Automatic tuning and training method and system of unmanned aerial vehicle smart city inspection model

Abstract

The application discloses an unmanned aerial vehicle (UAV) smart city inspection model automatic tuning and training method and system, and belongs to the technical field of artificial intelligence. The method comprises the following steps: acquiring real-time data flow and model inference data generated during UAV online inspection, performing multi-dimensional feature extraction based on the real-time data flow and the model inference data, and generating a multi-dimensional performance attenuation signal; inputting the multi-dimensional performance attenuation signal into a root cause analysis model for joint analysis to obtain a root cause analysis result; selecting and combining a dynamic optimization strategy combination from a strategy library based on the root cause analysis result; executing the dynamic optimization strategy combination to generate verification feedback data; and associating the verification feedback data with the root cause analysis result to update the root cause analysis model. The method adopts multi-dimensional performance attenuation signal sensing, root cause analysis based on causal inference, dynamic optimization strategy generation, and a closed-loop feedback updating mechanism, and can realize continuous self-learning and adaptive evolution of the model in a complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an automatic optimization and training method and system for a drone-based smart city inspection model. Background Technology

[0002] With the integration of deep learning and edge computing technologies, neural network-based computer vision models have been widely applied in the field of autonomous drone inspection. These models enable real-time defect detection in complex industrial scenarios or power facilities through inference engines deployed at the edge of the drone. To cope with complex inspection environments, existing inspection systems typically employ high-performance computing units to process multi-source sensor data and rely on pre-trained models to perform inference tasks. However, due to the highly dynamic and uncertain nature of drone inspection environments, such as sudden changes in lighting, severe weather conditions, and imaging blurring caused by drone movement, the data distribution at the inspection site often drifts from the distribution during model training. This leads to a non-linear degradation in model inference performance, manifested as decreased target detection accuracy, increased uncertainty, or fluctuating inference latency.

[0003] In related technologies, Chinese invention patent CN120298938A discloses an intelligent identification system and method for UAV inspection faults, including: constructing a dual-target path model through historical fault data; dynamically optimizing the UAV inspection path based on multi-source perception data using the dual-target path model to obtain the optimized path and perception area; performing hierarchical identification on the multi-source perception data to obtain the UAV's state prediction value; constructing an intelligent path guidance function based on the optimized path, perception area, and state prediction value; performing backpropagation correction on the intelligent path guidance function to obtain fault type features and detection accuracy; constructing a feature association weight matrix for the inspection path based on the fault type features and detection accuracy; and then driving the UAV to intelligently identify faults on the reconstructed inspection path based on the feature association weight matrix.

[0004] While existing technologies can improve fault identification accuracy through path optimization and feature weighting, they are essentially static feedback mechanisms for pre-defined tasks, lacking the ability to diagnose the root causes of performance degradation in model inference. Specifically, existing technologies struggle to quantify and integrate external dynamic environmental parameters with the model's internal inference behavior in real time, failing to accurately decouple the dominant and secondary features causing performance degradation. This makes it difficult to provide targeted dynamic optimization strategies such as data supplementation, specific layer fine-tuning, or cloud-edge load balancing when facing sudden environmental disturbances. Furthermore, existing technologies typically lack a complete self-learning loop of diagnosis-execution-verification-learning, unable to quantify reward scores based on historical optimization results and iteratively update diagnostic logic. When dealing with regional or common performance degradation issues, they struggle to leverage swarm intelligence to achieve global performance evolution. Therefore, how to build a system capable of automatically identifying the root causes of performance degradation and achieving dynamic closed-loop optimization of strategies in resource-constrained edge inspection scenarios remains a key technical challenge in the field of industrial artificial intelligence. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an automatic optimization and training method and system for UAV smart city inspection models. It employs multi-dimensional performance degradation signal perception, root cause analysis based on causal inference, dynamic optimization strategy generation, and a closed-loop feedback update mechanism, enabling the model to continuously learn and adaptively evolve in complex dynamic environments.

[0006] The above objectives can be achieved through the following approach: An automatic optimization and training method for a UAV smart city inspection model includes: acquiring real-time data streams and model inference data generated during online UAV inspections; extracting multi-dimensional features based on the real-time data streams and model inference data to generate multi-dimensional performance degradation signals; inputting the multi-dimensional performance degradation signals into a preset root cause analysis model based on causal inference for joint analysis to obtain root cause analysis results characterizing the root causes of model performance degradation; selecting and combining dynamic optimization strategy combinations from a strategy library storing data optimization strategies, model optimization strategies, and deployment optimization strategies based on the root cause analysis results; executing the dynamic optimization strategy combinations to generate verification feedback data; and associating the verification feedback data with the root cause analysis results to update the root cause analysis model.

[0007] Optionally, generating the multidimensional performance degradation signal includes: extracting environment-related parameters from the real-time data stream to generate a first-dimensional environment feature signal; extracting model behavior-related parameters from the model inference data to generate a second-dimensional model behavior feature signal; and fusing the first-dimensional environment feature signal and the second-dimensional model behavior feature signal to generate the multidimensional performance degradation signal.

[0008] Optionally, the root cause analysis model includes: analyzing the causal relationship between the feature signals in the multidimensional performance degradation signal, locating the dominant and secondary features; matching the degradation pattern library used to store the correspondence between performance degradation scenarios and feature combinations according to the combination of dominant and secondary features, and determining the specific performance degradation scenario; and outputting the performance degradation scenario and its corresponding dominant and secondary features as the root cause analysis result.

[0009] Optionally, generating a dynamic optimization strategy combination includes: parsing the root cause analysis results and determining the strategy selection priority based on the root cause analysis results; selecting a first optimization strategy for the dominant feature and a second optimization strategy for the secondary feature from the strategy library storing data optimization strategies, model optimization strategies, and deployment optimization strategies according to the strategy selection priority; combining the first optimization strategy and the second optimization strategy, and configuring the execution order and resource constraints between the strategies to form a dynamic optimization strategy combination.

[0010] Optionally, the method further includes: the data optimization strategy includes initiating a supplementary data collection instruction; the model optimization strategy includes instructions for selectively fine-tuning specific network layers of the model and instructions for initiating a lightweight architecture search oriented towards resource constraints; the deployment optimization strategy includes instructions for adjusting the computing load distribution ratio between the cloud and the edge and instructions for issuing incremental model parameters.

[0011] Optionally, generating verification feedback data includes: monitoring the execution indicators of the UAV inspection task in real time within a preset period after executing the dynamic optimization strategy combination; the execution indicators include target detection accuracy, inference latency, and edge resource utilization; comparing the performance indicator differences before and after executing the optimization strategy, and calculating the indicator improvement rate; and summarizing the execution indicators, indicator improvement rate, and real-time working condition data of the inspection site to generate verification feedback data.

[0012] Optionally, updating the root cause analysis model includes: associating the model validation feedback data with the root cause analysis results and the dynamic optimization strategy combination to form a complete optimization case record; obtaining multiple optimization case records and calculating the reward score corresponding to each optimization based on the multiple optimization case records; and using the optimization case records and the corresponding reward scores to periodically train the root cause analysis model to enhance its ability to identify the mapping relationship between performance degradation patterns and effective strategy combinations.

[0013] Optionally, the method further includes: maintaining a strategy effectiveness score and historical call record for each optimization strategy in the strategy library; updating the strategy effectiveness score of each optimization strategy in the executed dynamic optimization strategy combination based on the verification feedback data; sorting and filtering the strategies in the strategy library according to the updated strategy effectiveness score, eliminating inefficient strategies, and introducing new candidate optimization strategies for frequently occurring performance degradation scenarios.

[0014] Optionally, the method further includes: when multiple drones perform inspection tasks in a related geographical area, summarizing the multidimensional performance degradation signals and root cause analysis results of each drone, and performing cross-device collaborative analysis; if multiple drones identify the same or related performance degradation root causes, generating a global collaborative optimization strategy, wherein the global collaborative optimization strategy includes coordinating the data acquisition tasks of multiple drones and uniformly distributing shared incremental model parameters; executing the global collaborative optimization strategy, and updating the root cause analysis model and strategy library based on the comprehensive verification feedback data of multiple drones.

[0015] Based on the same inventive concept, this invention also provides an automatic optimization and training system for a UAV smart city inspection model, the system comprising: The multi-dimensional feature perception module is used to acquire real-time data streams and model inference data generated during UAV online inspection, and to extract multi-dimensional features based on the real-time data streams and model inference data to generate multi-dimensional performance degradation signals. The root cause joint analysis module is used to input the multidimensional performance degradation signal into a preset root cause analysis model based on causal inference for joint analysis, and obtain root cause analysis results that characterize the root causes of model performance degradation. The dynamic strategy generation module is used to select and combine dynamic optimization strategy combinations from a strategy library containing data optimization strategies, model optimization strategies, and deployment optimization strategies based on the root cause analysis results. An automated production execution module is used to execute the dynamic optimization strategy combination and generate verification feedback data; The closed-loop feedback update module is used to associate the verification feedback data with the root cause analysis results and update the root cause analysis model.

[0016] Compared with the prior art, the present invention has the following advantages: 1. By constructing a closed-loop self-learning system, proactive and intelligent maintenance of the UAV inspection model was achieved. It can detect performance degradation signs caused by environmental changes or data drift in real time and intervene before the problem worsens, transforming passive fault repair into proactive performance assurance. This ensures high reliability and continuity of inspection tasks and avoids the loss of critical information due to model failure.

[0017] 2. A root cause analysis mechanism based on causal inference is introduced, surpassing the traditional extensive monitoring that relies solely on performance metric thresholds. This mechanism can analyze multi-dimensional performance degradation signals, pinpoint the root causes of model performance decline, and make subsequent optimization strategy selection more targeted. This diagnostic capability avoids blind, global model retraining, improving problem-solving efficiency and resource utilization.

[0018] 3. A self-evolving model and strategy management framework is proposed. By quantifying the effect of each optimization into feedback data and using it to periodically update the root cause analysis model and strategy library, it can continuously learn from historical experience, thereby enhancing its ability to identify complex decay patterns and judge the effectiveness of optimization strategies.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the automatic optimization and training method of the UAV smart city inspection model according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram illustrating the performance degradation scenario library matching similarity distribution of the automatic optimization and training method and system for the UAV smart city inspection model according to an embodiment of the present invention.

[0023] Figure 3 The diagram illustrates the evolution curve of the effectiveness score of the automatic tuning and training method and system optimization strategy of the UAV smart city inspection model according to an embodiment of the present invention with the number of executions.

[0024] Figure 4 This is a schematic diagram of the automatic optimization and training system for the UAV smart city inspection model according to an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Reference Figure 1 One embodiment of the present invention proposes an automatic tuning and training method for a smart city inspection model using unmanned aerial vehicles (UAVs). It employs multi-dimensional performance degradation signal perception, root cause analysis based on causal inference, dynamic optimization strategy generation, and closed-loop feedback update mechanism, which enables the model to continuously learn and adapt to evolve in complex dynamic environments.

[0027] The method described in this embodiment specifically includes: The system acquires real-time data streams and model inference data generated during online inspections of drones, and performs multi-dimensional feature extraction based on these data streams and data to generate multi-dimensional performance degradation signals. The multidimensional performance degradation signal is input into a preset root cause analysis model based on causal inference for joint analysis to obtain root cause analysis results that characterize the root cause of model performance degradation. Based on the root cause analysis results, a dynamic optimization strategy combination is generated by selecting and combining strategies from a strategy library that stores data optimization strategies, model optimization strategies, and deployment optimization strategies. Execute the dynamic optimization strategy combination to generate verification feedback data; The verification feedback data is correlated with the root cause analysis results to update the root cause analysis model.

[0028] Specifically, by integrating real-time data from the UAV's external environment with inference behavior data from the model's internal processes, a multi-dimensional degradation signal capable of comprehensively characterizing performance status is constructed, enabling the perception of model health. The core of this approach is the introduction of a root cause analysis model based on causal inference. This model transcends traditional correlation analysis, aiming to explore the fundamental causes of performance degradation, thereby achieving problem diagnosis. Based on this diagnostic result, the most targeted optimization scheme is dynamically selected and combined from a knowledge base encompassing multiple dimensions of strategies, including data, models, and deployment. After executing this scheme, the optimization effect is quantitatively evaluated to generate validation feedback data. Crucially, this feedback data is correlated with the original root cause analysis results, forming a complete learning sample used for iterative updates and training of the root cause analysis model. This allows its diagnostic capabilities to continuously improve through ongoing practice, forming a self-learning, self-adaptive, and self-optimizing intelligent closed loop. This proactively and promptly detects model performance degradation, transforming passive fault repair into proactive performance maintenance, ensuring the continuity and reliability of inspection tasks. By inferring causal relationships, the root causes of problems can be identified, avoiding blind trial-and-error optimization and improving the accuracy of optimization decisions and resource utilization efficiency. Furthermore, dynamically generated optimization strategy combinations can flexibly address various complex performance degradation scenarios, providing customized, multi-layered solutions. Finally, closed-loop feedback and continuous learning mechanisms enable the entire system to self-evolve, constantly adapting to changes in the urban environment and emerging problems, extending the model's lifespan, reducing long-term manual maintenance costs, and continuously improving the overall efficiency of intelligent drone inspections.

[0029] Optionally, generating the multidimensional performance degradation signal includes: Environmental parameters are extracted from the real-time data stream to generate a first-dimensional environmental feature signal. Extract model behavior-related parameters from the model inference data to generate a second-dimensional model behavior feature signal; The first-dimensional environmental feature signal is fused with the second-dimensional model behavior feature signal to generate a multi-dimensional performance degradation signal.

[0030] Specifically, data is captured in real-time from image sensors, LiDAR, and weather sensors mounted on the UAV at frequencies of 10 to 30 Hz, forming a real-time data stream. Environmental parameters are extracted from this real-time data stream. For example, illumination uniformity parameters are obtained by calculating the grayscale histogram entropy of image frames, motion blur is quantified using the average amplitude of the optical flow field between consecutive frames, and atmospheric visibility and wind speed are obtained by directly reading sensor values. These parameters are assembled into a time-synchronized environmental feature vector, i.e., the first-dimensional environmental feature signal, typically with 5 to 15 dimensions. In parallel, the inference monitoring module captures the output results of each forward propagation of the model on the edge computing unit, i.e., the model inference data. The model behavior feature extraction unit calculates model behavior-related parameters from this data. For example, it counts the proportion of target boxes with a confidence level below a preset threshold of 0.5% in the target detection results, calculates the information entropy of the output category probability distribution to measure the uncertainty of the prediction, and records the mean and variance of the activation values ​​of key network layers. These parameters are constructed into a model behavior feature vector that is time-aligned with the environmental feature vector, i.e., the second-dimensional model behavior feature signal, whose dimension is typically between 4 and 10. These two heterogeneous feature signals are then fused. The signal fusion module performs a feature-level fusion operation to generate the final multidimensional performance degradation signal. This operation is implemented in engineering using vector concatenation to ensure lossless preservation of the original information and avoid artificial biases introduced by features of different dimensions in the weighted fusion. This fusion process can be characterized by the following formula: ; in, This represents the final generated multidimensional performance degradation signal, a high-dimensional time-series vector that serves as the direct input to the subsequent root cause analysis model. F represents the vector concatenation and fusion function, which will... and Connect along the vector dimension. It is an environmental feature vector obtained by processing drone sensor data streams in real time. It is a model behavior feature vector obtained by analyzing the output data of the model inference. This fusion operation is performed periodically at a fixed frequency of 1 Hz to ensure the continuity and real-time performance degradation signal.

[0031] For example, during a drone-based power line inspection, the data interface module is set to a sampling frequency of 20Hz, capturing 20 frames of images and associated sensor data per second. The feature extraction engine processes the image stream: the calculated grayscale histogram entropy value of the image frame is 4.2, representing the illumination distribution. The motion blur amplitude of consecutive frames was calculated to be 12 pixels using optical flow field analysis, and the visibility of 800 meters, wind speed of 5 m / s, and ambient temperature of 25°C were directly read from the sensor. These five parameters constitute the first-dimensional environmental feature vector. In parallel, the inference monitoring module acquires the forward propagation data of the edge model, statistically determining that the proportion of target boxes with a confidence level below 0.5 is 12%. The information entropy of the output class probability distribution is calculated to be 0.65 to measure uncertainty. The mean of the key layer activation values ​​is extracted to be 0.48, and the variance is 0.15, forming the second-dimensional model behavior feature vector. The signal fusion module executes the vector concatenation function F at a fixed frequency of 1Hz. The calculation process involves sequentially concatenating two low-dimensional vectors. The 9-dimensional time-series vector is used as input for subsequent analysis. It is spliced ​​rather than weighted to avoid information overload when different physical dimensions are summed, and to ensure that the original features of each dimension have independent interpretability in the root cause model.

[0032] Optionally, the root cause analysis model includes: Analyze the causal relationships between the characteristic signals in the multidimensional performance degradation signal to identify the dominant and secondary features; Based on the combination of dominant and secondary features, a degradation pattern library is matched to the correspondence between storage performance degradation scenarios and feature combinations to determine specific performance degradation scenarios. The output performance degradation scenarios and their corresponding dominant and secondary features are used as the root cause analysis results.

[0033] Specifically, the causal relationships between various feature signals in the multidimensional performance degradation signal are analyzed. The root cause analysis model, at its core, is an inference engine based on a dynamic Bayesian network or a structured causal model. Upon receiving the multidimensional performance degradation signal, it constructs or updates a local directed acyclic graph (DAG) in real time, representing the causal influence between features. In this graph, nodes represent various features in the multidimensional performance degradation signal, directed edges represent causal relationships, and edge weights quantify the causal strength. By calculating the total directed path weights of each node in the graph to nodes exhibiting abnormal model behavior, such as low confidence ratios and high inference latency, dominant and secondary features are identified. Dominant features are environmental or model features that directly and strongly cause performance degradation; they are typically located in key upstream positions in the causal graph, with total path weights exceeding a preset threshold of 0.7. Secondary features are those with strong causal relationships to dominant features or those that indirectly affect performance metrics. Based on the identified combinations of dominant and secondary features, a matching process is performed in a pre-defined degradation pattern library to determine the specific performance degradation scenario. The attenuation pattern library is a structured knowledge base that stores the correspondence between hundreds of predefined performance degradation scenarios and feature combinations. For example, the dominant feature corresponding to the nighttime glare interference scenario is the proportion of highlight areas in the image, and the secondary feature is the gradient blur of the target edges. The matching process is completed by calculating the similarity between the currently identified feature combination and the feature combinations of each scenario in the library. This similarity can be calculated using the following formula: ; Where J represents the Jaccard similarity function, used to calculate the similarity between sets. and It is the set of dominant and secondary features obtained from the current analysis. and This is the feature set corresponding to a specific scene entry in the decay pattern library. 'a' and 'b' are weight coefficients, summing to 1. Typically, 'a' takes a value between 0.6 and 0.8 to emphasize the decisive role of the dominant feature. (Selection) The scenario with the highest score is identified as the most likely scenario of performance degradation. Finally, the model encapsulates the identified performance degradation scenarios and their corresponding dominant and secondary features into a structured data object, which is output as the root cause analysis result, providing a precise decision-making basis for subsequent dynamic strategy generation. Figure 2 The diagram illustrates the process of scene matching in a pre-set attenuation pattern library, simulating the root cause analysis model. The horizontal axis represents the similarity score. The figure clearly shows the similarity distribution of different candidate scenes, such as nighttime glare and wind tremors. The wind tremor scene has the highest score, exceeding the decision threshold, intuitively demonstrating how to accurately pinpoint the root cause of performance degradation from multi-dimensional signals.

[0034] For example, the root cause analysis model receives Then, the inference engine constructs a local directed acyclic graph. Motion-fuzzy nodes are then computed. To low confidence proportion nodes If the total directed path weight has a value of 0.82 or higher than the preset threshold of 0.7, then motion blur is determined to be the dominant feature. The wind speed associated with it is a secondary characteristic. The threshold was set to 0.7 based on a significance test to ensure sufficient causal strength to exclude interfering factors. Subsequently, a high-wind imaging jitter scene was matched against the attenuation mode library, whose feature set for this scene is as follows: The system utilizes the Jaccard similarity function. Calculate the matching degree: , The intersection represents wind speed, and the union represents wind speed and illuminance unevenness. Weights are assigned. , Emphasizing the role of dominant characteristics, substitute into the formula Since 0.85 is the highest score in the current library, the final output is that this scenario is the definite root cause.

[0035] Optionally, the combination of dynamic optimization strategies includes: Analyze the root cause analysis results and determine the strategy selection priority based on the root cause analysis results; Based on the strategy selection priority, a first optimization strategy for the dominant feature and a second optimization strategy for the secondary feature are selected from the strategy library that stores data optimization strategies, model optimization strategies and deployment optimization strategies. The first optimization strategy and the second optimization strategy are combined, and the execution order and resource constraints between the strategies are configured to form a dynamic optimization strategy combination.

[0036] Specifically, the root cause analysis results are analyzed to determine the priority of strategy selection. Upon receiving the structured root cause analysis results, the dynamic strategy generation module immediately parses out the performance degradation scenario, dominant features, and secondary features. Based on the conclusions of causal inference, the strategy selection priority is automatically set: optimization measures for dominant features have the highest execution priority, followed by optimization measures for secondary features, to ensure that the root cause of the problem is addressed first. Based on this priority, optimization strategies are selected from the strategy library. The strategy library is a pre-built knowledge base where each strategy is associated with its applicable features, expected effects, resource costs, and historical effectiveness scores. For the dominant feature, the first optimization strategy is retrieved and selected from the strategy library. This selection process is not a simple matching but is determined through a comprehensive utility scoring function to ensure that the optimal strategy for the current situation is selected. Similarly, a second optimization strategy is selected for the secondary features. This selection process can be described as follows: ; Where O represents the candidate optimization strategy in the strategy library, These are the first and second optimization strategies, respectively. F represents the feature to be solved, i.e., the dominant or secondary feature. This represents the state vector of currently available computing and communication resources for the drone, including edge processor utilization, available memory, and downlink bandwidth. S is a utility scoring function that comprehensively evaluates the relevance of policy O to feature F, its historical success rate, and the resources required to execute the policy compared to currently available resources. The matching degree is considered. For example, when the dominant feature is image motion blur, the scoring function may assign a high score to the model optimization strategy of initiating a resource-constrained lightweight architecture search. However, when the secondary feature is generally low object detection confidence, the deployment optimization strategy of issuing incremental model parameters may be chosen. The selected first optimization strategy is combined with the second optimization strategy and fine-tuned. The combination configuration engine first determines the execution order between strategies; for example, data optimization strategies are usually executed before model optimization strategies. Then, specific resource constraints are configured for this strategy combination; for example, it is stipulated that the model fine-tuning task must be started when the UAV's computing load is below 75%, and the download rate of the incremental parameter package cannot exceed 2 megabits per second to avoid affecting normal inspection tasks. After the above steps, a dynamic optimization strategy combination containing specific operation instructions, execution order, and resource constraints is finally formed and issued to the automated production execution module.

[0037] For example, by analyzing root cause results, a strategy for motion blurring of dominant features can be determined. It is given the highest priority. The strategy library is retrieved and the utility scoring function is invoked. ,in The correlation is 0.9. With a historical success rate of 0.85, A resource utilization score of 1.2 is assigned, and the inverse proportionality is used to reflect the constraint. The calculation yields... Ultimately, the first strategy was selected. For lightweight architecture search, the second strategy Incremental parameters are issued. The combined configuration engine is configured with the following execution logic: lightweight model switching is prioritized. Simultaneously, the current drone status is considered. With a 60% CPU usage constraint, model fine-tuning can only begin when CPU usage is below 75%, and the incremental packet download rate is limited to 1.5 Mbps. This resource-constrained argmax search algorithm ensures the executability of the strategy and avoids interruptions to the inspection task due to excessive resource consumption during the optimization process.

[0038] Optionally, the method further includes: The data optimization strategy includes initiating supplementary data collection commands; The model optimization strategy includes instructions for selectively fine-tuning specific network layers of the model and instructions for initiating a resource-constrained lightweight architecture search. The deployment optimization strategy includes instructions to adjust the computing load distribution ratio between the cloud and the edge, and instructions to issue incremental model parameters.

[0039] Specifically, the core of the data optimization strategy is initiating a supplementary data acquisition command. When root cause analysis results indicate data distribution drift, such as identifying performance degradation scenarios as newly emerging urban facilities or target deformation under rare weather conditions, the dynamic strategy generation module generates this command. This command is a data structure encapsulating specific task requirements, including the geofence coordinates of the target acquisition area, the target category labels to be prioritized, and recommended acquisition time windows and sensor configuration parameters, such as the image sensor's exposure compensation value should be between -1.0 and -0.5 to cope with strong light. This command is sent to the UAV mission planning system, triggering a dedicated data acquisition flight mission to obtain high-value, difficult samples, directly supplementing the model's cognitive blind spots from the source.

[0040] The model optimization strategy comprises two key instructions. The first is an instruction for selectively fine-tuning specific network layers of the model. This instruction is triggered when root cause analysis results indicate a decline in the model's ability to recognize specific features, such as the dominant feature being small, ambiguous targets. It specifies freezing the parameters of the first 80% of the model's backbone network, retraining only the terminal network layers responsible for high-level semantic feature extraction and the classification head. This process uses a small amount of newly acquired supplementary data and is performed for 5 to 15 epochs on edge computing units or near-end servers, with a learning rate set at a low level of 1e-5 to 1e-4. The aim is to quickly adapt to new features with low resource consumption, rather than reconstructing the entire model's knowledge system. The second instruction is to initiate a resource-constrained lightweight architecture search. This high-cost strategy is initiated when root cause analysis reveals a long-term mismatch between the model's inherent complexity and edge computing power, leading to persistently excessive inference latency. This command triggers a cloud-based neural network architecture search service. Within a pre-defined lightweight operator library, it searches for the optimal sub-network structure, constrained by edge inference latency of less than 50 milliseconds and model size of less than 100 megabytes, ultimately generating a new model architecture.

[0041] The deployment optimization strategy focuses on runtime resource allocation and model delivery. The first type of instruction adjusts the computing load distribution ratio between the cloud and edge. This instruction is issued when root cause analysis identifies a sudden surge in high-density targets causing edge computing overload. The instruction includes dynamically adjusted load distribution coefficients; for example, downgrading the edge task from complete detection and recognition to only extracting high-recall target candidate boxes, and then batch-uploading the extracted image slices to the cloud server for subsequent refined recognition by a high-performance model in the cloud. This ensures task continuity while mitigating computational peaks at the edge. The second type of instruction issues incremental model parameters. This instruction is a delivery method after model fine-tuning or architecture updates. It calculates the difference matrix between the old and new model parameters, generating an incremental parameter package typically only 1% to 10% the size of the full model. This parameter package is then sent to the drone via this instruction. Upon receiving it, the deployment management module on the drone applies the incremental parameters to the existing model during non-task intervals, updating the model.

[0042] For example, when root cause analysis points to data drift caused by rare weather, the data optimization strategy generates supplementary collection instructions, specifying the geofence coordinates as... The range was adjusted, and the image sensor exposure compensation was set to -0.8 to handle strong light. For the model optimization strategy, specific network layers were fine-tuned, and the freeze ratio was calculated: for a 100-layer model, the first 80 layers (the first 80%) were frozen, and only the parameters of the last 20 layers were updated. The learning rate was set to... The process is executed for 10 cycles. During fine-tuning, if the inference latency continues to exceed the limit, a lightweight architecture search is initiated, requiring the new model size to be less than 100MB and the latency to be less than 50ms13. The deployment optimization strategy is to generate an incremental parameter package of 4MB for updating by calculating the difference matrix of the weights of the old and new models. These fine-tuned parameter settings ensure that the model can iterate quickly in a low-power environment at the edge.

[0043] Optionally, the generation of verification feedback data includes: Within a preset period after executing the dynamic optimization strategy combination, the execution indicators of the UAV inspection task are monitored in real time, including target detection accuracy, inference latency, and edge resource utilization. Compare the performance metrics before and after implementing the optimization strategy, and calculate the improvement rate of the metrics. The execution indicators, indicator improvement rates, and real-time operating data from the inspection site are aggregated to generate verification feedback data.

[0044] Specifically, within a preset cycle after executing the dynamic optimization strategy combination, the performance evaluation unit within the automated production execution module monitors the execution metrics of the UAV inspection task in real time. This preset cycle is typically set to cover 2 to 5 complete inspection tasks, or a cumulative run of 60 to 180 minutes, to ensure the collected data is statistically significant. The monitored execution metrics mainly include three dimensions: target detection accuracy, obtained through a bypass verification mechanism with a sampling rate of 0.5% to 1%, by sending some real-time images and model results to the cloud for comparison with a higher-precision benchmark model or manual annotation; inference latency, measured in milliseconds, recording the end-to-end time taken for the model to complete one forward propagation at the edge; and edge resource utilization, by capturing real-time GPU or NPU utilization, memory usage, and power consumption data of the edge computing unit. The performance metrics before and after executing the optimization strategy are compared to calculate the metric improvement rate. To this end, before executing the optimization, the moving average of the above execution metrics from the most recent evaluation cycle, such as the past 24 hours, is cached as a baseline. The metric improvement rate is calculated using the following formula: ; In this formula, This represents the improvement rate of the i-th performance indicator. It is the average value recorded by this indicator during the monitoring period after the optimization strategy is implemented. This is the baseline average of the corresponding metric cached before the optimization strategy is implemented. For example, for object detection accuracy, a positive... This indicates a performance improvement; however, for inference latency or resource utilization, a negative result indicates a performance improvement. This indicates that the optimization has achieved positive results, meaning that the resources or time consumed have been reduced. All evaluation information is aggregated to generate structured verification feedback data. The raw time-series data of the aforementioned performance indicators, the calculated improvement rates of each indicator, and the multi-dimensional performance degradation signals recorded synchronously with real-time operating data from the inspection site throughout the entire post-optimization monitoring cycle are packaged together. This complete verification feedback data is assigned a unique transaction ID and strongly correlated with the root cause analysis results that triggered this optimization and the dynamic optimization strategy combination executed, forming a complete and traceable record of optimization cases.

[0045] For example, evaluation metrics are obtained through 1% bypass validation sampling within a 120-minute monitoring period after strategy execution. The system retrieves the baseline average value from the 24 hours prior to optimization. Accuracy was 0.82, inference latency was 55ms, and resource usage was 60%. Optimized monitoring averages. Accuracy was 0.86, inference latency was 45ms, and resource usage was 65%. Improvement rate of computational metrics. Accuracy improvement rate Positive improvement; latency reduction rate A negative value indicates reduced time consumption, which is a positive effect; resource improvement rate The summary module will summarize these. Numerical values ​​along with real-time operating condition feature vectors Package it and assign it the transaction ID "TRANS-20251231" to provide a quantitative basis for subsequent closed-loop reward calculation.

[0046] Optionally, updating the root cause analysis model includes: The model validation feedback data is correlated with the root cause analysis results and the dynamic optimization strategy combination to form a complete optimization case record; Obtain multiple optimization case records, and calculate the reward score corresponding to each optimization based on the multiple optimization case records; The root cause analysis model is periodically trained using the optimized case records and corresponding reward scores to enhance its ability to identify the mapping relationship between performance degradation patterns and effective strategy combinations.

[0047] Specifically, scattered information is integrated into structured knowledge. The validation feedback data, the root cause analysis results that triggered the optimization, and the final dynamic optimization strategy are combined and strongly linked using a unique optimization event ID. This bound data set is formatted and stored in a long-term case database, forming a complete optimization case record, providing standardized input samples for subsequent quantitative analysis and model training. Based on the accumulated optimization case records, the quantitative effectiveness of each optimization intervention, i.e., the reward score, is calculated. When the number of newly added optimization case records in the database reaches a preset batch, such as 100, the reward calculation engine is triggered. This engine iterates through each new record, extracts the improvement rate of multi-dimensional indicators from the validation feedback data, and calculates a comprehensive reward score through weighted summation based on preset business objectives. The calculation process is as follows: ; Where R is the final reward score, a scalar value. , and The metrics representing the improvement rates of target detection accuracy, inference latency, and edge resource utilization are obtained directly from the validation feedback data. , and These are the corresponding weight coefficients, which sum to 1. These weights are configured by system operators based on the core requirements of the current inspection task, such as prioritizing high real-time performance or high accuracy. The negative sign in the formula is used to convert the reduction in latency and resource consumption—a negative improvement rate—into a positive reward contribution. The calculated reward score is backfilled into the corresponding optimized case record, serving as the final effect label for the complete diagnosis-decision-execution closed loop. Using these optimized case records with reward scores, the root cause analysis model is periodically trained. The training task is set to be executed periodically, for example, starting whenever 500 new case records with reward scores are collected. The training process uses a reinforcement learning framework, treating the root cause analysis model as an agent, multidimensional performance degradation signals as states, and the output root cause analysis results as actions. The objective function of the training is to maximize the long-term cumulative reward. Using algorithms such as policy gradient or Q-learning, the parameters of the causal inference network within the root cause analysis model are iteratively optimized using the optimized case records and their corresponding reward scores. The core of this training process lies in enhancing the model to accurately map specific performance degradation signal patterns to the probability of a root cause analysis result that has been historically verified to produce high reward scores, thereby continuously improving its ability to identify the deep mapping relationship between performance degradation patterns and effective strategy combinations.

[0048] For example, the reward calculation engine is triggered after accumulating 100 optimization cases. (Set business weights.) Prioritize characterization accuracy. , Substitute the data from the previous example to calculate the overall reward score. The calculation process is as follows: This is for... and The negative sign is used because a decrease in these two metrics represents a positive contribution. When 500 samples with reward labels have been accumulated, the training module updates the causal network parameters using the policy gradient algorithm. This process will... The model treats the state as a state, the model's diagnostic result as an action, and the R value as feedback. Through this reinforcement learning mechanism, the model continuously optimizes the weights of its internal nodes, enabling it to output root cause diagnostic conclusions that previously earned high reward scores with a higher probability when faced with similar signal patterns in the future.

[0049] Optionally, the method further includes: The strategy library maintains a strategy effectiveness score and historical call record for each optimization strategy. Based on the verification feedback data, update the strategy effectiveness score of each optimization strategy in the executed dynamic optimization strategy combination; Based on the updated strategy effectiveness score, the strategies in the strategy library are sorted and filtered, inefficient strategies are eliminated, and new candidate optimization strategies are introduced for frequently occurring performance degradation scenarios.

[0050] Specifically, in the strategy library database, two key maintenance fields are added for each optimization strategy: a strategy effectiveness score and a JSON-formatted historical call record. The strategy effectiveness score is a floating-point number, initialized to 0.5, used to quantify the long-term performance of the strategy; the historical call record stores the timestamp of each execution of the strategy, the associated performance degradation scenarios, and the final reward score. When a dynamic optimization strategy combination is executed and generates verification feedback data containing a comprehensive reward score, the strategy update service is triggered. This service parses each atomic optimization strategy in the combination and independently updates their respective strategy effectiveness scores. The update algorithm uses an exponential moving average method to smooth out the randomness of a single execution and more robustly reflect the long-term trend of the strategy. Its update formula is as follows: ; in, This is the updated strategy effectiveness score. This is the strategy's score before the update. R is the comprehensive reward score for this optimization event extracted from the validation feedback data, with its value normalized to between 0 and 1. 'a' is a learning rate constant between 0.1 and 0.3, which determines the weight of the latest execution result on the overall score. Simultaneously, detailed information about this execution is appended to the strategy's historical call record. Based on the updated score, the strategy library is periodically cleaned and optimized, for example, every 24 hours. First, all strategies in the library are sorted and filtered in descending order based on their strategy effectiveness scores. This sorting result directly affects the priority of the dynamic strategy generation module when selecting strategies, making high-scoring strategies more likely to be selected. Second, a cleanup process is executed, which marks and archives strategies whose effectiveness scores have remained below a preset threshold of 0.2 for 30 consecutive days and have been called more than 50 times, defining them as long-term inefficient strategies. Finally, by analyzing historical call records, the top 5% of frequently occurring performance degradation scenarios are identified. If the average reward score for these scenarios is below 0.4, a strategy optimization suggestion ticket is generated, domain experts are notified to intervene, failure cases are analyzed, new candidate optimization strategies are designed, and manually added to the strategy library, thereby achieving iterative expansion of strategy knowledge. For example... Figure 3 The diagram illustrates the evolution of the effectiveness score of a single optimization strategy in the strategy library as the execution rounds change, demonstrating how an adaptive scoring mechanism can automatically filter efficient strategies and eliminate long-term inefficient operation instructions.

[0051] For example, when an event containing a lightweight search strategy generates a normalized reward score R=0.75, the strategy update service updates the score using an exponential moving average method. Let the original strategy score be... With a learning rate a = 0.2, calculate the new score. The system executes a cleanup procedure every 24 hours. If a strategy's score is below 0.2 for 30 consecutive days and it is called more than 50 times, it is marked as an inefficient strategy and archived. Simultaneously, analysis revealed that the average reward for reflective interference in the top 5% of high-frequency scenes was only 0.35. A work order was then generated to notify experts to intervene and design new strategies, such as introducing polarization filtering for data acquisition, which were then added to the database. This dynamic maintenance mechanism ensures that the strategy database always retains the most effective solutions.

[0052] Optionally, the method further includes: When multiple drones perform inspection tasks in a related geographical area, the multi-dimensional performance degradation signals and root cause analysis results of each drone are aggregated for cross-device collaborative analysis. If multiple drones identify the same or related root causes of performance degradation, a global collaborative optimization strategy is generated. This strategy includes coordinating the data acquisition tasks of multiple drones and uniformly distributing shared incremental model parameters. The global collaborative optimization strategy is executed, and the root cause analysis model and strategy library are updated based on the comprehensive verification feedback data from multiple UAVs.

[0053] Specifically, when multiple drones, such as 3 to 5, perform inspection tasks within a geographically related area (e.g., a radius of 5 to 10 square kilometers) at the same time window, the collaborative analysis mechanism is activated. At this time, the multi-dimensional performance degradation signals of each drone and their respective root cause analysis results are aggregated and transmitted to the cloud-based collaborative analysis engine. The core task of this collaborative analysis engine is to perform cross-device collaborative analysis to determine if common problems exist. The engine first uses a spatiotemporal clustering algorithm to process the aggregated root cause analysis results. If multiple drones identify the same or semantically related performance degradation root causes, such as image noise under severe weather conditions or glare interference from specific types of buildings, a common degradation event is determined. This correlation determination is quantified by a collaborative similarity score. ; in, The co-similarity of the performance degradation events of the two drones. It is the semantic similarity of the root cause analysis results of the two, calculated by comparing the IDs or tags in the decay pattern library. It is the cosine similarity of the multidimensional performance degradation signal vectors of the two, reflecting the consistency of the underlying data features. and These are the weighting coefficients, which sum to 1, typically It is given a higher weight, approximately 0.7, to highlight the consistency of diagnostic conclusions. When the average weight of a drone swarm is... When the threshold of 0.8 is exceeded, a common root cause is identified, triggering the generation of a global collaborative optimization strategy. Once the common root cause is identified, a global collaborative optimization strategy is generated. This strategy aims to achieve a globally optimal solution through resource integration and task collaboration. For example, the strategy coordinates data collection tasks from multiple drones, avoiding redundant collection and instead assigning different collection angles or sensor parameters to each drone, collectively constructing a more diverse and higher-quality supplementary dataset targeting the common problem. Simultaneously, for model updates, this aggregated dataset is used for centralized model fine-tuning in the cloud, generating a shared incremental model parameter package, which is then uniformly distributed to all affected drones. The global collaborative optimization strategy is executed, and during the evaluation period following strategy execution, validation feedback data from all relevant drones is collected. This data is aggregated into comprehensive validation feedback data, used to calculate a global reward score representing the overall improvement effect of the cluster.

[0054] For example, four drones within a 7-kilometer radius simultaneously reported performance degradation. The root cause results and signal vectors of two drones were extracted. Semantic similarity was set. All were diagnosed as heavy fog; signal vector cosine similarity. Set weights , Calculate collaborative similarity .because The existence of regional common root causes was confirmed. A global optimization strategy was generated: drone A was coordinated to collect high-altitude distant views, while drones B, C, and D collected low-altitude details. After aggregating the data from all drones in the cloud, a unified incremental model package was generated through joint fine-tuning and distributed to the four drones. During the final evaluation period, the average improvement rate of each drone's indicators was calculated to obtain the global reward score. This cross-device collaboration reduced the computational redundancy of independent optimization by a single drone and improved the generalization accuracy of all models within the region by utilizing the consistency of group data.

[0055] Reference Figure 4 Based on the same inventive concept, this invention also provides an automatic optimization and training system for a UAV smart city inspection model, the system comprising: The multi-dimensional feature perception module is used to acquire real-time data streams and model inference data generated during UAV online inspection, and to extract multi-dimensional features based on the real-time data streams and model inference data to generate multi-dimensional performance degradation signals. The root cause joint analysis module is used to input the multidimensional performance degradation signal into a preset root cause analysis model based on causal inference for joint analysis, and obtain root cause analysis results that characterize the root causes of model performance degradation. The dynamic strategy generation module is used to select and combine dynamic optimization strategy combinations from a strategy library containing data optimization strategies, model optimization strategies, and deployment optimization strategies based on the root cause analysis results. An automated production execution module is used to execute the dynamic optimization strategy combination and generate verification feedback data; The closed-loop feedback update module is used to associate the verification feedback data with the root cause analysis results and update the root cause analysis model.

[0056] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0057] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. An automatic optimization and training method for a UAV smart city inspection model, characterized in that, include: The system acquires real-time data streams and model inference data generated during online inspections of drones, and performs multi-dimensional feature extraction based on these data streams and data to generate multi-dimensional performance degradation signals. The multidimensional performance degradation signal is input into a preset root cause analysis model based on causal inference for joint analysis to obtain root cause analysis results that characterize the root cause of model performance degradation. Based on the root cause analysis results, a dynamic optimization strategy combination is generated by selecting and combining strategies from a strategy library that stores data optimization strategies, model optimization strategies, and deployment optimization strategies. Execute the dynamic optimization strategy combination to generate verification feedback data; The verification feedback data is correlated with the root cause analysis results to update the root cause analysis model.

2. The automatic optimization and training method for the UAV smart city inspection model according to claim 1, characterized in that, The generation of the multidimensional performance degradation signal includes: Environmental parameters are extracted from the real-time data stream to generate a first-dimensional environmental feature signal. Extract model behavior-related parameters from the model inference data to generate a second-dimensional model behavior feature signal; The first-dimensional environmental feature signal is fused with the second-dimensional model behavior feature signal to generate a multi-dimensional performance degradation signal.

3. The automatic optimization and training method for the UAV smart city inspection model according to claim 1, characterized in that, The root cause analysis model includes: Analyze the causal relationships between the characteristic signals in the multidimensional performance degradation signal to identify the dominant and secondary features; Based on the combination of dominant and secondary features, a degradation pattern library is matched to the correspondence between storage performance degradation scenarios and feature combinations to determine specific performance degradation scenarios. The output performance degradation scenarios and their corresponding dominant and secondary features are used as the root cause analysis results.

4. The automatic optimization and training method for the UAV smart city inspection model according to claim 3, characterized in that, The combination of dynamic optimization strategies includes: Analyze the root cause analysis results and determine the strategy selection priority based on the root cause analysis results; Based on the strategy selection priority, a first optimization strategy for the dominant feature and a second optimization strategy for the secondary feature are selected from the strategy library that stores data optimization strategies, model optimization strategies and deployment optimization strategies. The first optimization strategy and the second optimization strategy are combined, and the execution order and resource constraints between the strategies are configured to form a dynamic optimization strategy combination.

5. The automatic optimization and training method for the UAV smart city inspection model according to claim 4, characterized in that, The method further includes: The data optimization strategy includes initiating supplementary data collection commands; The model optimization strategy includes instructions for selectively fine-tuning specific network layers of the model and instructions for initiating a resource-constrained lightweight architecture search. The deployment optimization strategy includes instructions to adjust the computing load distribution ratio between the cloud and the edge, and instructions to issue incremental model parameters.

6. The automatic optimization and training method for the UAV smart city inspection model according to claim 1, characterized in that, The generated verification feedback data includes: Within a preset period after executing the dynamic optimization strategy combination, the execution indicators of the UAV inspection task are monitored in real time, including target detection accuracy, inference latency, and edge resource utilization. Compare the performance metrics before and after implementing the optimization strategy, and calculate the improvement rate of the metrics. The execution indicators, indicator improvement rates, and real-time operating data from the inspection site are aggregated to generate verification feedback data.

7. The automatic optimization and training method for the UAV smart city inspection model according to claim 1, characterized in that, The update of the root cause analysis model includes: The model validation feedback data is correlated with the root cause analysis results and the dynamic optimization strategy combination to form a complete optimization case record; Obtain multiple optimization case records, and calculate the reward score corresponding to each optimization based on the multiple optimization case records; The root cause analysis model is periodically trained using the optimized case records and corresponding reward scores to enhance its ability to identify the mapping relationship between performance degradation patterns and effective strategy combinations.

8. The automatic optimization and training method for the UAV smart city inspection model according to claim 1, characterized in that, The method further includes: The strategy library maintains a strategy effectiveness score and historical call record for each optimization strategy. Based on the verification feedback data, update the strategy effectiveness score of each optimization strategy in the executed dynamic optimization strategy combination; Based on the updated strategy effectiveness score, the strategies in the strategy library are sorted and filtered, inefficient strategies are eliminated, and new candidate optimization strategies are introduced for frequently occurring performance degradation scenarios.

9. The automatic optimization and training method for the UAV smart city inspection model according to claim 1, characterized in that, The method further includes: When multiple drones perform inspection tasks in a related geographical area, the multi-dimensional performance degradation signals and root cause analysis results of each drone are aggregated for cross-device collaborative analysis. If multiple drones identify the same or related root causes of performance degradation, a global collaborative optimization strategy is generated. This strategy includes coordinating the data acquisition tasks of multiple drones and uniformly distributing shared incremental model parameters. The global collaborative optimization strategy is executed, and the root cause analysis model and strategy library are updated based on the comprehensive verification feedback data from multiple UAVs.

10. An automatic optimization and training system for a UAV smart city inspection model, applied to the automatic optimization and training method for a UAV smart city inspection model as described in any one of claims 1-9, characterized in that, The system includes: The multi-dimensional feature perception module is used to acquire real-time data streams and model inference data generated during UAV online inspection, and to extract multi-dimensional features based on the real-time data streams and model inference data to generate multi-dimensional performance degradation signals. The root cause joint analysis module is used to input the multidimensional performance degradation signal into a preset root cause analysis model based on causal inference for joint analysis, and obtain root cause analysis results that characterize the root causes of model performance degradation. The dynamic strategy generation module is used to select and combine dynamic optimization strategy combinations from a strategy library containing data optimization strategies, model optimization strategies, and deployment optimization strategies based on the root cause analysis results. An automated production execution module is used to execute the dynamic optimization strategy combination and generate verification feedback data; The closed-loop feedback update module is used to associate the verification feedback data with the root cause analysis results and update the root cause analysis model.

Images