Cargo emergency buffer method and system under logistics unmanned aerial vehicle rotor failure

Abstract

This invention discloses an emergency cargo buffering method and system for handling rotor failures in logistics drones, relating to the field of cargo protection technology under logistics drone malfunctions. The method includes: acquiring multi-source data after a cargo box is impacted; outputting a cargo damage probability value through a pre-trained damage probability inference model; calculating weights based on cargo attributes, environmental conditions, and data confidence levels, and weighted fusion to generate an initial threshold adjustment; fitting the short-term change slope of the damage probability using the least squares method, and combining this with a data confidence level screening mechanism to remove low-confidence historical data; dynamically correcting the threshold adjustment based on the deviation coefficient between the current data confidence level and the historical mean, and obtaining the final damage probability threshold through fusion; if the current damage probability value exceeds the threshold, executing the final protection strategy. This invention achieves multi-source information fusion evaluation and dynamic threshold adjustment, improving the accuracy and anti-interference capability of cargo protection.

Description

Technical Field

[0001] This invention relates to the field of cargo protection technology under logistics drone malfunctions, and more specifically, this application relates to an emergency cargo buffering method and system under the condition of a logistics drone rotor malfunction. Background Technology

[0002] With the development of drone logistics technology, its application in express delivery, medical supply distribution, and special transportation is becoming increasingly widespread. However, drones may lose control and crash during flight due to mechanical failures, severe weather, or external collisions. How to effectively protect the goods they carry and prevent them from being damaged by impacts has become a problem to be solved.

[0003] Most existing cargo protection solutions are emergency systems based on a single trigger condition. These systems involve installing accelerometers or altimeters on the drone. When the drone is detected to be in free fall (e.g., acceleration exceeding a threshold and at a high altitude), a simple protection mechanism is triggered, such as deploying a parachute or a single airbag within the cargo hold. The advantage of this type of existing technology is its simple system structure, rapid response, and ability to slow the descent of the drone and its cargo to some extent.

[0004] However, its protection mechanism is "one-time". Once triggered, regardless of the protection effect, the system cannot make subsequent adjustments and optimizations. This results in limited protection effect and poor system flexibility. It cannot cope with complex environments and cannot make subsequent precise adjustments based on the execution effect of the protection action. For high-value goods such as precision instruments and fragile items, the success rate of its buffer protection is still relatively low. Therefore, an emergency buffering method and system for cargo under the failure of the rotor of a logistics drone is proposed to solve this problem. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an emergency cargo buffering method and system for handling rotor failures in logistics drones. This technical solution resolves the issues raised in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this application provides an emergency cargo buffering method in the event of a rotor failure in a logistics drone, the method comprising:

[0008] Acquire multi-source data of the cargo container after it is impacted while it is currently in the initial protection strategy state. The multi-source data includes sensor data streams, cargo attributes and environmental status. The initial protection strategy is triggered by the cargo container's built-in system when it detects a drone malfunction.

[0009] The sensor data stream is preprocessed and features are extracted, then input into a pre-trained damage probability inference model, which outputs the damage probability value and confidence level of the goods in the current cargo container.

[0010] The weighting coefficients are calculated based on the cargo attributes, environmental conditions, and damage probability values, and the initial threshold adjustment amount is obtained through weighted fusion.

[0011] Obtain historical damage probability values ​​and corresponding confidence levels, and calculate the slope and mean confidence level representing short-term changes in historical damage probability values ​​based on least squares fitting.

[0012] If the slope is greater than a preset trend threshold and the mean confidence level is less than a preset confidence threshold, then after removing historical damage probability values ​​with a confidence level less than the confidence threshold, the mean confidence level of the remaining historical damage probability values ​​is recalculated.

[0013] Calculate the deviation between the current probability of damage and the recalculated mean confidence level, normalize it to obtain the deviation coefficient, and multiply the initial threshold adjustment by the difference between 1 and the deviation coefficient to obtain the final threshold adjustment.

[0014] The final threshold adjustment is added to the initial damage probability threshold to obtain the adjusted damage probability threshold;

[0015] If the damage probability value is greater than the adjusted damage probability threshold, then the final protection strategy is generated and executed.

[0016] Secondly, this application provides an emergency cargo buffering system for handling rotor failures of a logistics drone, used to implement the emergency cargo buffering method for handling rotor failures of a logistics drone as described in any of the above claims. The system is integrated within the cargo box and includes:

[0017] The data acquisition module is used to acquire multi-source data of the cargo box after it is impacted while it is currently in the initial protection strategy state. The multi-source data includes sensor data streams, cargo attributes and environmental status. The initial protection strategy is triggered by the cargo box's built-in system when it detects a drone malfunction.

[0018] The damage probability model processing module is used to preprocess and extract features from the sensor data stream, input it into the pre-trained damage probability inference model, and output the damage probability value and confidence level of the goods in the current cargo box.

[0019] The threshold adjustment amount acquisition module is used to calculate the weighting coefficients based on the cargo attributes, environmental conditions and damage probability values, and obtain the initial threshold adjustment amount through weighted fusion.

[0020] Obtain historical damage probability values ​​and corresponding confidence levels, and calculate the slope and mean confidence level representing short-term changes in historical damage probability values ​​based on least squares fitting.

[0021] If the slope is greater than a preset trend threshold and the mean confidence level is less than a preset confidence threshold, then after removing historical damage probability values ​​with a confidence level less than the confidence threshold, the mean confidence level of the remaining historical damage probability values ​​is recalculated.

[0022] Calculate the deviation between the current probability of damage and the recalculated mean confidence level, normalize it to obtain the deviation coefficient, and multiply the initial threshold adjustment by the difference between 1 and the deviation coefficient to obtain the final threshold adjustment.

[0023] The damage probability threshold adjustment module is used to add the final threshold adjustment amount to the initial damage probability threshold to obtain the adjusted damage probability threshold.

[0024] The final protection strategy generation module is used to generate and execute a final protection strategy if the damage probability value is greater than the adjusted damage probability threshold.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] This application solves the problem of traditional solutions relying on single data by integrating multi-source information such as sensor data streams, cargo attributes, and environmental conditions, and inputting it into a pre-trained damage probability inference model. It enables a comprehensive and accurate assessment of the risk of cargo damage after the cargo box is impacted, and provides a data foundation for solving the problem of setting fixed threshold values.

[0027] This application calculates weights based on multi-source data and performs weighted fusion to generate an initial threshold adjustment amount. It also introduces a short-term change slope fitting and data confidence discrimination mechanism based on the least squares method to effectively eliminate interference from low-confidence historical data. This solves the problems of high noise and poor reliability of sensor data during the crash, achieves stable calculation of the threshold adjustment amount, and enhances the anti-interference ability of the system.

[0028] This application dynamically generates a deviation coefficient by calculating the deviation coefficient between the current data confidence level and the historical confidence level after filtering, and then refines the threshold adjustment accordingly. This solution solves the problem that traditional fixed thresholds cannot respond to real-time risk changes, and improves the protection strategy triggering logic from "static preset" to "dynamic intelligent adjustment". Ultimately, it achieves the dual optimization goal of avoiding false triggers while executing the final protection strategy in high-risk situations, thereby improving the success rate of protecting cargo safety. Attached Figure Description

[0029] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0030] Figure 1This is a flowchart of the emergency cargo buffering method under rotor failure of a logistics drone proposed in this invention;

[0031] Figure 2 This is a flowchart of the method for obtaining the initial damage probability threshold in the emergency cargo buffering method under rotor failure of a logistics drone proposed in this invention;

[0032] Figure 3 This is a data flow diagram of the emergency cargo buffering method under rotor failure of a logistics drone proposed in this invention;

[0033] Figure 4 This is a structural block diagram of the emergency cargo buffer system for a logistics drone in case of rotor failure, as proposed in this invention. Detailed Implementation

[0034] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0035] In existing technologies, emergency cargo protection for logistics drones suffers from problems such as a single judgment dimension and a lack of dynamic feedback and adaptive optimization capabilities. It typically employs a fixed threshold triggering mechanism based on a single sensor. Once a protection action is triggered, regardless of the actual protection effect, the system cannot reassess or adjust its strategy based on subsequent impacts, cargo status, and environmental changes. This lag in decision-making not only reduces protection accuracy but may also lead to damage to high-value cargo due to the failure to optimize buffering strategies in real time.

[0036] To address the aforementioned issues, this application acquires multi-source data (including sensor data streams, cargo attributes, and environmental conditions) after a cargo container is impacted, inputs this data into a pre-trained damage probability inference model, and calculates the cargo damage probability value in real time. Based on data confidence, a weighted fusion is performed to generate a threshold adjustment amount, and this adjustment amount is dynamically optimized through historical data trend analysis and time-series confidence assessment. Finally, an adaptively adjusted damage probability threshold is generated through a multi-level threshold fusion mechanism. When the real-time damage probability value exceeds this threshold, the final protection strategy is generated and executed. This application, through multi-round iterative optimization of multi-source data fusion, real-time confidence assessment, and dynamic threshold adjustment, ensures the accuracy of the protection strategy triggering conditions, overcomes the shortcomings of fixed threshold mechanisms in complex environments, and achieves closed-loop control of dynamically optimizing the protection strategy based on real-time risk status, thereby improving the reliability and security of cargo protection.

[0037] Example 1

[0038] like Figure 1 As shown, this application introduces an emergency cargo buffering method in the event of rotor failure in a logistics drone. The method includes:

[0039] S1. Obtain multi-source data of the cargo box after being impacted while it is currently in the initial protection strategy state. The multi-source data includes sensor data stream, cargo attributes and environmental status. The initial protection strategy is triggered by the cargo box's built-in system when it detects a drone malfunction.

[0040] S2. Preprocess and extract features from the sensor data stream, input it into the pre-trained damage probability inference model, and output the damage probability value and confidence level of the goods in the current cargo box.

[0041] S3. Calculate the weighting coefficients based on the cargo attributes, environmental conditions, and damage probability values, and obtain the initial threshold adjustment amount through weighted fusion.

[0042] S4. Obtain historical damage probability values ​​and corresponding confidence levels, and calculate the slope and mean confidence level representing the short-term changes in historical damage probability values ​​based on the least squares method.

[0043] If the slope is greater than a preset trend threshold and the mean confidence level is less than a preset confidence threshold, then after removing historical damage probability values ​​with a confidence level less than the confidence threshold, the mean confidence level of the remaining historical damage probability values ​​is recalculated.

[0044] S5. Calculate the deviation between the current damage probability value and the recalculated confidence mean value, normalize it to obtain the deviation coefficient, and multiply the initial threshold adjustment amount by the difference between 1 and the deviation coefficient to obtain the final threshold adjustment amount.

[0045] The final threshold adjustment is added to the initial damage probability threshold to obtain the adjusted damage probability threshold;

[0046] S6. If the damage probability value is greater than the final damage probability threshold, then the final protection strategy is generated and executed.

[0047] Regarding step S1:

[0048] Acquire sensor data streams, cargo attributes, and environmental conditions collected by multimodal sensors built into the cargo container after a cushioning impact event, including:

[0049] Multimodal sensors include inertial measurement units, pressure distribution sensors, and acoustic sensors;

[0050] (1) Data acquisition:

[0051] The linear acceleration and angular velocity of the cargo box in three-dimensional space are synchronously collected by the inertial measurement unit to characterize the motion state and attitude changes of the cargo box at the moment of impact and in the subsequent process;

[0052] The pressure distribution matrix distributed on the inner surface of the airbag is collected by a pressure distribution sensor at a predetermined sampling frequency to characterize the spatial distribution of the impact force on different parts of the cargo box (the pressure distribution sensor is an array of piezoresistive thin film sensors, which are attached to the inner wall of the cushioning airbag; the pressure distribution matrix is ​​an array of data, whose rows and columns correspond to the physical distribution positions of the piezoresistive thin film sensors, and whose values ​​reflect the real-time pressure values ​​borne by each sensing unit).

[0053] Acoustic sensors collect acoustic vibration data generated by impact events to capture specific acoustic events caused by deformation or breakage of cargo or container structures (the acoustic vibration data is pre-emphasized, and a first-order finite impulse response filter algorithm is used to enhance high-frequency components to compensate for the attenuation of high-frequency components during signal acquisition).

[0054] The system retrieves cargo attributes characterizing the physical properties of the transported cargo from the non-volatile memory of the cargo container's built-in system. The data includes at least one or more of the following: cargo value class, fragile class, and cargo weight. The system also receives the environmental status of the current flight environment in real time. The data includes at least one or more of the following: UAV altitude, ground obstacle density assessment value, and current wind speed.

[0055] (2) Data processing:

[0056] The acquired linear acceleration and angular velocity are filtered and denoised (the filtering and denoising process is implemented by connecting a time-window-based moving average filter and a low-pass filter in series, wherein the cutoff frequency of the low-pass filter is set according to the highest frequency component of the effective impact signal in the linear acceleration) to eliminate high-frequency noise and interference caused by sensor drift.

[0057] The pressure distribution matrix is ​​standardized and calibrated to eliminate sensitivity differences between the sensing units.

[0058] The acoustic vibration data is pre-emphasized and framed to prepare for subsequent frequency domain feature extraction.

[0059] (3) Data output:

[0060] The filtered and denoised inertial data (linear acceleration and angular velocity) are combined into a synchronized six-axis inertial data stream;

[0061] The pressure data (pressure distribution matrix) after standardization and calibration is output as a standardized pressure distribution data stream;

[0062] The acoustic data (sound wave vibration data) after pre-emphasis and framing processing is output as a pre-processed acoustic data stream;

[0063] The six-axis inertial data stream, the standardized pressure distribution data stream, and the pre-processed acoustic data stream together constitute the original sensing data stream, which is then transmitted to subsequent modules.

[0064] Through the above technical solution, this application achieves multi-dimensional, high-precision perception and data preprocessing of impact events, providing comprehensive and reliable input for subsequent cargo damage risk assessment. By integrating multiple sensing modes such as inertial measurement, pressure distribution, and acoustic vibration, the system can comprehensively characterize the complete process of an impact event from multiple physical dimensions, including motion state, force distribution, and structural acoustic emission. Synchronous acquisition and standardized preprocessing of multi-source heterogeneous data effectively eliminates interference caused by sensor noise, drift, and unit differences, ensuring the quality and consistency of the raw data stream. This solution not only enhances the system's ability to perceive complex impact events, but its standardized data stream also lays the foundation for subsequent intelligent inference models, thereby comprehensively improving safety monitoring and emergency response capabilities during cargo transportation.

[0065] The steps to obtain the initial protection policy include:

[0066] Acquire the real-time health status and environmental attitude generated by the UAV flight control system;

[0067] The acquired health status and environmental attitude are normalized.

[0068] The normalized health status is numerically compared with a preset status threshold.

[0069] If the normalized health status and environmental attitude are both greater than the preset state threshold and attitude threshold, then an initial protection strategy is generated based on the health status and environmental attitude.

[0070] Specifically, it receives system health status and flight environment attitude data in real time from the UAV flight control system via the data bus;

[0071] The system health status includes at least one or more of the following: motor fault code, remaining battery power, and rotor speed deviation value;

[0072] The flight environment attitude includes at least one or more of the UAV’s current pitch angle, roll angle, yaw angle and their rate of change.

[0073] The acquired system health status and flight environment attitude are normalized and converted into dimensionless scalar values.

[0074] At the same time, the normalized environmental attitude is numerically compared with a preset attitude threshold.

[0075] The initial protection strategy is a digital command signal, which is used to trigger the cargo box to perform the initial buffer action.

[0076] For example, the system compares this data with preset decision threshold data. When both are determined to be "motor failure" and in a "roll angle" state, the system built into the cargo box is triggered upon receiving instructions from the drone flight control system.

[0077] Upon receiving the command, the cargo container system instantly activates the compressed air tank, inflating the flexible airbag within the container. The airbag absorbs kinetic energy through deformation and decompression, transforming the violent impact into a buffer, protecting the cargo and the ground.

[0078] Through the above technical solutions, this application achieves accurate assessment of the UAV's status and generation of high-confidence commands within an extremely short time window, providing decision-making time for subsequent buffer protection and triggering effective countermeasures in the early stages of rotor failure. Joint judgment of health status and environmental attitude reduces the risk of false triggering and missed triggering. Multi-parameter parallel comparison and collaborative judgment preserve the system's agility in responding to abnormal states while significantly improving the reliability of command generation. Furthermore, normalization processing allows sensor data with different physical meanings to be processed within the same decision logic framework, enhancing the system's scalability and adaptability.

[0079] Health status and environmental posture can also be weighted.

[0080] The steps of weighted fusion processing include:

[0081] Assign first and second weighting coefficients to the normalized health status and environmental attitude;

[0082] Multiply the first weighting coefficient by the health status to obtain the weighted health status;

[0083] Multiply the second weighting coefficient by the environmental pose to obtain the weighted environmental pose;

[0084] The weighted health status is added to the environmental posture to obtain a fusion judgment value;

[0085] If the fusion judgment value is greater than the preset judgment threshold, an initial protection strategy is generated based on the fusion judgment value.

[0086] Specifically, the normalized health status after normalization. With normalized environmental attitude .

[0087] To normalize health status Assign a first weight coefficient ;

[0088] For normalized environmental attitude Assign a second weighting coefficient ;

[0089] The values ​​of the first and second weighting coefficients are determined based on the current flight phase of the UAV and historical fault statistics, and satisfy the following conditions: ;

[0090] The fusion judgment value is calculated using a weighted fusion algorithm. :

[0091]

[0092] The calculated fusion judgment value With a preset judgment threshold Perform numerical comparisons.

[0093] If the fusion judgment value Greater than the judgment threshold Then, the initial protection strategy is generated and output;

[0094] The initial protection strategy is a digital command signal used to trigger the cargo box to perform initial buffering actions.

[0095] Through the above technical solution, this application achieves multi-dimensional judgment of the health status and flight environment of the UAV based on a dynamic weighted fusion judgment mechanism, improving the accuracy and timeliness of the initial protection strategy generation. By adaptively weighting and fusing the normalized health status and environmental attitude, a fusion judgment value that comprehensively reflects the overall risk level of the aircraft is constructed. The weight coefficients can be adaptively adjusted according to the fault type, flight stage, or external environmental conditions, thereby enhancing the system's adaptability to different scenarios. This application maps multi-source information into a unified risk quantification index through weighted fusion, effectively utilizing the complementarity between state parameters and solving the problem of dynamic changes in the reliability of different sensor data through a dynamic weight adjustment mechanism.

[0096] This application further proposes that, before preprocessing and feature extraction of the sensor data stream, validity verification and processing of the sensor data stream should also be included:

[0097] Acquire sensor data streams collected by redundant sensor arrays deployed at multiple different physical locations within the cargo container;

[0098] Determine whether the sensing data stream of redundant sensor arrays at different physical locations exceeds a preset tolerance threshold;

[0099] If yes, the redundant sensor array is deemed invalid; otherwise, the redundant sensor array is deemed valid.

[0100] The average value of the sensor data streams acquired by all valid redundant sensor arrays is calculated as the final sensor data stream.

[0101] Through the above technical solution, this application achieves hierarchical discrimination and fusion processing of the reliability of multi-source sensor data based on validity verification and redundancy fault tolerance mechanisms, improving the integrity and robustness of the data preprocessing stage. By cross-validating and judging tolerance thresholds on redundant sensor arrays deployed in different physical locations, failed sensors are dynamically identified and abnormal data is removed. Then, by weighted fusion to calculate the average value of valid data, a consistent sensor data stream that can comprehensively reflect the cargo container status is constructed. This mechanism not only enhances the reliability of data acquisition through redundancy design, but also solves the data distortion problem caused by local sensor failures through adaptive invalid data removal and fusion processing.

[0102] This application further proposes that when calculating the average value of the sensor data stream acquired by the effective redundant sensor array as the final sensor data stream, the deviation value from the initially acquired sensor data stream should also be considered, specifically including the following steps:

[0103] Obtain the average value of the sensor data stream acquired by the effectively redundant sensor array. and the initial sensor data stream. ;

[0104] calculate and deviation value and to Normalization is performed to obtain the normalized deviation value. ,in ;

[0105] based on The adjusted sensor data stream is calculated using the following formula. :

[0106]

[0107] Output As the final sensor data stream.

[0108] Through the above technical solution, this application achieves consistent fusion and reliability improvement of multi-source sensor data based on deviation perception dynamic calibration mechanism, enhancing the accuracy and stability of the final sensor data stream. By calculating the normalized deviation between the average value of the effective redundant sensor data and the initial acquisition value, and using this deviation value as a weighting coefficient for adaptive data fusion, an adjusted output value that reflects both the overall data trend and the reliability of the initial data is constructed. This mechanism, through dynamic weighting of the normalized deviation, achieves quantitative perception and compensation of the original deviation during the redundant data fusion process, thereby effectively suppressing output fluctuations caused by local sensor anomalies or environmental interference, and improving the robustness of the system and data quality.

[0109] Regarding step S2:

[0110] The raw sensor data stream is preprocessed and features are extracted to obtain a structured multidimensional feature vector representing the current impact event and the cargo container state, including:

[0111] (1) Data acquisition:

[0112] Acquire the synchronized six-axis inertial data stream, the standardized pressure distribution data stream, and the preprocessed acoustic data stream output from step S1.

[0113] (2) Data processing:

[0114] Processing of six-axis inertial data streams:

[0115] From the linear acceleration, the time point of the impact event is identified, and data within a predetermined time window centered on that time point is extracted;

[0116] Calculate the maximum value of the vector magnitude of the linear acceleration within the time window and record it as the peak impact force;

[0117] Calculate the time span during which the vector magnitude exceeds a preset impact force threshold and record it as the impact duration.

[0118] The angular velocity is integrated to calculate the maximum roll angle of the cargo box from the time of impact to the end of the time window (the calculation of the maximum roll angle includes: numerical integration of the angular velocity within a predetermined time window to obtain the real-time Euler angle change of the cargo box in three-dimensional space relative to the initial impact time, and taking the component with the largest absolute value of the Euler angle change as the maximum roll angle).

[0119] Processing of standardized pressure distribution data streams:

[0120] Near the time point corresponding to the occurrence of the peak impact force, obtain one or more frames of pressure distribution matrix;

[0121] Calculate the standard deviation of all pressure values ​​in a single frame of data and record this standard deviation as a pressure uniformity index. This index is used to characterize the degree of concentration or dispersion of impact force on the surface of the cargo box.

[0122] Processing of the preprocessed acoustic data stream:

[0123] The acoustic data frames after framing are transformed in the frequency domain (the frequency domain transformation is implemented using the Fast Fourier Transform algorithm, and the main frequency feature of the acoustic pattern is the frequency value with the largest amplitude in the spectrum obtained after the Fast Fourier Transform, where N is a natural number greater than or equal to 1), and the data is mapped from the time domain to the frequency domain to obtain the spectrum of each frame of data.

[0124] Analyze the spectrum, identify one or more frequency components with the largest amplitude, and record them as the main frequency features of the voiceprint.

[0125] (3) Data output:

[0126] The peak impact force (the pressure uniformity index is associated with the peak impact force and the maximum attitude roll angle, and is used together to determine whether the cargo box has overturned or suffered a single-point severe impact), impact duration, maximum attitude roll angle, pressure uniformity index, and main frequency characteristics of the acoustic signature are combined and spliced ​​in a predetermined order to form a fixed-dimensional, structured multi-dimensional feature vector, and this vector is output to the subsequent intelligent evaluation module.

[0127] Through the above technical solution, this application achieves in-depth processing and information extraction of multimodal sensing data of impact events, generating a set of high-dimensional feature vectors that can comprehensively and structurally characterize impact intensity, force distribution characteristics, and acoustic response. By extracting feature indicators with clear physical meaning from inertial, pressure, and acoustic data respectively, the system realizes a multi-angle quantitative description of the spatiotemporal and frequency domain characteristics of impact events. This structured feature vector effectively integrates the dynamic information of transient impact and continuous process, providing reliable and interpretable input for subsequent intelligent assessment modules to judge damage risk. This not only improves the ability to characterize complex impact scenarios but also enhances the discrimination ability and decision reliability of the entire assessment system through deep feature-level fusion.

[0128] Before inputting the above data into the damage probability inference model, the data is normalized using a min-max method.

[0129] Min-max processing, also known as deviation standardization, is a linear transformation of the original data, mapping the resulting values ​​to the range [0–1]. The transformation function is as follows:

[0130] x*=(x−min) / (max−min)

[0131] Where x* is the normalized value, x is the current sample data, max is the maximum value of the sample data, and min is the minimum value of the sample data.

[0132] The normalized multidimensional feature vector is input into a pre-trained damage probability inference model to obtain a damage probability value representing the risk of damage to goods inside the container, as well as the confidence level of the data, including:

[0133] (1) Data acquisition:

[0134] Receives a structured multidimensional feature vector, output by the feature extraction module, representing the current impact event and the cargo container state.

[0135] (2) Data processing:

[0136] The multidimensional feature vector is fed into a pre-built damage probability inference model deployed on the cargo container processing module. (The damage probability inference model is obtained through supervised training using a large amount of historical crash sensor data in the cloud. Each training sample contains a historical multidimensional feature vector extracted from sensor data and a binary label data corresponding to the actual damage state of the cargo in that historical event. The training objective of the model is to minimize the difference between its output prediction value and the binary label data.)

[0137] The model processes multidimensional feature vectors in real time, and its computational latency is strictly limited to the millisecond level to ensure that the overall processing time from sensor data input to output damage probability value meets the timing requirements of cargo container emergency response.

[0138] The damage probability inference model is based on a deeply compressed neural network structure, which contains multiple interconnected computational units. Each computational unit performs weighted summation and nonlinear transformation operations on the input data.

[0139] The model uses its internal computing units to express and calculate the complex nonlinear relationships between the peak impact force, impact duration, maximum attitude roll angle, pressure uniformity index, and acoustic signature main frequency features contained in the input multidimensional feature vector in a layer-by-layer and distributed manner.

[0140] Finally, the output layer computation unit of the model generates a continuous value between 0 and 1, along with the confidence level of that value.

[0141] (3) Data output:

[0142] The continuous values ​​generated by the output layer calculation unit of the damage probability inference model are output as the damage probability value and the confidence level of the data, representing the overall damage risk of the goods in the cargo container. (The physical meaning of the damage probability value is that its value is positively correlated with the confidence level of the goods in the cargo container being damaged. The closer the value is to 1, the higher the confidence level of the goods being judged as damaged.)

[0143] Through the above technical solution, this application achieves real-time intelligent assessment of cargo damage risk caused by complex impact events. By inputting structured multidimensional feature vectors into a pre-trained damage probability inference model, the system can deeply mine the complex nonlinear mapping relationship between impact force, attitude, pressure distribution, and acoustic features, and output a high-confidence damage probability value. This model, trained on a large amount of historical crash data, possesses strong generalization ability and real-time inference performance, ensuring the timeliness of emergency response procedures. The output continuous probability value intuitively reflects the overall damage risk confidence of the cargo, providing a reliable quantitative decision-making basis for the generation of subsequent protection instructions, and improving the automation and intelligence level of cargo safety status assessment.

[0144] This application further proposes the process for constructing the damage probability inference model, including:

[0145] (1) Data acquisition

[0146] Obtain the historical training dataset required for model training, where each training data sample contains a historical multidimensional feature vector. and its corresponding binarized damage label The historical multidimensional feature vector is generated from sensor data collected during multiple drone crashes or simulated crash tests throughout history through a feature extraction process, and the resulting binarized damage labels are used. This indicates that the collision resulted in damage to the cargo. 0 indicates that the goods are intact.

[0147] (2) Data processing

[0148] The core operation of constructing an initial deep neural network model involves mapping the input feature vector to a predicted probability value through multiple layers of nonlinear transformations. This model's first... Layer output Due to its input to the previous layer The following transformation is performed to obtain:

[0149]

[0150]

[0151] in, This is the weight matrix for this layer. For bias vectors, For the first The activation function of the layer (for example, the intermediate layer uses the ReLU function, i.e.) The output layer uses the Sigmoid function, i.e. This is to ensure that the output value is within the (0,1) range.

[0152] The final output of the model is ,in The total number of layers in the model. This is the model's predicted value for the probability of damage.

[0153] Minimize the predicted value using an optimization algorithm. With real labels The difference between them, the difference is determined by the loss function Quantization is performed using a binary cross-entropy loss function:

[0154]

[0155] Iteratively optimize the weight matrix on the entire training dataset. With bias vector The parameter values ​​are adjusted until the loss function converges.

[0156] in, The first one representing the neural network layer, For the first The linear output vector (or "net input") of a layer is an intermediate result before the activation function of that layer processes it, and its dimension is determined by the number of neurons in that layer. For the first The layer weight matrix is ​​one of the key parameters that the model needs to train. Each element in the matrix... Represents the previous level The output value of the nth neuron affects the current layer's nth neuron. The training process essentially involves finding the optimal weight matrix, which measures the influence of the linear output of each neuron (i.e., the connection strength).

[0157] (3) Data output

[0158] The final weight matrix obtained after training With bias vector The parameter set and the model structure definition are combined as the output of the trained model parameter set.

[0159] After model training is complete, perform model compression and lightweighting:

[0160] The trained model parameter set is subjected to weight quantization, converting the weight values ​​originally stored in 32-bit floating-point format. Quantize to a lower bit width (such as an 8-bit integer) format for storage to reduce model size and improve inference speed on embedded processors;

[0161] The quantized model is pruned, and the weight matrix is... Weight elements whose absolute value is below a preset threshold are set to zero, thereby generating a sparse weight matrix and further compressing the model size.

[0162] The final lightweight model parameter set and model structure definition, after quantization and pruning, are deployed to the non-volatile memory of the cargo box system via wired or wireless data transmission to form a pre-set damage probability inference model, which is then loaded and used for forward inference calculations during runtime.

[0163] Through the above technical solution, this application realizes the automated construction and deployment of a high-precision, lightweight damage probability inference model based on historical crash data. This solution utilizes a deep neural network to perform distributed learning and representation of the complex nonlinear relationship between multidimensional feature vectors and cargo damage labels, and optimizes the prediction output using a binary cross-entropy loss function to ensure high accuracy. Furthermore, through lightweight processing such as weight quantization and pruning, the model size and computational overhead are compressed while preserving model performance to the maximum extent, enabling deployment in resource-constrained cargo containers. This solution not only achieves end-to-end automation from data to a usable model, but also ensures that the final deployed model has real-time inference capabilities, providing an inference engine for online intelligent assessment of cargo damage risk and improving the decision-making reliability and response agility of the entire protection system.

[0164] Regarding step S3:

[0165] The process of obtaining the initial threshold adjustment amount includes:

[0166] The cargo attributes, environmental conditions, and damage probability values ​​are obtained, and the above data are normalized using a min-max method; wherein, the cargo attributes include the quantified value of the cargo's value level. Quantitative value of fragile item grade and normalized values ​​of cargo weight ;

[0167] The environmental conditions include the normalized value of the drone's ground altitude. Ground obstacle density assessment value and the current wind speed level quantification value ;

[0168] Based on the aforementioned cargo attributes, environmental conditions, and damage probability value The weight coefficients of each element are calculated through weight allocation, and the initial threshold adjustment is calculated using a weighted fusion algorithm. :

[0169]

[0170] in, , , , , , , , , The preset weighting coefficients, , , The above weighting coefficients were determined through optimization using historical data.

[0171] Through the above technical solution, this application realizes a dynamic threshold adaptive adjustment mechanism based on multi-source information fusion. This solution constructs an input vector that integrates cargo attributes, environmental conditions, and data confidence levels, and generates a threshold adjustment amount based on preset weights. Intrinsic factors such as cargo value, fragility, and weight, along with external environmental conditions such as ground clearance, obstacle density, and wind speed, as well as model inference confidence levels, are uniformly incorporated into the calculation framework. Through the synergistic effect of data and weights, the direction and magnitude of the threshold adjustment are jointly determined. This mechanism enhances the system's adaptability, robustness, and reliability in decision-making under different cargo types and complex flight environments, effectively avoiding potential misjudgments and omissions under traditional fixed threshold strategies, and providing an adaptive judgment benchmark for the accurate generation of subsequent protection commands.

[0172] Regarding step S4:

[0173] The process of processing historical damage probability values ​​and data confidence levels includes the following steps:

[0174] Get the reservation time slot The sequence of historical damage probability values, and the sequence of confidence levels corresponding to each historical damage probability value:

[0175] Get in the time window Historical damage probability value sequence collected internally and its corresponding data confidence sequence ,in For the current moment, The preset time window length, This represents the total number of data points within the window.

[0176] The historical damage probability value sequence is fitted using the least squares method to calculate the slope of its short-term trend; at the same time, the arithmetic mean of the data confidence sequence is calculated.

[0177] The least squares method was used to analyze the historical damage probability value sequence. With time series Perform a linear fit and calculate its short-term slope. The fitted model is:

[0178]

[0179] in, For the first Each damage probability value, The intercept is... For residuals; slope The calculation formula is:

[0180]

[0181] in, For the first That moment.

[0182] Simultaneously, calculate the data confidence sequence. arithmetic mean :

[0183]

[0184] in, For the first The confidence level of the data for each probability value of damage. This represents the total number of data points within the window.

[0185] If the slope value is greater than a preset trend threshold and the arithmetic mean is less than a preset confidence threshold, then all confidence data points less than the confidence threshold are removed from the data confidence sequence, and the corresponding data points in the historical damage probability value sequence are also removed simultaneously; based on the remaining data confidence sequence after removal, the arithmetic mean is recalculated.

[0186] If both conditions are met:

[0187]

[0188] in, To preset the trend threshold, To pre-set the reliability threshold, the following elimination and recalculation are performed:

[0189] Constructing an effective data index set :

[0190]

[0191] based on from Extracting effective confidence subsets And recalculate the arithmetic mean:

[0192]

[0193] Output the recalculated arithmetic mean as the final evaluation result of the confidence level of the valid data.

[0194] Output The final evaluation result is the mean of the effective confidence level.

[0195] Regarding step S5:

[0196] The process of obtaining the final threshold adjustment amount includes:

[0197] Get the confidence level of the current probability of failure. The recalculated confidence mean and initial threshold adjustment amount ;

[0198] Calculate confidence level with confidence level mean The degree of deviation is determined, and this degree of deviation is normalized to obtain the deviation coefficient. :

[0199] ;

[0200] Then, the adjustment factor is calculated. Adjustment coefficient for With deviation coefficient The difference;

[0201] ;

[0202] Adjust the initial threshold amount With adjustment coefficient Multiply by each other to obtain the final threshold adjustment amount.

[0203] The calculation formula is: .

[0204] like Figure 2 The flowchart shown illustrates the method for obtaining the initial damage probability threshold in the emergency cargo buffering method under the condition of rotor failure of a logistics drone.

[0205] This application further proposes a process for calculating the initial damage probability threshold:

[0206] (1) Data acquisition

[0207] A large number of historical crash data samples were obtained from historical crash case databases and simulation test databases;

[0208] Each data sample contains a historical multidimensional feature vector extracted from sensor data, and a binary true damage label, characterized by expert annotation or post-crash verification, representing the actual damage state of the cargo in that crash.

[0209] Define historical dataset ;

[0210] in, For the first The historical multidimensional feature vector of each sample For the corresponding binary real damage label, This indicates actual damage. The total number of historical data samples. For sample index, .

[0211] (2) Data processing:

[0212] Each historical multidimensional feature vector is input into a cloud-based training model with the same structure as the damage probability inference model built into the cargo box to obtain the corresponding historical damage probability prediction value.

[0213] Each feature vector is input into a cloud-based training model with the same structure as the cargo box model. :

[0214]

[0215] in The probability of damage predicted by the model. This is a function for training models in the cloud. For the input of the first Each sample feature vector For the model to the first The predicted probability of damage for each sample. This is the sample index.

[0216] Using the binarized true damage label as the gold standard and the historical damage probability prediction value as the test index, an ROC curve was plotted.

[0217] Construct the ROC curve by traversing all predicted values:

[0218]

[0219] Where the threshold The corresponding false positive rate and true positive rate are:

[0220]

[0221]

[0222] in, To be at the threshold The false positive rate, This is an indicator function that returns 1 if the condition is true, and 0 otherwise. For the first The predicted probability of a sample. For classification threshold, For the first The true label of each sample The total number of samples, For logical AND operator, To be at the threshold The true rate.

[0223] By iterating through all possible threshold points on the ROC curve, the false alarm rate and recall rate corresponding to each threshold point are calculated.

[0224] Based on the preset system performance requirements, an optimal threshold point is selected so that the overall protection performance of the system is optimal at this threshold. The overall protection performance is calculated by weighting the false alarm rate and the recall rate.

[0225] The optimal threshold is determined by optimizing the overall efficiency index:

[0226]

[0227] in, This is the recall rate weighting coefficient. To be the optimal basic threshold, The parameters that maximize the objective function. The classification threshold has a value range of [0,1]. This is the recall rate weighting coefficient. , To be at the threshold The actual rate, To be at the threshold The false positive rate.

[0228] (3) Data output:

[0229] The calculated optimal threshold point value is output as the basic threshold parameter.

[0230] The basic threshold parameters are sent to the cargo container management system and stored in a designated area of ​​its non-volatile memory as the calculation basis for the threshold management loop.

[0231] Output the optimal threshold as the base threshold parameter:

[0232]

[0233] in, Basic threshold parameters The optimal threshold obtained through optimization.

[0234] and through secure communication protocols The data is distributed to each cargo container management system and stored in non-volatile memory.

[0235] Algorithm constraints:

[0236] Data volume requirements: ;

[0237] Category balance: ;

[0238] Model consistency: .

[0239] in, To meet the minimum sample size requirement, , To avoid damaging the effective range of sample proportions, For the first The true label of each sample The total number of samples, To meet the minimum requirement of damaged sample ratio, To meet the maximum requirement of the damaged sample ratio, The sum of labels for all damaged samples.

[0240] Through the above technical solution, this application realizes a method for determining the optimal damage probability threshold based on historical data. This method utilizes a large amount of historical crash or simulation test data, employs an inference model with a structure consistent with edge models in the cloud for batch prediction, and plots ROC curves based on this to comprehensively evaluate the false alarm and recall performance at different threshold points. Finally, by optimizing the comprehensive performance index, the optimal basic threshold that can maximize the balance between false alarm and missed alarm risks is scientifically determined. This greatly improves the scientific nature, interpretability, and overall system performance of the threshold setting, laying the foundation for the operation of cargo container protection systems in various real-world scenarios.

[0241] This application further proposes that after obtaining the final damage probability threshold, the confidence level of the current damage probability value should also be considered, specifically including the following steps:

[0242] Obtain the initial damage probability threshold Adjusted damage probability threshold and the confidence level of the current probability of damage. ;

[0243] With the aforementioned confidence level As a weight, the initial damage probability threshold Compared with the adjusted damage probability threshold Perform weighted calculation:

[0244]

[0245] The weighted calculation result is used as the final damage probability threshold. .

[0246] Through the above technical solution, this application achieves dynamic optimization and credibility fusion of the cargo damage probability threshold based on a confidence-weighted dual-threshold fusion mechanism, improving the accuracy and environmental adaptability of threshold decisions. By using the real-time confidence level of the current damage probability value as a weight, the initial threshold and the adjusted threshold are adaptively weighted and fused to construct a final damage probability threshold that comprehensively reflects prior knowledge and real-time adjustment results. This mechanism, through dynamic adjustment of confidence, fully utilizes the stability of the initial threshold while introducing the scenario adaptability of the adjusted threshold, thereby effectively enhancing the robustness and reliability of the system's judgment under different data credibility conditions.

[0247] Regarding step S6:

[0248] Based on the adjusted damage probability threshold Output the final protection strategy execution command, which includes, but is not limited to: multi-level buffer mechanism actuation sequence, energy absorption device trigger parameters, cargo box attitude stabilization scheme, and emergency landing control command.

[0249] For example:

[0250] Activate the secondary buffer mechanism: Detonate the controllable compressed gas device to further inflate and enhance the rigidity of the buffer airbag; Trigger the directional pressure relief valve: Based on the pressure distribution data, open the pressure relief valve in the high-pressure area to balance the impact force distribution; Activate the attitude control system: Ignite the attitude adjustment micro rocket to suppress the cargo box's rolling motion; and Execute the emergency landing procedure: Release the friction-resistant skid device to reduce the impact during ground sliding.

[0251] This application further proposes that after the final protection strategy is executed, the multi-source data after the execution of the final protection strategy is received, the ratio of the damage probability value to the final damage probability threshold is calculated iteratively, and the final protection strategy is optimized until the damage probability value in the iterative calculation is less than the final damage probability threshold or the number of iterations is reached, and then the optimized final protection strategy is output.

[0252] Receive multi-source data after the final protection strategy is implemented, and iteratively calculate the ratio of the damage probability value to the final damage probability threshold: repeat S1-S5.

[0253] The damage probability value is compared with the final damage probability threshold;

[0254] If the comparison result shows that the damage probability value is greater than the final damage probability threshold, then the final protection strategy optimization logic is triggered.

[0255] The final protection strategy optimization logic is based on a preset optimization strategy. It adjusts the previously generated final protection strategy to generate an optimized final protection strategy. The optimization strategy includes, but is not limited to: enhancing the strength of the current protection measures, triggering additional protection mechanisms, or changing the timing of action between different protection mechanisms.

[0256] Alternatively, based on the number of iterations, a maximum iteration threshold can be set. Before generating the optimized final protection strategy each time, it is determined whether the current iteration count has reached the maximum iteration threshold. If it has, the iteration is terminated and the current final protection strategy is output. If it has not, the iteration counter is incremented by one, and the protection strategy optimization and output continue.

[0257] Record the final protection strategy executed in each iteration and the corresponding output damage probability value; in the current iteration, select the final protection strategy combination that has not been executed before and is expected to further reduce the impact as the optimized final protection strategy for output and testing.

[0258] The optimized final protection strategy generated in the last iteration that makes the damage probability value less than the final damage probability threshold; or, the optimized final protection strategy generated when the maximum number of iterations is reached that makes the damage probability value as small as possible.

[0259] Output and execute the optimized final protection strategy.

[0260] Through the above technical solution, this application implements an intelligent iterative optimization control mechanism based on closed-loop feedback for dynamically adjusting the final protection strategy. This scheme continuously compares the latest assessed damage probability value with the final damage probability threshold, triggering multi-round strategy optimization logic. Based on preset strategies or historical execution records, it adaptively adjusts the strength, combination, or timing of the protection strategy. By introducing iteration number constraints and optimization goal guidance, the system ensures that the optimization process converges to an effective solution within a finite number of steps. This improves the adaptability and intervention accuracy of the protection system in complex impact scenarios while meeting real-time requirements, ultimately ensuring the safety of the cargo.

[0261] like Figure 3 The diagram shows a data flow chart illustrating an emergency cargo buffering method in the event of a rotor failure in a logistics drone.

[0262] To facilitate understanding of the above embodiments, a specific application scenario of the above embodiments will be used as an example for illustration below:

[0263] Example: Emergency buffer for high-value medical equipment in the event of rotor failure in a logistics drone

[0264] (1) Scene description

[0265] A quadcopter logistics drone (model: CargoDrone200) was carrying out an emergency medical supplies delivery mission. Its cargo box contained a high-precision portable blood analyzer (value class: A, fragile class: 5, weight: 15kg). During flight, due to a sudden strong crosswind (wind speed: 12m / s) and aging motors, the drone's right front rotor experienced a sudden stoppage. The flight control system detected in real-time that the motor speed deviation exceeded the tolerance limit and, combined with the current roll angle (25°) and altitude (80m) data, determined that the drone had entered an abnormal state.

[0266] (2) Initial protection strategy triggered

[0267] The drone flight control system sends health status alarms (including motor fault code: E102, battery power: 45%, rotor speed deviation: 95%) and environmental attitude data (pitch angle: 5°, roll angle: 25°, yaw angle: 10°) to the internal system of the cargo box via the CAN bus.

[0268] The cargo container system normalizes the received data and compares it with preset thresholds (health status threshold: 0.7, attitude threshold: 0.6). A fusion judgment value of 0.85 is calculated (higher than the judgment threshold of 0.75), and an initial protection strategy is immediately generated.

[0269] Initial protection strategy triggered: The compressed air tank at the bottom of the cargo box is activated instantly, inflating the flexible airbag that wraps around the cargo box (initial cushioning action).

[0270] (3) Multi-source data acquisition and preprocessing

[0271] Inertial Measurement Unit (IMU): Acquired peak linear acceleration (X: 12.5g, Y: 8.2g, Z: 15.3g) and peak angular velocity (X: 45° / s, Y: 60° / s, Z: 30° / s) at the moment of impact.

[0272] Pressure distribution sensor: detected local pressure peak in the pressure distribution matrix of the inner surface of the airbag (location: lower right quadrant, peak value: 85 kPa), pressure uniformity standard deviation: 12.3 kPa.

[0273] Acoustic sensor: captures the sound wave signal generated by the impact (main frequency components: 1.2kHz, 3.5kHz), and outputs it after pre-emphasis and framing processing.

[0274] Environmental conditions: Current ground clearance: 30m (normalized value: 0.3), ground obstacle density: medium (assessed value: 0.6), wind speed: 10m / s (level: 4).

[0275] Goods attributes: Value grade: 0.9 (Grade A), Fragile grade: 0.95 (Grade 5), Weight normalization value: 0.5.

[0276] (4) Damage probability inference

[0277] The feature extraction module extracts structured feature vectors from multi-source data:

[0278] Peak impact force: 18.2g;

[0279] Impact duration: 120ms;

[0280] Maximum roll angle: 40°;

[0281] Pressure uniformity index: 12.3 kPa;

[0282] Voiceprint dominant frequency characteristics: [1.2kHz, 3.5kHz];

[0283] After the feature vectors are normalized by min-max, they are input into the pre-trained lightweight damage probability inference model (which has been deployed on the cargo box embedded processor).

[0284] Model output:

[0285] The probability of failure is 0.88.

[0286] The current confidence level is 0.92;

[0287] (5) Dynamic threshold adjustment and decision-making

[0288] Calculate the initial threshold adjustment amount :

[0289] Based on cargo attributes, environmental conditions, and damage probability values, a weighted fusion calculation is performed:

[0290]

[0291]

[0292] Historical trends and confidence optimization:

[0293] Obtain the historical damage probability sequence P=[0.82,0.85,0.87,0.86,0.88] and the confidence sequence C=[0.90,0.91,0.93,0.92,0.92] within the last 5 seconds.

[0294] The least squares fit slope is 0.015 (trend threshold 0.01), and the mean confidence level is 0.916 (confidence threshold 0.90).

[0295] Since the slope is greater than the trend threshold and the mean confidence level is less than the confidence threshold, the current confidence level sequence is used directly.

[0296] Calculate the final threshold adjustment amount:

[0297] The deviation coefficient D = |0.92 - 0.916| / 0.916 = 0.0044;

[0298] The adjustment factor K = 10.0044 = 0.9956;

[0299] Final threshold adjustment amount =0.312 * 0.9956 = 0.310;

[0300] Generate the final failure probability threshold:

[0301] Base threshold 0.75 (determined through optimization using historical ROC curves);

[0302] Adjusted threshold (If the value exceeds 1.0, it will be truncated to 1.0).

[0303] Confidence-weighted fusion:

[0304]

[0305] (6) Final protection strategy execution and iterative optimization

[0306] Due to the current probability of damage The system determined that the risk of the goods was controllable and therefore did not trigger the secondary buffer.

[0307] However, continuous monitoring data from the cargo box showed that after 2 seconds, the pressure sensor detected that the pressure in the lower right quadrant rose to 95 kPa (local stress concentration), and the acoustic sensor captured structural deformation noise (main frequency: 4.0 kHz).

[0308] The probability of failure was recalculated as C=0.94 (confidence level 0.90).

[0309] Compare again: Triggering the final protection policy:

[0310] Enhanced cushioning: The detonation of the secondary compressed gas tank increases the rigidity of the airbag.

[0311] Directional pressure relief: Open the pressure relief valve in the lower right quadrant to equalize the internal pressure.

[0312] Attitude stabilization: Triggers the attitude adjustment support component in the lower left quadrant to suppress cargo box rollover.

[0313] Emergency landing: Deploy the bottom friction-resistant skid device.

[0314] Post-execution data monitoring: The probability of failure dropped to 0.6 (confidence level 0.93), the system determined that the risk was eliminated, and the iteration was terminated.

[0315] Through the above technical solution, this application achieves the protection of electronic products inside the cargo container. Although the drone was damaged, the electronic products inside the cargo container only suffered minor vibrations, and the vast majority remained intact, successfully completing the delivery mission. Throughout the entire process, from the initial impact to final stabilization, the system automatically completed all sensing, decision-making, execution, and optimization steps in a very short time, without human intervention, demonstrating the effectiveness and advancement of the technical solution presented in this application.

[0316] Example 2

[0317] like Figure 4 As shown, this application proposes an emergency cargo buffering system for the event of a logistics drone rotor failure, used to implement the aforementioned emergency cargo buffering method for the event of a logistics drone rotor failure. The system is integrated into the cargo box and includes:

[0318] 100. Data acquisition module, which is used to acquire multi-source data of the cargo box after being impacted while it is currently in the initial protection strategy state. The multi-source data includes sensor data stream, cargo attributes and environmental status. The initial protection strategy is triggered by the cargo box's built-in system when it detects a drone malfunction.

[0319] 200. Damage probability model processing module, which is used to preprocess and extract features from the sensor data stream, input it into the pre-trained damage probability inference model, and output the damage probability value and confidence level of the goods in the current cargo box.

[0320] 300. Threshold adjustment amount acquisition module, which is used to calculate weight coefficients based on the cargo attributes, environmental conditions and damage probability values, and obtain the initial threshold adjustment amount through weighted fusion;

[0321] Obtain historical damage probability values ​​and corresponding confidence levels, and calculate the slope and mean confidence level representing short-term changes in historical damage probability values ​​based on least squares fitting.

[0322] If the slope is greater than a preset trend threshold and the mean confidence level is less than a preset confidence threshold, then after removing historical damage probability values ​​with a confidence level less than the confidence threshold, the mean confidence level of the remaining historical damage probability values ​​is recalculated.

[0323] Calculate the deviation between the current probability of damage and the recalculated mean confidence level, normalize it to obtain the deviation coefficient, and multiply the initial threshold adjustment by the difference between 1 and the deviation coefficient to obtain the final threshold adjustment.

[0324] 400. Damage probability threshold adjustment module, which is used to add the final threshold adjustment amount to the initial damage probability threshold to obtain the adjusted damage probability threshold;

[0325] 500. Final protection strategy generation module, which generates and executes a final protection strategy if the damage probability value is greater than the adjusted damage probability threshold.

[0326] The technical scope of this application is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this application, and all such modifications and variations should fall within the protection scope of this application.

Claims

1. An emergency cargo buffering method in case of rotor failure in a logistics drone, characterized in that, The method includes: Acquire multi-source data of the cargo container after it is impacted while it is currently in the initial protection strategy state. The multi-source data includes sensor data streams, cargo attributes and environmental status. The initial protection strategy is triggered by the cargo container's built-in system when it detects a drone malfunction. The sensor data stream is preprocessed and features are extracted, then input into a pre-trained damage probability inference model, which outputs the damage probability value and confidence level of the goods in the current cargo container. The weighting coefficients are calculated based on the cargo attributes, environmental conditions, and damage probability values, and the initial threshold adjustment amount is obtained through weighted fusion. Obtain historical damage probability values ​​and corresponding confidence levels, and calculate the slope and mean confidence level representing short-term changes in historical damage probability values ​​based on least squares fitting. If the slope is greater than a preset trend threshold and the mean confidence level is less than a preset confidence threshold, then after removing historical damage probability values ​​with a confidence level less than the confidence threshold, the mean confidence level of the remaining historical damage probability values ​​is recalculated. Calculate the deviation between the current probability of damage and the recalculated mean confidence level, normalize it to obtain the deviation coefficient, and multiply the initial threshold adjustment by the difference between 1 and the deviation coefficient to obtain the final threshold adjustment. The final threshold adjustment is added to the initial damage probability threshold to obtain the adjusted damage probability threshold; If the damage probability value is greater than the adjusted damage probability threshold, then a final protection strategy is generated and executed. The process of obtaining the initial damage probability threshold includes: Acquire historical crash data, which includes multidimensional feature vectors and damage labels. The multidimensional feature vectors are sensor data streams after preprocessing and feature extraction. The multidimensional feature vector is input into the damage probability inference model to obtain the damage probability prediction value; Using the damaged label as the standard and the predicted damage probability as the test index, an ROC curve was plotted. Calculate the false alarm rate and recall rate at each threshold point on the ROC curve, and calculate the overall protection effectiveness based on the weighted average of the false alarm rate and recall rate; The value that maximizes the overall protection effectiveness is used as the initial damage probability threshold.

2. The emergency cargo buffering method under rotor failure of a logistics drone according to claim 1, characterized in that, Before preprocessing and feature extraction of the sensor data stream, the validity of the sensor data stream is also verified and processed. Acquire sensor data streams collected by redundant sensor arrays deployed at multiple different physical locations within the cargo container; Determine whether the sensing data stream of redundant sensor arrays at different physical locations exceeds a preset tolerance threshold; If yes, the redundant sensor array is deemed invalid; otherwise, the redundant sensor array is deemed valid. The average value of the sensor data streams acquired by all valid redundant sensor arrays is calculated as the final sensor data stream.

3. The emergency cargo buffering method under rotor failure of a logistics drone according to claim 2, characterized in that, When calculating the average value of the sensor data stream acquired by the effective redundant sensor array as the final sensor data stream, the deviation value from the initially acquired sensor data stream is also considered. This process includes the following steps: Obtain the average value of the sensor data stream acquired by the effectively redundant sensor array. and the initial sensor data stream. ; calculate and deviation value and to Normalization is performed to obtain the normalized deviation value. ,in ; based on The adjusted sensor data stream is calculated using the following formula. : ； Output As the final sensor data stream.

4. The emergency cargo buffering method under rotor failure of a logistics drone according to claim 1, characterized in that, The steps for preprocessing and feature extraction of the sensor data stream include: The sensor data stream includes linear acceleration and angular velocity characterizing the motion state and attitude changes of the cargo box, pressure distribution matrix characterizing the spatial impact force on the cargo box, and acoustic vibration data characterizing the acoustic events of the cargo or cargo box. The linear acceleration and angular velocity are filtered and denoised using a time-window-based moving average filter and a low-pass filter in series. The pressure distribution matrix is ​​then standardized and calibrated. The acoustic vibration data is pre-emphasized and framed. From the linear acceleration, the time point of the buffer impact event is identified, and the linear acceleration, angular velocity, pressure distribution matrix and acoustic vibration data within a predetermined time window are extracted with the time point as the center. Calculate the maximum value of the linear acceleration vector magnitude within the predetermined time window and use it as the peak impact force. The duration for which the vector magnitude exceeds a preset impact force threshold is calculated and taken as the impact duration. The angular velocity is integrated to calculate the absolute value of the change in Euler angle of the cargo box from the time of impact to the end of the predetermined time window. The maximum value of the absolute value is taken as the maximum attitude roll angle. At the time point corresponding to the occurrence of the peak impact force, the processed pressure distribution matrix is ​​obtained; Calculate the standard deviation of the pressure distribution matrix and use it as an index of pressure uniformity. The processed acoustic vibration data is subjected to frequency domain transformation to obtain its spectrum; Analyze the spectrum and take the frequency with the largest amplitude as the main frequency feature of the voiceprint.

5. The emergency cargo buffering method under rotor failure of a logistics drone according to claim 1, characterized in that, The steps for constructing the damage probability inference model include: Obtain the training dataset, which contains multidimensional feature vectors. and its corresponding binarized damage label ; Construct an initial deep neural network model, the deep neural network model of which the first... Layer output Due to its input to the previous layer The following transformation is performed to obtain: ； ； in, For the first The linear output vector of the layer, for The weight matrix of the layer, For bias vectors, For the first Activation function of the layer; The final output of the deep neural network model for: ； in The total number of layers in the model. This is the model's predicted value for the probability of damage. For the Sigmoid function; Using the binary cross-entropy loss function Minimize the predicted value With damaged label Differences between them: ； Iteratively optimize the weight matrix on the training dataset. With bias vector The parameter values ​​are calculated until the loss function converges; The final weight matrix obtained after training With bias vector The parameter set is used as the parameter set of the deep neural network model; The trained deep neural network model is used as the damage probability inference model.

6. The emergency cargo buffering method under rotor failure of a logistics drone according to claim 1, characterized in that, The process of obtaining the initial threshold adjustment amount includes: Obtain cargo attributes, environmental conditions, and damage probability values; normalize the above data. The cargo attributes include a quantified value of the cargo's value level. Quantitative value of fragile item grade and normalized values ​​of cargo weight ; The environmental conditions include the normalized value of the drone's ground altitude. Ground obstacle density assessment value and the current wind speed level quantification value ; Based on the aforementioned cargo attributes, environmental conditions, and damage probability value The weight coefficients of each element are calculated through weight allocation, and the initial threshold adjustment is calculated using a weighted fusion algorithm. : ； in, , , , , , , , , The preset weighting coefficients, , , .

7. The emergency cargo buffering method under rotor failure of a logistics drone according to claim 2, characterized in that, After obtaining the adjusted damage probability threshold, the confidence level of the current damage probability value is also considered, specifically including the following steps: Obtain the initial damage probability threshold Adjusted damage probability threshold and the confidence level of the current probability of damage. ; With the aforementioned confidence level As a weight, the initial damage probability threshold Compared with the adjusted damage probability threshold Perform weighted calculation: ； The weighted calculation result is used as the final damage probability threshold. .

8. The emergency cargo buffering method under rotor failure of a logistics drone according to claim 7, characterized in that, After executing the final protection strategy, the system receives multi-source data after the final protection strategy is executed, iteratively calculates the ratio of the damage probability value to the final damage probability threshold, and optimizes the final protection strategy until the damage probability value in the iterative calculation is less than the final damage probability threshold or the number of iterations is reached, and then outputs the optimized final protection strategy.

9. An emergency cargo buffering system for a logistics drone rotor failure, characterized in that it is used to implement the emergency cargo buffering method for a logistics drone rotor failure as described in any one of claims 1-8, the system being integrated into the cargo box, comprising: The data acquisition module is used to acquire multi-source data of the cargo box after it is impacted while it is currently in the initial protection strategy state. The multi-source data includes sensor data streams, cargo attributes and environmental status. The initial protection strategy is triggered by the cargo box's built-in system when it detects a drone malfunction. The damage probability model processing module is used to preprocess and extract features from the sensor data stream, input it into the pre-trained damage probability inference model, and output the damage probability value and confidence level of the goods in the current cargo box. The threshold adjustment amount acquisition module is used to calculate the weighting coefficients based on the cargo attributes, environmental conditions and damage probability values, and obtain the initial threshold adjustment amount through weighted fusion. Obtain historical damage probability values ​​and corresponding confidence levels, and calculate the slope and mean confidence level representing short-term changes in historical damage probability values ​​based on least squares fitting. If the slope is greater than a preset trend threshold and the mean confidence level is less than a preset confidence threshold, then after removing historical damage probability values ​​with a confidence level less than the confidence threshold, the mean confidence level of the remaining historical damage probability values ​​is recalculated. Calculate the deviation between the current probability of damage and the recalculated mean confidence level, normalize it to obtain the deviation coefficient, and multiply the initial threshold adjustment by the difference between 1 and the deviation coefficient to obtain the final threshold adjustment. The damage probability threshold adjustment module is used to add the final threshold adjustment amount to the initial damage probability threshold to obtain the adjusted damage probability threshold. The final protection strategy generation module is used to generate and execute a final protection strategy if the damage probability value is greater than the adjusted damage probability threshold.

Images