Deep learning-driven outdoor hangar equipment fault prediction method and system

Abstract

The invention discloses a deep learning-driven outdoor hangar equipment fault prediction method and system. The method comprises the steps of A1, collecting vibration frequency time sequence data and temperature-humidity time sequence data of unmanned aerial vehicle hangar components and temperature-humidity time sequence data of an environment, and performing preprocessing; a2, through a local feature enhancement convolution unit, vibration frequency initial time sequence features and component temperature-humidity initial time sequence features are extracted respectively; a3, noise separation is carried out through a noise-fault distinguishing gating mechanism; a4, early weak fault features are extracted; a5, environment temperature-humidity time sequence features are extracted, and then fault development track features are calculated; a6, calculating a risk quantized value of the unmanned aerial vehicle hangar and judging a risk level; and A7, judging a processing strategy corresponding to the current risk level, generating fault early warning information matched with the risk level, and pushing graded reminding. According to the method, the problem that fault pre-judgment is difficult to realize in an outdoor complex scene by a traditional method can be solved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) hangar technology, and in particular to a deep learning-driven method and system for predicting faults in outdoor hangar equipment. Background Technology

[0002] As the core infrastructure of the drone operation and maintenance system, the stable operation of outdoor hangar equipment directly affects the efficiency and continuity of operation and maintenance. The outdoor hangar equipment fault prediction technology refers to a technical system that predicts potential faults of key components of the equipment by collecting various parameters during the operation of the equipment and combining them with specific analysis methods. It aims to identify fault precursors in advance and trigger maintenance intervention to avoid operation and maintenance interruptions caused by sudden shutdowns.

[0003] Traditional methods for preventing and controlling equipment failures in drone hangars mainly rely on two modes: regular preventive maintenance and reactive maintenance. Regular preventive maintenance sets fixed maintenance cycles based on equipment operating time or experience, which cannot adapt to the different impacts of complex operating conditions such as temperature fluctuations and humidity changes on equipment components in outdoor environments, and is prone to over-maintenance or untimely maintenance. Reactive maintenance, on the other hand, is carried out after the equipment has failed and stopped, which leads to the interruption of drone operation and maintenance, increasing operation and maintenance costs and downtime losses.

[0004] With the expansion of deep learning technology in the field of fault prediction, some solutions have begun to use deep learning models to analyze equipment operation data to achieve fault prediction. Existing technologies are mostly based on models such as convolutional neural networks (CNN) and recurrent neural networks (RNN) to extract features and identify faults for single types of equipment operation parameters, or to conduct fault prediction research on equipment in indoor fixed environments. However, for outdoor hangar equipment, which has strong temporal correlation of core component operation parameters and large parameter fluctuations due to complex outdoor environments during long-term operation, existing deep learning solutions generally suffer from insufficient temporal feature mining, insufficient sensitivity to identify weak fault precursors in complex environments, and difficulty in predicting fault risk levels. They cannot effectively solve the core pain points of difficult prediction of fault precursors of key components in outdoor hangar equipment and the impact of sudden downtime on operation and maintenance efficiency. Therefore, there is an urgent need for a deep learning-driven fault prediction technology that can accurately capture multi-dimensional temporal operation parameter correlation features and adapt to complex outdoor working conditions to improve the accuracy and timeliness of fault prediction. Summary of the Invention

[0005] In view of this, the present invention aims to provide a deep learning-driven method and system for predicting faults in outdoor hangar equipment, in order to solve the problem that traditional methods are difficult to predict faults in complex outdoor scenarios.

[0006] A deep learning-driven method for predicting equipment failures in outdoor hangars includes:

[0007] A1: Collect vibration frequency time-series data, temperature-humidity time-series data, and ambient temperature-humidity time-series data of the drone hangar components and preprocess them to obtain preprocessed component vibration frequency time-series data, preprocessed component temperature-humidity time-series data, and preprocessed ambient temperature-humidity time-series data.

[0008] A2: Based on the preprocessed component vibration frequency time series data and the preprocessed temperature-humidity time series data, preliminary vibration frequency time series features and preliminary component temperature-humidity time series features are extracted respectively through local feature enhancement convolutional units.

[0009] A3: Based on the preliminary time-series characteristics of vibration frequency, the preliminary time-series characteristics of component temperature and humidity, and the pre-processed environmental temperature and humidity time-series data, noise separation is performed through a noise-fault differentiation gating mechanism to obtain the denoised vibration frequency time-series characteristics and the denoised component temperature and humidity time-series characteristics.

[0010] A4: By fusing the denoised vibration frequency timing characteristics and the denoised component temperature-humidity timing characteristics, early weak fault characteristics are obtained;

[0011] A5: Based on the preprocessed ambient temperature-humidity time series data and early weak fault characteristics, extract the ambient temperature-humidity time series features; then calculate the fault development trajectory features through an improved LSTM network.

[0012] A6: Based on the early weak fault characteristics and fault development trajectory characteristics, calculate the early weak fault characteristic weights and fault development trajectory characteristic weights, and fuse them to obtain fault fusion characteristics; then calculate the drone hangar risk quantification value and determine the drone hangar risk level;

[0013] A7: Based on the risk quantification value and risk level of the drone hangar, determine the handling strategy corresponding to the current risk level; based on the handling strategy corresponding to the current risk level, generate fault warning information adapted to the risk level and push tiered reminders.

[0014] Furthermore, step A1 also includes:

[0015] A11: Vibration frequency time-series data of the UAV hangar components are collected by piezoelectric vibration sensors. The data type is time-series waveform data, including the vibration frequency of the motor shaft and the vibration frequency of the telescopic mechanism guide rail. The vibration frequency time-series data is preprocessed using wavelet threshold denoising, three-times standard deviation criterion, and maximum-minimum standardization to obtain the pre-processed component vibration frequency time-series data.

[0016] A12: Temperature-humidity time-series data of the UAV hangar components are collected using a platinum resistance temperature sensor. The data type is time-series numerical data, including motor winding temperature and telescopic mechanism drive module temperature. Temperature-humidity time-series data of the environment are then collected using a temperature and humidity sensor. The data type is time-series numerical data. The temperature-humidity time-series data of the UAV hangar components and the temperature-humidity time-series data of the environment are preprocessed by linear interpolation to complete missing values ​​and Z-score normalization, respectively, to obtain pre-processed component temperature-time-series data and pre-processed ambient temperature-humidity time-series data.

[0017] A13: The time-series data of component vibration frequency, component temperature and humidity after preliminary preprocessing, and ambient temperature and humidity after preliminary preprocessing are time-stamp aligned and data length is unified to obtain the preprocessed time-series data of component vibration frequency, component temperature and humidity, and ambient temperature and humidity.

[0018] Furthermore, the calculation method for step A2 includes:

[0019]

[0020]

[0021] in, Preliminary temporal characteristics of vibration frequency, Convolutional units for enhancing local features For Hadama accumulation, For ReLU function, For batch normalization, This is the preprocessed component vibration frequency time series data. Preliminary time-series characteristics of component temperature and humidity. This is the preprocessed component temperature-humidity time series data. To sum element by element, For global average pooling;

[0022] The calculation method for the local feature enhancement convolutional unit is as follows:

[0023]

[0024]

[0025] in, It is a one-dimensional convolutional layer. The input to the convolutional unit is used to enhance local features. Let x be the adaptive adjustment coefficient of the local receptive field when the input is x. For the Sigmoid function, For average pooling, The weight matrix is ​​adaptively adjusted for the local receptive field. The bias vector is adaptively adjusted for the local receptive field.

[0026] It should be further explained that in the scenario of fault prediction for outdoor hangar equipment, the operating parameters of the core components of the equipment have strong temporal correlations, and the complex working conditions such as temperature fluctuations and humidity changes in the outdoor environment will cause large fluctuations in parameters. At the same time, the signs of failure often show weak characteristics, making it extremely difficult to accurately capture the signs of failure. It is difficult to effectively identify them through conventional feature extraction methods, which in turn affects the accuracy and timeliness of fault prediction.

[0027] This invention constructs a local feature enhancement convolutional unit: First, after receiving preprocessed time-series data (vibration frequency time-series data, temperature-humidity time-series data), the local feature enhancement convolutional unit performs preliminary feature extraction on the input data through a one-dimensional convolutional layer, mining local correlation features in the time-series data to lay the foundation for subsequent feature enhancement and solve the problem of effective information being buried when directly processing the original time-series data; then, it performs preliminary global information integration on the input data through average pooling, and then calculates the adaptive adjustment coefficient of the local receptive field by combining the local receptive field adaptive adjustment weight matrix and bias vector, and outputs the adaptive adjustment coefficient of the local receptive field through the Sigmoid function. This adaptive adjustment coefficient can dynamically adjust the focus range of the local receptive field according to the fluctuation characteristics of the input data, solving the problem of inaccurate local feature capture caused by large parameter fluctuations under complex working conditions, and ensuring that effective features can be extracted adaptively for data with different fluctuation levels; finally, the output of the one-dimensional convolutional layer and the adaptive adjustment coefficient are subjected to a Hadamard product operation to complete the local feature enhancement. Strong processing; for different types of time-series data, the local feature enhancement convolutional unit also adopts differentiated post-processing strategies: for vibration frequency time-series data, the enhanced local features are combined with the input data processed by batch normalization and ReLU function to perform Hadamard product operation. Batch normalization can alleviate the problem of data distribution offset, while the ReLU function realizes the non-linear transformation of features. The combination of the two and fusion with the enhanced features can further improve the expressive power of vibration frequency features and solve the problem of insufficient feature stability caused by large time-series fluctuations in vibration frequency data; for temperature-humidity time-series data, the enhanced local features are summed element-wise with the input data processed by global average pooling. Global average pooling can supplement the global correlation information of temperature-humidity data. After fusion with local enhanced features, the complementarity of local and global features can be realized, solving the problem of incomplete extraction of single local features caused by strong coupling of temperature-humidity data. The collaborative cooperation of the core technology parts finally realizes the feature extraction of the two key time-series data.

[0028] Existing technologies for extracting time-series features related to outdoor hangar equipment mostly rely on conventional convolutional neural networks and recurrent neural networks, and primarily focus on extracting features from single-type equipment operating parameters. They fail to consider the large parameter fluctuations and multi-parameter coupling characteristics under complex outdoor operating conditions. Furthermore, the fixed receptive field of conventional convolutional models cannot adaptively adjust to data fluctuations, resulting in insufficient time-series feature mining and inadequate sensitivity in identifying subtle fault precursors. In contrast, the local feature enhancement convolutional unit of this invention offers significant advantages: First, it possesses the ability to adaptively adjust the local receptive field, dynamically adapting to parameter fluctuations under complex operating conditions, thus overcoming the limitation of fixed receptive fields in existing technologies. Second, it employs differentiated feature fusion strategies to adapt to different types of time-series data, achieving targeted feature extraction from multi-dimensional time-series data and overcoming the limitations of single-parameter processing in existing technologies. Third, through multi-stage feature enhancement and optimization design, it strengthens the local feature expression corresponding to subtle fault precursors, improving the accuracy of feature extraction. This addresses the insufficient sensitivity of existing technologies in identifying subtle fault precursors, providing core support for the accuracy of subsequent fault prediction.

[0029] Furthermore, the calculation method for the noise-fault differentiation gating mechanism in step A3 includes:

[0030]

[0031]

[0032]

[0033] in, For noise-fault differentiation, gating weights, It is a multilayer perceptron. For splicing operations, This represents the mean of the preprocessed ambient temperature-humidity time series data. The vibration frequency time sequence characteristics after denoising. The temperature-humidity time-series characteristics of the component after noise reduction.

[0034] It should be further noted that the vibration frequency time-series characteristics and component temperature-humidity time-series characteristics of outdoor drone hangars are easily mixed with noise signals caused by environmental fluctuations. Moreover, such noise signals have certain similarities in time-series distribution with the characteristics of equipment failure precursors, making it difficult to accurately separate the two. At the same time, the interference of environmental factors on the two core characteristics is coupled, and denoising processing of a single feature dimension cannot completely eliminate noise interference, which can easily lead to the problem of noise being misjudged as a fault or the fault characteristics being masked by noise.

[0035] This invention constructs a noise-fault differentiation gating mechanism: First, a splicing operation is used to fuse the preliminary temporal features of vibration frequency, the preliminary temporal features of component temperature and humidity, and the mean of ambient temperature and humidity time-series data. This fully utilizes the environmental baseline information contained in the mean of environmental parameters, providing a reference for distinguishing noise and fault features. Since the generation of outdoor environmental noise is directly related to fluctuations in ambient temperature and humidity, fusing environmental baseline information with two types of core operational features constructs a correlation between operational features and the environmental baseline, solving the problem that a single operational feature dimension cannot effectively define the source of noise. Subsequently, the fused features are input into a multilayer perceptron for nonlinear mapping, and then the noise-fault differentiation gating weights are output through a Sigmoid function. The multilayer perceptron can learn complex correlation patterns between different feature dimensions, capturing the differential distribution of noise and fault features in the fused feature space. The Sigmoid function then normalizes the mapping results. In the 0-1 range, the quantitative distinction between noise and fault features is achieved. The generation process of the gating weight is essentially an adaptive judgment of noise and fault components in the fused features, solving the drawbacks of traditional denoising methods that rely on manually set thresholds and cannot adapt to dynamic environmental changes. Finally, the preliminary time-series features of vibration frequency and the preliminary time-series features of component temperature and humidity are respectively subjected to Hadamard product operation with the noise-fault distinction gating weight to complete the denoising process. In this step, the gating weight can adaptively filter the information of each dimension in the two types of features—for feature components determined to be related to faults, the gating weight outputs a high value to retain effective information; for components determined to be environmental noise, the gating weight outputs a low value to suppress interference information. This element-wise weighted filtering method realizes the stripping of noise and the preservation of fault features, while ensuring the consistency and synergy of the two core feature denoising processes, avoiding the feature imbalance problem caused by single feature denoising.

[0036] Existing technologies for handling noise interference from outdoor equipment operation characteristics often employ mean filtering, wavelet filtering, or rely directly on the model's generalization ability to suppress noise. Conventional filtering algorithms use fixed filtering parameters, which cannot adapt to the dynamic fluctuations of the outdoor environment, easily leading to over-filtering (false deletion of fault features) or incomplete filtering (residual noise). Methods relying on model generalization ability do not establish a clear distinction between noise and faults, resulting in limited noise suppression. Compared to existing technologies, the noise-fault differentiation gating mechanism of this invention has the following advantages: First, it introduces the mean of ambient temperature and humidity as a differentiation reference, constructing a distinction between noise and faults through multi-feature fusion. Second, it uses adaptive adjustment of gating weights to achieve noise filtering, dynamically adapting to changes in noise characteristics caused by environmental fluctuations. This solves the drawback of existing technologies with fixed parameter filtering that cannot adapt to dynamic environments, improves the effectiveness of denoised features, and ensures the accuracy of subsequent fault prediction. This effectively addresses the core problems of poor noise suppression and the tendency to falsely delete fault features in existing technologies.

[0037] Furthermore, the calculation method for step A4 includes:

[0038]

[0039] in, These are early, minor fault characteristics. For L2 regularization, The vibration frequency weighting matrix is... This is the component temperature-humidity weighting matrix. This is an attention mechanism.

[0040] Furthermore, step A5 also includes:

[0041] A51: Based on the preprocessed ambient temperature-humidity time-series data, the difference between adjacent time steps is calculated and combined in chronological order to obtain ambient temperature-humidity time-series difference data. Then, combined with early weak fault characteristics, ambient temperature-humidity time-series features are extracted. The calculation method is as follows:

[0042]

[0043]

[0044]

[0045] in, Preliminary environmental temperature-humidity time-series characteristics, This is the preprocessed ambient temperature-humidity time series data. For max pooling, For gated loop unit, This is the time-series difference data of ambient temperature and humidity. The environment-fault correlation weight matrix, The time-series characteristics of ambient temperature and humidity;

[0046] A52: The preprocessed ambient temperature-humidity time-series data is input into an improved LSTM network for time-series evolution modeling to calculate the fault development trajectory characteristics. The calculation method is as follows:

[0047]

[0048]

[0049]

[0050]

[0051] in, The characteristics of fault development and fluctuation. For environmental fluctuation adaptive correction coefficients, For covariance calculation, To take the absolute value, To avoid the minimum value where the denominator is 0, The fault is in a hidden state. For Long Short-Term Memory (LSTM) networks, This describes the characteristics of the fault development trajectory.

[0052] It should be further explained that the fault characteristics of drone outdoor hangar equipment will gradually evolve with the dynamic fluctuations of ambient temperature and humidity. Sudden changes in ambient temperature and humidity may temporarily mask the fault evolution trend or accelerate the fault development rate, resulting in the fault trajectory exhibiting non-linear and non-stationary change characteristics. At the same time, the fault development trajectory depends on the cumulative influence of historical environmental data and is closely related to early weak fault characteristics. Single-dimensional data modeling is difficult to capture this evolution pattern.

[0053] To address the aforementioned problems, this invention extracts the temporal features of ambient temperature and humidity and constructs an improved LSTM network: First, in the A51 step, the difference between adjacent time steps is calculated on the preprocessed ambient temperature and humidity temporal data and combined to obtain temporal difference data, capturing the dynamic fluctuation information of ambient temperature and humidity; then, the original environmental data is processed through a combination of one-dimensional convolution and max pooling. One-dimensional convolution can uncover the local temporal correlation features of the environmental data, while max pooling can retain key fluctuation peak information. The two work together to achieve the initial extraction of static environmental features and key fluctuation features; simultaneously, a gated recurrent unit is used to process the temporal difference data and output it through the Sigmoid function. The gated recurrent unit can capture the temporal evolution law in the difference data, and the Sigmoid function can capture the temporal evolution law in the difference data. The `id` function normalizes and enhances the fluctuation features, and then performs a Hadamard product operation with the output of convolution and max pooling to further highlight the effective information in environmental fluctuations and suppress meaningless fluctuation interference. Next, an attention mechanism is used to calculate the environment-fault correlation weight matrix, focusing on the parts of environmental features that are strongly correlated with early weak fault features, achieving alignment between environmental features and fault features, and solving the problem of redundant information in environmental data interfering with fault correlation modeling. Finally, the early weak fault features and preliminary environmental features are concatenated and input into a multilayer perceptron, and then a Hadamard product operation is performed with the correlation weight matrix. The multilayer perceptron achieves nonlinear fusion of the two features, and the correlation weight matrix ensures that the fused environmental features always revolve around the fault-related dimension, ultimately obtaining more targeted environmental temperature-humidity time-series features.

[0054] In step A52, this invention first processes the time-series difference data through a gated recurrent unit and transforms it through a multilayer perceptron to obtain fault development fluctuation characteristics, converting the correlation between environmental fluctuations and fault development into quantifiable fluctuation characteristics. Subsequently, by calculating the absolute value of the covariance between the fault development fluctuation characteristics and the early weak fault characteristics, and combining it with the minimum value, an adaptive correction coefficient for environmental fluctuations is obtained. The covariance calculation is used to quantify the correlation strength between the two, and the correction coefficient enables dynamic assessment of the impact of environmental fluctuations, solving the problem that fixed modeling parameters cannot adapt to changes in environmental fluctuations. Next, the original environmental data and the time-series difference data are concatenated and input into a long short-term memory network, while... By introducing a correction coefficient and weighting it according to the deviation of environmental data and environmental mean, the Long Short-Term Memory (LSTM) network is able to capture long-term temporal dependencies. The spliced ​​data provides complete static and dynamic environmental information, while the weighting effect of the correction coefficient can dynamically adjust the weight of the input information according to the degree of influence of environmental fluctuations on the fault, so as to offset the interference of sudden environmental changes on fault trajectory modeling and ensure that the LSM network can focus on the evolution law of the fault itself. Finally, the hidden state of the LSM network is fused with the early weak fault features through an attention mechanism to obtain the fault development trajectory features, ensuring that the fault development trajectory features can accurately reflect the law of environmental influence and fault evolution.

[0055] Existing technologies often employ conventional Long Short-Term Memory (LSTM) networks to directly process single-type time-series data. Even when environmental data is introduced, some solutions simply concatenate it before inputting it into the model, failing to consider the dynamic characteristics of environmental fluctuations and their correlation with early fault features. Furthermore, the fixed input parameters of conventional LSTM networks prevent adaptive adjustment of modeling weights based on environmental fluctuations. This makes the model susceptible to environmental noise interference during periods of significant environmental volatility, hindering the accurate capture of the true evolution trajectory of the fault. Additionally, the lack of targeted fusion of environmental and fault features leads to discrepancies between the modeling results and the actual fault development patterns. In contrast, the improved LSTM network of this invention introduces adaptive correction coefficients for environmental fluctuations, dynamically adjusting modeling weights based on the correlation strength between the environment and the fault. This overcomes the drawbacks of fixed parameters in conventional LSTM networks, which cannot adapt to dynamic environmental changes. Compared to the simple data concatenation or single-modeling methods of existing technologies, it significantly improves the accuracy and reliability of fault development trajectory features, providing crucial support for the timeliness and accuracy of subsequent fault prediction.

[0056] Furthermore, step A6 also includes:

[0057] A61: Calculate the weights of early weak fault characteristics and fault development trajectory characteristics based on the characteristics of early weak faults and the characteristics of fault development trajectory characteristics.

[0058] A62: Based on the weights of early weak fault features and fault development trajectory features, the early weak fault features and fault development trajectory features are fused to obtain fault fusion features;

[0059] A63: Based on fault fusion characteristics, calculate the risk quantification value of the drone hangar and determine the risk level of the drone hangar.

[0060] This invention also discloses a deep learning-driven outdoor hangar equipment fault prediction system, comprising:

[0061] Data acquisition module: Collects vibration frequency time-series data, temperature-humidity time-series data, and ambient temperature-humidity time-series data of the drone hangar components and performs preprocessing to obtain preprocessed component vibration frequency time-series data, preprocessed component temperature-humidity time-series data, and preprocessed ambient temperature-humidity time-series data.

[0062] Feature extraction module: Based on the preprocessed component vibration frequency time series data and the preprocessed temperature-humidity time series data, the module extracts the preliminary time series features of vibration frequency and the preliminary time series features of component temperature-humidity through the local feature enhancement convolution unit.

[0063] Feature denoising module: Based on the preliminary time-series characteristics of vibration frequency, the preliminary time-series characteristics of component temperature and humidity, and the pre-processed environmental temperature and humidity time-series data, noise separation is performed through a noise-fault differentiation gating mechanism to obtain the denoised vibration frequency time-series characteristics and the denoised component temperature and humidity time-series characteristics.

[0064] Feature fusion module: fuses the denoised vibration frequency time-series features and the denoised component temperature-humidity time-series features to obtain early weak fault features;

[0065] Fault development trajectory feature extraction module: Based on the preprocessed ambient temperature-humidity time series data and early weak fault features, the ambient temperature-humidity time series features are extracted; then, the fault development trajectory features are calculated through an improved LSTM network.

[0066] Risk quantification module: Based on the early weak fault characteristics and fault development trajectory characteristics, calculate the weights of the early weak fault characteristics and the fault development trajectory characteristics, and fuse them to obtain the fault fusion characteristics; then calculate the risk quantification value of the drone hangar and determine the risk level of the drone hangar;

[0067] Early warning module: Based on the risk quantification value and risk level of the drone hangar, determine the handling strategy corresponding to the current risk level; based on the handling strategy corresponding to the current risk level, generate fault early warning information adapted to the risk level and push tiered reminders.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] (1) This invention effectively solves the problems of poor adaptability of traditional maintenance to complex outdoor working conditions, and the interruption of operation and maintenance and the surge in costs caused by post-event repairs. At the same time, it makes up for the shortcomings of existing deep learning solutions in the prediction of equipment failure in outdoor hangars, such as insufficient mining of time series features and insufficient sensitivity in the identification of weak fault precursors. It first collects multi-dimensional time series data and preprocesses it, then extracts key operating features through local feature enhancement, removes environmental interference through the noise-fault differentiation mechanism, and fuses early weak fault features. Subsequently, it combines environmental features with an improved LSTM network to capture the fault development trajectory, and finally fuses features to quantify the risk level and push graded early warnings to form a complete fault prediction link. It can accurately adapt to the fluctuation characteristics of the outdoor environment, improve the accuracy and timeliness of fault prediction, identify potential fault risks in advance, effectively avoid sudden shutdown of UAV hangars, and reduce operation and maintenance costs and losses.

[0070] (2) In response to the problem of large parameter fluctuations and weak fault precursors that are difficult to capture in the complex working conditions of outdoor hangar equipment fault prediction, this invention innovatively constructs a local feature enhancement convolutional unit. It extracts local correlation features through one-dimensional convolution, and combines average pooling, weight bias calculation and Sigmoid function to generate adaptive adjustment coefficients to dynamically adjust the receptive field. It adopts a differentiated fusion strategy for vibration frequency and temperature and humidity data, which can accurately adapt to parameter fluctuation characteristics, realize targeted feature extraction of multi-dimensional data, strengthen the expression of weak fault precursors, significantly improve the accuracy of feature extraction, and provide core support for the accuracy of fault prediction.

[0071] (3) In view of the problem that the vibration frequency and component temperature-humidity time sequence characteristics of outdoor UAV hangars are easily mixed with environmental noise, this invention innovatively constructs a noise-fault differentiation gating mechanism, splices three features to construct the correlation dimension, generates adaptive gating weights through multilayer perceptron nonlinear mapping and Sigmoid function, and then uses Hadamard product to denoise the two core features in a coordinated manner, establishes a clear noise-fault differentiation basis, realizes adaptive screening that dynamically adapts to environmental fluctuations, avoids over-filtering or residual noise, ensures complete preservation of fault features, and improves the effectiveness of features after denoising.

[0072] (4) In view of the fact that the fault characteristics of outdoor UAV hangar equipment will gradually evolve with the dynamic fluctuation of ambient temperature and humidity, this invention innovatively extracts environmental time series features and constructs an improved LSTM network. By calculating the covariance, an adaptive correction coefficient for environmental fluctuation is generated. The data is spliced ​​into a long short-term memory network and weighted for correction. This accurately captures the correlation between the environment and fault evolution, dynamically adapts to environmental fluctuations, and offsets the interference of sudden environmental changes. It ensures that the fault trajectory features accurately reflect the evolution law, improves the modeling accuracy, and provides support for the effectiveness of fault prediction. Attached Figure Description

[0073] Figure 1 A flowchart illustrating a deep learning-driven method for predicting faults in outdoor hangar equipment provided by this invention.

[0074] Figure 2 The graph shows the trend of the loss value during the training process of the model provided by this invention as a function of iteration rounds. Detailed Implementation

[0075] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0076] Example 1: A deep learning-driven method for predicting faults in outdoor hangar equipment, such as... Figure 1 As shown, it includes the following steps:

[0077] A1: Collect and preprocess the vibration frequency time-series data, temperature-humidity time-series data, and ambient temperature-humidity time-series data of the UAV hangar components to obtain preprocessed component vibration frequency time-series data, preprocessed component temperature-humidity time-series data, and preprocessed ambient temperature-humidity time-series data, including:

[0078] A11: Vibration frequency time-series data of the UAV hangar components are collected by piezoelectric vibration sensors. The data type is time-series waveform data, including the vibration frequency of the motor shaft and the vibration frequency of the telescopic mechanism guide rail. The vibration frequency time-series data is preprocessed using wavelet threshold denoising, three-times standard deviation criterion, and maximum-minimum standardization to obtain the pre-processed component vibration frequency time-series data.

[0079] A12: Temperature-humidity time-series data of the UAV hangar components are collected using a platinum resistance temperature sensor. The data type is time-series numerical data, including motor winding temperature and telescopic mechanism drive module temperature. Temperature-humidity time-series data of the environment are then collected using a temperature and humidity sensor. The data type is time-series numerical data. The temperature-humidity time-series data of the UAV hangar components and the temperature-humidity time-series data of the environment are preprocessed by linear interpolation to complete missing values ​​and Z-score normalization, respectively, to obtain pre-processed component temperature-time-series data and pre-processed ambient temperature-humidity time-series data.

[0080] A13: The time-series data of component vibration frequency, component temperature and humidity after preliminary preprocessing, and ambient temperature and humidity after preliminary preprocessing are time-stamp aligned and data length is unified to obtain the preprocessed time-series data of component vibration frequency, component temperature and humidity, and ambient temperature and humidity.

[0081] A2: Based on the preprocessed component vibration frequency time-series data and the preprocessed temperature-humidity time-series data, preliminary vibration frequency time-series features and preliminary component temperature-humidity time-series features are extracted using a local feature enhancement convolutional unit, including:

[0082]

[0083]

[0084] in, Preliminary temporal characteristics of vibration frequency, Convolutional units for enhancing local features For Hadama accumulation, For ReLU function, For batch normalization, This is the preprocessed component vibration frequency time series data. Preliminary time-series characteristics of component temperature and humidity. This is the preprocessed component temperature-humidity time series data. To sum element by element, For global average pooling;

[0085] The calculation method for the local feature enhancement convolutional unit is as follows:

[0086]

[0087]

[0088] in, It is a one-dimensional convolutional layer. The input to the convolutional unit is used to enhance local features. Let x be the adaptive adjustment coefficient of the local receptive field when the input is x. For the Sigmoid function, For average pooling, The weight matrix is ​​adaptively adjusted for the local receptive field. The bias vector is adaptively adjusted for the local receptive field.

[0089] Specifically, for scenarios where severe weather conditions such as strong winds and heavy rain cause drastic fluctuations in equipment operating parameters and significantly mask early signs of failure, this invention also provides a multi-scale enhanced local feature enhancement convolutional unit calculation method to replace the traditional calculation method for local feature enhancement convolutional units. The calculation method is as follows:

[0090]

[0091]

[0092] in, This is a one-dimensional convolutional layer of the first scale. It is a one-dimensional convolutional layer of the second scale. For adaptive average pooling.

[0093] A3: Based on the preliminary time-series characteristics of vibration frequency, the preliminary time-series characteristics of component temperature-humidity, and the pre-processed ambient temperature-humidity time-series data, noise separation is performed through a noise-fault differentiation gating mechanism to obtain the denoised vibration frequency time-series characteristics and the denoised component temperature-humidity time-series characteristics, including:

[0094]

[0095]

[0096]

[0097] in, For noise-fault differentiation, gating weights, It is a multilayer perceptron. For splicing operations, This represents the mean of the preprocessed ambient temperature-humidity time series data. The vibration frequency time sequence characteristics after denoising. The temperature-humidity time-series characteristics of the component after noise reduction.

[0098] A4: By fusing the denoised vibration frequency timing characteristics and the denoised component temperature-humidity timing characteristics, early subtle fault characteristics are obtained, including:

[0099]

[0100] in, These are early, minor fault characteristics. For L2 regularization, The vibration frequency weighting matrix is... This is the component temperature-humidity weighting matrix. This is an attention mechanism.

[0101] A5: Based on the preprocessed ambient temperature-humidity time-series data and early weak fault characteristics, extract the ambient temperature-humidity time-series features; then, using an improved LSTM network, calculate the fault development trajectory features, including:

[0102] A51: Based on the preprocessed ambient temperature-humidity time-series data, the difference between adjacent time steps is calculated and combined in chronological order to obtain ambient temperature-humidity time-series difference data. Then, combined with early weak fault characteristics, ambient temperature-humidity time-series features are extracted. The calculation method is as follows:

[0103]

[0104]

[0105]

[0106] in, Preliminary environmental temperature-humidity time-series characteristics, This is the preprocessed ambient temperature-humidity time series data. For max pooling, For gated loop unit, This is the time-series difference data of ambient temperature and humidity. The environment-fault correlation weight matrix, The time-series characteristics of ambient temperature and humidity;

[0107] A52: The preprocessed ambient temperature-humidity time-series data is input into an improved LSTM network for time-series evolution modeling to calculate the fault development trajectory characteristics. The calculation method is as follows:

[0108]

[0109]

[0110]

[0111]

[0112] in, The characteristics of fault development and fluctuation. For environmental fluctuation adaptive correction coefficients, For covariance calculation, To take the absolute value, To avoid the minimum value where the denominator is 0, The fault is in a hidden state. For Long Short-Term Memory (LSTM) networks, This describes the characteristics of the fault development trajectory.

[0113] In this embodiment, the parameter settings for the related neural network module of the improved LSTM network are as follows:

[0114] The GRU unit is configured as a single-layer structure with 64 hidden layer units, using tanh as the activation function, a dropout probability of 0.2, and an input sequence length of 32; the weight parameters of the GRU unit are initialized using the Xavier method.

[0115] The MLP consists of three fully connected layers. The number of neurons in the input layer matches the output dimension of the GRU unit (64 dimensions). The number of neurons in the hidden layer is 128 and 64 respectively, and the number of neurons in the output layer is 32. ReLU is used as the activation function in each hidden layer, and there is no activation function in the output layer. Batch normalization is used between layers, and He initialization is used for weight initialization.

[0116] The LSTM network is configured with a 2-layer structure, with 128 hidden units in each layer. The forget gate, input gate, and output gate all use Sigmoid as the activation function, the cell state update uses the tanh activation function, the dropout probability is 0.2, the recursive dropout probability is 0.1, and the weight parameters are initialized using Xavier.

[0117] The attention mechanism employs an additive attention structure, with the query vector dimension and early weak fault features. The dimensions are consistent with those of the LSTM network (64 dimensions), and the dimensions of the key vector and value vector are consistent with those of the hidden state H of the LSTM network (128 dimensions). The attention score is calculated using the Softmax function for normalization, and the weights are initialized using Xavier initialization.

[0118] The value of ε is 1e-6 to ensure that the denominator is non-zero.

[0119] A6: Based on the characteristics of early minor faults and the characteristics of fault development trajectory, calculate the weights of early minor fault characteristics and fault development trajectory characteristics, and fuse them to obtain the fault fusion characteristics; then calculate the quantified risk value of the drone hangar and determine the risk level of the drone hangar, including:

[0120] A61: Based on the characteristics of early-stage weak faults and the characteristics of fault development trajectory, calculate the weights of early-stage weak fault characteristics and fault development trajectory characteristics. The calculation method is as follows:

[0121]

[0122]

[0123]

[0124] in, The attention matrix contributes to the fault features. Weights for early, weak fault characteristics. Calculated for the mean. Weights for fault development trajectory features;

[0125] A62: Based on the weights of early weak fault features and fault development trajectory features, the early weak fault features and fault development trajectory features are fused to obtain fault fusion features. The calculation method is as follows:

[0126]

[0127] in, This is a fault fusion feature;

[0128] A63: Based on fault fusion characteristics, calculate the risk quantification value of the drone hangar and determine the risk level of the drone hangar. The calculation method is as follows:

[0129]

[0130]

[0131] in, Quantifying the risk of drone hangars. Risk level of drone hangar These are the first risk threshold and the second risk threshold, respectively.

[0132] A7: Based on the risk quantification value and risk level of the drone hangar, determine the handling strategy corresponding to the current risk level; based on the handling strategy corresponding to the current risk level, generate fault warning information adapted to the risk level and push tiered reminders.

[0133] The neural network modules involved in this invention, such as the local feature enhancement convolutional unit, noise-fault differentiation gating mechanism, improved LSTM network, multilayer perceptron, and attention mechanism, all employ existing common end-to-end training methods for model optimization. The training process uses preprocessed time-series data as input, dividing the dataset into training, validation, and test sets according to a preset ratio. The training set is used for iterative updates of model parameters, the validation set is used to monitor overfitting and adjust hyperparameters during training, and the test set is used for final model performance evaluation. Model training uses a combination of cross-entropy and mean squared error loss functions to calculate losses for both the classification task of fault risk level determination and the regression task of risk quantification prediction. The loss values ​​are propagated back to each neural network module via backpropagation to solve for parameter gradients. The Adam optimizer is used, adaptively adjusting the learning rate to balance training speed and convergence stability, accelerating the convergence of model parameters to the optimal solution. During training, overfitting is suppressed using methods such as L2 regularization, and training terminates when the validation set loss value does not decrease for a preset number of consecutive rounds. The change process of the loss function during training is as follows: Figure 2 As shown.

[0134] This embodiment is applied to an outdoor drone hangar deployed in the southern coastal area. The monitoring object is the hangar telescopic mechanism, and the core components are the telescopic guide rail and the drive motor. There are 24 consecutive hours of rainy and windy weather with ambient temperature fluctuations of 18-25℃ and relative humidity fluctuations of 85%-95%. The strong winds cause the hangar structure to sway slightly.

[0135] The preprocessed data is shown below:

[0136] Table 1. Various types of data after preprocessing

[0137] Time step Preprocessed component vibration frequency timing data Pre-processed component temperature-humidity time series data Preprocessed ambient temperature-humidity time series data 1 0.12 0.25 / 0.88 0.30 / 0.92 4 0.15 0.28 / 0.89 0.45 / 0.95 8 0.18 0.30 / 0.90 0.20 / 0.93 12 0.23 0.35 / 0.91 0.50 / 0.94 16 0.31 0.42 / 0.92 0.35 / 0.90 20 0.45 0.55 / 0.93 0.40 / 0.88 24 0.62 0.68 / 0.91 0.32 / 0.85

[0138] This embodiment sets a risk threshold: =0.3, =0.7, the risk quantification value and risk level are calculated, and the specific results are as follows:

[0139] Table 2 Risk Quantification Values ​​and Risk Levels

[0140] Time step Risk Quantification Risk level 1 0.15 Low risk 4 0.18 Low risk 8 0.22 Low risk 12 0.33 Medium risk 16 0.48 Medium risk 20 0.65 Medium risk 24 0.78 High risk

[0141] Traditional methods use a fixed 72-hour maintenance cycle. In this scenario, the drone has already entered a high-risk state after 24 hours, but traditional methods require waiting until 72 hours before maintenance is carried out. During this period, the telescopic mechanism may jam or malfunction, causing the drone to be unable to leave the warehouse for normal operation, resulting in maintenance interruption and losses.

[0142] This invention monitors the dynamic changes of three types of time-series data in real time, accurately captures the "fault evolution process masked by environmental interference", and issues a medium-risk warning within 12 hours by pushing a reminder to "strengthen the status monitoring of the telescopic mechanism". It also issues a high-risk warning within 24 hours by pushing a reminder to "stop the machine for maintenance immediately", thus enabling early identification and graded intervention of fault precursors.

[0143] Example 2: This invention also discloses a deep learning-driven outdoor hangar equipment fault prediction system, comprising:

[0144] Data acquisition module: Collects vibration frequency time-series data, temperature-humidity time-series data, and ambient temperature-humidity time-series data of the drone hangar components and performs preprocessing to obtain preprocessed component vibration frequency time-series data, preprocessed component temperature-humidity time-series data, and preprocessed ambient temperature-humidity time-series data.

[0145] Feature extraction module: Based on the preprocessed component vibration frequency time series data and the preprocessed temperature-humidity time series data, the module extracts the preliminary time series features of vibration frequency and the preliminary time series features of component temperature-humidity through the local feature enhancement convolution unit.

[0146] Feature denoising module: Based on the preliminary time-series characteristics of vibration frequency, the preliminary time-series characteristics of component temperature and humidity, and the pre-processed environmental temperature and humidity time-series data, noise separation is performed through a noise-fault differentiation gating mechanism to obtain the denoised vibration frequency time-series characteristics and the denoised component temperature and humidity time-series characteristics.

[0147] Feature fusion module: fuses the denoised vibration frequency time-series features and the denoised component temperature-humidity time-series features to obtain early weak fault features;

[0148] Fault development trajectory feature extraction module: Based on the preprocessed ambient temperature-humidity time series data and early weak fault features, the ambient temperature-humidity time series features are extracted; then, the fault development trajectory features are calculated through an improved LSTM network.

[0149] Risk quantification module: Based on the early weak fault characteristics and fault development trajectory characteristics, calculate the weights of the early weak fault characteristics and the fault development trajectory characteristics, and fuse them to obtain the fault fusion characteristics; then calculate the risk quantification value of the drone hangar and determine the risk level of the drone hangar;

[0150] Early warning module: Based on the risk quantification value and risk level of the drone hangar, determine the handling strategy corresponding to the current risk level; based on the handling strategy corresponding to the current risk level, generate fault early warning information adapted to the risk level and push tiered reminders.

[0151] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0152] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0153] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A deep learning-driven method for predicting faults in outdoor hangar equipment, characterized in that, Includes the following steps: A1: Collect vibration frequency time-series data, temperature-humidity time-series data, and ambient temperature-humidity time-series data of the drone hangar components and preprocess them to obtain preprocessed component vibration frequency time-series data, preprocessed component temperature-humidity time-series data, and preprocessed ambient temperature-humidity time-series data. A2: Based on the preprocessed component vibration frequency time series data and the preprocessed component temperature-humidity time series data, preliminary vibration frequency time series features and preliminary component temperature-humidity time series features are extracted respectively through local feature enhancement convolutional units. A3: Based on the preliminary time-series characteristics of vibration frequency, the preliminary time-series characteristics of component temperature and humidity, and the pre-processed environmental temperature and humidity time-series data, noise separation is performed through a noise-fault differentiation gating mechanism to obtain the denoised vibration frequency time-series characteristics and the denoised component temperature and humidity time-series characteristics. A4: By fusing the denoised vibration frequency timing characteristics and the denoised component temperature-humidity timing characteristics, early weak fault characteristics are obtained; A5: Extract the ambient temperature-humidity time series features based on the preprocessed ambient temperature-humidity time series data and early weak fault characteristics; Then, the fault development trajectory characteristics are calculated using an improved LSTM network; A6: Based on the early weak fault characteristics and fault development trajectory characteristics, calculate the early weak fault characteristic weights and fault development trajectory characteristic weights, and fuse them to obtain fault fusion characteristics; then calculate the drone hangar risk quantification value and determine the drone hangar risk level; A7: Determine the handling strategy corresponding to the current risk level based on the quantified risk value of the drone hangar and the risk level of the drone hangar; Based on the handling strategy corresponding to the current risk level, generate fault warning information that is appropriate to the risk level and push tiered reminders.

2. The deep learning-driven outdoor hangar equipment fault prediction method according to claim 1, characterized in that, Step A1 includes: A11: Vibration frequency time-series data of the UAV hangar components are collected by piezoelectric vibration sensors. The data type is time-series waveform data, including the vibration frequency of the motor shaft and the vibration frequency of the telescopic mechanism guide rail. The vibration frequency time-series data is preprocessed using wavelet threshold denoising, three-times standard deviation criterion, and maximum-minimum standardization to obtain the pre-processed component vibration frequency time-series data. A12: Temperature-humidity time-series data of the UAV hangar components are collected using a platinum resistance temperature sensor. The data type is time-series numerical data, including motor winding temperature and telescopic mechanism drive module temperature. Temperature-humidity time-series data of the environment are then collected using a temperature and humidity sensor. The data type is time-series numerical data. The temperature-humidity time-series data of the UAV hangar components and the temperature-humidity time-series data of the environment are preprocessed by linear interpolation to complete missing values ​​and Z-score normalization, respectively, to obtain pre-processed component temperature-time-series data and pre-processed ambient temperature-humidity time-series data. A13: The time-series data of component vibration frequency, component temperature and humidity after preliminary preprocessing, and ambient temperature and humidity after preliminary preprocessing are time-stamp aligned and data length is unified to obtain the preprocessed time-series data of component vibration frequency, component temperature and humidity, and ambient temperature and humidity.

3. The deep learning-driven outdoor hangar equipment fault prediction method according to claim 1, characterized in that, The calculation method for step A2 includes: in, Preliminary temporal characteristics of vibration frequency, Convolutional units for enhancing local features For Hadama accumulation, For ReLU function, For batch normalization, This is the preprocessed component vibration frequency time series data. Preliminary time-series characteristics of component temperature and humidity. This is the preprocessed component temperature-humidity time series data. To sum element by element, For global average pooling; The calculation method for the local feature enhancement convolutional unit is as follows: in, It is a one-dimensional convolutional layer. The input to the convolutional unit is used to enhance local features. Let x be the adaptive adjustment coefficient of the local receptive field when the input is x. For the Sigmoid function, For average pooling, The weight matrix is ​​adaptively adjusted for the local receptive field. The bias vector is adaptively adjusted for the local receptive field.

4. The deep learning-driven outdoor hangar equipment fault prediction method according to claim 3, characterized in that, The calculation method for the noise-fault differentiation gating mechanism in step A3 includes: in, Apply noise-fault differentiation gating weights. It is a multilayer perceptron. For splicing operations, This represents the mean of the preprocessed ambient temperature-humidity time series data. The vibration frequency time sequence characteristics after denoising. The temperature-humidity time-series characteristics of the component after noise reduction.

5. The deep learning-driven outdoor hangar equipment fault prediction method according to claim 4, characterized in that, The calculation method for step A4 includes: in, These are early, minor fault characteristics. For L2 regularization, The vibration frequency weighting matrix is... This is the component temperature-humidity weighting matrix. This is an attention mechanism.

6. The deep learning-driven outdoor hangar equipment fault prediction method according to claim 5, characterized in that, Step A5 includes: A51: Based on the preprocessed ambient temperature-humidity time-series data, the difference between adjacent time steps is calculated and combined in chronological order to obtain ambient temperature-humidity time-series difference data. Then, combined with early weak fault characteristics, ambient temperature-humidity time-series features are extracted. The calculation method is as follows: in, Preliminary environmental temperature-humidity time-series characteristics, This is the preprocessed ambient temperature-humidity time series data. For max pooling, For gated loop unit, This is the time-series difference data of ambient temperature and humidity. The environment-fault correlation weight matrix, The time-series characteristics of ambient temperature and humidity; A52: The preprocessed ambient temperature-humidity time-series data is input into an improved LSTM network for time-series evolution modeling to calculate the fault development trajectory characteristics. The calculation method is as follows: in, The characteristics of fault development and fluctuation. For environmental fluctuation adaptive correction coefficients, For covariance calculation, To take the absolute value, To avoid the minimum value where the denominator is 0, The fault is in a hidden state. For Long Short-Term Memory (LSTM) networks, This describes the characteristics of the fault development trajectory.

7. The deep learning-driven outdoor hangar equipment fault prediction method according to claim 6, characterized in that, Step A6 includes: A61: Calculate the weights of early weak fault characteristics and fault development trajectory characteristics based on the characteristics of early weak faults and the characteristics of fault development trajectory characteristics. A62: Based on the weights of early weak fault features and fault development trajectory features, the early weak fault features and fault development trajectory features are fused to obtain fault fusion features; A63: Based on fault fusion characteristics, calculate the risk quantification value of the drone hangar and determine the risk level of the drone hangar.

8. A deep learning-driven fault prediction system for outdoor hangar equipment, characterized in that, include: Data acquisition module: Collects vibration frequency time-series data, temperature-humidity time-series data, and ambient temperature-humidity time-series data of the drone hangar components and performs preprocessing to obtain preprocessed component vibration frequency time-series data, preprocessed component temperature-humidity time-series data, and preprocessed ambient temperature-humidity time-series data. Feature extraction module: Based on the preprocessed component vibration frequency time series data and the preprocessed temperature-humidity time series data, the module extracts the preliminary time series features of vibration frequency and the preliminary time series features of component temperature-humidity through the local feature enhancement convolution unit. Feature denoising module: Based on the preliminary time-series characteristics of vibration frequency, the preliminary time-series characteristics of component temperature and humidity, and the pre-processed environmental temperature and humidity time-series data, noise separation is performed through a noise-fault differentiation gating mechanism to obtain the denoised vibration frequency time-series characteristics and the denoised component temperature and humidity time-series characteristics. Feature fusion module: fuses the denoised vibration frequency time-series features and the denoised component temperature-humidity time-series features to obtain early weak fault features; Fault development trajectory feature extraction module: Extracts ambient temperature-humidity time series features based on preprocessed ambient temperature-humidity time series data and early weak fault features; Then, the fault development trajectory characteristics are calculated using an improved LSTM network; Risk quantification module: Based on the early weak fault characteristics and fault development trajectory characteristics, calculate the weights of the early weak fault characteristics and the fault development trajectory characteristics, and fuse them to obtain the fault fusion characteristics; then calculate the risk quantification value of the drone hangar and determine the risk level of the drone hangar; Early warning module: Based on the risk quantification value and risk level of the drone hangar, determine the corresponding handling strategy for the current risk level; Based on the processing strategy corresponding to the current risk level, generate fault warning information adapted to the risk level and push graded reminders; so as to realize the deep learning-driven outdoor hangar equipment fault prediction method as described in any one of claims 1-7.

Images