A fault prediction system and a prediction method for a helium recharge system of a helium mass spectrometer

Abstract

The application discloses a fault prediction system and method for a helium recharge system of a helium mass spectrometer device, comprising a data acquisition module connected to the PLC of the helium mass spectrometer device through an industrial communication protocol to acquire multi-source raw data in real time; a data fusion processing module connected to the data acquisition module to perform data cleaning, data segmentation, data correlation and data integration processing on the multi-source raw data, and intelligently classify the raw data based on a preset algorithm; and a fault prediction module connected to the data fusion processing module to intelligently analyze the data, including a decision tree prediction unit, a convolutional neural network prediction unit and a model switching unit to output a fault probability or a fault type. The application can not only accurately predict potential faults, but also easily adapt to different application requirements and technical upgrades through a flexible modular architecture, thereby significantly enhancing the reliability and maintenance efficiency of the helium mass spectrometer device.

Description

Technical Field

[0001] The invention belongs to the technical field of vacuum leak detection and relates to a fault prediction system and a prediction method for a helium filling and recovery system of a helium mass spectrometry device. Background Art

[0002] Helium mass spectrometry equipment is a high-precision, high-sensitivity device for detecting the air tightness of objects. The helium filling and recovery system, as a key detection system in the helium mass spectrometry equipment, uses helium as a leak-indicating gas and mass spectrometry analysis technology to detect tiny leaks in workpieces. This detection system has the advantages of high sensitivity, strong reliability, wide applicability, and precise positioning.

[0003] Helium mass spectrometry leak detection equipment has demonstrated its irreplaceable value in numerous fields. Primarily used in core industries such as aerospace, semiconductor manufacturing, and scientific research, this equipment ensures the leak tightness of gas in engine systems, high-pressure equipment, and steam systems, effectively preventing safety incidents caused by leaks. Helium mass spectrometry leak detection equipment is also crucial in the pharmaceutical, communications, and electronics industries, ensuring the leak tightness of medical devices (such as pacemakers and blood transfusion bags) and pharmaceutical packaging, as well as the reliability of electronic components. Furthermore, related equipment has also penetrated widely into fields such as refrigeration, vacuum processing, and new energy development, meeting diverse demands for high-precision and high-standard leak detection.

[0004] During long-term operation, helium mass spectrometer leak detection equipment may encounter various failure scenarios, such as decreased sensitivity, false alarms caused by laboratory environmental contamination, and helium leaks. These issues can lead to inaccurate test results, thereby compromising the safety and reliability of the equipment or system being tested. In particular, the helium filling and recovery system, as the core component of the helium mass spectrometer leak detection equipment, is responsible for accurately detecting leaks in the workpiece under test. It comprises key components such as the gas tank, pump assembly, leak detector, pneumatic valves, and piping. Failures in any of these components can significantly disrupt the entire leak detection process.

[0005] Currently, this type of equipment is typically equipped with only basic fault alarm mechanisms, providing limited analysis of the cause of the failure. Due to the high cost of helium mass spectrometer leak detection equipment, companies often do not stock additional spare units. Consequently, once a failure occurs, not only does it require the reliance on specialized personnel for troubleshooting and repair, but these technicians struggle to respond immediately, inevitably causing production line delays during repairs. This can lead to significant financial losses and technical risks for manufacturers.

[0006] Therefore, in order to improve the overall performance and production efficiency of helium mass spectrometry equipment, there is an urgent need for a helium filling and recovery system fault prediction solution that integrates high precision, multifunctionality, high scalability and modular design. Summary of the Invention

[0007] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a fault prediction system and prediction method for the helium filling and recovery system of helium mass spectrometry equipment. The system can not only accurately predict potential faults, but also easily adapt to different application requirements and technical upgrades through a flexible modular architecture, which can significantly enhance the reliability and maintenance efficiency of helium mass spectrometry equipment.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A fault prediction system for a helium filling and recovery system of a helium mass spectrometer, comprising:

[0010] The data acquisition module is used to connect to the PLC of the helium mass spectrometer through the industrial communication protocol to collect raw data in real time;

[0011] A data fusion processing module, connected to the data acquisition module, is used to perform data cleaning, data segmentation, data association, and data integration processing on the above raw data, and to intelligently classify the raw data based on a preset algorithm;

[0012] A fault prediction module, connected to the data fusion processing module, is used to perform intelligent analysis on the raw data, including a decision tree prediction unit, a convolutional neural network prediction unit and a model switching unit to output a fault probability or fault type;

[0013] The communication and interaction module is connected to the fault prediction module and is used to display the prediction results through the local human-machine interface and remotely push alarm information.

[0014] As a preferred embodiment, the data acquisition module includes:

[0015] Communication configuration submodule, used to set PLC communication parameters and map variable tags;

[0016] The data stream processing submodule is used to periodically poll the PLC for data and store the original data in a first-in-first-out queue;

[0017] The exception handling submodule is used to automatically trigger reconnection and record exception logs when communication interruption is detected.

[0018] As a preferred embodiment, the data fusion processing module includes:

[0019] The data cleaning submodule is used to fill in the gaps in the intermittent sensor signals using interpolation and remove outliers;

[0020] The data segmentation submodule is used to divide the data interval according to the equipment workflow and dynamically segment each stage using a sliding window;

[0021] The data association submodule is used to establish association rules between sensor data and device status, and to fuse pressure sensor and flow meter data through the Kalman filter algorithm.

[0022] As a preferred embodiment, the convolutional neural network prediction unit sets the input layer size to a two-dimensional tensor of [num_sensors, sequence_length], where num_sensors represents the number of sensors and sequence_length corresponds to the length of the time series of each sensor;

[0023] The hidden layer uses Rectified Linear Unit as the activation function, and the output layer uses Sigmoid activation function.

[0024] As a preferred embodiment, the fault prediction system further includes a data management unit.

[0025] The data management unit is used for storing, backing up and exporting data.

[0026] The present invention also discloses a prediction method for a fault prediction system of a helium filling and recovery system of a helium mass spectrometry device, comprising the following steps:

[0027] S1. Data acquisition: Establish a connection with the PLC through the data acquisition module to collect the original data of the high-pressure tank, low-pressure tank and workpiece sensor, as well as the valve switch status signal in real time;

[0028] S2. Data cleaning: Perform cubic spline interpolation on the collected raw data to fill in missing points and remove outliers;

[0029] S3, data segmentation: Divide the collected data processed by S2 into three stages: vacuuming, helium filling, and recovery, and use a sliding window to segment the data of each stage;

[0030] S4, data association: Kalman filtering is used to fuse the pressure sensor and flow meter data in the same stage;

[0031] S5, model selection: determine whether to call the decision tree prediction unit or the CNN prediction unit;

[0032] S6, Decision tree prediction: When the decision tree conditions are met, analysis is performed based on the decision tree algorithm to output the fault category and location information;

[0033] S7, CNN prediction: When the CNN conditions are met, analysis is performed based on the neural network algorithm and the failure probability is output;

[0034] S8. Warning output: The prediction results are transmitted to the communication and interaction module, and warnings are issued to users through local HMI display and Wi-Fi remote push.

[0035] The present invention has the following advantages:

[0036] (1) By combining a unique combination of neural network algorithms and decision tree algorithms, the present invention is able to perform in-depth analysis of multi-sensor data. This method is not only applicable to situations with abundant historical data and clear fault types, but also can handle scenarios with insufficient historical data records and complex fault patterns. During equipment operation, the system can accurately identify abnormal conditions and changes in the characteristic values ​​of key components, achieving early fault prediction and timely warning. This greatly improves the reliability and safety of the system, reduces downtime caused by sudden equipment failures, and reduces maintenance costs.

[0037] (2) The present invention adopts a modular system design, which ensures a low degree of coupling between modules and supports program expansion and deletion functions, facilitating maintenance and troubleshooting. In particular, the fault prediction module can be flexibly replaced and upgraded according to actual needs, whether using a decision tree algorithm or convolutional neural network technology. In addition, the system also supports functions such as a multi-functional user interactive touch screen, wireless WiFi remote control and information display, further enhancing the user experience and system adaptability. Overall, this flexibility not only reduces hardware costs but also improves the system's scalability and maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is the overall schematic diagram of the helium mass spectrometer equipment failure prediction system;

[0039] Figure 2 This is the architecture diagram of the data fusion processing module;

[0040] Figure 3 Draw an example diagram for a fault prediction decision tree;

[0041] Figure 4 This is the structure diagram of the convolutional neural network for multi-sensor fusion;

[0042] Figure 5 This is the value change diagram of the high-pressure tank sensor;

[0043] Figure 6 This is the value change diagram of the low-pressure tank sensor;

[0044] Figure 7 This is the value change diagram of the workpiece sensor;

[0045] Figure 8 This is a graph showing the change in model training loss under different numbers of training rounds;

[0046] Figure 9 This is a graph showing the changes in model test set accuracy under different numbers of training rounds. DETAILED DESCRIPTION

[0047] The present invention will be described in detail below with reference to the accompanying drawings.

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] like Figures 1 to 9 As shown, the present invention provides a fault prediction system for a helium mass spectrometer filling and recovery system, including: a data acquisition module for connecting to the PLC of the helium mass spectrometer via an industrial communication protocol to collect multi-source raw data in real time, which is specifically implemented as follows:

[0050] Communication configuration: Set the PLC's communication parameters (such as baud rate and station address) and map variable tags (such as "Valve1_Status" to the PLC's MB100 register) through the graphical configuration interface.

[0051] Data stream processing: adopts multi-threading mechanism, with independent threads responsible for periodic polling of PLC to ensure that the data acquisition frequency reaches above 100Hz; the data buffer area adopts FIFO queue management to avoid packet loss;

[0052] Exception handling: When a communication interruption is detected, the reconnection mechanism is automatically triggered and the exception log is recorded. For data packets that time out or fail verification, data consistency is ensured through sliding window retransmission.

[0053] A data fusion processing module, connected to the data acquisition module, is used to perform data cleaning, data segmentation, data association, and data integration processing on the multi-source raw data, and to intelligently classify the raw data based on a preset algorithm;

[0054] Data cleaning: Interpolation methods (such as cubic spline interpolation) are used to fill gaps in intermittent sensor signals. Abnormal data missing for five consecutive sampling periods is marked as invalid. The sampling times of different sensors are synchronized through timestamps and aligned to a unified time base using linear interpolation, with an error control within ±1ms.

[0055] Data segmentation: Divide data intervals according to the equipment workflow. For example, the "vacuuming stage," "helium filling stage," and "helium recovery stage" of mass spectrometry detection can be segmented separately, and dynamic segmentation can be performed using a sliding window.

[0056] Data association: Establish association rules between sensor data and device status, such as the causal relationship between valve switch closure and flow change, and use Kalman filtering to fuse pressure sensor and flow meter data to eliminate noise interference.

[0057] The fault prediction module, connected to the data fusion processing module, performs intelligent data analysis and includes a decision tree prediction unit, a convolutional neural network prediction unit, and a model switching unit to output fault probabilities or fault types. This module first preprocesses the received multi-sensor data through the data fusion processing unit, including signal filtering, noise reduction, and format conversion, to ensure the effective operation of subsequent algorithms. At the algorithmic level, the fault prediction module utilizes a unique combination of neural networks and improved decision trees. For cases with abundant historical data and clear fault types, a decision tree algorithm is used for classification or regression to construct a tree-like model based on the feature space. This not only simplifies the data analysis process but also provides intuitive rule interpretation, helping to understand the relationship between different features and faults. For cases where there is insufficient historical data and the fault modes are complex and difficult to define, convolutional neural network (CNN) technology is used to capture complex spatiotemporal feature patterns, thereby enhancing the ability to identify potential fault conditions.

[0058] Communication and Interaction Module: To achieve efficient data transmission and user interaction experience, this embodiment includes a comprehensive communication and interaction module; this module mainly includes the following parts:

[0059] Wireless communication interface: integrated Wi-Fi module, supports remote control and information display functions, allowing users to monitor equipment operation status and receive early warning information anytime and anywhere;

[0060] Wired communication interface: Provides multiple standard communication interfaces such as RS232, RS485, USB, etc. for connecting different external devices and internal components, such as vacuum gauges, molecular pumps, etc., to ensure the stability and reliability of data transmission.

[0061] Human-Machine Interface (HMI): A multifunctional user-interactive touch screen is used as the primary human-machine interface, supporting graphical data display, parameter setting, real-time monitoring, and other functions. In addition, a set of status indicators are also equipped to intuitively display the system's operating status.

[0062] Data Management Unit: Responsible for data storage, backup, and export, ensuring that important data is not lost even in the event of a sudden power outage. It also supports rapid data export via the USB interface for further analysis and processing.

[0063] In the embodiments of the present application, data cleaning aims to identify and remove erroneous and incomplete data records to ensure the accuracy and integrity of the data set; data segmentation involves accurate segmentation of data collected by sensors according to the workflow of the equipment to facilitate subsequent analysis; the data association step focuses on integrating data from different key components and establishing effective links between these data to support comprehensive analysis; finally, the data integration step is committed to efficiently merging relevant data in each workflow together to form a comprehensive data view and lay the foundation for in-depth data mining and analysis. This systematic data preprocessing method significantly improves the effectiveness and accuracy of data analysis.

[0064] In the embodiments of the present application, the fault prediction module is used for intelligent analysis of data, including a decision tree prediction unit, a convolutional neural network prediction unit and a model switching unit, to output a fault probability or a fault type. When historical data is abundant and the fault type is clear, a decision tree algorithm is recommended. This method classifies or regresses by constructing a tree model based on a feature space. Its advantages are easy to understand and implement, and it can directly process symbolic attributes and numerical data without excessive data preprocessing. In addition, decision trees can provide intuitive rule explanations, which helps to understand the relationship between different features and faults. Therefore, under sufficient labeled data support, especially when the fault type is clear, the decision tree algorithm is an efficient and transparent choice.

[0065] For those cases that lack sufficient historical data records and have complex and difficult-to-define fault patterns, a multi-sensor fusion convolutional neural network (CNN) technology is recommended. This deep learning method is particularly suitable for processing high-dimensional, nonlinear sensor data sets and can work effectively even with a small number of labeled samples. By designing a special CNN architecture to integrate information from multiple sensors, complex spatiotemporal feature patterns can be captured, thereby enhancing the ability to identify potential fault states. Especially in the face of difficult-to-detect fault situations, the powerful feature extraction ability and generalization performance of CNN can to some extent make up for the challenges brought by data scarcity.

[0066] The steps for establishing a multi-sensor fusion convolutional neural network fault prediction model are as follows:

[0067] 1) Data preprocessing

[0068] First, the sensor data needs to be preprocessed by normalizing the data, and the formula is as follows:

[0069]

[0070] in, express The minimum value of express The maximum value of .

[0071] 2) Fusion of data based on fault type

[0072] In this study, three types of sensor data closely related to failures were selected as input features: high-pressure tank sensors, low-pressure tank sensors, and workpiece sensors in the helium recovery system. Each sensor collected 1,000 data points during a complete operating cycle, and the sensor data was normalized to eliminate dimensional differences and improve model adaptability.

[0073] The three sets of normalized sensor data are combined into a 3×1000 two-dimensional time series vector according to the channel dimension as the input feature of the model. Figure 5 、 Figure 6 and Figure 7 As shown in Figure 2, the changing trends of the values ​​of these three types of sensors within a cycle are displayed.

[0074] To ensure the consistency and integrity of data sampling, all sensors are trained backwards, capturing 1,000 consecutive data points, starting with the last time point. This method effectively covers the dynamic range of sensor changes within a complete cycle, ensuring that the time series features relied upon for model training and prediction are representative and stable.

[0075] 3) Constructing a convolutional neural network prediction architecture for multi-sensor fusion

[0076] When designing the convolutional neural network architecture for multi-sensor data fusion, we set the input layer to a two-dimensional tensor of size [num_sensors, sequence_length], where num_sensors represents the number of sensors and sequence_length corresponds to the length of the time series of each sensor.

[0077] The network architecture uses a 3×3 convolution kernel to effectively capture local spatiotemporal features, which helps improve the model's ability to recognize complex patterns. For the hidden layers, ReLU (Rectified Linear Unit) is selected as the activation function. Known for its efficient computation and ability to prevent the vanishing gradient problem, the ReLU function can significantly accelerate training and improve model performance.

[0078] The output layer uses the Sigmoid activation function. The Sigmoid function maps real values ​​to the range (0, 1), providing a natural way to interpret the probability of the output belonging to a particular class. This choice is particularly suitable for binary classification scenarios where clear class affiliation is required, as it not only provides an intuitive probabilistic interpretation but also facilitates optimization using the standard binary cross-entropy loss function.

[0079] 4) Training process of convolutional neural network with multi-sensor fusion

[0080] For the training dataset, we collected and annotated time series data from multiple sensors, with each sample having the dimensions [num_sensors, sequence_length]. The data was normalized and time-aligned, and then divided into training, validation, and test sets in a ratio of 7:2:1. In this example, num_sensors and sequence_length were set to 3 and 1000, respectively. The data distribution of the training, validation, and test sets is shown in Table 1. This type of fault data contains 200 records, including 150 normal samples and 50 abnormal samples. These samples were divided into training, validation, and test sets in a ratio of 7:2:1.

[0081] Table 1 Data distribution table

[0082]

[0083] The training of convolutional neural networks for multi-sensor fusion follows these steps:

[0084] Step 1: Initialize the network. Set hyperparameters such as the learning rate, optimizer (such as Adam or SGD), and loss function (binary cross entropy). Table 2 lists the relevant hyperparameters. Different hyperparameters are required for different fault types.

[0085] Table 2 Key hyperparameters of multi-sensor fusion CNN

[0086]

[0087] Step 2: Forward propagation. The training sample is input and passes through the convolution layer, pooling layer, activation function, feature fusion and fully connected layer to calculate the model's predicted output.

[0088] Step 3: Calculate the loss. Calculate the loss based on the predicted value and the true label. Use the binary cross entropy loss function to calculate the error between the predicted output and the true label:

[0089]

[0090] in, is the true label, is the model predicted value.

[0091] Step 4: Back propagation. Calculate the gradient based on the loss value and update the network parameters through the optimizer.

[0092] Step 5: Repeat the above steps until the predetermined training rounds (epochs) are reached or the loss value converges.

[0093] Figure 8 The performance evaluation results of the model at different training epochs are presented. By analyzing key metrics such as precision, recall, and F1 score at each epoch, we can identify the optimal training epoch range, thereby guiding the selection of hyperparameter adjustments and optimization strategies, ensuring that the final model has the best generalization ability and performance.

[0094] 5) Verification and tuning of prediction models

[0095] After each round of training, the validation set is used to evaluate model performance (such as accuracy, recall, and F1 score). Hyperparameters (such as learning rate, batch size, and network depth) are adjusted based on the validation results to optimize model performance.

[0096] For classification problems related to fault detection, the most direct way to evaluate a classifier model is by measuring its classification accuracy (Acc). This is the ratio of correctly classified samples to the total number of samples. Here, TP refers to the number of positive samples predicted as positive by the model, FP refers to the number of negative samples predicted as positive by the model, FN refers to the number of positive samples predicted as negative by the model, and TN refers to the number of negative samples predicted as negative by the model.

[0097] There are three other common model evaluation indicators: Precision (Precision, Pre) represents the proportion of predicted positive categories that are actually positive; Recall represents the proportion of predicted positive categories that are actually positive; F1 Score represents the harmonic mean of precision and recall.

[0098]

[0099]

[0100]

[0101]

[0102] Table 3 Detection results of multi-sensor fusion CNN model

[0103]

[0104] Table 3 shows the model's detection results for four common fault types in helium recovery systems. All performance metrics are evaluated using the test set data. As can be seen from the table, the model demonstrates high detection accuracy and stability for Class 1, Class 2, and Class 3, demonstrating good overall performance. However, detection performance for Class 4 is relatively weak, with particularly low accuracy. This may be due to factors such as inappropriate feature selection, an uneven distribution of training samples, or model overfitting.

[0105] To further enhance the model's classification capabilities for Class 3 and Class 4, the following optimization measures were implemented: increasing the number of samples in the relevant categories to alleviate data imbalance, adjusting the model's hyperparameters using a cross-validation strategy, and introducing a more efficient feature extraction method to enhance the model's ability to learn key features. These improvements are expected to significantly improve the model's accuracy and recall for identifying low-frequency or complex fault types, thereby enhancing the overall system's fault diagnosis reliability.

[0106] Table 4 Detection results after improvement of multi-sensor fusion CNN model

[0107]

[0108] Table 4 shows the improved model test results. By increasing training samples, adjusting hyperparameters, and optimizing feature extraction methods, the model's performance on Class 4 has been significantly improved, with a significant breakthrough in precision. The recall rate for Class 3 has also improved.

[0109] The present invention is not limited to the aforementioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A fault prediction system for a helium mass spectrometer filling and recovery system, characterized in that: include: The data acquisition module is used to connect to the PLC of the helium mass spectrometer through an industrial communication protocol to collect raw data in real time, including raw data from the high-pressure tank, low-pressure tank, and workpiece sensors, as well as valve switch status signals; A data fusion processing module, connected to the data acquisition module, is used to perform data cleaning, data segmentation, data association, and data integration processing on the above raw data, and to intelligently classify the raw data based on a preset algorithm; The data fusion processing module includes: a data cleaning submodule for filling gaps in intermittent sensor signals using interpolation and removing outliers; a data segmentation submodule for dividing data intervals according to the three stages of the equipment workflow: vacuuming, helium filling, and recovery, and dynamically segmenting each stage using a sliding window; a data association submodule for establishing a correspondence between valve opening and closing states and flow rate changes, and fusing pressure sensor and flow meter data using a Kalman filter algorithm; A fault prediction module, connected to the data fusion processing module, is used to perform intelligent analysis on the raw data, including a decision tree prediction unit, a convolutional neural network prediction unit and a model switching unit to output a fault probability or fault type; The communication and interaction module is connected to the fault prediction module and is used to display the prediction results through the local human-machine interface and remotely push alarm information.

2. The fault prediction system for the helium filling and recovery system of a helium mass spectrometer according to claim 1, characterized in that: The data acquisition module includes: Communication configuration submodule, used to set PLC communication parameters and map variable tags; The data stream processing submodule is used to periodically poll the PLC for data and store the original data in a first-in-first-out queue; The exception handling submodule is used to automatically trigger reconnection and record exception logs when communication interruption is detected.

3. The fault prediction system for the helium filling and recovery system of a helium mass spectrometer according to claim 1, characterized in that: The convolutional neural network prediction unit sets the input layer to a two-dimensional tensor of size [num_sensors, sequence_length], where num_sensors represents the number of sensors and sequence_length corresponds to the length of the time series of each sensor; The hidden layer uses Rectified Linear Unit as the activation function, and the output layer uses Sigmoid activation function.

4. The fault prediction system for the helium filling and recovery system of a helium mass spectrometer according to claim 1, characterized in that: The fault prediction system further includes a data management unit, The data management unit is used for storing, backing up and exporting data.

5. A prediction method for a fault prediction system of a helium mass spectrometry equipment helium filling and recovery system according to any one of claims 1 to 4, characterized in that: The following steps are involved: S1. Data acquisition: Establish a connection with the PLC through the data acquisition module to collect the original data of the high-pressure tank, low-pressure tank and workpiece sensor, as well as the valve switch status signal in real time; S2. Data cleaning: Perform cubic spline interpolation on the collected raw data to fill in missing points and remove outliers; S3, data segmentation: Divide the collected data processed by S2 into three stages: vacuuming, helium filling, and recovery, and use a sliding window to segment the data of each stage; S4, data association: Kalman filtering is used to fuse the pressure sensor and flow meter data in the same stage; S5. Model selection: Based on the richness of historical data and the clarity of the fault type, the decision tree prediction unit or the CNN prediction unit is used. S6, Decision Tree Prediction: When historical data is abundant and the fault type is clear, analysis is performed based on the decision tree algorithm to output fault category and location information; S7, CNN prediction: When historical data is insufficient or the failure mode is complex and difficult to define, it is analyzed based on the neural network algorithm and outputs the failure probability; S8. Warning output: The prediction results are transmitted to the communication and interaction module, and warnings are issued to users through local HMI display and Wi-Fi remote push.

Images