Boarding bridge fault diagnosis method, system and device based on physical constraint neural network and storage medium

Abstract

The invention discloses a boarding bridge fault diagnosis method, system and device based on a physical constraint neural network, and a storage medium, and the method comprises the steps: collecting the multi-source sensor data of a boarding bridge in real time, and carrying out the preprocessing, and obtaining the preprocessed sensor data; determining a physical constraint equation based on the basic physical law and a physical quantity range which can be measured by the multi-source sensor data; constructing a physical constraint neural network model which comprises a data feature extraction network and a physical constraint fusion layer; the data feature extraction network adopts a parallel combination structure of a convolutional neural network and a long-short-term memory network; the physical constraint fusion layer fuses the extracted data features with a physical constraint equation to construct a joint loss function; historical operation data of the boarding bridge are collected and input into the neural network model for training; and inputting sensor data acquired in real time into the trained neural network model, and outputting a boarding bridge fault diagnosis result. And intelligent and precise fault diagnosis of the boarding bridge equipment is realized.

Description

Technical Field

[0001] This invention relates to a boarding bridge fault diagnosis method, specifically to a boarding bridge fault diagnosis method, system, device, and storage medium based on a physical constraint neural network. Background Technology

[0002] Passenger boarding bridges (also known as jet bridges or boarding bridges) are essential ground support equipment in modern airports, connecting the terminal building to aircraft and providing passengers with a safe and convenient boarding and disembarking route. The boarding bridge system integrates multiple subsystems, including mechanical, hydraulic, and electrical systems. It has a complex structure, operates in harsh environments, and undergoes high-frequency, heavy-load operation, resulting in a relatively high failure rate. Boarding bridge malfunctions not only affect normal flight operations and cause delays but can also endanger passenger safety, leading to significant losses for airport operations.

[0003] Existing boarding bridge fault diagnosis technologies primarily rely on manual inspections, with maintenance personnel periodically observing equipment operation to determine the presence of faults. This approach is not only inefficient and unable to achieve real-time monitoring, but also heavily dependent on the experience level of maintenance personnel, resulting in high labor costs. Current methods employ threshold-based automatic alarm systems, installing sensors in key areas to trigger alarms when monitored parameters exceed preset thresholds. However, this method can only detect single-parameter anomalies, cannot analyze multi-parameter coupled faults, and threshold setting is difficult, easily leading to false alarms or missed alarms.

[0004] With the development of machine learning technology, algorithms such as support vector machines and random forests have been introduced into the field of equipment fault diagnosis. However, the feature engineering of these methods relies on manual design and is difficult to model the coupling relationships between multiple subsystems such as mechanical, hydraulic, and electrical components of boarding bridges.

[0005] In recent years, deep learning has demonstrated powerful feature learning capabilities in fault diagnosis. However, pure data-driven black-box models have significant limitations in the specific application of boarding bridges: First, the model may output results that violate physical laws, leading to low reliability of the diagnostic results; Second, boarding bridge fault samples are scarce, and pure data-driven models are prone to overfitting under small sample conditions, resulting in poor generalization ability; Third, the model's decision-making process is not transparent, making it difficult for maintenance personnel to understand its diagnostic basis, thus making it difficult to trust and adopt.

[0006] Therefore, it is of great significance to provide a method, system, device and storage medium for boarding bridge fault diagnosis based on physical constraint neural networks. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to provide a boarding bridge fault diagnosis method, system, device and storage medium based on physical constraint neural network, which aims to solve the problems of low accuracy, poor real-time performance and weak interpretability of existing boarding bridge fault diagnosis methods, and realize intelligent and accurate fault diagnosis of boarding bridge equipment.

[0008] Technical solution: The present invention provides a boarding bridge fault diagnosis method based on a physical constraint neural network, comprising the following steps:

[0009] (1) Collect multi-source sensor data of boarding bridge in real time and preprocess it to obtain preprocessed sensor data. The multi-source sensor data includes mechanical motion data, hydraulic system data, electrical system data and environmental monitoring data.

[0010] (2) Based on the basic physical laws and the range of physical quantities that can be measured by the multi-source sensor data, determine the physical constraint equations, which include kinematic constraint equations, dynamic constraint equations and energy conservation constraint equations;

[0011] (3) Construct a physical constraint neural network model, which includes a data feature extraction network and a physical constraint fusion layer; the data feature extraction network adopts a parallel combination structure of convolutional neural network and long short-term memory network to extract features from preprocessed sensor time series data; the physical constraint fusion layer fuses the extracted data features with the physical constraint equation to construct a joint loss function; the joint loss function includes a data fitting loss term, a physical constraint loss term and a regularization term;

[0012] (4) Collect historical operation data of the boarding bridge and input it into the physical constraint neural network model for training;

[0013] (5) Input the preprocessed sensor data collected in real time into the trained physical constraint neural network model and output the boarding bridge fault diagnosis results, including fault type, fault degree and fault location;

[0014] (6) Generate fault alarms and maintenance suggestions based on the diagnostic results, and display the health status of the equipment through a visual interface.

[0015] Furthermore, the mechanical motion data in step (1) includes the extension and retraction displacement, rotation angle, lifting height, movement speed and acceleration of the boarding bridge; the hydraulic system data includes hydraulic oil pressure, flow rate, temperature and oil level; the electrical system data includes motor current, voltage, power and temperature; and the environmental monitoring data includes ambient temperature, humidity and wind speed.

[0016] Furthermore, the structure of the physical constraint neural network specifically includes:

[0017] Input layer: Used to receive time-series data from sensors and output the data to convolutional neural networks and long short-term memory networks;

[0018] Convolutional Neural Network Layer: Processes input data in parallel with Long Short-Term Memory Network Layer to extract local spatial features of sensor time-series data. It employs multiple one-dimensional convolutional layers, with batch normalization and max pooling layers added after each convolutional layer. It outputs the spatial feature vector of the convolutional neural network branch through global average pooling.

[0019] Long Short-Term Memory (LSTM) network layer: Processes input data in parallel with the convolutional neural network layer to extract long-term dependency features of sensor time-series data. It adopts a bidirectional LSM network and outputs the feature vectors of the LSM network branches through temporal pooling.

[0020] Attention mechanism layer: Receives spatial features extracted by the convolutional neural network layer and temporal features from the long short-term memory network layer, and performs weighted fusion through adaptive attention weights;

[0021] Fully connected layer: Consists of multiple fully connected neurons, used to map the fused features to a high-dimensional feature space and perform nonlinear transformations.

[0022] Output layer: Used to output the probability distribution of fault types.

[0023] Furthermore, the physical constraint fusion layer fuses the extracted data features with the physical constraint equations, including:

[0024] A regression branch is derived from the fully connected layer in the data feature extraction network to predict the value of physical quantities, including but not limited to displacement, velocity, acceleration, and force. The output dimension is the quantity of the corresponding physical quantity.

[0025] Based on the physical constraint equations, calculate the residuals of the physical constraints and convert them into loss values.

[0026] Furthermore, the joint loss function is expressed as:

[0027]

[0028] in, For data fitting loss, For physical constraint loss, For regularization terms, , , These are weighting coefficients used to balance the importance of different loss terms.

[0029] Furthermore, step (4), the process of training the physical constraint neural network model, includes:

[0030] Historical operational data is divided into training set, validation set and test set. Historical operational data includes normal operation data and failure data.

[0031] The neural network model is initialized using Xavier initialization, orthogonal initialization, and He initialization.

[0032] The Adam optimizer is used for parameter optimization, and a learning rate decay strategy is employed.

[0033] During training, an early stopping mechanism is set to prevent overfitting, and the model parameters with the minimum loss on the validation set are saved as the final model.

[0034] Furthermore, the diagnostic results output in step (5) include:

[0035] Fault types include, but are not limited to, hydraulic system leakage faults, motor overload faults, mechanical jamming faults, sensor faults, and control system faults;

[0036] The severity of the fault is categorized into three levels: minor, moderate, and severe.

[0037] Location of the fault: pinpoint the subsystem and specific component where the fault occurred.

[0038] The present invention discloses a boarding bridge fault diagnosis system based on a physical constraint neural network, comprising:

[0039] Data acquisition module: Real-time acquisition of multi-source sensor data from the boarding bridge, and preprocessing to obtain preprocessed sensor data. The multi-source sensor data includes mechanical motion data, hydraulic system data, electrical system data, and environmental monitoring data.

[0040] Constraint equation determination module: Based on fundamental physical laws and the range of physical quantities that can be measured by the multi-source sensor data, physical constraint equations are determined, including kinematic constraint equations, dynamic constraint equations and energy conservation constraint equations;

[0041] Model building module: Constructs a physical constraint neural network model, which includes a data feature extraction network and a physical constraint fusion layer. The data feature extraction network adopts a parallel combination structure of convolutional neural network and long short-term memory network to extract features from preprocessed sensor time-series data. The physical constraint fusion layer fuses the extracted data features with the physical constraint equations to construct a joint loss function. The joint loss function includes a data fitting loss term, a physical constraint loss term, and a regularization term.

[0042] Model training module: Collects historical operational data of the boarding bridge, inputs it into the physical constraint neural network model for training, and optimizes the network parameters through the backpropagation algorithm to minimize the joint loss function;

[0043] Diagnostic result output module: Input the pre-processed sensor data collected in real time into the trained physical constraint neural network model, and output the boarding bridge fault diagnosis results, including fault type, fault degree and fault location;

[0044] Visualization module: Generates fault alarms and maintenance suggestions based on diagnostic results, and displays the health status of the equipment through a visual interface.

[0045] The present invention provides a computer device comprising one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of the boarding bridge fault diagnosis method based on physical constraint neural networks.

[0046] The present invention discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the boarding bridge fault diagnosis method based on a physical constraint neural network.

[0047] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0048] (1) The present invention embeds the physical equations of the mechanical motion data, hydraulic system data and electrical system data of the boarding bridge as constraints into the neural network training process, so that the diagnostic results conform to physical laws while being data-driven, and can still maintain high performance diagnosis under small sample conditions.

[0049] (2) Based on the multi-dimensional physical constraints extracted by the general basic physical laws, the present invention directly relates to the operating parameters of mechanical motion, hydraulic system and electrical subsystem, and can capture multi-system coupled faults that are difficult to detect by traditional single-parameter monitoring methods, and realize early fault warning.

[0050] (3) The present invention constructs a joint loss function to balance data fitting and physical constraints through weight coefficients, ensuring that the output of the neural network does not violate physical laws and enhancing the generalization ability of the model under different working conditions.

[0051] (4) The present invention uses a parallel structure of convolutional neural network and long short-term memory network to automatically extract spatial and temporal features, avoiding the manual feature engineering of traditional methods and reducing the dependence on expert experience.

[0052] (5) The diagnostic results output by the method of the present invention not only include the fault type, but also assess the fault degree, locate the fault location, and automatically generate maintenance suggestions, thereby improving the pertinence and efficiency of maintenance work. Attached Figure Description

[0053] Figure 1 This is a flowchart of the boarding bridge fault diagnosis method of the present invention.

[0054] Figure 2 This is a schematic diagram of the physical constraint neural network model of the present invention. Detailed Implementation

[0055] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0056] Example 1

[0057] like Figure 1 As shown, this embodiment provides a boarding bridge fault diagnosis method based on a physical constraint neural network, including the following steps:

[0058] Step S1: Collect multi-source sensor data from the boarding bridge in real time.

[0059] The boarding bridge mainly consists of a fixed end, a telescopic section, a rotating section, and a head platform. Multiple sensors are installed at key locations on the boarding bridge to achieve comprehensive monitoring of its operational status. Specifically, these include:

[0060] (1) Mechanical motion data acquisition: Displacement sensors are installed on the telescopic section to measure the telescopic displacement of the boarding bridge, with a measurement range of 0-30 meters and an accuracy of ±10 mm; angle sensors are installed on the rotary joint to measure the rotation angle of the boarding bridge, with a measurement range of 0-360 degrees and an accuracy of ±0.5 degrees; displacement sensors are installed on the lifting mechanism to measure the lifting height of the boarding bridge, with a measurement range of 0-8 meters and an accuracy of ±5 mm. The motion velocity and acceleration are calculated by differentiating the displacement and angle signals.

[0061] (2) Hydraulic system data acquisition: Pressure sensors are installed at the outlet of the hydraulic pump and the inlet of each hydraulic cylinder to measure the hydraulic oil pressure. The measurement range is 0-25MPa and the accuracy is ±0.1MPa. Flow sensors are installed in the hydraulic pipeline to measure the hydraulic oil flow rate. The measurement range is 0-100L / min and the accuracy is ±1L / min. Temperature sensors are installed in the hydraulic oil tank to measure the hydraulic oil temperature. The measurement range is -20~100℃ and the accuracy is ±0.5℃. Level sensors are installed in the oil tank to measure the hydraulic oil level.

[0062] (3) Environmental monitoring data collection: Install an ambient temperature sensor at the fixed end of the boarding bridge to measure the ambient temperature, with a measurement range of -40~60℃; install a humidity sensor to measure the ambient humidity; install a wind speed sensor to measure the ambient wind speed.

[0063] All sensor data is acquired via a data acquisition card at a sampling frequency of 100Hz and transmitted to a data processing server via industrial Ethernet. The data acquisition system employs a redundant design, with backups configured for critical sensors, ensuring the reliability of data acquisition.

[0064] Step S2: Preprocess the multi-source sensor data.

[0065] Includes the following steps:

[0066] (1) Data cleaning: including the following steps:

[0067] (a) Outlier detection: Using the 3σ principle, outlier data points exceeding the mean ± 3 times the standard deviation are deleted;

[0068] (b) Missing value handling: Missing data is filled using forward imputation or linear interpolation methods;

[0069] (c) Noise filtering: A median filter is used to remove high-frequency noise.

[0070] (2) Data normalization: The min-max normalization method is used to scale the data to the [0,1] interval.

[0071]

[0072] in This is the original data. and These are the minimum and maximum values ​​of the data from this type of sensor, respectively.

[0073] (3) Timing alignment: Since there may be slight time deviations in the sampling of different sensors, timing alignment is required. Based on the timestamp of the data acquisition system, a linear interpolation method is used to align all sensor data to a unified time grid with a time step of 10ms.

[0074] (4) Sliding window processing: In order to construct the input samples of the neural network, the time series data is segmented using a sliding time window method. The time window length is set to 10 seconds (1000 sampling points), and the sliding step size is set to 1 second (100 sampling points). The data in each time window constitutes an input sample, and the dimension of the sample is (number of sensors × time window length).

[0075] After preprocessing, a standardized input dataset is obtained for subsequent model training and diagnosis.

[0076] Step S3: Based on the fundamental physical laws and the range of physical quantities that can be measured by the multi-source sensor data, determine the physical constraint equations.

[0077] Based on fundamental physical laws (such as Newtonian mechanics, kinematics, and the law of conservation of energy) and the range of physical quantities measurable by the multi-source sensor data in step S1, physical constraint equations suitable for embedding into neural networks are directly determined and formalized into computable mathematical expressions. These constraints include:

[0078] (1) Kinematic constraint equations: describe the differential relationship between displacement, velocity and acceleration.

[0079] Speed ​​constraints:

[0080]

[0081] in, This represents the predicted velocity at time t. This represents the predicted displacement at time t, where t represents time.

[0082] The derivative of the displacement predicted by the neural network with respect to time is the predicted velocity.

[0083] Acceleration constraints:

[0084]

[0085] in, This represents the predicted acceleration at time t. This represents the predicted velocity at time t.

[0086] The derivative of the velocity predicted by the neural network with respect to time is the predicted acceleration.

[0087] (2) Dynamic constraint equations: describe the relationship between force, mass and acceleration, and follow Newton's second law.

[0088] Force balance constraints:

[0089]

[0090] in, As the driving force, For frictional resistance, For system quality, It is acceleration.

[0091] The forces, friction, mass, and acceleration predicted by the neural network satisfy the force balance relationship.

[0092] Torque balance constraint:

[0093]

[0094] in, For electromagnetic torque, For load torque, For rotational inertia, Angular velocity, ω is angular acceleration.

[0095] The electromagnetic torque, load torque, moment of inertia, and angular acceleration predicted by the neural network satisfy the torque balance relationship.

[0096] (3) Energy conservation constraint equation: describes the energy conversion and conservation relationship of the system.

[0097] Energy conservation in motors:

[0098]

[0099] in, For input electrical power, For mechanical output power, This is for power loss.

[0100] Input electrical power equals the sum of mechanical output power and power loss.

[0101] Energy conservation in hydraulic systems:

[0102]

[0103] in, For the output power of the hydraulic pump, Power consumed by the hydraulic cylinder This refers to the power loss in the hydraulic system.

[0104] The output power of a hydraulic pump is equal to the sum of the power consumed by the hydraulic cylinder and the power lost due to hydraulic pressure.

[0105] Total energy of the system is conserved:

[0106]

[0107] in, The change in the system's kinetic energy. The change in the system's potential energy. For input power, This is for energy loss.

[0108] The changes in the system's kinetic and potential energy are equal to the difference between the input work and the loss work.

[0109] The aforementioned physical constraint equations will be used as soft constraints during the training of the neural network and optimized through the loss function.

[0110] Step S4: Construct a physically constrained neural network model

[0111] like Figure 2 As shown, the physical constraint neural network model includes a data feature extraction network and a physical constraint fusion layer. The data feature extraction network adopts a parallel combination structure of convolutional neural network (CNN) and long short-term memory network (LSTM).

[0112] The structure of a physically constrained neural network includes:

[0113] Input layer 201: Receives preprocessed multi-source sensor time-series data, including mechanical motion data, hydraulic system data, and electrical system data. The input data dimension is represented as (N, T, M). Here, N represents the batch size, set to 32; T represents the time step, set to 1000 (corresponding to a 10-second time window); and M represents the number of sensors, set to 20. The input layer simultaneously outputs the data to a Convolutional Neural Network (CNN) layer and a Long Short-Term Memory (LSTM) layer to achieve parallel feature extraction.

[0114] CNN layer 202: Processes input data in parallel with the LSTM layer to extract spatial domain features from sensor data. It employs three one-dimensional convolutional layers: the first layer has a kernel size of 7, 64 kernels, a stride of 1, and uses ReLU activation; the second layer has a kernel size of 5, 128 kernels, a stride of 1, and uses ReLU activation; the third layer has a kernel size of 3, 256 kernels, a stride of 1, and uses ReLU activation. Batch normalization is added after each convolutional layer to accelerate training, and max pooling layers (pooling window size of 2, stride of 2) are added to reduce feature dimensionality and prevent overfitting. Finally, global average pooling is used to output the spatial feature vector of the CNN branch.

[0115] LSTM layer 203: Processes input data in parallel with CNN layers to extract long-term dependency features from temporal data. It employs a Bidirectional Long Short-Term Memory (Bi-LSTM) network, a two-layer structure with 256 hidden neurons and an output dimension of (Batch_size, Time_steps, 512). The feature vectors of the LSTM branches are output through temporal pooling.

[0116] Attention mechanism layer 204: Receives spatial features extracted by the CNN layer and temporal features extracted by the LSTM layer, and performs weighted fusion using adaptive attention weights, automatically focusing on the features most important for fault diagnosis. After concatenating the feature vectors from the two branches, a self-attention mechanism is used to calculate the attention weights. The formula for calculating the attention weights is:

[0117]

[0118] in, This is the attention weight matrix. For querying the matrix, The key matrix, This is the transpose of the key matrix. The dimension (scaling factor) of the key vector.

[0119] The output is:

[0120]

[0121] in, The output features of the attention mechanism layer For attention weights, It is a value matrix.

[0122] in , , These are the query, key, and value matrices, respectively, which are obtained from the LSTM output through linear transformation.

[0123] Fully connected layer 205: This layer contains multiple fully connected neurons used to map the fused features to a high-dimensional feature space for nonlinear transformation. A three-layer fully connected network is used. The first layer has 512 neurons with ReLU activation and a Dropout layer (0.5 dropout rate) to prevent overfitting. The second layer has 256 neurons with ReLU activation and a Dropout layer. A physical quantity regression branch is derived from the second layer to predict the numerical values ​​of physical quantities such as displacement, velocity, acceleration, force, torque, and power. This branch consists of two fully connected layers, with the output dimension being the number of physical quantities.

[0124] Output layer 206: The third layer is the output layer, with the number of neurons equal to the number of fault types (6 in this embodiment, including 5 types of faults and normal states). The activation function is Softmax, and the output layer is the probability distribution of each fault type.

[0125] The physical constraint fusion layer fuses the output of the data feature extraction network with the physical constraint equations.

[0126] Feature Regression Branch: A regression branch is derived from the penultimate layer of the fully connected layers to predict the numerical values ​​of physical quantities, such as displacement, velocity, acceleration, force, torque, and power. This branch consists of two fully connected layers, and the output dimension is the number of physical quantities.

[0127] Physical constraint calculation module: Calculates the residuals of the physical constraints based on the physical constraint equations determined in step S3. For example, for velocity constraints...

[0128]

[0129] The residuals are calculated as follows:

[0130]

[0131] in and These are the velocity and displacement predicted by the neural network, respectively. Calculated using numerical differentiation.

[0132] Constraint Loss Calculation Module: Converts physical constraint residuals into loss values ​​for subsequent loss function optimization.

[0133] Through the physical constraint fusion layer, the neural network not only learns data-driven feature representations during training, but is also constrained by physical laws, ensuring that the prediction results conform to the physical operation of the boarding bridge, thereby improving the accuracy and reliability of the diagnosis.

[0134] Step S5: Construct the joint loss function

[0135] The joint loss function includes a data fitting loss term, a physical constraint loss term, and a regularization term, and its expression is as follows:

[0136]

[0137] (1) Data fitting loss

[0138] The difference between the predicted fault type and the true label is calculated using the cross-entropy loss function:

[0139]

[0140] in One-hot encoding of the real label. The probability of the fault type output by the neural network. This represents the number of samples.

[0141] For the physical quantities predicted by the regression branch, the mean squared error (MSE) loss is used:

[0142]

[0143] in The true value of the physical quantity measured by the sensor. For predicted values, This represents the number of samples.

[0144] The total data fitting loss is:

[0145]

[0146] in The balance coefficient is set to 0.5.

[0147] (2) Physical constraint loss

[0148] Physical constraint losses include kinematic constraint losses, dynamic constraint losses, and energy conservation constraint losses:

[0149] Kinematic constraint loss:

[0150]

[0151] in, This represents the kinematic constraint loss value. The number of training samples; The speed value predicted by the neural network; The displacement value predicted by the neural network; To predict the derivative of displacement with respect to time; The acceleration value predicted by the neural network; To predict the derivative of velocity with respect to time.

[0152] Dynamic constraint loss:

[0153]

[0154] in, This represents the loss value due to dynamic constraints. The number of samples; The driving force behind neural network prediction; Frictional resistance; For equipment quality; The acceleration predicted by the neural network; The electromagnetic torque predicted by the neural network; This is the load torque; It is the moment of inertia; To predict the derivative of angular velocity with respect to time.

[0155] Energy conservation constraint loss:

[0156]

[0157] in, Energy conservation constraint loss value; The number of samples; Input power; This refers to the output power. For power loss; This represents the change in the system's kinetic energy. This represents the change in the system's potential energy. For input power; This is for energy loss.

[0158] The total physical constraint loss is:

[0159]

[0160] in, This represents the total physical constraint loss. , and These are the calculated kinematic, dynamic, and energy conservation constraint loss terms, respectively.

[0161] (3) Regularization term

[0162] L2 regularization is used to prevent overfitting:

[0163]

[0164] in These are the weight parameters of the neural network.

[0165] (4) Weighting coefficient settings

[0166] Weighting coefficient , , This is used to balance the importance of different loss terms. Based on experience and experimental tuning, the following settings are configured: , , . The smaller value is because the physical constraint loss plays an auxiliary role, while the main purpose is to guide the neural network to learn feature representations that conform to physical laws.

[0167] Step S6: Collect historical operational data of the boarding bridge and input it into the physical constraint neural network model for training.

[0168] Specifically, it includes:

[0169] (1) Dataset preparation

[0170] Two years of operational data, totaling approximately 5000 hours of runtime, was collected from two boarding bridges at a major international airport. Data was acquired at a frequency of 100Hz and included both normal operation and fault data. Fault samples were manually labeled and categorized into five types: hydraulic system leakage (300 samples), motor overload (250 samples), mechanical jamming (200 samples), sensor faults (150 samples), and control system faults (100 samples). 5000 normal operation samples were also collected.

[0171] The dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for model parameter optimization, the validation set for hyperparameter tuning and early stopping detection, and the test set for final performance evaluation.

[0172] To address the imbalance of fault samples, data augmentation methods are employed for a minority of fault samples, including time stretching, adding Gaussian noise, and time shifting, to achieve a relative balance in the number of fault samples across different categories.

[0173] (2) Model initialization

[0174] The weights of fully connected layers are initialized using the Xavier initialization method; for LSTM layers, the weight matrix is ​​initialized using the orthogonal initialization method to ensure that the gradients do not vanish or explode during backpropagation; the convolutional layers are initialized using the He initialization method, which is adapted to the ReLU activation function.

[0175] (3) Optimizer settings

[0176] The Adam optimizer was used for parameter optimization, with the initial learning rate set to 0.001 and the batch size set to 32.

[0177] A learning rate decay strategy is adopted. When the validation set loss does not decrease for 5 consecutive epochs, the learning rate is multiplied by a decay factor of 0.5, and the minimum learning rate is set to 1e-6.

[0178] (4) Training process

[0179] Set the maximum number of training epochs to 200. Implement an early stopping mechanism: stop training when the validation set loss does not decrease for 10 consecutive epochs to prevent overfitting. Save the model parameters with the minimum validation set loss as the final model.

[0180] The training process is accelerated using a GPU, specifically an NVIDIA Tesla V100 GPU. The total training time is approximately 8 hours, and an early stop mechanism is triggered after 120 epochs.

[0181] (5) Model performance evaluation

[0182] The trained model is evaluated on the test set, and the performance metrics include: accuracy, precision, recall, F1 score, and confusion matrix.

[0183] Test results show that the overall accuracy of the model reaches 95.8%, and the F1 score of various faults is above 90%. Compared with the traditional threshold alarm method, the accuracy is improved by 23.5%, and compared with the pure data-driven deep learning method, the accuracy is improved by 4.5%, which verifies the effectiveness of incorporating physical constraints.

[0184] Step S7: Input the preprocessed sensor data collected in real time into the trained physical constraint neural network model and output the boarding bridge fault diagnosis results.

[0185] Specifically, it includes:

[0186] (1) Model Deployment

[0187] The trained model was deployed to the local monitoring workstation on the boarding bridge, configured with an Intel Xeon processor, 16GB of RAM, and an NVIDIA GTX 1080 Ti GPU. The model was converted to ONNX format and loaded, then remained resident in memory.

[0188] The time for a single diagnostic inference is approximately 50ms, which meets the real-time requirements.

[0189] (2) Real-time data processing

[0190] The real-time acquired sensor data undergoes a preprocessing step S2, including data cleaning, normalization, and sliding window processing. The real-time data stream generates input samples at 1-second intervals (100 sampling point steps), with each sample containing historical data for a 10-second time window (1000 sampling points). These samples are continuously input into the model for diagnosis. The data preprocessing delay is less than 10ms, which does not affect real-time performance.

[0191] (3) Fault diagnosis output

[0192] The model output consists of three parts:

[0193] Fault Type: Outputs the probability distribution of each fault type, selecting the category with the highest probability as the diagnostic result. When the probability of all fault types is below the threshold of 0.6, it is judged as a normal state or an unknown fault, triggering manual review.

[0194] Fault severity: The severity of the fault is assessed based on the degree to which the physical quantity deviates from the normal value, and is divided into three levels: minor, moderate, and severe. Minor fault: Physical quantity deviates from the normal range by 10%-30%; Moderate fault: Deviates by 30%-60%; Severe fault: Deviates by more than 60%.

[0195] Fault Location: By analyzing the degree of anomaly in the data from each sensor and the residuals of physical constraints, the subsystem and component where the fault occurred can be located. For example, when the hydraulic cylinder pressure sensor data is abnormal but the flow sensor is normal, the fault is determined to be in the hydraulic cylinder itself; when both are abnormal, the fault is determined to be in the hydraulic pump or pipeline.

[0196] (4) Diagnostic confidence assessment

[0197] To assess the reliability of the diagnostic results, a diagnostic confidence index is calculated:

[0198] Classification confidence score: This is the maximum probability of the fault type, reflecting the certainty of the classification. A confidence score above 0.9 is considered high confidence, 0.6-0.9 is considered medium confidence, and below 0.6 is considered low confidence.

[0199] Physical consistency: This involves calculating the magnitude of the physical constraint residuals. The smaller the physical constraint residuals, the more the diagnostic results conform to physical laws, and the higher the confidence level. The physical consistency index is defined as follows:

[0200]

[0201] The closer the value is to 1, the better the physical consistency.

[0202] Overall confidence level: Combining classification confidence level and physical consistency, the overall confidence level is defined as:

[0203]

[0204] in This represents the maximum probability of the fault type.

[0205] The diagnostic result will only be automatically output when the overall confidence level is higher than 0.75; otherwise, a manual review process will be triggered.

[0206] Step S8: Generate fault alarms and maintenance suggestions based on the diagnostic results, and display the equipment health status through a visual interface.

[0207] Specifically, it includes:

[0208] (1) Fault alarm

[0209] When a fault is diagnosed, the system immediately generates an alarm message, including: alarm time, fault type, fault severity, fault location, and diagnostic confidence level. The alarm message is sent through multiple methods: monitoring interface pop-ups, SMS notifications to maintenance personnel, email notifications to management personnel, and integration into the airport operations management system.

[0210] Alarm levels are set according to the severity of the fault: minor faults are indicated by a blue warning, and it is recommended to check during the next scheduled maintenance; moderate faults are indicated by a yellow warning, and it is recommended to check within 24 hours; severe faults are indicated by a red alarm, and it is recommended to stop the machine immediately for repair.

[0211] (2) Maintenance suggestion generation

[0212] Based on expert knowledge base and diagnostic results, maintenance recommendations are automatically generated. These recommendations include: fault cause analysis, suggested inspection items, components that may need replacement, estimated maintenance duration, and maintenance priority.

[0213] The specific maintenance recommendations are as follows:

[0214] Hydraulic system leakage fault:

[0215] For minor faults, it is recommended to check the hydraulic line seals, hydraulic cylinder seals, and pipe joints for looseness within 48 hours, locate possible leaks, and replace aged seals if necessary. The estimated maintenance time is 1-2 hours. For moderate or severe faults, it is recommended to immediately stop the machine for repair, replace leaking parts, replenish hydraulic oil, and check for hydraulic oil contamination. The estimated maintenance time is 3-5 hours.

[0216] Motor overload fault:

[0217] It is recommended to check for abnormal load, any jamming in the mechanical transmission parts, wear on the motor bearings, and the lubrication of the reducer. Measure the motor's insulation resistance. If the load is normal but the motor is still overloaded, the motor may need to be replaced. The estimated maintenance time is 3-6 hours.

[0218] Mechanical jamming fault:

[0219] It is recommended to check the lubrication of the mechanical transmission parts, check for any foreign objects stuck in them, check for deformation of the guide rails, and check for loose connections. Clean and lubricate the transmission parts, and replace worn parts if necessary. The estimated maintenance time is 4 hours.

[0220] Sensor malfunction:

[0221] We recommend checking if the sensor connection lines are loose or damaged, checking if the sensor power supply is normal, and measuring if the sensor output signal is within the normal range. If the wiring is normal but the signal is abnormal, the sensor needs to be replaced. The estimated maintenance time is 1-2 hours.

[0222] Control system malfunction:

[0223] It is recommended to check whether the controller hardware is functioning properly, whether the control program is malfunctioning, and whether the control signal transmission is normal. If necessary, restart the controller or update the control program. If there is a hardware failure, the control module needs to be replaced. The estimated maintenance time is 2-4 hours.

[0224] (3) Visualization

[0225] The system provides a visual interface to display the health status of the boarding bridge in real time:

[0226] Equipment Overview Interface: Displays the real-time status of all boarding bridges in the airport, using colors to indicate their health status: green for normal, yellow for warning, orange for alert, and red for alarm. Clicking on an individual boarding bridge displays detailed information.

[0227] Real-time monitoring interface: Displays real-time curves of key parameters for the selected boarding bridge, including displacement, speed, pressure, and current. Abnormal parameters are highlighted in red. The right side displays the current diagnostic results and fault probability distribution.

[0228] History Records: Displays the boarding bridge's historical fault records, including fault occurrence time, fault type, fault handling status, and maintenance records. Users can filter and search by time range and fault type.

[0229] Health rating interface: Based on historical operating data and failure frequency, calculates the boarding bridge's health score (0-100 points) to assess the overall health of the equipment. Equipment with a health score below 60 points is recommended for a comprehensive overhaul.

[0230] Statistical analysis interface: Displays fault statistics and analysis, including fault type distribution, fault frequency trend, fault downtime statistics, maintenance cost statistics, etc., providing data support for equipment management decisions.

[0231] Example 2

[0232] This embodiment provides a boarding bridge fault diagnosis system based on a physical constraint neural network, including the following modules:

[0233] Data acquisition module: Real-time acquisition of multi-source sensor data from the boarding bridge, and preprocessing to obtain preprocessed sensor data. The multi-source sensor data includes mechanical motion data, hydraulic system data, electrical system data, and environmental monitoring data.

[0234] Constraint equation determination module: Based on fundamental physical laws and the range of physical quantities that can be measured by the multi-source sensor data, physical constraint equations are determined, including kinematic constraint equations, dynamic constraint equations and energy conservation constraint equations;

[0235] Model building module: Constructs a physical constraint neural network model, which includes a data feature extraction network and a physical constraint fusion layer. The data feature extraction network adopts a parallel combination structure of convolutional neural network and long short-term memory network to extract features from preprocessed sensor time-series data. The physical constraint fusion layer fuses the extracted data features with the physical constraint equations to construct a joint loss function. The joint loss function includes a data fitting loss term, a physical constraint loss term, and a regularization term.

[0236] Model training module: Collects historical operational data of the boarding bridge, inputs it into the physical constraint neural network model for training, and optimizes the network parameters through the backpropagation algorithm to minimize the joint loss function;

[0237] Diagnostic result output module: Input the pre-processed sensor data collected in real time into the trained physical constraint neural network model, and output the boarding bridge fault diagnosis results, including fault type, fault degree and fault location;

[0238] Visualization module: Generates fault alarms and maintenance suggestions based on diagnostic results, and displays the health status of the equipment through a visual interface.

[0239] This embodiment is based on the same inventive concept as Embodiment 1, and will not be described again here.

[0240] Example 3

[0241] This embodiment provides a computer device, including one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of a boarding bridge fault diagnosis method based on a physical constraint neural network.

[0242] Example 4

[0243] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a boarding bridge fault diagnosis method based on a physically constrained neural network.

Claims

1. A method for diagnosing boarding bridge faults based on a physical constraint neural network, characterized in that, Includes the following steps: (1) Collect multi-source sensor data of boarding bridge in real time and preprocess it to obtain preprocessed sensor data. The multi-source sensor data includes mechanical motion data, hydraulic system data, electrical system data and environmental monitoring data. (2) Based on the basic physical laws and the range of physical quantities that can be measured by the multi-source sensor data, determine the physical constraint equations, which include kinematic constraint equations, dynamic constraint equations and energy conservation constraint equations; (3) Construct a physical constraint neural network model, which includes a data feature extraction network and a physical constraint fusion layer; the data feature extraction network adopts a parallel combination structure of convolutional neural network and long short-term memory network to extract features from preprocessed sensor time series data; the physical constraint fusion layer fuses the extracted data features with the physical constraint equation to construct a joint loss function; the joint loss function includes a data fitting loss term, a physical constraint loss term and a regularization term; (4) Collect historical operation data of the boarding bridge and input it into the physical constraint neural network model for training; (5) Input the preprocessed sensor data collected in real time into the trained physical constraint neural network model and output the boarding bridge fault diagnosis results, including fault type, fault degree and fault location; (6) Generate fault alarms and maintenance suggestions based on the diagnostic results, and display the health status of the equipment through a visual interface.

2. The boarding bridge fault diagnosis method based on physical constraint neural network according to claim 1, characterized in that, The mechanical motion data in step (1) includes the extension and retraction displacement, rotation angle, lifting height, movement speed and acceleration of the boarding bridge; the hydraulic system data includes hydraulic oil pressure, flow rate, temperature and oil level; the electrical system data includes motor current, voltage, power and temperature; and the environmental monitoring data includes ambient temperature, humidity and wind speed.

3. The boarding bridge fault diagnosis method based on physical constraint neural network according to claim 1, characterized in that, The structure of the physical constraint neural network specifically includes: Input layer: Used to receive time-series data from sensors and output the data to convolutional neural networks and long short-term memory networks; Convolutional Neural Network Layer: Processes input data in parallel with Long Short-Term Memory Network Layer to extract local spatial features from sensor time-series data. It employs multiple one-dimensional convolutional layers, with batch normalization and max pooling layers added after each convolutional layer. It outputs the spatial feature vector of the convolutional neural network branch through global average pooling. Long Short-Term Memory (LSTM) network layer: Processes input data in parallel with the convolutional neural network layer to extract long-term dependency features of sensor time-series data. It adopts a bidirectional LSM network and outputs the feature vectors of the LSM network branches through temporal pooling. Attention mechanism layer: Receives spatial features extracted by convolutional neural network layers and temporal features from long short-term memory network layers, and performs weighted fusion through adaptive attention weights; Fully connected layer: Consists of multiple fully connected neurons, used to map the fused features to a high-dimensional feature space and perform nonlinear transformations. Output layer: Used to output the probability distribution of fault types.

4. The boarding bridge fault diagnosis method based on physical constraint neural network according to claim 1, characterized in that, The physical constraint fusion layer fuses the extracted data features with the physical constraint equations, including: A regression branch is derived from the fully connected layer in the data feature extraction network to predict the value of physical quantities, including but not limited to displacement, velocity, acceleration, and force. The output dimension is the quantity of the corresponding physical quantity. Based on the physical constraint equations, calculate the residuals of the physical constraints and convert them into loss values.

5. The boarding bridge fault diagnosis method based on physical constraint neural network according to claim 1, characterized in that, The joint loss function is expressed as: in, For data fitting loss, For physical constraint loss, For regularization terms, , , These are weighting coefficients used to balance the importance of different loss terms.

6. The boarding bridge fault diagnosis method based on physical constraint neural network according to claim 1, characterized in that, The process of training the physical constraint neural network model in step (4) includes: Historical operational data is divided into training set, validation set and test set. Historical operational data includes normal operation data and failure data. The neural network model is initialized using Xavier initialization, orthogonal initialization, and He initialization. The Adam optimizer is used for parameter optimization, and a learning rate decay strategy is employed. During training, an early stopping mechanism is set to prevent overfitting, and the model parameters with the minimum loss on the validation set are saved as the final model.

7. The boarding bridge fault diagnosis method based on physical constraint neural network according to claim 1, characterized in that, The diagnostic results output in step (5) include: Fault types include, but are not limited to, hydraulic system leakage faults, motor overload faults, mechanical jamming faults, sensor faults, and control system faults; The severity of the fault is categorized into three levels: minor, moderate, and severe. Location of the fault: pinpoint the subsystem and specific component where the fault occurred.

8. A boarding bridge fault diagnosis system based on a physical constraint neural network, characterized in that, include: Data acquisition module: Real-time acquisition of multi-source sensor data from the boarding bridge, and preprocessing to obtain preprocessed sensor data. The multi-source sensor data includes mechanical motion data, hydraulic system data, electrical system data, and environmental monitoring data. Constraint equation determination module: Based on fundamental physical laws and the range of physical quantities that can be measured by the multi-source sensor data, physical constraint equations are determined, including kinematic constraint equations, dynamic constraint equations and energy conservation constraint equations; Model building module: Constructs a physical constraint neural network model, which includes a data feature extraction network and a physical constraint fusion layer. The data feature extraction network adopts a parallel combination structure of convolutional neural network and long short-term memory network to extract features from preprocessed sensor time-series data. The physical constraint fusion layer fuses the extracted data features with the physical constraint equations to construct a joint loss function. The joint loss function includes a data fitting loss term, a physical constraint loss term, and a regularization term. Model training module: Collects historical operational data of the boarding bridge, inputs it into the physical constraint neural network model for training, and optimizes the network parameters through the backpropagation algorithm to minimize the joint loss function; Diagnostic result output module: Input the pre-processed sensor data collected in real time into the trained physical constraint neural network model, and output the boarding bridge fault diagnosis results, including fault type, fault degree and fault location; Visualization module: Generates fault alarms and maintenance suggestions based on diagnostic results, and displays the health status of the equipment through a visual interface.

9. A computer device, characterized in that, The method includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of the boarding bridge fault diagnosis method based on physical constraint neural networks as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the boarding bridge fault diagnosis method based on physical constraint neural networks as described in any one of claims 1-7.

Images