Unmanned aerial vehicle communication fault diagnosis method based on Bayesian decision and man-machine cooperation

Abstract

The invention discloses an unmanned aerial vehicle communication fault diagnosis method based on Bayesian decision and man-machine cooperation, and belongs to the technical field of unmanned aerial vehicle fault prediction and health management. The method comprises the following steps: firstly, acquiring three types of multivariate operation parameters including a link state, a hardware parameter and a flight state through a real-time acquisition module, and preprocessing by adopting a multi-mode adaptive filtering and parameter discretization unit; secondly, a fault knowledge base is constructed, the prior probability and the likelihood probability are calculated based on the labeled samples, Gaussian kernel density estimation is adopted for continuous parameters, and polynomial distribution modeling is adopted for discrete parameters; further, constructing a maximum posterior probability decision engine based on the Bayesian theorem, and outputting a fault diagnosis result and a confidence level; finally, through a human-in-the-loop mechanism, expert intervention is started when the confidence coefficient is insufficient or manual rechecking errors exist, and the probability model is dynamically updated by utilizing a correction sample, so that system self-optimization is realized. According to the method, the three problems of insufficient multi-source parameter coupling analysis, poor static rule base adaptability and low manual intervention efficiency are effectively solved, and the fault diagnosis accuracy and the system adaptive capacity in a complex environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of unmanned aerial vehicle (UAV) technology, communication technology, and artificial intelligence, and particularly to a UAV communication fault diagnosis method based on Bayesian decision-making and human-machine collaboration, capable of high-precision, robust, and self-optimizing UAV communication links. This invention integrates dynamic Bayesian networks, robust state estimation, and human-in-the-loop mechanism, aiming to solve the problem of rapid and accurate root cause localization and continuous optimization of UAV communication link faults in complex environments. Background Technology

[0002] With the increasingly widespread application of unmanned aerial vehicles (UAVs), the reliability of their communication links has become crucial to ensuring the successful completion of missions and flight safety. Traditional UAV fault diagnosis methods mainly fall into three categories: probe-based, model-based, and learning-based. However, these methods face numerous challenges when dealing with increasingly complex UAV communication networks.

[0003] Probe-based diagnostic methods collect fault data by inserting probes into the network, but this consumes valuable network resources, easily causes network congestion, and leads to a decline in service quality. This is unacceptable for UAV communication links, which have extremely high real-time requirements.

[0004] Traditional learning-based diagnostic methods, such as the Naive Bayes classifier, while showing some advantages in handling classification tasks, typically assume that feature variables are independent of each other. This is inconsistent with the strong correlation between parameters in the UAV communication link (such as signal strength, bit error rate, and latency). Furthermore, these models often struggle with high-dimensional symptom data with weak correlations, limiting their diagnostic performance. In practical applications, UAV log data contains numerous parameters, potentially exceeding 200, much of which have weak relevance to fault diagnosis. Without effective processing, this can severely impact the model's training and inference efficiency.

[0005] Existing technologies have significant limitations in handling system uncertainties. The traditional Kalman filter is a commonly used optimal filtering tool, but its performance is highly dependent on the deterministic system dynamics and the zero-mean white noise assumption. During UAV flight, noise characteristics are often uncertain and nonlinear due to external disturbances (such as electromagnetic interference and environmental factors) and model perturbations, which can cause the Kalman filter to diverge, thus affecting the accuracy of state estimation. Similarly, while traditional PID controllers are easy to implement, they cannot guarantee optimal performance and robustness of the closed-loop system in the presence of numerous uncertain disturbances.

[0006] Existing diagnostic methods generally lack the ability to handle small sample failures and to continuously evolve. Because UAV communication link failures are low-frequency events, it is difficult to collect large-scale failure data, which limits the application of traditional data-driven methods. Furthermore, once a model is trained, its performance becomes fixed, and it cannot learn and update itself based on new, unseen failure modes, resulting in insufficient generalization ability.

[0007] To address the shortcomings of the existing technologies, this invention proposes an innovative solution that integrates multiple technologies to provide a highly accurate, robust, and self-optimizing method for diagnosing UAV communication link faults. Summary of the Invention

[0008] In view of this, this invention aims to address three core problems in UAV communication link fault diagnosis: insufficient multi-source parameter coupling analysis, poor adaptability of static rule bases, and low efficiency of manual intervention. It proposes a fault diagnosis method for UAV communication links based on multi-parameter fusion analysis, a probability-driven fault knowledge base, a Bayesian decision engine, and human intervention in the loop dynamic optimization. By real-time acquisition of continuous parameters (such as RSSI, SNR, and RF temperature) and discrete status codes (such as antenna open circuit and equipment malfunction), combined with physically meaning-driven non-uniform hierarchical classification and multi-mode adaptive filtering, a dynamically updatable fault probability model is constructed, enabling accurate fault location and system self-evolution in complex environments.

[0009] The technical solution adopted in this invention is as follows:

[0010] A method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration includes the following steps:

[0011] Step 1, Data Acquisition and Preprocessing: The real-time acquisition module acquires multi-dimensional operational parameters of the UAV ground station link, including link status indicators, hardware parameter indicators, and flight status indicators, categorized into continuous and discrete parameters. A preprocessing module processes the acquired data, comprising an adaptive Kalman filter unit and a parameter discretization unit. The adaptive Kalman filter unit eliminates environmental noise, while the parameter discretization unit performs hierarchical quantization of continuous parameters and standardization of discrete parameters.

[0012] Step 2, Fault Knowledge Base Construction: Establish a structured fault knowledge base containing at least N fault types and storing M labeled samples of historical faults. Each labeled sample is a multi-dimensional feature vector containing observed values ​​of various operating parameters. Based on the stored labeled samples, calculate the fault information for each type. Prior probability and each running parameter In a given fault Conditional probability distribution For continuous operating parameters Modeling is performed using Gaussian kernel density estimation for discrete operating parameters. Multinomial distribution modeling is used; where M and N are set values;

[0013] Step 3, Build a Bayesian decision engine: for real-time collected operating parameters The fault type is calculated using the maximum a posteriori probability estimation method, and the calculation formula is as follows:

[0014]

[0015] Output the posterior probability for each fault category; output the fault type. and posterior probability confidence level ;

[0016] Step 4, Human-Loop Update: When the posterior probability confidence of the diagnosis result... If the values ​​fall below a preset threshold or an error is confirmed by manual review, manual correction is initiated; the corrected labeled samples are used to dynamically update the prior probabilities in the knowledge base. and conditional probability distribution .

[0017] Furthermore, the multivariate operating parameters in step 1 include two main categories: continuous parameters and discrete parameters;

[0018] Among them, continuous parameters include:

[0019] 1) Link status indicators: signal strength, signal-to-noise ratio, bit error rate, correlation peak, automatic gain control value, intermediate frequency signal-to-noise ratio, angle deviation, error voltage, and AD sampling value; among which, angle deviation includes azimuth deviation and elevation deviation;

[0020] 2) Hardware parameters: RF input and output levels, RF front-end temperature, power amplifier temperature, and power amplifier output power;

[0021] 3) Flight status indicators: aircraft attitude angles, flight altitude, and aircraft latitude and longitude; among which, aircraft attitude angles include pitch angle, roll angle, and heading angle;

[0022] Discrete parameters include:

[0023] 1) Hardware parameters: RF front-end status, antenna status, ADDA chip status, and modem status;

[0024] 2) Link status indicators: frequency consistency, power setting mode, and antenna selection identifier.

[0025] Furthermore, in step 1, the adaptive Kalman filter unit eliminates environmental noise by specifically including the following steps:

[0026] Step 101: Real-time detection of signal-to-noise ratio fluctuations using sliding window variance analysis. :

[0027]

[0028] In the formula, T is the window length, k represents the index at the current time, and is the right boundary time of the sliding window. and Let be the signal-to-noise ratio at time t and time t-1;

[0029] Step 102, when When the signal-to-noise ratio is less than or equal to the threshold, it is considered a stable flight scenario. When the signal-to-noise ratio is greater than the threshold, it is a maneuvering flight scenario; when the interference power ratio is greater than the power threshold, it is a strong electromagnetic interference scenario.

[0030] Step 103: Under stable flight conditions, standard Kalman filtering is used to eliminate environmental noise, and the process noise covariance is set to... In maneuvering flight scenarios, a strong-tracking Kalman filter is used to eliminate environmental noise, and the process noise covariance is set to... In scenarios with strong electromagnetic interference, a robust H∞ filtering algorithm is used to eliminate environmental noise.

[0031] Furthermore, a robust H∞ filtering algorithm is used to eliminate environmental noise. The specific process includes:

[0032] The Kalman gain matrix is ​​dynamically adjusted using weight γ. :

[0033]

[0034]

[0035]

[0036] In the formula, As the baseline value, To adjust the amplitude, Environmental sensitivity coefficient, and These are the estimated interference power and signal power, respectively. ∈ Where u is the measurement dimension and v is the state dimension. This represents the true state at time k. This means that at time k, based only on information from time k-1 and earlier, the true state is... The predicted value made;

[0037] Simultaneously, the observation noise covariance matrix... Dynamic robust constraints are applied, and the Kalman gain matrix is ​​then dynamically adjusted to achieve an adaptive optimal balance of performance under various environments.

[0038] ;

[0039] In the formula, To observe the noise covariance matrix; The environmental disturbance sensitivity coefficient is a decimal less than 1.

[0040] Furthermore, the sample sources for establishing the structured fault knowledge base in step 2 include:

[0041] Historical Fault Database: Stores multiple sets of labeled samples of various faults, including fault types such as manual operation faults, hardware equipment faults, and space physical factor faults;

[0042] Real-time sample injection: Through a hardware-in-the-loop simulation test platform, preset fault signals or simulation environment parameters are injected into the real communication hardware components of the UAV in a programmable control manner to actively simulate various fault states, including device overheating, signal interference and hardware failure, and simultaneously collect system response data under the corresponding states to form a fault sample set.

[0043] Furthermore, the conditional probability distribution in step 2 Distinguished by whether the parameter is continuous or discrete:

[0044] (1) Continuous parameters are estimated using Gaussian kernel density:

[0045] For any fault category ,when When, conditional probability distribution for:

[0046]

[0047] In the formula, the number of samples Fault Category The total number of valid labeled samples under, The minimum sample threshold, Fault Category The specific value of a certain continuous feature of the k-th sample in the historical effective fault knowledge base;

[0048] when When, conditional probability distribution for:

[0049]

[0050] In the formula, These are the weighting coefficients. , For eigenvalues Joint kernel density estimation for all fault categories, Is it only using fault categories? The local conditional probability density estimate is obtained by calculating n samples of a small sample using the Gaussian kernel density estimation method;

[0051]

[0052] In the formula, For eigenvalues Global probability density estimation at; The total number of samples across all fault categories; The total number of fault categories defined in the fault knowledge base; For the first Class of faults The number of valid historical samples; For the first Class of faults Feature values ​​of each sample; For bandwidth Gaussian kernel function;

[0053] (2) Discrete parameters are modeled using a multinomial distribution:

[0054] Suppose a certain discrete feature Total One possible value:

[0055]

[0056] In the formula, Fault Lower features Number of times it appears; Fault The total number of samples.

[0057] Furthermore, in step 3, the posterior probability confidence level The calculation is as follows:

[0058]

[0059] In the formula, The decision entropy value. This represents the total number of fault categories. For each fault category The posterior probability, = 1, 2, ..., N.

[0060] Furthermore, step 4 includes:

[0061] Confidence triggering module: Receives diagnostic results in real time. and their corresponding posterior probability confidence levels ;Will Compare with a preset confidence threshold θ; when When < θ, a manual review request is automatically triggered;

[0062] Manual verification module: Receives the corrected fault label and updates the label only when the original diagnosis is confirmed to be incorrect by a human.

[0063] Incremental update module: provides the diagnostic results. posterior probability confidence level and related feature data; receive the corrected fault label input. ;

[0064] Knowledge base incremental update module: This module updates manually verified new samples (x, Inject the fault knowledge base and dynamically update the relevant probability model to update the fault. Prior probability:

[0065]

[0066] In the formula, The number of historical samples for the corrected fault category. This represents the current total number of samples. This represents the total number of fault categories. It is an additive smoothing factor.

[0067] Furthermore, dynamic updates include two methods:

[0068] (1) Intelligent batch retraining strategy: When any of the following conditions are met, batch retraining of the probabilistic models of all fault categories in the entire fault knowledge base is automatically triggered:

[0069] a) The cumulative number of manually corrected samples received reaches a preset threshold;

[0070] b) The KL divergence between the current probability distribution estimated based on the new received samples and the existing distribution in the fault knowledge base exceeds a set threshold δ; the threshold δ is determined based on the typical distribution offset of the UAV communication link fault modes.

[0071] (2) Real-time incremental fine-tuning strategy: For samples generated by a single manual correction ( , Sample x refers to the set of all relevant operating parameter values ​​observed at the time of a failure. A lightweight update is performed, and the update objects are limited to:

[0072] a) Fault Category The prior probability P( );

[0073] b) In the eigenvector In the context of diagnostic decision-making, all feature dimensions are included. The corresponding conditional probability distribution .

[0074] The advantages of this invention compared to the prior art are:

[0075] This invention achieves significant improvements in accuracy, robustness, and self-evolution capabilities in UAV communication link fault diagnosis through the deep integration of a Bayesian decision framework and human-in-the-loop mechanism. Traditional methods rely on fixed thresholds or static models, which are ill-suited for complex electromagnetic environments and rare faults. This invention first utilizes adaptive H∞ filtering technology, dynamically adjusting robustness parameters and observation noise covariance to enable the system to sense interference intensity and optimize state estimation, thus maintaining extremely high stability under strong noise. Second, it employs a probabilistic diagnostic model, integrating historical prior knowledge with real-time data to output diagnostic results with confidence levels, resulting in higher accuracy and supporting uncertainty management. When the system's self-diagnosis confidence is insufficient, it automatically triggers manual review, and expert confirmation results are injected into the knowledge base as gold samples, forming a closed loop of "diagnosis-verification-learning," enabling the system to continuously evolve from practice and effectively solve the diagnostic challenges of small samples and unknown faults. Furthermore, by injecting fault samples through hardware-in-the-loop simulation, it overcomes the bottleneck of high cost and long cycle in acquiring real fault data. This invention ultimately provides an advanced fault diagnosis solution that integrates intelligent perception, autonomous adaptation, human-machine collaboration, and continuous learning, significantly enhancing the reliability and safety of UAVs in complex tasks. Attached Figure Description

[0076] Figure 1 This is a flowchart of the UAV communication fault diagnosis method based on Bayesian decision-making and human-machine collaboration of the present invention. Detailed Implementation

[0077] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0078] like Figure 1 As shown, a method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration is presented. The specific steps of this method are as follows:

[0079] Step 1, Data Acquisition and Preprocessing: The real-time acquisition module acquires multi-dimensional operational parameters of the UAV ground station link, including link status indicators, hardware parameter indicators, and flight status indicators, categorized into continuous and discrete parameters. A preprocessing module is used to process the acquired data. This preprocessing module includes an adaptive Kalman filter unit and a parameter discretization unit. The adaptive Kalman filter unit is used to eliminate environmental noise, while the parameter discretization unit is used to perform hierarchical quantization of continuous parameters and standardization of discrete parameters.

[0080] Step 1.1: Through the data bus and dedicated monitoring interface of the UAV ground station, collect the following three categories of operating parameters in real time at a sampling frequency of no less than 100Hz. The specific parameter categories, names, and classification rules are as follows:

[0081] (1) Link status indicators. This type of parameter is used to characterize the quality and stability of the communication link, and includes 8 key parameters:

[0082] Signal Strength Index (RSSI): ≥-60dBm indicates excellent signal strength, corresponding to a strong and stable signal transmission state; -61dBm to -70dBm indicates good signal strength, corresponding to an acceptable signal attenuation state; -71dBm to -80dBm indicates moderate signal strength, corresponding to a signal weakening state requiring warning; below -80dBm indicates poor signal strength, corresponding to a state with a high risk of communication interruption.

[0083] Signal-to-noise ratio (SNR): >20dB indicates excellent SNR level, corresponding to high noise immunity; between 15dB and 20dB indicates good SNR level, corresponding to normal noise immunity; between 10dB and 15dB indicates medium SNR level, corresponding to the system being sensitive to noise; <10dB indicates poor SNR level, corresponding to the communication quality being on the verge of collapse.

[0084] Bit Error Rate (BER): When the logarithm to base 10 is less than -4 (i.e., BER < 0.0001), the bit error rate is considered normal; when it is between -4 and -3.5 (i.e., BER is 0.0001 to 0.0003), the bit error rate is considered warning; when it is greater than -3.5 (i.e., BER > 0.0003), the bit error rate is considered abnormal.

[0085] Correlation peaks: >0.8 indicates excellent synchronization quality, corresponding to a strong signal synchronization state; between 0.6 and 0.8 indicates good synchronization quality, corresponding to a stable synchronization state; between 0.4 and 0.6 indicates medium synchronization quality, corresponding to a state of synchronization fluctuation; <0.4 indicates poor synchronization quality, corresponding to a state of risk of losing synchronization.

[0086] Automatic Gain Control (AGC): When the AGC value is between 30% and 80%, the signal is normal, corresponding to the optimal gain range; when it is less than 30%, the signal is weak, corresponding to the need to increase receiver sensitivity; when it is greater than 80%, the signal is at risk of saturation, corresponding to the possibility of signal overload and distortion.

[0087] Intermediate frequency signal-to-noise ratio (IFSNR): >25dB indicates excellent intermediate frequency signal-to-noise ratio, corresponding to a high signal-to-noise ratio state; between 15dB and 25dB indicates good intermediate frequency signal-to-noise ratio, corresponding to an acceptable signal-to-noise ratio state; between 5dB and 15dB indicates medium intermediate frequency signal-to-noise ratio, corresponding to a state with significant noise interference; <5dB indicates poor intermediate frequency signal-to-noise ratio, corresponding to a state where the signal may be submerged by noise.

[0088] Error voltage: <1V indicates normal performance and corresponds to a low-error stable state; 1V to 2V indicates a warning and corresponds to an error accumulation state; 2V to 3V indicates a severe performance and corresponds to a possible control instability state; >3V indicates a performance fault and corresponds to a system out of control state.

[0089] AD sampling values: When the sampling value is in the range of 800 to 3500, the representation state is normal, corresponding to the effective dynamic range; when it is <800, the representation state is weak, corresponding to insufficient quantization accuracy; when it is >3500, the representation state is saturated, corresponding to signal clipping distortion.

[0090] (2) Hardware parameter indicators. This category of parameters is used to monitor the working status and health of the communication system hardware, and includes 8 key parameters:

[0091] RF input / output levels: When the level is in the range of -60dBm to -30dBm, the signal is normal and corresponds to the safe input power range; when it is <-60dBm, the signal is weak and requires gain compensation; when it is >-30dBm, the signal is overloaded and there is a risk of damage to the front-end devices.

[0092] RF front-end temperature: <70℃ indicates normal performance and corresponds to the safe temperature range; 70℃ to 80℃ indicates a warning and potential performance degradation; 80℃ to 85℃ indicates a severe performance issue and potential heat dissipation failure; >85℃ indicates a performance failure and a risk of chip burnout.

[0093] Amplifier temperature: <80℃ indicates normal operation and stable working condition; 80℃ to 90℃ indicates warning and requires derating output; 90℃ to 100℃ indicates severe condition and triggers over-temperature protection; >100℃ indicates fault and a risk of permanent damage to the amplifier.

[0094] Amplifier output power: When the output power is in the range of 5W to 20W, the characteristic is normal and corresponds to the rated power operation state; when it is <5W, the characteristic is derating and corresponds to insufficient output power; when it is >20W, the characteristic is overload risk and corresponds to the power device has a breakdown risk state.

[0095] Device status parameters (RF front-end status, antenna status, ADDA chip status, modem status): Status code 0 (normal) is mapped to encoding [1, 0]; status code 1 (fault / abnormal) is mapped to encoding [0, 1]. This encoding method transforms discrete states into numerical vectors, facilitating subsequent probabilistic model calculations.

[0096] (3) Flight status indicators. This type of parameter is used to provide contextual information about the aircraft platform and includes three key parameters:

[0097] Aircraft attitude angle changes (pitch / roll / heading): When the attitude angle change is <10°, it indicates a stable state, corresponding to a smooth flight state; when it is in the range of 10° to 30°, it indicates a fluctuating state, corresponding to a controllable state with attitude fluctuations; when it is in the range of 30° to 45°, it indicates a high-risk state, corresponding to an attitude loss of control warning state; when it is >45°, it indicates a risk of loss of control, corresponding to a state that requires entering emergency recovery mode.

[0098] Flight latitude, longitude, and altitude: Based on the latitude and longitude coordinates (WGS84 coordinate system) transmitted back by the aircraft in real time and the fixed latitude and longitude coordinates known by the ground station, the final relative straight-line distance is calculated using the three-dimensional Euclidean distance formula. Then calculate the maximum reliable communication distance of the ground station communication system. Normalization is performed based on the baseline. ≤0.5 (Short Range): Indicates the aircraft is in a reliable communication area. Signal strength is good, path loss is negligible, and link reliability is high. 0.5 < ≤ 0.8 (Medium Range): Indicates the aircraft is in the attenuation transition zone. Path loss due to distance effects becomes significant, and signal strength gradually decreases with increasing distance. Link margin and signal quality changes need to be monitored. 0.8 < ≤ 1.0 (Long-range): Indicates that the aircraft is in the critical communication zone. Signal attenuation is significant, and link quality is extremely sensitive to equipment performance, environmental changes, and minor interference, placing it on the edge of reliable communication. > 1.0 (Overrange): Indicates that the aircraft is in the beyond-visual-range risk zone. The communication distance has exceeded the system's rated reliability range, and the impact of Earth's curvature, atmospheric refraction, and depth path loss is aggravated, resulting in an extremely high risk of communication interruption.

[0099] After all parameters are discretized according to the above rules, the original data with different dimensions and types are transformed into a unified discrete level or numerical vector with clear physical meaning.

[0100] Step 1.2 performs flight scene recognition and filtering mode switching. The preprocessing module uses the sliding window variance analysis method to calculate the fluctuation intensity of the signal-to-noise ratio in real time, which is used as a feature quantity to distinguish flight scenes.

[0101] First, set the sliding window length T = 500 milliseconds. At each sampling time k, calculate the mean absolute difference of the signal-to-noise ratio within the window as an indicator of fluctuation.

[0102]

[0103] In the formula, T is the window length, k represents the index at the current time, and is the right boundary time of the sliding window. and Let be the signal-to-noise ratio at time t and time t-1;

[0104] Based on the quantization index of ΔSNR, the system automatically switches to the optimal filtering mode. For stable flight scenarios (ΔSNR ≤ 3dB), the standard Kalman filter algorithm is enabled. Its process noise covariance matrix is ​​set to... In the maneuvering flight scenario (ΔSNR > 3dB), the Strong Tracking Kalman Filter (STF) algorithm is enabled, and its process noise covariance matrix is ​​set to... In scenarios with strong electromagnetic interference (interference power ratio > 3), the system switches to the robust H∞ filtering algorithm.

[0105] The Kalman gain matrix is ​​dynamically adjusted using weight γ.

[0106]

[0107]

[0108]

[0109] In the formula, and These are the estimated interference power and signal power, respectively. ∈ Where u is the measurement dimension and v is the state dimension. This represents the true state at time k. This means that at time k, based only on information from time k-1 and earlier, the true state is... The predicted value made;

[0110] Weight γ varies with disturbance intensity ( The value smoothly approaches 1.0 from 0.92 as the interference increases. When the interference is exceptionally strong ( When > 5), the upper limit of γ is locked at 1.0 to ensure the stability of the algorithm under extreme disturbances.

[0111] At the same time The observation noise covariance matrix is ​​dynamically and robustly constrained, and then the Kalman gain matrix is ​​dynamically adjusted to achieve an adaptive optimal balance of performance under various environments.

[0112]

[0113] In the formula, To observe the noise covariance matrix.

[0114] Step 2, Fault Knowledge Base Construction: Establish a structured fault knowledge base containing at least N fault types and storing M labeled samples of historical faults. Each labeled sample is a multi-dimensional feature vector containing observed values ​​of various operating parameters. Based on the stored labeled samples, calculate the fault information for each type. Prior probability and each running parameter In a given fault Conditional probability distribution For continuous operating parameters Modeling is performed using Gaussian kernel density estimation for discrete operating parameters. Multinomial distribution modeling is used; where M and N are set values;

[0115] For each type of fault defined in the fault library (e.g., power amplifier overheating fault) and characteristics of each dimension A corresponding likelihood probability model was constructed for power amplifier temperatures ∈ [0, 80]℃. This model is used to quantify the probability of observing a certain characteristic value when a specific fault occurs. Different modeling strategies are adopted for the different characteristics of continuous and discrete parameters.

[0116] (1) For continuous parameters such as signal strength and equipment temperature, their values ​​are complex and usually do not satisfy the simple normal distribution assumption. Therefore, this invention uses a non-parametric Gaussian kernel density estimation method for modeling. This method does not require prior distribution assumptions and can directly learn and reconstruct the true probability density function from the sample data.

[0117] For fault category Let a certain continuous feature have n valid historical samples. Then a new observation value The likelihood probability density is calculated as follows:

[0118]

[0119] Among them, the number of samples Fault Category The total number of valid labeled samples under the given conditions, whose values ​​satisfy the condition that for any fault category... , must meet (Minimum sample threshold, typical value) );when At that time, a global mixed distribution is used as a priori:

[0120]

[0121] in (Weighting coefficient) For eigenvalues The joint kernel density estimation for all fault categories is insufficient for a given fault type due to inadequate historical case data, failing to meet the minimum requirements for reliable statistical modeling. Is it only using fault categories? The local conditional probability density estimate is obtained by calculating n samples of a small sample using the Gaussian kernel density estimation method;

[0122]

[0123] The total number of samples across all fault categories; The total number of fault categories defined in the fault knowledge base; For the first Class of faults The number of valid historical samples; For the first Class of faults Feature values ​​of each sample; Bandwidth is Gaussian kernel function;

[0124] (2) Multinomial distribution modeling is used for discrete parameters:

[0125]

[0126] In the formula, Fault Lower features Number of times it appears; Fault The total number of samples (e.g., the number of times "antenna open circuit" occurs in hardware failures); Fault The total number of samples.

[0127] Step 3: For each predefined fault category in the fault knowledge base ( Given a set of integers (e.g., 1, 2, ..., N), calculate its posterior probability. :

[0128]

[0129] Subsequently, maximum a posteriori probability estimation is performed:

[0130] .

[0131] Posterior probability confidence level The calculation is as follows:

[0132]

[0133] In the formula, The decision entropy value. This represents the total number of fault categories. Output the fault types. and posterior probability confidence level .

[0134] Step 4, regarding the diagnostic results and its confidence level A threshold judgment is performed. The system presets a confidence threshold θ of 0.7. When When the value is less than θ, it indicates that the algorithm has low certainty regarding the current diagnostic result and there is a risk of misjudgment. The system will automatically trigger a manual verification request. After the system initiates the manual verification interface, the interface clearly displays the real-time collected feature vectors to the operator. Automatic diagnostic results from the system This includes the confidence level and the relevant raw data waveforms or parameter lists. Based on their professional knowledge and experience, operators analyze and judge the actual situation, and input the corrected fault label through the interface. The system will use the current feature vector Correct labeling by hand Combined, they form a new, validated labeled sample. This sample has extremely high reliability. Then, the manually corrected labeled sample is injected into the fault knowledge base to dynamically update the prior probability parameters.

[0135]

[0136] Update bug Prior probability:

[0137]

[0138] This represents the number of samples for this type of fault. The number of historical samples for the corrected fault category. This represents the current total number of samples. It is an additive smoothing factor, and its value is 1 when solving zero-probability problems.

[0139] Dynamic updates include two methods:

[0140] (1) Intelligent batch retraining strategy: When any of the following conditions are met, batch retraining of the probabilistic models of all fault categories in the entire fault knowledge base is automatically triggered:

[0141] a) The cumulative number of manually corrected samples received reaches a preset threshold;

[0142] b) The KL divergence between the current probability distribution estimated based on the new received samples and the existing distribution in the fault knowledge base exceeds a set threshold δ; the threshold δ is determined based on the typical distribution offset of the UAV communication link fault modes.

[0143] (2) Real-time incremental fine-tuning strategy: For samples generated by a single manual correction ( , Sample x refers to the set of all relevant operating parameter values ​​observed at the time of a failure. A lightweight update is performed, and the update objects are limited to:

[0144] a) Fault Category The prior probability P( );

[0145] b) In the eigenvector In the context of diagnostic decision-making, all feature dimensions are included. The corresponding conditional probability distribution .

[0146] In summary, in the field of intelligent diagnostics for UAV communication links, this invention proposes a fault diagnosis method based on multivariate parameter fusion analysis, probability-driven approaches, and human-in-the-loop dynamic optimization. This method achieves synergistic optimization through multi-source parameter coupling analysis, Bayesian decision-making, and system self-evolution, significantly improving the accuracy of fault location and system adaptability in complex environments. Furthermore, the probabilistic model constructed in this invention possesses clear physical meaning and interpretability, and the related calculation process is efficient and standardized, compatible with the computing power constraints of embedded platforms, and has the potential for engineering deployment and real-time application.

[0147] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0148] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for diagnosing communication faults in unmanned aerial vehicles (UAVs) based on Bayesian decision-making and human-machine collaboration, characterized in that, Includes the following steps: Step 1, Data Acquisition and Preprocessing: The real-time acquisition module acquires multi-dimensional operational parameters of the UAV ground station link, including link status indicators, hardware parameter indicators, and flight status indicators, categorized into continuous and discrete parameters. A preprocessing module processes the acquired data, comprising an adaptive Kalman filter unit and a parameter discretization unit. The adaptive Kalman filter unit eliminates environmental noise, while the parameter discretization unit performs hierarchical quantization of continuous parameters and standardization of discrete parameters. Step 2, Fault Knowledge Base Construction: Establish a structured fault knowledge base containing at least N fault types and storing M labeled samples of historical faults. Each labeled sample is a multi-dimensional feature vector containing observed values ​​of various operating parameters. Based on the stored labeled samples, calculate the fault information for each type. Prior probability and each running parameter In a given fault Conditional probability distribution For continuous operating parameters Modeling is performed using Gaussian kernel density estimation for discrete operating parameters. Multinomial distribution modeling is used; where M and N are set values; Step 3, Build a Bayesian decision engine: for real-time collected operating parameters The fault type is calculated using the maximum a posteriori probability estimation method, and the calculation formula is as follows: Output the posterior probability for each fault category; output the fault type. and posterior probability confidence level ; Step 4, Human-Loop Update: When the posterior probability confidence of the diagnosis result... If the values ​​fall below a preset threshold or an error is confirmed by manual review, manual correction is initiated; the corrected labeled samples are used to dynamically update the prior probabilities in the knowledge base. and conditional probability distribution .

2. The method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration according to claim 1, characterized in that, The multivariate operating parameters in step 1 include two main categories: continuous parameters and discrete parameters. Among them, continuous parameters include: 1) Link status indicators: signal strength, signal-to-noise ratio, bit error rate, correlation peak, automatic gain control value, intermediate frequency signal-to-noise ratio, angle deviation, error voltage, and AD sampling value; among which, angle deviation includes azimuth deviation and elevation deviation; 2) Hardware parameters: RF input and output levels, RF front-end temperature, power amplifier temperature, and power amplifier output power; 3) Flight status indicators: aircraft attitude angles, flight altitude, and aircraft latitude and longitude; among which, aircraft attitude angles include pitch angle, roll angle, and heading angle; Discrete parameters include: 1) Hardware parameters: RF front-end status, antenna status, ADDA chip status, and modem status; 2) Link status indicators: frequency consistency, power setting mode, and antenna selection identifier.

3. The method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration according to claim 1, characterized in that, In step 1, the adaptive Kalman filter unit eliminates environmental noise by including the following steps: Step 101: Real-time detection of signal-to-noise ratio fluctuations using sliding window variance analysis. : In the formula, T is the window length, k represents the index at the current time, and is the right boundary time of the sliding window. and Let be the signal-to-noise ratio at time t and time t-1; Step 102, when When the signal-to-noise ratio is less than or equal to the threshold, it is considered a stable flight scenario. When the signal-to-noise ratio is greater than the threshold, it is a maneuvering flight scenario; when the interference power ratio is greater than the power threshold, it is a strong electromagnetic interference scenario. Step 103: Under stable flight conditions, standard Kalman filtering is used to eliminate environmental noise, and the process noise covariance is set to... In maneuvering flight scenarios, a strong-tracking Kalman filter is used to eliminate environmental noise, and the process noise covariance is set to... In scenarios with strong electromagnetic interference, a robust H∞ filtering algorithm is used to eliminate environmental noise.

4. The method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration according to claim 3, characterized in that, The robust H∞ filtering algorithm is used to eliminate environmental noise. The specific process includes: The Kalman gain matrix is ​​dynamically adjusted using weight γ. : In the formula, As the baseline value, To adjust the amplitude, Environmental sensitivity coefficient, and These are the estimated interference power and signal power, respectively. ∈ Where u is the measurement dimension and v is the state dimension. This represents the true state at time k. This means that at time k, based only on information from time k-1 and earlier, the true state is... The predicted value made; Simultaneously, the observation noise covariance matrix... Dynamic robust constraints are applied, and the Kalman gain matrix is ​​then dynamically adjusted to achieve an adaptive optimal balance of performance under various environments. ； In the formula, To observe the noise covariance matrix; The environmental disturbance sensitivity coefficient is a decimal less than 1.

5. The method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration according to claim 1, characterized in that, The sample sources for establishing the structured fault knowledge base in step 2 include: Historical Fault Database: Stores multiple sets of labeled samples of various faults, including fault types such as manual operation faults, hardware equipment faults, and space physical factor faults; Real-time sample injection: Through a hardware-in-the-loop simulation test platform, preset fault signals or simulation environment parameters are injected into the real communication hardware components of the UAV in a programmable control manner to actively simulate various fault states, including device overheating, signal interference and hardware failure, and simultaneously collect system response data under the corresponding states to form a fault sample set.

6. The method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration according to claim 1, characterized in that, Conditional probability distribution in step 2 Distinguished by whether the parameter is continuous or discrete: (1) Continuous parameters are estimated using Gaussian kernel density: For any fault category ,when When, conditional probability distribution for: In the formula, the number of samples Fault Category The total number of valid labeled samples under, The minimum sample threshold, Fault Category The specific value of a certain continuous feature of the k-th sample in the historical effective fault knowledge base; when When, conditional probability distribution for: In the formula, These are the weighting coefficients. , For eigenvalues Joint kernel density estimation for all fault categories, Is it only using fault categories? The local conditional probability density estimate is obtained by calculating n samples of a small sample using the Gaussian kernel density estimation method; In the formula, For eigenvalues Global probability density estimation at; The total number of samples across all fault categories; The total number of fault categories defined in the fault knowledge base; For the first Class of faults The number of valid historical samples; For the first Class of faults Feature values ​​of each sample; For bandwidth Gaussian kernel function; (2) Discrete parameters are modeled using a multinomial distribution: Suppose a certain discrete feature Total One possible value: In the formula, Fault Lower features Number of times it appears; Fault The total number of samples.

7. The method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration according to claim 6, characterized in that, posterior probability confidence level in step 3 The calculation is as follows: In the formula, The decision entropy value. This represents the total number of fault categories. For each fault category The posterior probability, = 1, 2, ..., N.

8. The method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration according to claim 1, characterized in that, Step 4 includes: Confidence triggering module: Receives diagnostic results in real time. and their corresponding posterior probability confidence levels ;Will Compare with a preset confidence threshold θ; when When < θ, a manual review request is automatically triggered; Manual verification module: Receives the corrected fault label and updates the label only when the original diagnosis is confirmed to be incorrect by a human. Incremental update module: provides the diagnostic results. posterior probability confidence level and related feature data; receive the corrected fault label input. ; Knowledge base incremental update module: This module updates manually verified new samples (x, Inject the fault knowledge base and dynamically update the relevant probability model to update the fault. Prior probability: In the formula, The number of historical samples for the corrected fault category. This represents the current total number of samples. This represents the total number of fault categories. It is an additive smoothing factor.

9. The method for diagnosing UAV communication faults based on Bayesian decision-making and human-machine collaboration according to claim 8, characterized in that, Dynamic updates include two methods: (1) Intelligent batch retraining strategy: When any of the following conditions are met, batch retraining of the probabilistic models of all fault categories in the entire fault knowledge base is automatically triggered: a) The cumulative number of manually corrected samples received reaches a preset threshold; b) The KL divergence between the current probability distribution estimated based on the new received samples and the existing distribution in the fault knowledge base exceeds the set threshold δ; The threshold δ is determined based on the typical distribution offset of the UAV communication link failure modes; (2) Real-time incremental fine-tuning strategy: For samples generated by a single manual correction ( , Sample x refers to the set of all relevant operating parameter values ​​observed at the time of a failure. A lightweight update is performed, and the update objects are limited to: a) Fault Category The prior probability P( ); b) In the eigenvector In the context of diagnostic decision-making, all feature dimensions are included. The corresponding conditional probability distribution .

Images