Introduction

The pressure chamber constitutes a pivotal component of an aircraft, primarily functioning to sustain appropriate air pressure and oxygen levels within the cabin, thereby ensuring the safety and comfort of passengers and crew during high-altitude flights. At elevated altitudes, the marked reduction in external air pressure necessitates that the pressure chamber, via the pressurization system, maintains air pressure approximating that at ground level. Additionally, the pressure chamber must possess thermal, acoustic, and impact resistance capabilities to provide a secure and comfortable flight environment1. The optimal performance of the pressure chamber is paramount, as any malfunction in sustaining proper chamber pressure can result in severe safety risks, including hypoxia and decompression sickness. Segmentation control is a sophisticated strategy frequently employed in the management of complex systems, predicated on the principle of dividing the entire system into relatively autonomous subsystems or modules, each with distinct control objectives and strategies. This approach allows for precise regulation of system components, thereby enhancing overall system stability and response time. Within the aerospace domain, segmented control is extensively applied across various vehicle subsystems, such as engine control, air pressure control, and environmental control2. Specifically regarding the management of the pressure chamber, segmented control facilitates the independent monitoring and adjustment of air pressure across different zones, ensuring that the pressure within each zone remains within safe parameters, thus augmenting aircraft safety and passenger comfort.

The significance of troubleshooting the pressure chamber lies in its capacity to detect and rectify potential failures promptly. Given the pressure chamber's critical role in high-altitude flight, an effective fault diagnosis system is imperative to identify and address any anomalies within the pressurization system. Early fault detection can avert catastrophic failures, thereby ensuring the safety of passengers and crew3. Fault detection can primarily be approached as a state identification problem based on time series signals4. Traditional machine learning methods, such as Support Vector Machines and Random Forests, can efficiently identify and classify fault types by learning fault patterns from historical data. SVMs maximize inter-category spacing by constructing hyperplanes, thus achieving efficient classification performance and are particularly suitable for processing linearly separable data. However, SVMs exhibit high computational complexity and limited scalability when handling large-scale datasets and nonlinear data. Conversely, Random Forests, as an ensemble learning method, enhance model robustness and resistance to overfitting. Random Forests possess significant advantages in handling high-dimensional data and addressing missing values and can provide feature importance scores to aid in data comprehension. Nonetheless, Random Forests may perform suboptimally when faced with highly correlated features, and their model interpretability is relatively low. Signal processing methods are equally crucial in fault diagnosis. Common methods include Fourier Transform (FT)5, Wavelet Transform (WT)6, and Empirical Mode Decomposition (EMD)7. FT is suitable for frequency analysis of periodic signals, but has significant limitations when dealing with non-stationary signals; WT can provide time–frequency localization analysis, which improves the accuracy of fault feature extraction; EMD can better capture fault characteristics in nonlinear and non-stationary signals by adaptively decomposing signals, but its noise resistance is weak. Empirical Wavelet Transform (EWT) combines the advantages of WT and EMD, and constructs wavelet bases through adaptive frequency band partitioning to achieve more refined signal decomposition. EWT has particular advantages in fault diagnosis, as it can perform high-resolution decomposition of complex signals, effectively capture fault features, and exhibit strong noise resistance, making it widely recognized in fault recognition applications in complex mechanical systems8.

As the volume of data increases and the complexity of failure modes escalates, deep learning methods are progressively becoming a focal point of research. Deep learning models, particularly Convolutional Neural Networks (CNNs) and LSTM, excel in processing high-dimensional data and extracting intricate features. CNNs autonomously extract features and capture hierarchical information through their multilayered network structure, significantly enhancing classification performance. LSTM, a specialized form of recurrent neural network (RNN), excels in processing and predicting long-distance dependencies in time-series data. LSTM efficiently resolves the gradient vanishing and exploding gradient problems faced by traditional RNNs when managing long-term dependencies by incorporating memory units and a gating mechanism. This makes LSTM particularly adept at processing time-series data, such as barometric pressure monitoring data, by capturing the hierarchical information and improving classification performance. Furthermore, the advantages of deep learning methods in managing nonlinear and high-dimensional data enable them to adapt to more complex fault diagnosis tasks. The combination of deep learning models, especially CNNs and LSTMs, allows for the processing of both time and frequency domain features, offering more comprehensive and accurate fault diagnosis capabilities9. This approach not only enhances the accuracy and real-time performance of fault detection but also demonstrates strong adaptability and robustness in complex environments. In addition to traditional time-domain signals, which reflect the system state during fault diagnosis, frequency-domain signal analysis can effectively reveal the periodicity and frequency characteristics of signals, aiding in the identification and differentiation of various fault modes. By transforming time-domain signals into the frequency domain, fault features obscured by noise can be more readily detected, thereby improving the accuracy and reliability of fault diagnosis.

Therefore, this paper investigates the fault diagnosis problem of the pressure chamber by integrating time–frequency domain signal characteristics and establishes a fault diagnosis network based on EWTLM-FNN. The specific contributions of this paper are as follows:

  1. 1.

    Addressing the denoising requirements of barometric pressure monitoring data, a denoising filtering method is proposed, which significantly mitigates noise interference. This method employs EWT to extract modal information from the barometric pressure data, thereby enhancing data quality.

  2. 2.

    A three-layer LSTM network model is developed to accommodate dual-mode data features in the time and frequency domains. This model outputs the corresponding feature time series through the extraction of time and frequency domain features.

  3. 3.

    The EWTLM-FNN model is constructed by integrating the feature sequences extracted by LSTM with the original signals and the EWT-extracted feature signals. The results demonstrate that the proposed methodological framework performs exceptionally well, optimizing fault classification outcomes.

The remainder of the paper is structured as follows: Section "Related works" presents related works and methods for fault detection. Section "Methodology" introduces the establishment process of EWTLM-FNN. Experimental results and related analysis are detailed in Section "Experiment and analysis". Finally, the conclusion is drawn.

Related works

Dynamic fault diagnosis and machinery health management

As a vital component of spacecraft, the pressure module plays a crucial role. It not only maintains appropriate air pressure and oxygen levels within the capsule to ensure the astronauts' survival and working environment in space but also protects against dangers such as external radiation and micrometeorites10. NASA has conducted extensive research on health management technologies for intelligent aircraft components, encompassing distributed fault-tolerant control11. The primary objectives of intelligent control include: (1) model-based control and diagnostic strategies; (2) control strategies to extend the lifespan of components; (3) control strategies to mitigate the degradation of component performance12. Active component control focuses on: (1) flow control strategies to enhance compressor efficiency, (2) combustion chamber control strategies to reduce emissions, and (3) high-bandwidth turbine gap control strategies. Current research emphasizes optimizing each spacecraft part to achieve high performance through segmented control of overall components13. Gou et al. proposed a control and fault diagnosis method for a pressure sensor-based brake control system, achieving fault detection, isolation, and identification through feed-forward and feedback control, and demonstrated its feasibility in an electrohydraulic braking system14. Wei et al. investigated the dynamic effects of air supply failures on cabin pressure control, particularly for fighter aircraft flying at high altitudes, finding that higher cabin pressure drop rates are necessary to ensure safety15. Yu et al. reviewed the application of model-based fault detection methods in the cabin pressure systems of airliners, detailing the process modeling and fault parameter estimation16. Mukhachev et al. explored model and knowledge-based process fault diagnosis methods, emphasizing fault detection through residual equations and observers despite the scarcity of fault samples17. Xu et al. presented a fault injection and diagnosis methodology for cabin pressure control systems in civil aircraft, utilizing BP neural networks to diagnose typical faults, and verified the effectiveness of the methodology through Rhapsody18.

Fault diagnosis prediction based on neural network approach

Fault diagnosis methods utilizing traditional signal processing techniques typically encompass three steps: signal acquisition, feature extraction, and fault classification. Common signal processing techniques include Fourier transform19, wavelet transform20, and time–frequency analysis21. These methods offer rapid computation speed but require manual feature extraction, introducing a degree of subjectivity. Gu et al.22 established a joint fault diagnosis model incorporating Variational Modal Decomposition (VMD), Continuous Wavelet Transform (CWT), CNN, and Support Vector Machine (SVM). In this model, VMD, CWT, and CNN handle feature extraction from the original data, while SVM is employed for classification, addressing the issue of small sample sizes in rolling bearing faults. Jin et al.23 introduced a Light Neural Network (LiNet) to enhance CNN computational efficiency, achieving nearly 100% recognition accuracy of normal signals in gearbox and bearing datasets. Xin24 proposed a multi-object oriented CNN for fault diagnosis of bearings and gearboxes, utilizing training data from the original vibration signals in multi-domain. Xiao et al.25 proposed an Improved Variational Mode Decomposition (IVMD) combined with CNN for fault diagnosis, addressing the challenge of feature extraction in the time domain inherent in traditional methods. Wang et al.26 developed an improved one-dimensional CNN for bearing fault diagnosis, using collected vibration and sound signals as input data. Experiments demonstrated that one-dimensional convolution enhances feature extraction for temporal signals, and dual-source data input yields higher accuracy than models trained on single-sensor data. Zhang et al.27 proposed an Ensemble Domain Adaptation (EDA) method for rotating machinery fault diagnosis, which effectively removes redundant information, achieving test accuracies of 100%, 99.69%, and 99.92% on three common datasets, respectively.

Through the aforementioned research, it is evident that for the pressure chamber, a spacecraft component with high precision, optimizing control to achieve distributed control of various functional components is essential. With the continuous advancement of neural network technology, integrating diverse information and employing corresponding neural network methods can effectively facilitate fault diagnosis. For the pressure chamber, the primary failure point is pressure leakage, characterized by abnormal dynamic changes in chamber air pressure. By fusing the time sequence characteristics of different information, the corresponding diagnosis can be accomplished using various network models, thereby ensuring safety.

Methodology

EWT

Empirical Wavelet Transform (EWT) is an adaptive signal processing method that integrates the benefits of both Fourier transform and wavelet transform. EWT decomposes a signal by adaptively segmenting its spectrum to construct a set of wavelet basis functions tailored to the signal's characteristics. This approach is particularly effective for processing signals with complex spectra28. In a typical EWT process, an initial Fourier transform is performed to obtain the corresponding spectrum of the signal.

$$ \hat{X}\left( f \right) = {\mathcal{F}}\left\{ {x\left( t \right)} \right\} $$
(1)

where \(\hat{X}\left( f \right)\) is the Fourier transform result of the signal \(x\left( t \right)\) . Then according to the energy distribution of the signal spectrum, the segmentation point is adaptively selected to segment the spectrum into \(N\) sub-bands. After the segmentation of the frequency bands is completed, the corresponding wavelet functions need to be constructed. These wavelet functions are localized Fourier inverse transform results and can be expressed as.

$$ \psi_{i} \left( t \right) = {\mathcal{F}}^{ - 1} \left\{ {W_{i} \left( f \right)} \right\} $$
(2)

Where \(W_{i} \left( f \right)\) is the window function in the first i sub-band. The window function is typically chosen as a tightly supported function, such as a wavelet mother function or a Gaussian function. Upon completing the wavelet construction, the signal is reconstructed using the inverse Fourier transform. The process is illustrated in Eqs. (3) and (4):

$$ x_{i} \left( t \right) = {\mathcal{F}}^{ - 1} \left\{ {\hat{X}\left( f \right) \cdot W_{i} \left( f \right)} \right\} $$
(3)
$$ x\left( t \right) = \mathop \sum \limits_{i = 1}^{N} x_{i} \left( t \right) $$
(4)

For the above the overall algorithmic flow can be represented by Algorithm1:

Algorithm 1
figure a

Empirical Wavelet Transform for the pressure data.

Full size image

LSTM

Considering the barometric pressure monitoring process, LSTM is adopted to extract the sequence feature of the pressure monitoring data. The primary components of LSTM include four key elements: memory cell, forget gate, input gate, and output gate. The forget gate determines which information in the memory cell needs to be discarded. The calculation process for the forget gate is shown in Eq. (5):

$$ f_{t} = \sigma \left( {W_{f} \cdot \left[ {h_{t - 1} ,x_{t} } \right] + b_{f} } \right) $$
(5)

Where ft denotes the output of the forgetting gate, \(\sigma\) is the Sigmoid activation function, Wf is the weight matrix, ht−1 is the hidden state, xt is the input of the current moment, and bf is the bias. The input gate then decides which new information needs to be added to the memory cell, which is calculated as shown in Eq. (6):

$$ i_{t} = \sigma \left( {W_{i} \cdot \left[ {h_{t - 1} ,x_{t} } \right] + b_{i} } \right) $$
(6)

where \(i_{t}\) denotes the output of the input gate, and the other symbols are the same as the forgetting gate. And for the candidate memory state unit, the process is shown in Eq. (7):

$$ \tilde{C}_{t} = {\text{tanh}}\left( {W_{C} \cdot \left[ {h_{t - 1} ,x_{t} } \right] + b_{C} } \right) $$
(7)

where \(\tilde{C}_{t}\) denotes the state of the candidate memory cell,\({\text{tanh}}\) is activation function,\(W_{C}\) is the weight matrix, and \(b_{C}\) is the bias. Then we need to update the memory cell according to the above information to get Eq. (8):

$$ C_{t} = f_{t} \cdot C_{t - 1} + i_{t} \cdot \tilde{C}_{t} $$
(8)

where \(C_{t}\) is the current state of the memory cell and \(C_{t - 1}\) is the previous state of the memory cell. The corresponding hidden state updated as follows:

$$ o_{t} = \sigma \left( {W_{o} \cdot \left[ {h_{t - 1} ,x_{t} } \right] + b_{o} } \right) $$
(9)
$$ h_{t} = o_{t} \cdot {\text{tanh}}\left( {C_{t} } \right) $$
(10)

Where ot denotes the output. ht is the hidden state.

For pressure chamber air pressure monitoring and its fault diagnosis, the active learning of time series information can significantly enhance feature performance by LSTM.

EWTLM-FNN

After completing the corresponding feature extraction, we need to identify and diagnose the fault. In this paper, considering the inherent fluctuations in barometric data, a FNN is selected for the analysis and processing of this data. FNN is a hybrid model that combines fuzzy logic and neural networks, aiming to utilize fuzzy logic to handle uncertainty and imprecise information while leveraging the learning and generalization capabilities of neural networks. FNN excels in managing fuzzy and uncertain data, modeling complex nonlinear systems, and other applications29. The fuzzy neural network primarily consists of four components: fuzzification layer, fuzzy rule layer, inference layer, and defuzzification layer. The calculation of the fuzzification layer is shown in Eq. (11):

$$ \mu_{{A_{i} }} \left( {x_{i} } \right) = \frac{1}{{1 + \left( {\frac{{x_{i} - c_{i} }}{{\sigma_{i} }}} \right)^{2} }} $$
(11)

where \(c_{i}\) is the center of the fuzzy set \(A_{i}\) and \(\sigma_{i}\) is the width parameter. Then, the calculation of various types of weights in the inference layer can be carried out based on the fuzzification layer and the corresponding fuzzy rules. The calculation process is as follows:

$$ \omega_{j} = \mu_{{A_{1j} }} \left( {x_{1} } \right) \cdot \mu_{{A_{2j} }} \left( {x_{2} } \right) $$
(12)

where \(\mu_{{A_{1j} }} \left( {x_{1} } \right)\) represents the corresponding affiliation function. After completing the calculation of the inference weights, the weights are normalized thus obtaining the defuzzification layer to achieve a more accurate output, which is shown in Eqs. (13) and (14):

$$ \overline{\omega }_{j} = \frac{{\omega_{j} }}{{\mathop \sum \nolimits_{k = 1}^{R} \omega_{k} }} $$
(13)
$$ y = \mathop \sum \limits_{j = 1}^{R} \overline{\omega }_{j} \cdot y_{j} $$
(14)

where \(\overline{\omega }_{j}\) is the normalized weights and y is the final output. Based on the time–frequency domain analysis, as well as the deep features of LSTM and the classification performance of FNN, we constructed EWTLM-FNN for monitoring and analyzing the faults of the pressure chamber. The overall flow of the framework is shown in Fig. 1:

Fig. 1
figure 1

The framework for the pressure fault diagnosis.

Full size image

In Fig. 1, we first denoise and filter the barometric pressure monitoring data. Subsequently, we obtain the modal information in the relevant frequency domain of the data using EWT. On this basis, a three-layer LSTM network is employed to extract features in both time and frequency domains, yielding the corresponding sequence of features. These feature sequences, along with the original signal and the EWT-extracted features, are fed into the FNN network for fault identification and diagnosis. The output is the identification of the current state of the pressure chamber based on the barometric pressure signal over a specified period.

To better analyze the recognition performance of pressure chamber faults, it is essential to further examine the timing of the input data and discuss the pressure chamber characteristics under different temporal lengths. After completing the model determination, we designed and built the model parameters. In the EWT spectrum partitioning, we used a frequency band partitioning method based on local extremum, and designed three hidden layers in the LSTM section, each containing 64 nodes. In the FNN design, we also used three layers, each with 64 nodes.

Experiment and analysis

Dataset and experiment setup

After completing the analysis of the corresponding data, we need to further validate the model's performance by testing it with a public dataset. In this paper, we choose the "Sensor Fault Detection Data" from Kaggle30. This dataset includes time series measurements from sensors uniquely identified by their sensor IDs. During these measurements, the sensors may disconnect or fail. This helps us achieve a final and rapid diagnosis using the time series data from the sensors. The dataset contains over 60,000 pieces of sensor monitoring data. The main structure of the data is illustrated in Fig. 2:

Fig. 2
figure 2

The sensor data from the public dataset.

Full size image

Figure 2 shows a period of relatively smooth data, indicating that when the machinery is running smoothly, the sensor data exhibits stable performance. This stability allows for effective data analysis and accurate fault diagnosis, as the timing data remains relatively smooth, facilitating precise fault identification.

After selecting the data, we proceeded with constructing the experimental environment. Since the analysis involves LSTM and related methodologies, specific settings for the experimental and data environments are required. The experimental environment for this paper is detailed in Table 1:

Table 1 The experiment environment.
Full size table

After constructing the experimental environment for the model, we analyzed the experimental data and compared the results with classical fault diagnosis methods, including SDAE31, DSAE32, LSTM, FNN, and the improved EWTLM-FNN framework proposed. These methods were evaluated using classical metrics: accuracy, precision, recall, and F1-score. The comparative analysis is detailed as follows.

Experiment result and analysis

We conducted performance analysis and comparison experiments after validating the model and the related dataset. The fault recognition results of different models on the public dataset are illustrated in Fig. 3:

Fig. 3
figure 3

The fault identification results.

Full size image

In Fig. 3, it can be observed that there are significant differences in recognition accuracy among the various methods. For data not utilizing EWT multimodal feature fusion, the overall performance is relatively general; whether based on LSTM or FNN, the fault diagnosis accuracy exceeds 90%. The current methods commonly used in fault diagnosis, such as SDAE and DSAE, which involve time series anomaly analysis, display traditional method performance with main recognition accuracies of 0.873 and 0.891, respectively. In contrast, the EWTLM-FNN method proposed achieves the highest recognition accuracy of 0.912, effectively accomplishing the structural component inspection process of fault recognition and detection tasks.

In the previous subsection, we noted that the time step in the recognition process of time series data significantly impacts the overall task. Therefore, this paper analyzes the recognition results at different time steps. Recognition accuracies are analyzed at 2, 4, 8, 16, 32, 64, 128, and 256 steps, respectively, and the results are shown in Fig. 4:

Fig. 4
figure 4

The fault identification results under different time steps.

Full size image

In Fig. 4, we can see that the four comparison methods selected all exhibit good stability, with overall recognition accuracy fluctuations under different time steps being less than 0.1. The EWTLM-FNN method shows superior recognition performance, with an average recognition accuracy of 0.907 across different time steps, consistently exceeding 0.9. This indicates that the method can effectively diagnose faults with a smaller amount of data and a narrower time window. These results provide robust methodological support for the application of EWTLM-FNN in future real-time monitoring scenarios.

The practical test for the model

After completing the tests on the public dataset, we proceeded to test the model under actual working conditions. We simulated the air pressure state in a real environment and introduced faults through perturbations. The data obtained during the actual test is shown in Fig. 5. On the left side, the figure displays the air pressure data of the oxygen, while on the right side, it shows the air pressure data of the pressure chamber as a whole. We conducted tests on the small fluctuations in air pressure based on the usage state to facilitate corresponding fault diagnosis.

Fig. 5
figure 5

The self-established dataset for the pressure and oxygen pressure.

Full size image

In Fig. 5, it can be observed that to better simulate the actual application context of the pressure chamber, the overall air pressure fluctuation is set within a smaller range. This approach increases the difficulty of the actual fault detection process, aiming to provide a more rigorous performance evaluation of the model. After data collection, we proceeded with the analysis of the corresponding EWT variations. The main fitted wavelet characteristics during the oxygen barometric pressure variations are illustrated in Fig. 6.

Fig. 6
figure 6

The primary pressure wavelet after EWT.

Full size image

Through Fig. 6, we can observe that the pressure chamber primarily exhibits low-frequency characteristics in its frequency domain, and its overall noise distribution is more regular. This regularity aids in accurately identifying its fluctuation status during actual applications. Similarly, we calculated the main EWT modes of the actual pressure chamber air pressure, and the results are shown in Fig. 7:

Fig. 7
figure 7

The primary oxygen pressure wavelet after EWT.

Full size image

In Fig. 7, it is evident that both oxygen pressure and actual air pressure share similar frequency domain characteristics, with a balanced overall performance. The similarity in frequency domain distribution between oxygen and air pressure indicates that the pressure chamber's overall change characteristics are consistent. This suggests that effective monitoring can be achieved by observing both oxygen and overall air pressure.

After completing the time–frequency domain analysis of the data, we further analyzed the identification accuracy, precision, and other metrics under different models. The results are shown in Fig. 8:

Fig. 8
figure 8

The fault identification results on self-established dataset.

Full size image

In Fig. 8, it is evident that the recognition accuracy of all methods has improved due to the small data size, achieving better results overall. The recognition accuracies of the classical SDAE and DSAE models exceed 0.9, while the recognition accuracy of the EWTLM-FNN model proposed reaches 0.935, indicating a superior recognition effect. The stability of the EWTLM-FNN method at different time steps further demonstrates its advantage.

Therefore, in future applications, this method can effectively realize a more real-time pressure chamber fault diagnosis task, whether it involves monitoring oxygen-related air pressure or actual air pressure monitoring.

Discussion

In this paper, we propose an EWTLM-FNN-based framework for pressure chamber fault diagnosis to address the segmented control of high-precision components such as those in spacecraft. Initially, by denoising and filtering the barometric pressure monitoring data, we effectively reduce noise interference, ensuring data cleanliness and reliability for subsequent analysis. Then, we employ the EWT to obtain the frequency-domain modal information of the data, enriching the feature set by incorporating both time-domain and frequency-domain information. This feature extraction process leverages the adaptive band segmentation capability of EWT, allowing for detailed capture and analysis of frequency domain information. In the feature extraction stage, we use a three-layer LSTM network, renowned for its strong capability in modeling temporal data, effectively capturing temporal dependencies in barometric monitoring data. By extracting features from both domains, the LSTM network outputs the corresponding time series of features. Compared to using LSTM alone for feature extraction, our method provides a richer set of features, enhancing the comprehensiveness and accuracy of the data support for fault diagnosis. Our method combines EWT and FNN to fully utilize the advantages of multiple features and different models, resulting in significant improvements in the accuracy and reliability of fault diagnosis. While the traditional SDAE method has certain advantages in feature extraction, it often struggles with time-series data and frequency-domain features. In contrast, our method enhances the capability of capturing time-series features by introducing LSTM, while integrating the frequency-domain features of EWT and the fuzzy processing capability of FNN. This combination allows for a more comprehensive analysis and diagnosis of pressure chamber faults.

As a crucial component of an aircraft, the primary function of the pressure chamber is to maintain appropriate air pressure and oxygen levels within the chamber, ensuring the survival environment and comfort of passengers and crew during high-altitude flights. However, prolonged operation in complex and extreme environments makes the pressure chamber susceptible to malfunctions, which, if not detected and addressed promptly, can lead to serious safety hazards. The EWTLM-FNN method not only achieves accurate diagnosis of pressure chamber malfunctions but also provides a significant guarantee of safety and reliability in the aerospace field. Through segmented control and monitoring, our research method is capable of analyzing the state of the pressure chamber and predicting faults in real-time. The EWTLM-FNN method effectively minimizes noise influence through denoising filtering in the data preprocessing stage. It then extracts frequency-domain features using EWT and combines these with the deep learning capabilities of LSTM to analyze both time-domain and frequency-domain features comprehensively. Ultimately, the fusion of these features and fault identification using the FNN network allows for precise fault localization and prediction of the pressure chamber's state. By enhancing the early detection capability of pressure chamber faults, we can effectively prevent potential risks and improve the safety and reliability of the aircraft. This advancement safeguards the lives of passengers and crew members, demonstrating significant practical application value and social importance.

Conclusion

In this paper, we propose a fault diagnosis network, EWTLM-FNN, to process barometric pressure monitoring data. Initially, we utilize denoising filtering to process the barometric pressure monitoring data. We then extract modal information in the frequency domain through the EWT and complete the extraction of features in the time and frequency domains using a three-layer Long LSTM network. Finally, we identify and diagnose faults using a FNN. The EWTLM-FNN module captures changes in barometric data across multiple dimensions by integrating feature information from different sources and then employs LSTM for deep fusion of features, enhancing diagnostic accuracy. We tested the EWTLM-FNN model on several barometric pressure monitoring datasets. While the model did not achieve state-of-the-art results, it outperformed traditional fault diagnosis methods across several metrics. In tests using a self-built dataset, the model surpassed other traditional methods, achieving recognition rates exceeding 90% in both the single-sensor-based fault recognition task in the public dataset and the self-built pressure chamber fault simulation detection process. These results demonstrate that the framework makes a significant contribution to future technical support for aircraft pressure chamber faults and sensor data-based fault diagnosis and maintenance.

In future research, we aim to extend the generalization performance of the current model, applying it to a broader range of pressure chamber monitoring data and attempting to fuse additional sensor information, such as temperature and humidity. Although the EWTLM-FNN model performs well on the pressure chamber dataset, further testing is necessary for other types of vehicle fault monitoring.