flight, including:

[0028] Based on the load test data from the drone flight, the baseline load range distribution of the drone's wings after carrying the target cargo is obtained;

[0029] During the flight of the drone, the distribution of flight loads is monitored and acquired through sensors configured on the wings. The distribution of flight loads includes loads in multiple wing regions.

[0030] First, based on the load test data recorded by the UAV during historical or experimental flights, the baseline load range distribution of the UAV's wings after carrying the target cargo is obtained through processing and calculation.

[0031] Specifically, based on the load test data from the UAV flight, the baseline load range distribution of the UAV's wings after carrying the target cargo is obtained, including:

[0032] Based on the wing loads tested during normal flight of similar drones carrying similar target cargo, obtain the test load distribution set;

[0033] Based on the endpoint values ​​of multiple wing regions within the test load distribution set, a baseline load interval distribution is constructed.

[0034] First, wing load data recorded during multiple known normal flights of similar UAVs carrying similar target cargo were collected and organized to obtain a test load distribution set. This test load distribution set consists of multiple sets of historical flight data samples. Each set of samples fully records the load values ​​of multiple predefined key monitoring areas on the wing at the same time or time period during a normal flight mission. It is important to note that the collection of this test load distribution set is premised on ensuring a stable flight process without abnormal environmental interference or changes in cargo status, so as to ensure that the data can reflect ideal baseline operating conditions.

[0035] Secondly, based on the test load distribution set, a baseline load range distribution for the wing is constructed. Specifically, for each monitoring area on the wing, all load values ​​corresponding to that area are extracted from all samples in the test load distribution set. Statistical analysis of all load values ​​for that area determines the upper and lower limits of its value range. These upper and lower limits form the reasonable load range for that monitoring area. This process is repeated for all monitoring areas to determine an independent load range for each area. Finally, the complete set of load ranges for all monitoring areas is defined as the baseline load range distribution for the wing after the UAV carries the target cargo. This baseline load range distribution characterizes the allowable range of load fluctuations in various parts of the wing under normal, stable conditions.

[0036] For example, suppose the wing is divided into four monitoring areas: left wingtip, left wing middle, right wing middle, and right wingtip. From 100 normal tests, the load value of the left wingtip area is extracted. Its minimum value is 85 kPa and its maximum value is 115 kPa. Then the reference pressure range of the left wingtip area can be defined as [85 kPa, 115 kPa].

[0037] Furthermore, during the actual flight mission of the UAV, multiple sensors pre-configured on the wing structure continuously and synchronously monitor and collect real-time force data at each monitoring point. The real-time data acquired by each sensor, after processing, collectively constitutes the flight load distribution. This flight load distribution is a multi-dimensional dataset containing real-time load values ​​for multiple wing regions of the UAV at the current moment. The purpose of monitoring the flight load distribution is to obtain the actual force condition of the wing under flight conditions in real time and compare it with the aforementioned baseline load range distribution, thus serving as the primary data basis for determining whether the flight condition is abnormal.

[0038] S20: When the flight load distribution exceeds the reference load range distribution, retrieve the historical environmental data sequence;

[0039] When the real-time monitored flight load distribution exceeds the pre-established baseline load range, it indicates that the current load state of the wing has deviated from the normal and reasonable range, posing a potential risk of anomalies. However, abnormal load fluctuations may be caused by two main factors: first, sudden changes in the external flight environment, such as encountering unexpected strong gusts or turbulence; and second, abnormalities in the condition of the onboard cargo itself, such as cargo displacement, loosening, or changes in shape.

[0040] To effectively distinguish the root causes of anomalies and to reasonably interpret and deeply analyze load changes, it is necessary to obtain environmental information related to the events before and after the load anomaly, i.e., to retrieve historical environmental data sequences. These historical environmental data sequences record the environmental conditions experienced by the aircraft before the load anomaly event, particularly airflow data, providing a crucial data foundation for subsequent steps to determine whether the anomaly was primarily caused by environmental factors and to predict the cargo's condition.

[0041] Specifically, when the flight load distribution exceeds the baseline load range, historical environmental data sequences are retrieved, including:

[0042] Determine whether any wing region within the flight load distribution exceeds the corresponding reference load range;

[0043] If yes, then process and obtain the historical environmental data sequence within the preset time range; otherwise, continue monitoring the flight payload distribution. The historical environmental data sequence includes environmental airflow data within the preset time range.

[0044] First, the flight load distribution acquired through real-time monitoring is analyzed region by region. Specifically, the real-time load value of each wing monitoring area in the flight load distribution is compared with the corresponding reasonable load range predefined in the baseline load range distribution for that area. The judgment criterion is: whether there is any wing area in the flight load distribution whose real-time load value exceeds the upper limit of the baseline load range set for that area, or falls below its lower limit.

[0045] If the comparison results show that the load value of at least one wing region exceeds its corresponding baseline load range, it is determined that the current flight state has triggered an abnormal load condition. The system will immediately initiate a data retrieval process to process and obtain historical environmental data sequences recorded within a preset time range. The preset time range can be configured according to the UAV's flight speed and system response requirements, for example, retrieving environmental data from tens of seconds to several minutes before the abnormality was triggered.

[0046] For example, if the reference load range for the left wingtip region is [85kPa, 115kPa], and the real-time monitoring value is 125kPa, then the load in this region is determined to be outside the reference range, and historical environmental data sequences within a preset time range need to be obtained.

[0047] The historical environmental data sequence mainly includes environmental airflow data continuously collected and recorded by airborne environmental sensors within this time period. This environmental airflow data can specifically include multi-dimensional parameters such as the direction, angle, speed, and turbulence intensity of the airflow relative to the UAV body, accurately depicting the external aerodynamic environment experienced by the aircraft before the occurrence of load anomalies.

[0048] Furthermore, if the comparison results show that the load values ​​of all wing regions in the flight load distribution are within their respective reference load ranges, the current flight status is determined to be normal, and no abnormal conditions have been triggered. In this case, the system will not retrieve historical environmental data sequences, but will instead return to and continue real-time monitoring of the flight load distribution, continuously collecting data and periodically comparing and judging it.

[0049] S30: Based on the historical environmental data sequence, predict the abnormal state of the target cargo and obtain the probability distribution of the abnormal state;

[0050] After acquiring historical environmental data sequences, in order to deeply diagnose the underlying causes of load anomalies, it is necessary to intelligently infer the state of the airborne target cargo based on these historical environmental data sequences. Specifically, there is a correlation between airflow disturbance information included in the historical environmental data sequences and possible state changes of the cargo. For example, continuous turbulence in a specific direction may cause stress concentration at fixed points on the cargo, thereby inducing cargo displacement or attitude changes.

[0051] Specifically, based on the historical environmental data sequence, anomaly prediction of the target cargo is performed to obtain anomaly probability distribution, including:

[0052] Obtain a cargo status predictor, wherein the cargo status predictor includes multiple cargo anomaly prediction branches corresponding to various cargo anomaly states.

[0053] The historical environmental data sequence is input into the cargo status predictor, which outputs the probability of various cargo abnormal states and obtains the probability distribution of abnormal states.

[0054] First, a pre-built and trained cargo state predictor is obtained. This cargo state predictor is a machine learning model whose structure includes multiple parallel cargo anomaly prediction branches. Each cargo anomaly prediction branch specifically corresponds to and is responsible for predicting a particular cargo anomaly state. Specifically, a cargo anomaly state refers to the undesirable shape or positional changes that may occur to the cargo carried by the cargo drone during flight, such as the cargo slipping, tilting, rotating, packaging damage, or loosening of fasteners, either entirely or partially. There is a learnable statistical correlation between specific historical airflow environmental sequence patterns and the occurrence of specific cargo anomalies; for example, continuous and severe lateral turbulence may be associated with a high risk of cargo tipping over laterally.

[0055] Specifically, acquiring the cargo status predictor includes:

[0056] Based on test data of similar drones carrying similar cargo, a set of sample flight environment data sequences was collected, and the probability of cargo exhibiting various abnormal states under different sample flight environment data sequences was obtained. Multiple sample abnormal state probability sets were then labeled.

[0057] Based on machine learning, multiple cargo anomaly prediction branches are constructed to correspond to various cargo anomaly states.

[0058] Using the sample flight environment data sequence set as training input, and using the multiple sample abnormal state probability sets as supervision labels, supervised training and testing are performed on multiple cargo anomaly prediction branches. After convergence, a cargo state predictor is obtained.

[0059] First, based on test data of similar drones carrying similar cargo, multiple sample flight environment data sequences are systematically collected, forming a sample flight environment data sequence set. Each sample flight environment data sequence corresponds to historical environmental observation data within a specific time window. Simultaneously, through experimental observation, sensor monitoring, or post-event analysis, the probability of the cargo actually experiencing various preset abnormal states during the flight period corresponding to each sample flight environment data sequence is determined. This is then quantitatively evaluated in probabilistic form, for example, by assigning a probability vector to each sample flight environment data sequence through expert annotation, statistical analysis of repeated trials, or simulation calculations based on physical models. This probability vector indicates the probability value of each abnormal state of the cargo under that specific environmental sequence. A separate probability label value is compiled for each cargo abnormal state, thereby obtaining multiple sample abnormal state probability sets, each sample abnormal state probability set specifically corresponding to one abnormal state.

[0060] Among them, the preset abnormal state refers to the undesirable physical state change of the cargo carried by the cargo drone during flight due to environmental disturbances, fixation failure or other factors. Examples include linear slippage of the cargo as a whole relative to the cargo hold, attitude instability of the cargo tilting or rotating around its center of mass, relative displacement of internal components of the cargo due to vibration, and damage or loosening of the outer packaging or binding structure of the cargo.

[0061] Secondly, based on a machine learning framework, the basic structure of a cargo state predictor is constructed. Specifically, a machine learning model architecture with multiple parallel outputs is designed, where each output branch is defined as a cargo anomaly prediction branch, and each cargo anomaly prediction branch explicitly corresponds to a cargo anomaly state to be predicted.

[0062] Furthermore, the sample flight environment data sequence set is used as unified training input data. During training, the previously obtained multiple sample abnormal state probability sets are used as supervision signals. Specifically, for each cargo anomaly prediction branch, the sample flight environment data sequence set is used as the input feature of that branch, and the sample abnormal state probability set of the cargo abnormal state corresponding to that branch is used as the training target label for that branch. Supervised training is performed on all branches, and the model parameters are adjusted through optimization algorithms.

[0063] After training, the model performance is evaluated and validated using an independent test dataset. Training terminates when the model exhibits stable predictive performance on the test set, i.e., the loss function converges to a preset criterion and the prediction accuracy meets the requirements. The convergence criterion is set according to the training objectives and engineering application needs. For example, it can be set as follows: over 10 consecutive training epochs, the average decrease in predictive loss on the test dataset is less than 1‰, and the average cross-entropy between the probability distributions of various abnormal states predicted and the true labeled distributions is less than 0.05. Ultimately, a machine learning model capable of predicting the probabilities of various abnormal cargo states from environmental data sequences is obtained—that is, a cargo state predictor.

[0064] For example, since the mapping relationship between flight environment data sequences and cargo anomaly states exhibits temporal dependence and nonlinear coupling characteristics, and Long Short-Term Memory (LSTM) networks have advantages in processing temporal data and capturing long-term dependencies, LSTM networks can be selected as the core model for constructing a cargo state predictor. LSTM networks can effectively analyze the dynamic patterns at different time steps in environmental airflow data sequences and associate them with potential cargo anomaly risks.

[0065] Specifically, the constructed cargo state predictor employs a multi-output branch Long Short-Term Memory (LSTM) network architecture. This network mainly consists of an input layer, a temporal feature encoding layer, a feature fusion layer, and multiple parallel state probability output layers. The input layer receives normalized sample flight environment data sequences. The temporal feature encoding layer uses a single-layer or multi-layer LSM network, with the number of hidden units configured according to the input sequence length and feature dimension. Tanh and Sigmoid activation functions are used internally to control information flow transmission and forgetting. Following the temporal feature encoding layer is a feature fusion layer, which flattens and compresses the encoded temporal feature vectors to extract the overall pattern representation of the environment sequence. Finally, the network connects multiple independent output branches, each corresponding to a cargo anomaly state. A Sigmoid activation function maps the features output by the fusion layer to a probability value between 0 and 1, representing the predicted probability of the corresponding anomaly state occurring.

[0066] During training, key hyperparameters included an initial learning rate of 0.005, a cosine annealing strategy for dynamic learning rate adjustment, a total of 200 training epochs, and a batch size of 32. The initial learning rate was set to balance convergence speed and stability in the early stages of training, while the cosine annealing strategy facilitated fine-tuning of model parameters in the later stages. The selection of the number of training epochs and batch size considered both model convergence requirements and computational resource efficiency.

[0067] The specific training data organization and partitioning are as follows: The previously obtained set of sample flight environment data sequences is used as unified training input data, and the previously obtained sets of multiple sample abnormal state probabilities are used as supervision signals. The paired input data and label data are randomly sampled and divided into training set, validation set, and test set in a 6:2:2 ratio.

[0068] Furthermore, a supervised learning approach is employed during training. The sample flight environment data sequences from the training set are input into the model, and the predicted probability distributions for various anomalous states are calculated using forward propagation. The training objective is to minimize the difference between the predicted probability distribution and the true label probability distribution. Binary cross-entropy is chosen as the loss function for each output branch, and the weighted sum of the losses from all branches is used as the total loss. The backpropagation algorithm, combined with the Adam optimizer, iteratively updates the weight parameters of the Long Short-Term Memory network and subsequent layers. During training, the model performance is continuously monitored using a validation set. When the average predicted loss on the test dataset decreases by less than 1‰ over 10 consecutive training epochs, and the average cross-entropy between the predicted probability distributions for various anomalous states and the true labeled distributions is below 0.05, the model is considered converged, and training is terminated.

[0069] The cargo state predictor trained through the above process can effectively model the intrinsic relationship between complex flight environment sequences and the probabilities of various cargo abnormal states, providing a probabilistic basis for intelligent diagnosis of flight load anomalies.

[0070] Furthermore, the retrieved historical environmental data sequence is used as input data and fed into the aforementioned cargo status predictor to calculate and output the probability of various cargo anomaly states, thus obtaining the anomaly state probability distribution. This output explicitly provides an estimated probability of each of the various preset cargo anomaly states occurring under the environmental conditions represented by the currently input historical environmental data sequence.

[0071] The resulting probability distribution of abnormal states quantifies the risk level of environmental factors that may induce abnormal internal states of various goods, and can provide prior information on the state of goods based on probability inference for subsequent comprehensive analysis.

[0072] S40: Based on the historical environmental data sequence and the probability distribution of abnormal states, analyze and obtain the distribution of influencing loads, calculate the similarity with the flight load distribution, and calculate the wing load anomaly rate as the load anomaly identification result.

[0073] Furthermore, after obtaining the probability distribution of abnormal states characterizing the likelihood of cargo status, it needs to be combined with historical environmental data sequences to complete a quantitative assessment of the current flight load anomaly. Specifically, this step aims to compare the theoretical load effect under the combined influence of environmental factors and cargo status factors with the measured load data, thereby achieving accurate measurement of the degree of anomaly and identification of its root cause.

[0074] Specifically, based on the historical environmental data sequence and the probability distribution of abnormal states, an influencing load distribution set is obtained through analysis. The similarity to the flight load distribution is calculated, and the wing load anomaly rate is calculated as the load anomaly identification result, including:

[0075] An impact load predictor is obtained, wherein the impact load predictor includes multiple impact load prediction branches corresponding to various cargo anomaly states, and each impact load prediction branch is trained using a sample environmental data sequence set and a sample impact load distribution set tested under different cargo anomaly states;

[0076] The historical environmental data sequence is input into the impact load predictor, and multiple predicted impact load distributions are output.

[0077] Based on the abnormal state probability distribution, the multiple predicted influence load distributions are calculated and processed to obtain the influence load distribution;

[0078] Calculate the similarity between the influence load distribution and the flight load distribution, and calculate the wing load anomaly rate as the load anomaly identification result.

[0079] First, a pre-trained impact load predictor is obtained. This impact load predictor adopts a multi-branch model architecture, including multiple impact load prediction branches, each of which corresponds to a specific cargo anomaly state.

[0080] Specifically, the construction and training of each influence load prediction branch relies on a specially configured dataset. The acquisition process for this dataset is as follows: During controlled flight testing, the UAV is pre-set to be in a specific and known abnormal cargo state, such as the cargo undergoing horizontal rotation. During the test flight maintaining this abnormal state, historical environmental data sequences recorded by environmental sensors are simultaneously collected, along with actual load data for each monitored area of ​​the wing within the corresponding time window, recorded by wing load sensors. The latter constitutes the sample influence load distribution set.

[0081] During training, historical environmental data sequences are used as input features to the model, and the sample impact load distribution set is used as the target output for model training. Each impact load prediction branch is independently trained under supervised supervision using its corresponding dedicated dataset, paired in the manner described above. Through this training process, each impact load prediction branch learns the quantitative mapping relationship between the environmental data sequence and the final wing load distribution when the cargo is in its corresponding specific abnormal state.

[0082] Secondly, the current historical environmental data sequence is input into the influence load predictor. The influence load predictor generates multiple prediction results, each representing the expected wing load distribution under the assumption that the cargo is in a specific abnormal state, thus obtaining multiple predicted influence load distributions.

[0083] Furthermore, the obtained multiple predicted impact load distributions are fused to obtain a single comprehensive impact load distribution. This fusion process uses the aforementioned abnormal state probability distribution as the weighting basis. Specifically, the probability value of each state in the abnormal state probability distribution determines the proportion of the load distribution predicted by the corresponding branch in the final fusion result. Through weighted calculation, the predicted load information under multiple hypothetical states can be integrated into a unified theoretical impact load distribution that reflects the most likely combination of states.

[0084] Specifically, based on the probability distribution of the abnormal state, the multiple predicted influence load distributions are calculated and processed to obtain the influence load distribution, including:

[0085] Multiple state weights are assigned based on the probability of occurrence of various cargo abnormal states within the abnormal state probability distribution.

[0086] Multiple state weights are used to perform weighted calculations on the predicted influence loads of the same wing region within the multiple predicted influence load distributions to obtain the influence load distribution.

[0087] First, based on the probability distribution of abnormal states, a corresponding state weight is assigned to each type of cargo abnormal state. Specifically, for each cargo abnormal state in the probability distribution, its state weight is calculated based on its predicted probability value through normalization. The normalization process involves dividing the probability value of each cargo abnormal state by the sum of the probability values ​​of all cargo abnormal states.

[0088] The multiple state weights obtained through the above calculations satisfy the condition that each weight value is greater than or equal to zero, and the sum of all state weights is one. This weight allocation method ensures that the contribution of each cargo anomaly state in the final fusion calculation is strictly proportional to its relative probability in the anomaly state probability distribution, thereby quantitatively and normally expressing the relative importance of various anomalies in explaining the possibility of current load anomalies.

[0089] Secondly, using multiple pre-assigned state weights, data fusion is performed on the multiple predicted influence load distributions previously output by the influence load predictor. Specifically, for each monitoring area on the wing, the predicted load value corresponding to that area is extracted from each predicted influence load distribution. Using the state weights corresponding to various abnormal states as coefficients, a weighted summation is performed on the predicted load values ​​from different distributions but for the same area. This weighted calculation process is performed sequentially for all monitoring areas of the wing.

[0090] For example, suppose there are two abnormal states, A and B, with state weights of 0.7 and 0.3, respectively. For a specific area of ​​the wing, the prediction model for state A gives an impact load of 100 kPa, while the prediction model for state B gives 150 kPa. The fused impact load value for this area is then: 0.7 × 100 kPa + 0.3 × 150 kPa = 115 kPa. This calculation process is performed sequentially on all monitored areas of the wing to generate the final impact load distribution.

[0091] Finally, the new load values ​​obtained after weighted calculations of all regions are combined in the original region order to form a new and complete load distribution, namely the influence load distribution. This influence load distribution is essentially the result of a weighted average of theoretical load predictions under various possible abnormal cargo states, based on their respective probabilities. It represents the theoretical load pattern expected to be observed on the wing, taking into account the current environmental conditions and the probabilities of various possible cargo states.

[0092] Finally, the similarity between the theoretical influence load distribution obtained from the aforementioned steps and the actual flight load distribution monitored in real time is calculated, and the wing load anomaly rate is further calculated as the load anomaly identification result.

[0093] By comparing the theoretical load distribution derived from environmental data and cargo state probabilities with the actual observed flight load distribution, the degree of deviation between the theoretically expected load pattern and the actual observed load pattern can be accurately quantified. The calculated high similarity indicates that the currently observed load fluctuations are mainly attributable to the environmental factors already considered and the cargo state changes predicted based on data. Such fluctuations fall within the scope of physical responses or known risks that the system can reasonably explain, and do not necessarily indicate a structural or control failure in the UAV flight platform itself. Conversely, the calculated low similarity clearly indicates that there are significant load components in the actual flight load distribution that cannot be explained by the currently input environmental data and cargo state probability assumptions.

[0094] The wing load anomaly rate calculated based on this similarity effectively filters out load fluctuation signals caused solely by external airflow disturbances or explainable changes in cargo condition. Therefore, the final load anomaly identification result focuses on anomaly characteristics caused by potential internal faults within the UAV itself, such as latent damage to the wing structure, flight control system anomalies, sensor malfunctions, or other internal failure modes not yet represented in the model.

[0095] Specifically, the similarity between the influencing load distribution and the flight load distribution is calculated, and the wing load anomaly rate is calculated as the load anomaly identification result, including:

[0096] Calculate the load similarity between the influencing load distribution and the flight load distribution;

[0097] Based on the load similarity, the wing load anomaly rate is calculated and used as the result of load anomaly identification.

[0098] First, the load similarity between the influencing load distribution and the flight load distribution is calculated. Specifically, for all monitoring areas defined on the wing, load values ​​for corresponding areas in both the influencing load distribution and the flight load distribution are extracted, forming two load vectors. A preset similarity metric function, such as cosine similarity or normalized cross-correlation coefficient, is used to calculate the quantified load similarity value between the two vectors. This load similarity characterizes the degree of consistency between the theoretically predicted load pattern and the actual observed load pattern in terms of spatial distribution. Its value typically ranges from 0 to 1, with higher values ​​indicating stronger consistency.

[0099] Secondly, based on the calculated load similarity, the wing load anomaly rate is further calculated. The calculation of the wing load anomaly rate aims to convert the load similarity index into a standardized metric that intuitively reflects the severity of the anomaly. Preferably, the wing load anomaly rate can be set to 1 minus the load similarity. Through this conversion, the lower the load similarity, the higher the corresponding wing load anomaly rate; the two are inversely correlated numerically. The finally calculated wing load anomaly rate is defined as the system's load anomaly identification result. This load anomaly identification result not only indicates whether an anomaly has occurred but also, through its specific numerical value, quantifies the severity of the deviation of the current load state from the normal theoretical expectation explained by the environment and cargo state, providing data support for subsequent decision-making and response.

[0100] In summary, the embodiments of this application have at least the following technical effects:

[0101] First, this invention employs a dynamic baseline load range generated based on historical data statistics, replacing the traditional fixed threshold method. This enhances the adaptability to different flight conditions and cargo loading, effectively reducing the false alarm rate. Second, by introducing cargo anomaly state prediction based on environmental data, it achieves correlation analysis between load anomalies and potential causes, enabling a preliminary distinction between environmental interference and cargo's own state anomalies, thus enhancing the diagnostic focus. Third, this invention constructs a multi-branch prediction model, integrates environmental and state information to generate a theoretical load distribution, and compares it with measured data to calculate a quantified anomaly rate, achieving an objective and accurate assessment of the severity of anomalies.

[0102] Ultimately, through progressive analysis, this invention effectively filters out the influence of explainable environmental and cargo factors, focusing the diagnostic core on abnormal load components that may characterize UAV malfunctions, thereby providing a more accurate and reliable basis for flight safety decisions and predictive maintenance.

[0103] Example 2, as Figure 2 As shown, based on the same inventive concept as the data-driven UAV wing dynamic load anomaly identification method provided in Embodiment 1, this embodiment of the invention also provides a data-driven UAV wing dynamic load anomaly identification system, including:

[0104] The load monitoring module 11 is used to obtain the reference load range distribution of the wing of the UAV carrying the target cargo, and to monitor the flight load distribution through sensors during flight.

[0105] Data retrieval module 12 is used to retrieve historical environmental data sequences when the flight load distribution exceeds the reference load range distribution;

[0106] The anomaly prediction module 13 is used to predict the abnormal state of the target cargo based on the historical environmental data sequence and obtain the probability distribution of the abnormal state.

[0107] The anomaly identification module 14 is used to analyze and obtain the influence load distribution based on the historical environmental data sequence and the probability distribution of the abnormal state, calculate the similarity with the flight load distribution, and calculate the wing load anomaly rate as the load anomaly identification result.

[0108] The load monitoring module 11 is specifically used for:

[0109] Obtain the baseline load range distribution of the drone's wings after it carries the target cargo. During flight, monitor the flight load distribution using sensors, including:

[0110] Based on the load test data from the drone flight, the baseline load range distribution of the drone's wings after carrying the target cargo is obtained;

[0111] During the flight of the drone, the distribution of flight loads is monitored and acquired through sensors configured on the wings. The distribution of flight loads includes loads in multiple wing regions.

[0112] Specifically, based on the load test data from the UAV flight, the baseline load range distribution of the UAV's wings after carrying the target cargo is obtained, including:

[0113] Based on the wing loads tested during normal flight of similar drones carrying similar target cargo, obtain the test load distribution set;

[0114] Based on the endpoint values ​​of multiple wing regions within the test load distribution set, a baseline load interval distribution is constructed.

[0115] Specifically, the data retrieval module 12 is used for:

[0116] When the flight load distribution exceeds the baseline load range, historical environmental data sequences are retrieved, including:

[0117] Determine whether any wing region within the flight load distribution exceeds the corresponding reference load range;

[0118] If yes, then process and obtain the historical environmental data sequence within the preset time range; otherwise, continue monitoring the flight payload distribution. The historical environmental data sequence includes environmental airflow data within the preset time range.

[0119] The anomaly prediction module 13 is specifically used for:

[0120] Based on the historical environmental data sequence, anomaly prediction of the target cargo is performed to obtain the probability distribution of the anomaly state, including:

[0121] Obtain a cargo status predictor, wherein the cargo status predictor includes multiple cargo anomaly prediction branches corresponding to various cargo anomaly states.

[0122] The historical environmental data sequence is input into the cargo status predictor, which outputs the probability of various cargo abnormal states and obtains the probability distribution of abnormal states.

[0123] Specifically, acquiring the cargo status predictor includes:

[0124] Based on test data of similar drones carrying similar cargo, a set of sample flight environment data sequences was collected, and the probability of cargo exhibiting various abnormal states under different sample flight environment data sequences was obtained. Multiple sample abnormal state probability sets were then labeled.

[0125] Based on machine learning, multiple cargo anomaly prediction branches are constructed to correspond to various cargo anomaly states.

[0126] Using the sample flight environment data sequence set as training input, and using the multiple sample abnormal state probability sets as supervision labels, supervised training and testing are performed on multiple cargo anomaly prediction branches. After convergence, a cargo state predictor is obtained.

[0127] The anomaly detection module 14 is specifically used for:

[0128] Based on the historical environmental data sequence and the probability distribution of abnormal states, an influencing load distribution set is obtained through analysis. The similarity to the flight load distribution is calculated, and the wing load anomaly rate is calculated as the load anomaly identification result, including:

[0129] An impact load predictor is obtained, wherein the impact load predictor includes multiple impact load prediction branches corresponding to various cargo anomaly states, and each impact load prediction branch is trained using a sample environmental data sequence set and a sample impact load distribution set tested under different cargo anomaly states;

[0130] The historical environmental data sequence is input into the impact load predictor, and multiple predicted impact load distributions are output.

[0131] Based on the abnormal state probability distribution, the multiple predicted influence load distributions are calculated and processed to obtain the influence load distribution;

[0132] Calculate the similarity between the influence load distribution and the flight load distribution, and calculate the wing load anomaly rate as the load anomaly identification result.

[0133] Specifically, based on the probability distribution of the abnormal state, the multiple predicted influence load distributions are calculated and processed to obtain the influence load distribution, including:

[0134] Multiple state weights are assigned based on the probability of occurrence of various cargo abnormal states within the abnormal state probability distribution.

[0135] Multiple state weights are used to perform weighted calculations on the predicted influence loads of the same wing region within the multiple predicted influence load distributions to obtain the influence load distribution.

[0136] Specifically, the similarity between the influencing load distribution and the flight load distribution is calculated, and the wing load anomaly rate is calculated as the load anomaly identification result, including:

[0137] Calculate the load similarity between the influencing load distribution and the flight load distribution;

[0138] Based on the load similarity, the wing load anomaly rate is calculated and used as the result of load anomaly identification.

Claims

1. A data-driven method for identifying dynamic load anomalies on UAV wings, characterized in that, The method includes: Obtain the baseline load range distribution of the wing of the UAV carrying the target cargo, and monitor the flight load distribution through sensors during flight; When the flight load distribution exceeds the reference load range distribution, historical environmental data sequences are retrieved; Based on the historical environmental data sequence, the abnormal state of the target cargo is predicted to obtain the probability distribution of the abnormal state. Based on the historical environmental data sequence and the probability distribution of abnormal states, the distribution of influencing loads is analyzed, the similarity with the flight load distribution is calculated, and the wing load anomaly rate is calculated as the load anomaly identification result.

2. The data-driven method for identifying dynamic load anomalies in UAV wings according to claim 1, characterized in that, Obtain the baseline load range distribution of the drone's wings after it carries the target cargo. During flight, monitor the flight load distribution using sensors, including: Based on the load test data from the drone flight, the baseline load range distribution of the drone's wings after carrying the target cargo is obtained; During the flight of the drone, the distribution of flight loads is monitored and acquired through sensors configured on the wings. The distribution of flight loads includes loads in multiple wing regions.

3. The data-driven method for identifying dynamic load anomalies in UAV wings according to claim 2, characterized in that, Based on the payload test data from the UAV flight, the baseline payload range distribution of the UAV's wings after carrying the target cargo is obtained, including: Based on the wing loads tested during normal flight of similar drones carrying similar target cargo, obtain the test load distribution set; Based on the endpoint values ​​of multiple wing regions within the test load distribution set, a baseline load interval distribution is constructed.

4. The data-driven method for identifying dynamic load anomalies in UAV wings according to claim 1, characterized in that, When the flight load distribution exceeds the baseline load range, historical environmental data sequences are retrieved, including: Determine whether any wing region within the flight load distribution exceeds the corresponding reference load range; If yes, then process and obtain the historical environmental data sequence within the preset time range; otherwise, continue monitoring the flight payload distribution. The historical environmental data sequence includes environmental airflow data within the preset time range.

5. The data-driven method for identifying dynamic load anomalies in UAV wings according to claim 1, characterized in that, Based on the historical environmental data sequence, anomaly prediction of the target cargo is performed to obtain the probability distribution of the anomaly state, including: Obtain a cargo status predictor, wherein the cargo status predictor includes multiple cargo anomaly prediction branches corresponding to various cargo anomaly states. The historical environmental data sequence is input into the cargo status predictor, which outputs the probability of various cargo abnormal states and obtains the probability distribution of abnormal states.

6. The data-driven method for identifying dynamic load anomalies in UAV wings according to claim 5, characterized in that, Obtain a cargo status predictor, including: Based on test data of similar drones carrying similar cargo, a set of sample flight environment data sequences was collected, and the probability of cargo exhibiting various abnormal states under different sample flight environment data sequences was obtained. Multiple sample abnormal state probability sets were then labeled. Based on machine learning, multiple cargo anomaly prediction branches are constructed to correspond to various cargo anomaly states. Using the sample flight environment data sequence set as training input, and using the multiple sample abnormal state probability sets as supervision labels, supervised training and testing are performed on multiple cargo anomaly prediction branches. After convergence, a cargo state predictor is obtained.

7. The data-driven method for identifying dynamic load anomalies in UAV wings according to claim 1, characterized in that, Based on the historical environmental data sequence and the probability distribution of abnormal states, an influencing load distribution set is obtained through analysis. The similarity to the flight load distribution is calculated, and the wing load anomaly rate is calculated as the load anomaly identification result, including: An impact load predictor is obtained, wherein the impact load predictor includes multiple impact load prediction branches corresponding to various cargo anomaly states, and each impact load prediction branch is trained using a sample environmental data sequence set and a sample impact load distribution set tested under different cargo anomaly states; The historical environmental data sequence is input into the impact load predictor, and multiple predicted impact load distributions are output. Based on the abnormal state probability distribution, the multiple predicted influence load distributions are calculated and processed to obtain the influence load distribution; Calculate the similarity between the influence load distribution and the flight load distribution, and calculate the wing load anomaly rate as the load anomaly identification result.

8. The data-driven method for identifying dynamic load anomalies in UAV wings according to claim 7, characterized in that, Based on the abnormal state probability distribution, the multiple predicted impact load distributions are calculated and processed to obtain the impact load distribution, including: Multiple state weights are assigned based on the probability of occurrence of various cargo abnormal states within the abnormal state probability distribution. Multiple state weights are used to perform weighted calculations on the predicted influence loads of the same wing region within the multiple predicted influence load distributions to obtain the influence load distribution.

9. The data-driven method for identifying dynamic load anomalies in UAV wings according to claim 7, characterized in that, Calculate the similarity between the influence load distribution and the flight load distribution, and calculate the wing load anomaly rate as the load anomaly identification result, including: Calculate the load similarity between the influencing load distribution and the flight load distribution; Based on the load similarity, the wing load anomaly rate is calculated and used as the result of load anomaly identification.

10. A data-driven UAV wing dynamic load anomaly identification system, characterized in that, A method for performing the data-driven UAV wing dynamic load anomaly identification method according to any one of claims 1-9 includes: The load monitoring module is used to obtain the baseline load range distribution of the wing of the UAV carrying the target cargo. During flight, the load distribution is monitored by sensors. The data retrieval module is used to retrieve historical environmental data sequences when the flight load distribution exceeds the reference load range distribution. An anomaly prediction module is used to predict the abnormal state of the target cargo based on the historical environmental data sequence and obtain the probability distribution of the abnormal state. The anomaly identification module is used to analyze and obtain the influence load distribution based on the historical environmental data sequence and the probability distribution of the abnormal state, calculate the similarity with the flight load distribution, and calculate the wing load anomaly rate as the load anomaly identification result.

Images