Railway brake valve fault prediction model construction method and railway brake valve fault prediction method

Abstract

The invention discloses a railway brake valve fault prediction model construction method and a railway brake valve fault prediction method, and relates to the field of railway safety, the railway brake valve fault prediction model construction method comprises the steps that a first neural network model is constructed, and the first neural network model comprises an input layer, a feature extraction layer and a first output layer; training the first neural network model through fault data of the 120-type truck air control valve, freezing network parameters of an input layer and a feature extraction layer after training, and replacing a first output layer with a second output layer to obtain a second neural network model; training the second neural network model through the fault data of the 104 type passenger car air distribution valve to obtain a railway brake valve fault prediction model; when the fault data of the 104 type passenger car air distribution valve is less, the railway brake valve fault prediction model which can accurately predict the fault risk of the 104 type passenger car air distribution valve can be obtained through training.

Description

Technical Field

[0001] This application relates to the field of railway safety, and in particular to a method for constructing a railway brake valve failure prediction model and a method for predicting railway brake valve failures. Background Technology

[0002] The railway braking system is a core component ensuring train operation safety, and the brake valve, as a key component of the braking system, directly affects the braking performance and driving safety of the train. Currently, railway brake valves are mainly divided into two categories: the Type 104 passenger car air distribution valve (mostly used in passenger cars) and the Type 120 freight car air control valve (mostly used in freight cars). Due to differences in application scenarios (passenger / freight), operating environments, and safety requirements, their sensitivity to failure risks differs significantly: passenger cars (Type 104 passenger car air distribution valve) require strict control of failure risks to ensure passenger safety, and have more stringent requirements for failure warning thresholds; freight cars (Type 120 freight car air control valve), while also requiring safety, have a relatively higher risk tolerance.

[0003] Traditional methods typically involve building separate models for the air distribution valve of the Type 104 bus and the air control valve of the Type 120 freight car, requiring the collection of a large amount of labeled data and separate training for each type of valve. However, in actual operation, there is often a shortage of fault sample data for the air distribution valve of the Type 104 bus (due to the low failure rate and high data collection costs), resulting in insufficient model training and difficulty in guaranteeing prediction accuracy. Summary of the Invention

[0004] The purpose of this application is to provide a railway brake valve fault prediction model construction and a railway brake valve fault prediction method. Even when there is limited fault data for the air distribution valve of the Type 104 passenger car, a railway brake valve fault prediction model that can accurately predict the fault risk of the air distribution valve of the Type 104 passenger car can still be trained.

[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for constructing a fault prediction model for railway brake valves, including: Construct a first neural network model, which includes an input layer, a feature extraction layer, and a first output layer; The first neural network model is trained using fault data from the air control valve of a Type 120 truck. After training, the network parameters of the input layer and the feature extraction layer are frozen, and the first output layer is replaced with the second output layer to obtain the second neural network model. The second neural network model is trained using fault data from the air distribution valve of the 104-type passenger car to obtain a railway brake valve fault prediction model; wherein, the fault data of the air control valve of the 120-type freight car and the air distribution valve of the 104-type passenger car have multiple identical target fault influencing factors. A risk assessment threshold is set for the air distribution valve of the Type 104 passenger car. The risk assessment threshold is used to classify the output of the railway brake valve failure prediction model to determine the risk level of the air distribution valve of the Type 104 passenger car.

[0006] Secondly, this application provides a method for predicting railway brake valve failures, including: The fault data of the air distribution valve of the target 104 passenger car is input into the railway brake valve fault prediction model, and the risk level of the air distribution valve of the target 104 passenger car is determined according to the output of the railway brake valve fault prediction model; wherein, the railway brake valve fault prediction model is obtained by the railway brake valve fault prediction model construction method according to any one of claims 1-6.

[0007] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the railway brake valve fault prediction model construction method or the railway brake valve fault prediction method described in any one of the above.

[0008] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the railway brake valve fault prediction model construction method or the railway brake valve fault prediction method described in any one of the above.

[0009] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the railway brake valve fault prediction model construction method or the railway brake valve fault prediction method described in any one of the above.

[0010] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a railway brake valve fault prediction model construction and method. A transfer learning strategy is employed in the model training. Fault data from the air control valve of a Type 120 freight car and the air distribution valve of a Type 104 passenger car share some common target fault influencing factors, thus providing a basis for transfer learning. First, a large amount of fault data from the Type 120 freight car air control valve is used to train a first neural network model (source domain training). Then, fault data from the Type 104 passenger car air distribution valve is used to train a second neural network model (target domain training). This reduces the data volume requirement for the Type 104 passenger car air distribution valve fault data, meaning that even with less fault data for the Type 104 passenger car air distribution valve, a railway brake valve fault prediction model capable of accurately predicting fault risks can still be trained. The second neural network model is obtained by freezing the network parameters of the input layer and feature extraction layer in the first neural network (preserving its learned general fault feature extraction capabilities) and replacing the first output layer with the second output layer (making the output layer more adaptable to the Type 104 passenger car air distribution valve). Therefore, this invention solves the problem that the prediction accuracy of traditional railway brake valve fault prediction model construction methods is low due to the scarcity of fault data for the air distribution valve of the 104-type passenger car. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a method for constructing a railway brake valve fault prediction model in one embodiment of this application; Figure 2 This is a schematic diagram of a railway brake valve fault prediction model and its training parameters in one embodiment of this application; Figure 3 This is a schematic diagram of a railway brake valve fault prediction system according to an embodiment of this application; Figure 4 This is a bar chart showing the contribution of each fault factor (factor) of valve 120 to the diagnosis of valve 104 in one embodiment of this application; Figure 5 This is a comparison chart showing the prediction accuracy of valve 104 failure with and without the migration model in one embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] The method for constructing a railway brake valve fault prediction model provided in this application embodiment can be applied to a terminal or server.

[0016] The terminal can be, but is not limited to, various desktop computers, laptops, smartphones, and tablets. The server can be a standalone server, a server cluster consisting of multiple servers, or a cloud server.

[0017] In one exemplary embodiment, such as Figure 1 As shown, a method for constructing a fault prediction model for railway brake valves is provided. This method is executed by computer equipment, specifically by a computer device such as a terminal or a server alone, or by both a terminal and a server. In this embodiment, the method is described using a terminal as an example, including the following steps 110 to 140.

[0018] Step 110: Construct a first neural network model, which includes an input layer, a feature extraction layer, and a first output layer.

[0019] Step 120: The first neural network model is trained using fault data from the air control valve of a Type 120 freight car (referred to as the Type 120 valve). After training, the network parameters of the input layer and feature extraction layer are frozen, and the first output layer is replaced with the second output layer to obtain the second neural network model. The purpose of replacing the output layer is to make it more suitable for the air distribution valve of a Type 104 bus. The number of neurons and the output dimension of the second output layer match the number of target variables for the fault prediction task of the Type 104 bus air distribution valve. The parameters are fine-tuned and optimized using fault data from the Type 104 bus air distribution valve.

[0020] Step 130: Train the second neural network model using the fault data of the 104-type passenger car air distribution valve (hereinafter referred to as the 104 valve) to obtain the railway brake valve fault prediction model; among them, the fault data of the 120-type freight car air control valve and the 104-type passenger car air distribution valve have multiple identical target fault influence factors.

[0021] Step 140: Set the risk assessment threshold for the air distribution valve of the Type 104 passenger car. The risk assessment threshold is used to classify the output of the railway brake valve failure prediction model to determine the risk level of the Type 104 passenger car air distribution valve. The output of the railway brake valve failure prediction model is the comprehensive risk score of the Type 104 passenger car air distribution valve.

[0022] The fault data includes fault impact factor data (sample features) and its corresponding fault status data (sample labels).

[0023] By implementing steps 201 to 208 above, a railway brake valve fault prediction model can be obtained that can accurately predict the risk level of the air distribution valve of the Type 104 passenger car. Its main logic is to improve the accuracy of fault identification and prediction for the Type 104 passenger car air distribution valve by utilizing the similarity of fault factors between the Type 120 freight car air control valve and the Type 104 passenger car air distribution valve. First, a basic model (first neural network model) is constructed using the air control valve of a Type 120 freight car as the object. This model is trained using fault data (including corresponding fault influencing factors and faults) to capture the correlation features between "fault influencing factors" and "fault states" (i.e., training enables the feature extraction layer to learn the correlation mapping rules between "fault influencing factors" and "fault states" for the Type 120 freight car air control valve). Then, the network parameters of the input layer and feature extraction layer in the first neural network model are frozen (preserving its learned general fault feature extraction capabilities), and the first output layer is replaced with a second output layer adapted to the air distribution valve of a Type 104 passenger car (the second output layer is used to match the fault score distribution and model specificity of the Type 104 passenger car air distribution valve), forming a transfer model (second neural network model). Subsequently, the second neural network model is fine-tuned using fault data from the Type 104 passenger car air distribution valve to optimize its adaptability to the specific fault influencing factors of the Type 104 passenger car air distribution valve, ultimately forming a railway brake valve fault prediction model for the Type 104 passenger car air distribution valve. Finally, a relatively low risk judgment threshold is set for the Type 104 passenger car air distribution valve to meet the high-risk sensitivity requirements of passenger cars.

[0024] In the aforementioned method for constructing a railway brake valve fault prediction model, a transfer learning strategy is employed during model training. First, a large amount of fault data from the air control valves of Type 120 freight cars is used for initial model training (source domain training). Then, fault data from the air distribution valves of Type 104 passenger cars is used for further model training (target domain training). This reduces the data volume requirement for the fault data of the Type 104 passenger car air distribution valves, meaning that even with limited fault data for the Type 104 passenger car air distribution valves, a relatively accurate railway brake valve fault prediction model can still be trained to predict the fault risk of the Type 104 passenger car air distribution valves. Therefore, this invention solves the problem of low prediction accuracy in traditional railway brake valve fault prediction model construction methods due to the limited fault data for the Type 104 passenger car air distribution valves.

[0025] In this embodiment, the method for constructing a railway brake valve fault prediction model further includes: conducting an ablation experiment on the trained railway brake valve fault prediction model with respect to target fault influencing factors, obtaining the influence weights of each target fault influencing factor on the output of the railway brake valve fault prediction model, thereby verifying the effectiveness of transfer learning. This step can quantify the impact of different target fault influencing factors on the risk comprehensive score of the air distribution valve of the Type 104 passenger car, thus facilitating maintenance personnel to optimize the fault influencing factors with greater impact in a targeted manner. It can also verify the effectiveness of transfer learning. For example, if a certain target fault influencing factor has a significant impact on the risk comprehensive score of the air distribution valve of the Type 104 passenger car, it indicates that the faults of both the air control valve of the Type 120 freight car and the air distribution valve of the Type 104 passenger car are affected by this target fault influencing factor. Therefore, the knowledge learned by the railway brake valve fault prediction model from the fault data of the air control valve of the Type 120 freight car can be transferred to the fault prediction of the air distribution valve of the Type 104 passenger car.

[0026] Specifically, the multiple target fault influencing factors include at least: temperature and humidity difference influencing factors, skill influencing factors, sealing performance influencing factors, and parts replacement influencing factors. Therefore, the fault data for the air control valve of the Type 120 freight car and the air distribution valve of the Type 104 bus both include corresponding target fault influencing factor data and fault status data.

[0027] It should be noted that the aforementioned target fault influencing factor data may not be directly available and requires preprocessing of the original fault data. Therefore, the method for constructing a railway brake valve fault prediction model may further include: preprocessing fault data for the air control valve of the 120-type freight car and the air distribution valve of the 104-type passenger car. This preprocessing includes: weighted calculation of the temperature and humidity difference influencing factor based on the indoor and outdoor temperature and humidity difference; modeling the skill influencing factor based on the negative correlation between the rework rate and the average repair time; weighted fusion of the sealing performance influencing factor based on the pressure value and the sealing performance test value; and calculation of the component replacement influencing factor based on the correlation between the number of replacements and the rated life.

[0028] For example, first, historical operating data (fault data before preprocessing) of the air control valve of a Type 120 freight car and the air distribution valve of a Type 104 bus are obtained. The data fields include: Environmental characteristics: indoor temperature, outdoor temperature, indoor humidity, outdoor humidity.

[0029] Inspection characteristics: repair success rate, average repair time.

[0030] Performance characteristics: pressure value, sealing test value.

[0031] Life characteristics: average service life, number of times seals / springs / valve cores / connectors need to be replaced.

[0032] Derivative characteristics: temperature and humidity difference influencing factors, skill influencing factors, sealing influencing factors, replacement influencing factors.

[0033] Through preprocessing steps, the field values ​​of derived characteristics can be calculated based on the field values ​​of environmental characteristics, maintenance characteristics, performance characteristics, and lifespan characteristics. Specifically: the field value of the temperature and humidity difference influence factor is obtained based on the weighted calculation of the indoor and outdoor temperature and humidity difference; the field value of the skill influence factor is obtained based on the negative correlation modeling between the rework rate and the average repair time; the field value of the sealing performance influence factor is obtained based on the weighted fusion of pressure value and sealing performance test value; and the field value of the component replacement influence factor is obtained based on the correlation calculation between the number of replacements and the rated lifespan.

[0034] For setting the risk assessment threshold for the air distribution valve of the Type 104 passenger car, based on the difference in safety requirements between the air distribution valve of the Type 104 passenger car and the air control valve of the Type 120 freight car, the fault risk assessment thresholds for the two types of valves are calculated separately. The calculation of the fault risk assessment threshold needs to take into account the statistical distribution characteristics of their respective fault scores, and use the quantile as the fault risk assessment threshold, so that the risk assessment threshold of the air distribution valve of the Type 104 passenger car meets its higher risk sensitivity requirements. Using the trained railway brake valve fault prediction model and the adapted risk assessment threshold of the air distribution valve of the Type 104 passenger car, the fault risk prediction of the air distribution valve of the Type 104 passenger car is realized.

[0035] Specifically, the low-risk threshold quantile of the air distribution valve of the Type 104 bus is lower than that of the air control valve of the Type 120 freight car, and the high-risk threshold quantile of the air distribution valve of the Type 104 bus is lower than that of the air control valve of the Type 120 freight car, in order to match the more stringent early warning requirements of buses for fault risk; before calculating the threshold, outliers need to be removed from the historical fault scoring data, and outliers are determined based on the mean and standard deviation of the data.

[0036] Specifically, the feature extraction layer includes a fully connected layer and a dropout layer. The fully connected layer uses a non-linear activation function to enhance the feature fitting ability, while the dropout layer is used to suppress model overfitting.

[0037] Furthermore, the first and second neural network models employ regression-type loss functions and adaptive learning rate optimizers during training, which can be combined with an early stopping strategy to monitor validation set performance and avoid overtraining of the models.

[0038] In this embodiment, an ablation experiment is conducted on the trained second neural network model regarding the target fault influence factors. Specifically, this includes: setting any target fault influence factor to zero in the second neural network, determining the change in the prediction accuracy index of the second neural network after the target fault influence factor is set to zero; and determining the weight of each target fault influence factor in the second neural network based on the change in the prediction accuracy index corresponding to each target fault influence factor.

[0039] The predictive contribution of the corresponding target fault influencing factors can be measured by the absolute value of the change in the prediction accuracy index. According to the relationship between the predictive contributions of each target fault influencing factor, the weight of each target fault influencing factor in the second neural network is determined proportionally.

[0040] The following is a specific example to illustrate the method for constructing a railway brake valve fault prediction model in this embodiment.

[0041] I. Environmental preparation for implementation.

[0042] 1. Hardware environment: Processor: CPUs (such as Intel Core i7 and above) or GPUs (such as NVIDIA RTX series) that support multi-core computing to meet the computing power requirements for model training and real-time prediction.

[0043] Memory: ≥16GB (to ensure memory stability during data loading and model training).

[0044] Storage: ≥500GB SSD (used to store training data, model files and logs, improving data read and write speed).

[0045] 2. Software environment: Operating system: Windows 10 / 11 or Linux (Ubuntu 20.04 and above).

[0046] Programming language: Python 3.8 or above.

[0047] Core libraries: Data processing: pandas 1.4.0+, numpy 1.22.0+.

[0048] Machine learning: scikit-learn 1.0.0+ (data partitioning, standardization).

[0049] Deep learning: TensorFlow 2.10.0+ (model building and training).

[0050] Visualization: matplotlib 3.5.0+ (training curve plotting).

[0051] Interactive interface: tkinter 8.6 (real-time prediction GUI).

[0052] Model storage: joblib1.1.0+ (saves pre-processed assets).

[0053] II. Data Preparation and Preprocessing.

[0054] The core of this step is to transform the raw data into features and target values ​​that can be used for model training.

[0055] 1. Obtain the raw data: The collected data is stored in Excel format, containing historical operating data for the air control valves of the Type 120 truck (truck) and the air distribution valves of the Type 104 bus (bus). The fields include: Environmental characteristics: indoor temperature, outdoor temperature, indoor humidity, outdoor humidity.

[0056] Inspection characteristics: repair success rate, average repair time.

[0057] Performance characteristics: pressure value, sealing test value.

[0058] Life characteristics: average service life, number of times seals / springs / valve cores / connectors need to be replaced.

[0059] Derivative features: temperature and humidity difference influencing factors, skill influencing factors, sealing performance influencing factors, and parts replacement influencing factors; if not pre-calculated, they need to be dynamically generated.

[0060] 2. Data preprocessing: The field values ​​of the derived features are dynamically generated, specifically, four target fault influencing factors are generated in real time based on the input parameters (environmental features, maintenance features, performance features, and lifespan features).

[0061] The generation process of the temperature and humidity difference influencing factor is: (|indoor temperature - outdoor temperature| / 20)×0.6+(|indoor humidity - outdoor humidity| / 100)×0.4.

[0062] The process of generating the skill impact factor is: (Repair success rate / 100)×0.6+(1-Average repair time / 120)×0.4.

[0063] The process for generating the sealing performance influencing factor is: (pressure value / 500) × 0.5 + sealing performance test value × 0.5, where 500 is the rated pressure.

[0064] The process of generating the component replacement influencing factor is as follows: calculate "[1-(number of replacements / maximum number of replacements)×0.5+(average service life / 730)×0.5]" for 4 types of components and take the average value, where 730 is the rated service life.

[0065] Based on the four target fault impact factors (sample features), a corresponding comprehensive fault score is generated as the sample label: The four target fault impact factors are calculated by weighting based on preset weights (temperature and humidity difference impact factor 0.3, skill impact factor 0.2, sealing impact factor 0.3, and parts replacement impact factor 0.2). The formula is: Comprehensive fault score = 0.3 × temperature and humidity difference impact factor + 0.2 × skill impact factor + 0.3 × sealing impact factor + 0.2 × parts replacement impact factor.

[0066] Through the above preprocessing steps, the corresponding fault impact factor data (field values ​​of fault impact factors) and fault status data (comprehensive fault score) can be obtained.

[0067] At the same time, data cleaning is also required to automatically filter invalid values ​​(such as null values ​​and outliers) and ensure that the feature data range is consistent with reality (such as temperature -40℃~50℃, humidity 0%~100%).

[0068] III. Training of the first neural network model.

[0069] 1. Data partitioning and standardization: Feature and target definition: The four target fault influencing factors are used as the input features of the model (X_120), and the comprehensive fault score is the target value (y_120).

[0070] Dataset splitting: The dataset was divided into a training set (X_train_120, y_train_120) and a test set (X_test_120, y_test_120) in an 8:2 ratio, with a random seed of 42 to ensure reproducibility.

[0071] Standardization: The features are standardized (mean is 0, standard deviation is 1) to eliminate the influence of dimensions and improve the convergence speed of the model.

[0072] 2. Model structure and training parameters: Reference Figure 3 In the network hierarchy: the input layer receives 4-dimensional features (including temperature and humidity difference factors, skill factors, sealing factor, and parts replacement factor); the first fully connected layer (hidden layer 1) has 16 neurons and uses the ReLU activation function (to enhance nonlinear fitting ability); the dropout layer has a dropout rate of 0.2 (randomly deactivating 20% ​​of neurons to prevent overfitting); the second fully connected layer (hidden layer 2) has 8 neurons and uses the ReLU activation function (to further extract higher-order features); the output layer has 1 neuron (outputting a comprehensive fault score for risk assessment).

[0073] Training configuration: The optimizer uses Adam adaptive learning rate, which is suitable for regression tasks; the loss function uses MeanSquaredError to measure the deviation between the predicted value and the true value; the evaluation metric uses MAE (mean absolute error), which intuitively reflects the prediction accuracy; training parameters: 100 iterations, batch size 32, combined with EarlyStopping (monitoring validation loss, patience value 10), training stops when the validation loss does not decrease for 10 consecutive iterations, and the optimal parameters are retained.

[0074] 3. Model Evaluation and Storage: Performance evaluation: Calculate the loss (loss_120) and MAE (mae_120) on the test set, requiring MAE ≤ 0.05 (in actual operation, the output is similar to "test set MAE: 0.042 for air control valve of type 120 truck").

[0075] Model saving: Saves the model, normalizer, and dynamic threshold to a specified directory for easy loading and use later.

[0076] IV. Transfer learning training of the second neural network model.

[0077] 1. Construction of the second neural network model: Layer Freeze: Freeze the parameters from the input layer to the second fully connected layer in the first neural network model, retain the learned general feature extraction capability, and only allow the newly added parameters of the second output layer to be trained.

[0078] Second output layer: An output layer with one added neuron, whose initial parameters are randomly initialized using a normal distribution, is used to adapt to the fault score distribution characteristics of the air distribution valve of the 104-type bus.

[0079] 2. Fine-tune the training configuration: Data processing: The fault data of the air distribution valve of the 104-type passenger car is standardized in the same way as that of the air control valve of the 120-type freight car, and a separate normalizer and division (8:2 ratio) is used.

[0080] Training parameters: 50 iterations, batch size 32, using the Adam optimizer and MeanSquaredError loss function, and EarlyStopping strategy (patience value 10) to prevent overfitting.

[0081] Learning rate strategy: Based on the pre-trained parameters of the air control valve of the Type 120 truck, a low initial learning rate (0.001) is adopted to avoid destroying the existing feature extraction capabilities.

[0082] 3. Verification of migration effect: Performance requirements: The test set MAE must be ≤0.06 (in actual operation, the output is similar to "test set MAE of air distribution valve of type 104 bus: 0.057"). If it does not meet the standard, the structure of the second output layer needs to be readjusted.

[0083] Model saving: Save to the specified directory, and store it separately from the first neural network model after training.

[0084] V. Set risk assessment thresholds.

[0085] 1. Threshold calculation logic: Core principle: The air distribution valve of the Type 104 bus is more sensitive to risk. Compared with the air control valve of the Type 120 freight car, it needs to be judged as a higher risk under the same comprehensive fault score. Therefore, its risk judgment threshold is lower than that of the air control valve of the Type 120 freight car.

[0086] Quantile selection: The low-risk threshold for the air distribution valve of the Type 104 bus is the 50th percentile of the historical scores (50% of the historical data scores are higher than this value), and the high-risk threshold for the air distribution valve of the Type 104 bus is the 65th percentile; the low-risk threshold for the air control valve of the Type 120 freight car is the 70th percentile of the historical scores, and the high-risk threshold for the air control valve of the Type 120 freight car is the 85th percentile.

[0087] Outlier handling: Extreme scores exceeding the mean ± 3 standard deviations are removed before calculation to ensure threshold stability.

[0088] 2. Threshold application rules: Overall fault score < low risk threshold → risk level "low"; low risk threshold ≤ overall fault score < high risk threshold → risk level "medium"; overall fault score ≥ high risk threshold → risk level "high".

[0089] VI. Application of railway brake valve failure prediction model.

[0090] During the application process, a graphical user interface (GUI) can be used to realize real-time data input, prediction, and result display.

[0091] 1. Real-time data acquisition and verification: Input data: 13 parameters can be manually entered through the GUI interface or automatically collected by the sensor, namely: indoor temperature, outdoor temperature, indoor humidity, outdoor humidity, repair success rate, average repair time, pressure value, sealing test value, average service life, number of times the sealing ring is replaced, number of times the spring is replaced, number of times the valve core is replaced, and number of times the connecting parts are replaced.

[0092] Validity verification: The system automatically checks the data range (e.g., pressure value 440kPa~520kPa), triggers a pop-up alarm for outliers, and replaces them with the average of the three most recent valid data.

[0093] 2. Dynamic calculation of impact factor: Four target fault impact factors are generated in real time based on the input data.

[0094] 3. Prediction and Result Output: Model selection: Users select the valve type (120 type truck air control valve or 104 type bus air distribution valve) through a pop-up window, and the system automatically loads the corresponding model (the first neural network model and the second neural network model after training, respectively) and threshold.

[0095] Score prediction: After standardizing the target fault influencing factors, input them into the model to obtain a comprehensive fault score.

[0096] Risk assessment: The risk level is output based on the corresponding threshold. Based on the ablation experiment, the two target failure impact factors with the greatest impact on the current risk are marked by "feature importance analysis".

[0097] Results Display: The GUI text box displays the influencing factors, comprehensive score, risk level, and key influencing factors. Example output: "The failure prediction results of the air distribution valve of the Type 104 bus are as follows: the temperature and humidity difference influencing factor is 0.18, the skill influencing factor is 0.86, the sealing influencing factor is 0.89, and the parts replacement influencing factor is 0.72. The comprehensive failure score is 0.61, the risk level is low, and among the main target failure influencing factors, the sealing influencing factor is 0.09 and the skill influencing factor is 0.07."

[0098] Based on the above example, the method for constructing a railway brake valve fault prediction model in this embodiment is summarized and explained below.

[0099] This embodiment addresses the core problem of low prediction accuracy in cross-model fault diagnosis of the air control valve of the 120-type freight car and the air distribution valve of the 104-type bus: "the scarcity of fault samples of the air distribution valve of the 104-type bus leads to low prediction accuracy." It innovatively designs a transfer learning algorithm system of "common feature-driven, hierarchical transfer modeling, and contribution quantification verification." Using the air control valve of the 120-type freight car as the source domain, a reusable fault feature knowledge base is constructed. Through parameter isolation and adaptation strategies, efficient transfer of knowledge to the air distribution valve of the 104-type bus is achieved as the target domain. Simultaneously, the transfer value of each fault factor is quantified, ultimately overcoming the bottleneck of small sample modeling for the air distribution valve of the 104-type bus and improving its fault identification and prediction accuracy.

[0100] The feasibility of transfer learning relies on the commonality of fault mechanisms between the source and target domains. The core fault influencing factors of the two types of valves (temperature and humidity difference influencing factor, skill influencing factor, sealing performance influencing factor, and component replacement influencing factor) have the characteristics of "consistent calculation logic, overlapping correlation trends, and fine-tunable parameters," providing a basic support for knowledge transfer. The four core fault influencing factors of the two types of valves all adopt the same quantitative framework. The temperature and humidity difference influencing factor is based on a weighted calculation of "indoor and outdoor temperature difference weight 0.6 + humidity difference weight 0.4," with only minor adjustments to the fault trigger threshold due to differences in valve body structure: 0.85 for the air control valve of a 120-type truck vs. 0.65 for the air distribution valve of a 104-type bus. As mentioned earlier, due to the differences in safety requirements between valve 120 and valve 104, it is necessary to set risk thresholds for them separately. A comprehensive score is calculated using various influencing factors, and the comprehensive score is used to determine low, medium, and high risks. Here, the low-risk threshold for valve 120 is set to 0.70, and the high-risk threshold is set to 0.85. The low-risk threshold for valve 104 is set to 0.50, and the high-risk threshold is set to 0.65. However, the core correlation logic of "expanded temperature and humidity difference → low-temperature medium phase change → seal failure" is completely consistent; the skill impact factor is modeled based on "rework rate negative correlation weight 0.6 + average repair time negative correlation weight 0.4", and the correlation law of "every 10% increase in rework rate → 6% increase in failure risk" learned for the air control valve of the 120 truck can be directly reused; the sealing performance impact factor is calculated based on the fusion of "actual pressure value / rated pressure weight 0.5 + sealing performance test value weight 0.5", and the rated pressure (500kPa) has no model difference with the test standard; the component replacement impact factor adopts a unified mathematical model: calculate "[1 - (replacement times / maximum replacement times) × 0.5 + (average service life / 730) × 0.5]" for the four types of components and take the average value. Only the "maximum replacement times" needs to be adjusted (e.g., the maximum replacement times of the air control valve of the 120 truck is 10 times vs. the maximum replacement times of the air distribution valve of the 104 bus is 9 times) to adapt, and the trend of "increased replacement times → lifespan reduction → increased failure risk" is completely consistent. Meanwhile, the fault types of the two types of valves (valve body freezing, internal passage icing, seal aging, minor leakage, etc.) do not have model-specific categories, and the correlation strength trend of "factor-fault" is consistent. For example, the rule that "for every 10°C increase in temperature difference → the risk of icing fault increases by 30%" in the air control valve of the 120-type truck can be used as the basis for diagnosing similar faults in the air distribution valve of the 104-type bus. Only the risk trigger factor threshold needs to be finely adjusted to match the characteristics of the target domain.

[0101] This embodiment innovatively proposes a three-layer progressive transfer learning algorithm. The core of the algorithm is to achieve efficient reuse of source domain knowledge and accurate adaptation to the target domain through a parameter isolation strategy of "feature extraction layer freezing + output layer fine-tuning".

[0102] The first layer is the pre-training of the source domain basic model (first target neural network). Using the air control valve of a Type 120 truck as the source domain, the basic model is trained using ample fault samples. The focus is on strengthening the mapping ability of "fault influencing factors → higher-order features → comprehensive fault score," forming a transferable feature extraction capability. The basic model adopts a neural network architecture of "input layer - feature extraction layer - first output layer." The input layer has a dimension of 4 (corresponding to four core fault influencing factors) and receives standardized factor data. The feature extraction layer contains a "16-neuron ReLU fully connected layer + Dropout (0.2) layer + 8-neuron ReLU fully connected layer." The ReLU activation function is used to capture the nonlinear correlation between factors, and the Dropout layer suppresses overfitting to ensure feature generalization. After training with a large number of samples (≥500) of the Type 120 truck air control valve, this layer can stably output high-order fusion features of "temperature and humidity difference - sealing performance - component status." The first output layer has one neuron, outputting the comprehensive fault score of the Type 120 truck air control valve. The training objective is to minimize the mean squared error (MSE). During training, the Adam adaptive learning rate optimizer is used, combined with an early stopping strategy (monitoring and validating loss, patience value 10) to avoid overtraining, ensuring that the feature extraction layer learns "general fault association rules" rather than source domain-specific noise. Finally, the base model must meet the following requirements: mean absolute error (MAE) ≤ 0.05 on the test set, convergence of feature extraction layer parameters, and achievement of generalization standard.

[0103] The second layer is the construction of the target domain transfer model (the second neural network model, which becomes the railway brake valve fault prediction model after training). To address the problem of scarce samples (≤200) of the air distribution valve of the 104-type passenger car, the transfer model is built based on the basic model of the air control valve of the 120-type freight car. The core is to achieve knowledge reuse by "freezing the general layer and fine-tuning the special layer" to avoid overfitting caused by insufficient data in the target domain. The adjustment of the transfer model architecture follows the principle of "preserving general features and adapting to model differences". The parameters of the "input layer + feature extraction layer" of the basic model of the air control valve of the 120-type truck are frozen to preserve the learned general fault feature mapping ability (such as the correlation weight of "temperature and humidity difference → icing risk" and "replacement number → sealing failure risk"), so as to avoid small samples in the target domain from destroying general knowledge. At the same time, a dedicated output layer (second output layer, with initial parameters randomly initialized) for the air distribution valve of the 104-type bus is added after the frozen feature extraction layer. The parameters of this layer are fine-tuned only through the air distribution valve samples of the 104-type bus to adapt to the mapping differences of "high-order features → comprehensive fault score" in the target domain (such as the score weight of "sealing factor 0.7" of the air distribution valve of the 104-type bus needs to be higher than that of the air control valve of the 120-type truck to match the higher safety sensitivity of the bus). Fine-tuning training employs a low learning rate (initial learning rate 0.001) and a small batch size (batch_size=32). The air distribution valve data for the Type 104 bus uses the same standardized logic as the air control valve data for the Type 120 freight car (based on its own sample fitting to StandardScaler) to ensure the adaptability of the input feature distribution to the frozen layer. The MSE loss function is used in conjunction with an early stop strategy (patience value 10) to monitor and validate the loss, avoiding overfitting of the output layer to the small sample of the Type 104 bus air distribution valve. The transfer model must satisfy the test set MAE≤0.06, and the correlation trend of "factor-score" must be consistent with the actual fault patterns of the Type 104 bus air distribution valve (e.g., a decrease of 0.1 in the sealing factor → a decrease of 0.08 in the score).

[0104] The third layer is a transfer contribution quantification algorithm. To accurately evaluate the contribution value of each fault influence factor in the source domain to the prediction performance of the target domain, a quantification algorithm of "single factor ablation + performance comparison" is designed to clarify the core value source of transfer knowledge. In the ablation experiment, individual fault influence factors originating from the air control valve of a Type 120 freight car in the transfer model are sequentially "zeroed out" (simulating a scenario where "transfer knowledge is not used for this factor"). The change in the fault prediction accuracy of the air distribution valve of a Type 104 bus before and after ablation is compared. The larger the change, the higher the transfer contribution of that factor. Specifically, the high-risk fault identification accuracy of the transfer model on the Type 104 bus air distribution valve test set is first determined (baseline value 92%). Then, the "temperature and humidity difference influence factor, skill influence factor, sealing influence factor, and parts replacement influence factor" are zeroed out and the prediction accuracy is recalculated. Finally, the contribution is calculated using "Contribution = (Baseline accuracy - Ablation accuracy) / Baseline accuracy × 100%". (Refer to...) Figure 4 Experiments verified that the migration contribution of the four types of fault influencing factors showed significant differences. The temperature and humidity difference factor had the highest contribution (accuracy dropped to 75% after ablation, a decrease of 17%), because the low-temperature environment samples of the air control valve of the 120-type freight car were sufficient, which made up for the lack of low-temperature fault samples of the air distribution valve of the 104-type bus. The sealing factor had the second highest contribution (a decrease of 14%), and the "pressure fluctuation of 30kPa → leakage risk warning" standard learned by the air control valve of the 120-type freight car can be directly reused. The component replacement factor had a contribution of 11%, and the unified "replacement frequency - life decay" model reduced the need for data collection of components of the air distribution valve of the 104-type bus. The skill factor had a contribution of 7%, which, although low, still provided a basic framework for the maintenance quality assessment of the air distribution valve of the 104-type bus. The four types of fault influencing factors together provided more than 70% of the knowledge support for the fault diagnosis of the air distribution valve of the 104-type bus, greatly reducing the dependence of the target domain on its own samples.

[0105] To further enhance the practicality of the transfer model, a "differentiated risk threshold" was designed and a scalable generalization framework was constructed to address the higher safety sensitivity requirements of the air distribution valve in the Type 104 bus. The dynamic threshold adaptation strategy, based on the differences in safety requirements between the two types of valves, employs a calculation logic of "quantile statistics + model adaptation." For the air control valve in the Type 120 freight car, the low-risk threshold is set at the 70th quantile and the high-risk threshold at the 85th quantile (higher risk tolerance). For the air distribution valve in the Type 104 bus, the low-risk threshold is set at the 50th quantile and the high-risk threshold at the 65th quantile (more stringent risk assessment). Before threshold calculation, extreme scores exceeding "mean ± 3 standard deviations" are removed to ensure threshold stability. The design of the model's generalization capability enables the transfer learning framework to have the characteristic of "rapid adaptation to new valve types". For other brake valves (such as valve No. 13), it is only necessary to repeat the process of "freezing the feature extraction layer of the air control valve of the 120 truck → fine-tuning the exclusive output layer → calculating the adaptation threshold", without redesigning the network structure. This can shorten the model development cycle of new valve types by more than 60%, while reducing the collection requirements for fault samples of new valve types (only 200 samples are needed to meet the standard).

[0106] This railway brake valve fault prediction model construction method (transfer model construction method) solves the core pain point of small-sample modeling of the air distribution valve of the Type 104 passenger car through the logic of "common reuse-difference adaptation-contribution quantification": data cost is reduced by 60%, and the number of samples required for training the air distribution valve model of the Type 104 passenger car is reduced from 500 in the traditional non-transfer model construction method to 200, reducing the workload of data collection and annotation; refer to Figure 5 The migration model improved prediction accuracy by 20%, increasing the accuracy of identifying high-risk faults in the air distribution valve of the Type 104 bus from 72% to 92% compared to the traditional model, and reducing the MAE from 0.12 to 0.06, thus avoiding missed detection of high-risk faults. The operation and maintenance efficiency was significantly improved. By calling the migration model through the GUI interface and inputting 13 basic parameters (indoor and outdoor temperature and humidity, return rate, etc.), it can output "four types of factor values ​​+ comprehensive score + risk level + the top two influencing factors" in real time, providing operation and maintenance personnel with accurate directions for fault diagnosis.

[0107] In summary, the railway brake valve fault prediction model construction method in this embodiment is not a simple model reuse, but rather a comprehensive design of "commonality verification - hierarchical transfer - contribution quantification - scenario adaptation" to achieve efficient and controllable transfer of source domain knowledge to the target domain, and to construct a general technical framework for cross-model brake valve fault prediction.

[0108] In an exemplary embodiment, a railway brake valve fault prediction method is also provided, which includes: inputting fault data of the air distribution valve of a target 104 passenger car into a railway brake valve fault prediction model, and determining the risk level of the air distribution valve of the target 104 passenger car based on the output of the railway brake valve fault prediction model; wherein, the railway brake valve fault prediction model is obtained by the railway brake valve fault prediction model construction method provided in the above embodiment.

[0109] In one exemplary embodiment, a railway brake valve failure prediction system is also provided. (Refer to...) Figure 2 It comprises a data input layer, a preprocessing layer, a 120-valve basic model layer, a 104-valve migration model layer, a fault prediction and analysis layer, and a visualization and interaction layer. The data input layer receives basic data from the 120-valve system (initial fault data from the air control valves of the 120-type freight cars) and target data from the 104-valve system (initial fault data from the air distribution valves of the 104-type passenger cars). The preprocessing layer performs data clarification, standardization, and feature alignment on the initial fault data, generating data such as temperature and humidity difference influence factors, skill influence factors, sealing performance influence factors, and component replacement influence factors. The 120-valve basic model layer is used to mount the trained first neural network, and the 104-valve migration model layer is used to mount the trained second neural network. The fault prediction and analysis layer schedules either the first or second neural network to perform corresponding fault prediction and analysis tasks, including comprehensive fault score calculation, risk level classification, and feature importance analysis. The results are then displayed to the user through the visualization and interaction layer.

[0110] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0111] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0112] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0113] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0114] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0115] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0116] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0117] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for constructing a fault prediction model for railway brake valves, characterized in that, include: Construct a first neural network model, which includes an input layer, a feature extraction layer, and a first output layer; The first neural network model is trained using fault data from the air control valve of a Type 120 truck. After training, the network parameters of the input layer and the feature extraction layer are frozen, and the first output layer is replaced with the second output layer to obtain the second neural network model. The second neural network model is trained using fault data from the air distribution valve of the 104-type passenger car to obtain a railway brake valve fault prediction model; wherein, the fault data of the air control valve of the 120-type freight car and the air distribution valve of the 104-type passenger car have multiple identical target fault influencing factors. A risk assessment threshold is set for the air distribution valve of the Type 104 passenger car. The risk assessment threshold is used to classify the output of the railway brake valve failure prediction model to determine the risk level of the air distribution valve of the Type 104 passenger car.

2. The method for constructing a railway brake valve fault prediction model according to claim 1, characterized in that, The various target failure influencing factors include: temperature and humidity difference influencing factor, skill influencing factor, sealing performance influencing factor, and component replacement influencing factor.

3. The method for constructing a railway brake valve fault prediction model according to claim 1, characterized in that, Also includes: An ablation experiment was conducted on the trained railway brake valve fault prediction model with respect to the target fault influencing factors to obtain the influence weight of each target fault influencing factor on the output of the railway brake valve fault prediction model.

4. The method for constructing a railway brake valve fault prediction model according to claim 1, characterized in that, The feature extraction layer includes a fully connected layer and a dropout layer. The fully connected layer uses a non-linear activation function to enhance the feature fitting ability, and the dropout layer is used to suppress model overfitting. The first neural network model and the second neural network model employ a regression-type loss function and an adaptive learning rate optimizer during training.

5. The method for constructing a railway brake valve fault prediction model according to claim 2, characterized in that, Also includes: The fault data of the air control valve of the 120-type freight car and the air distribution valve of the 104-type bus are preprocessed. The preprocessing includes: obtaining the temperature and humidity difference influence factor based on the weighted calculation of the indoor and outdoor temperature and humidity difference; obtaining the skill influence factor based on the negative correlation modeling of the return rate and the average repair time; obtaining the sealing influence factor based on the weighted fusion of the pressure value and the sealing test value; and obtaining the component replacement influence factor based on the correlation calculation of the number of replacements and the rated life.

6. The method for constructing a railway brake valve fault prediction model according to claim 1, characterized in that, An ablation experiment was conducted on the trained second neural network model regarding the target fault influencing factors, specifically including: In the second neural network, any of the target fault influence factors is set to zero, and the change in the prediction accuracy index of the second neural network after the target fault influence factors are set to zero is determined. The weights of each of the target fault influencing factors in the second neural network are determined based on the changes in the prediction accuracy index corresponding to each of the target fault influencing factors.

7. A method for predicting railway brake valve failures, characterized in that, include: The fault data of the air distribution valve of the target 104 passenger car is input into the railway brake valve fault prediction model, and the risk level of the air distribution valve of the target 104 passenger car is determined according to the output of the railway brake valve fault prediction model; wherein, the railway brake valve fault prediction model is obtained by the railway brake valve fault prediction model construction method according to any one of claims 1-6.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the railway brake valve fault prediction model construction method according to any one of claims 1-6 or the railway brake valve fault prediction method according to claim 7.

9. 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 railway brake valve fault prediction model construction method according to any one of claims 1-6 or the railway brake valve fault prediction method according to claim 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the railway brake valve fault prediction model construction method according to any one of claims 1-6 or the railway brake valve fault prediction method according to claim 7.

Images