A multi-modal neural network computing method for ion tube residual life prediction

Abstract

The application discloses a kind of multi-modal neural network calculation methods for ion tube remaining life prediction, comprising the following steps: analyzing ion tube operating state characteristic data, extracting data modal features;Multi-modal neural network is constructed and trained;The remaining life of ion tube is predicted using multi-modal neural network.The designed method realizes the prediction calculation of the remaining life of ion tube, with the advantages of matching actual state and accurate calculation.

Description

Technical Field

[0001] This invention relates to a multimodal neural network calculation method for predicting the remaining life of ion tubes. Background Technology

[0002] An ion tube is a dielectric barrier discharge device used to generate atmospheric pressure low-temperature plasma, primarily for applications in air pollution control, air purification, and sterilization. In practical applications, the discharge process of ion tubes is gentle and stable, making it difficult to observe the corona discharge with the naked eye, and consequently, difficult to obtain their operational status through direct observation. From the user's perspective, this lack of observability leads to widespread "lifespan anxiety" regarding continuously operating ion tubes. From an ion tube operation and maintenance management perspective, timely and accurate acquisition of the remaining lifespan of the ion tube is crucial for efficient and precise operation and maintenance, and can also provide effective data support for alleviating users' "lifespan anxiety."

[0003] The Lissajous figure is a fundamental image characterizing the operating state of an ion tube. Generally, the QV method is used to obtain the Lissajous figure of an ion tube, that is, by obtaining the total voltage and the measured voltage of the ion tube during operation, and then drawing the Lissajous figure from the total voltage and the measured voltage. Therefore, the total voltage, the measured voltage, and their relationship can characterize the operating state of the ion tube. Based on the above analysis, to calculate the remaining lifespan of the ion tube, an existing scheme (patent application number: 202411253281.1) proposes a scheme using a long short-term memory neural network to predict the measured voltage of the ion tube and calculate the remaining lifespan. However, this scheme only predicts and calculates the measured voltage of the ion tube, ignoring the total voltage and the relationship between the total voltage and the measured voltage. Furthermore, due to the nonlinear characteristics of the ion tube aging process, the calculation scheme using a single neural network is difficult to adapt to the calculation of each stage of the ion tube's entire lifespan. Therefore, in the implementation process, there are problems such as a short effective prediction time, low long-term prediction accuracy, and failure to consider the "accelerated depreciation" and "delayed aging" of the ion tube. Based on this, in order to improve the accuracy of the remaining lifetime prediction of the ion tube, this scheme adopts a multimodal neural network that matches the state of the ion tube to perform segmented relay calculation of the remaining lifetime of the ion tube. The designed method has the advantages of matching the actual state and calculating accurately. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a multimodal neural network calculation method for the remaining lifetime of an ion tube that accurately matches the actual state of the ion tube.

[0005] The technical solution of this invention to solve the above problems is: a multimodal neural network calculation method for predicting the remaining life of ion tubes, comprising the following steps:

[0006] S1. Analyze the operating status characteristics data of the ion tube and extract the data modal features;

[0007] S2. Construct and train a multimodal neural network;

[0008] S3. Use a multimodal neural network to predict the remaining lifespan of the ion tube.

[0009] The advantages of this invention are as follows: Addressing issues such as short effective prediction time, low long-term prediction accuracy, and failure to consider the "accelerated depreciation" and "delayed aging" of ion tubes in the calculation of remaining lifespan, this invention provides a multimodal neural network calculation method that matches the actual state of the ion tube and offers high calculation accuracy, compared to traditional methods. The advantages of this method are: It constructs a multimodal neural network calculation model based on the modal characteristics of the ion tube's operating state data, and matches the neural network model to the ion tube's state data. By combining short-term and long-term prediction calculations, the remaining lifespan of the ion tube is calculated, improving the accuracy of the remaining lifespan calculation and reducing "lifespan anxiety" during the use of the ion tube. Attached Figure Description

[0010] Figure 1 A flowchart of the present invention.

[0011] Figure 2 The principle diagram of multimodal neural network construction in this invention.

[0012] Figure 3 The multimodal neural network structure diagram in this invention.

[0013] Figure 4 A comparison chart of the predicted total voltage calculation of the ion tube in this invention.

[0014] Figure 5 Comparison chart of voltage prediction calculation for ion tube measurement in this invention.

[0015] Figure 6 Flowchart for calculating the remaining lifespan of the ion tube in this invention.

[0016] Figure 7 A schematic diagram illustrating the calculation process of the remaining lifespan of the ion tube in this invention.

[0017] Figure 8 The diagram shows the calculation results of the ion tube state prediction in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0019] The present invention provides a multimodal neural network calculation method for predicting the remaining life of ion tubes, the flowchart of which is shown below. Figure 1As shown, it includes the following steps:

[0020] S1. Analyze the operating status characteristics data of the ion tube and extract the data modal features.

[0021] An ion tube is a dielectric barrier discharge device used to generate atmospheric pressure cryogenic plasma. This device employs a sealed glass tube structure, with a thin metal sheet lining the inner tube as the inner electrode and a stainless steel mesh surrounding it as the outer electrode. A stable corona discharge is generated by applying a high voltage between the inner and outer electrodes. Ion tube operating state characteristic data refers to data characterizing the continuous operating state of the ion tube, including the measured voltage and total voltage. The measured voltage and total voltage represent the discharge charge and operating voltage during ion tube operation, respectively. A Lissajous figure can be plotted using the measured and total voltages, which characterizes the operating state of the ion tube. Extracting modal features involves analyzing the time-series data of the total and measured voltages during ion tube operation to extract trend features, frequency features, and correlation features, thus obtaining the multimodal characteristics of the data.

[0022] S2. Construct and train a multimodal neural network.

[0023] Multimodal neural networks refer to Long Short-Term Memory (LSTM) neural networks that use the total voltage, measured voltage, and modal characteristics of the ion tube as inputs to predict and calculate the time series signals of the total voltage and measured voltage. Figure 2 The diagram shown illustrates the construction principle of the multimodal neural network in this invention. The structure of the constructed long short-term memory neural network is as follows. Figure 3 As shown, the system comprises an input layer, an LSTM layer, a fully connected layer, and an output layer. The input layer has a 500×6 feature matrix with 6 features and a sequence stride of 500 points. The LSTM layer consists of three LSTM neurons with 64, 32, and 16 neurons per layer, respectively, and a dropout rate of 20% for each neuron. The fully connected layer has 32 neurons with a dropout rate of 20%. The output layer has two neurons, outputting the total voltage and measured voltage signals, respectively. The training process for the constructed multimodal neural network is as follows: A sample dataset is constructed using the acquired ion tube operating state data (total voltage and measured voltage time series signals) and its modal feature data. The sample dataset contains 1.08 million data points over a 15-day time span. The constructed LSTM neural network is trained and tested using this sample dataset. The testing process involves predicting the state 360 ​​hours (15 days) after the current state data using the trained neural network. The test results are shown below. Figure 4 , Figure 5 As shown. Among them. Figure 4 This is a comparison chart of total voltage prediction calculations. Figure 5The comparison chart for voltage prediction calculation shows that the constructed multimodal LSTM neural network can predict the state of the ion tube within the next 15 days based on the current state of the tube, and the prediction calculation accuracy is high.

[0024] S3. Use a multimodal neural network to predict the remaining lifespan of the ion tube.

[0025] The prediction of the remaining lifetime of the ion tube using a trained multimodal LSTM neural network includes two processes: short-term prediction calculation and long-term prediction calculation. The remaining lifetime of the ion tube is measured in hours. The effective duration of the short-term prediction calculation is 360 hours (15 days). The long-term prediction calculation adopts a segmented relay calculation method, consisting of multiple short-term prediction calculation processes. Figure 6 The flowchart for calculating the remaining lifespan of the ion tube in this invention is as follows: ① Obtain the current operating status data of the ion tube; ② Identify the ion tube status based on the data; ③ Determine whether the ion tube has failed based on the status. If it has failed, the remaining lifespan is given based on the cumulative short-term prediction calculation time; if it has not failed, the multimodal LSTM neural network model participating in the prediction calculation is located based on the current status; ④ Perform short-term prediction calculation based on the multimodal LSTM neural network model to obtain the status prediction data; finally, return to the ion tube status identification process in step ② to realize the iterative calculation. Figure 7 This diagram illustrates the process of calculating the remaining life of an ion-tube. First, based on the current operating status data of the ion-tube, the model is located using a multimodal neural network (LSTM_2), which performs two short-term prediction calculations. Next, based on the state determined by the prediction calculations, another multimodal neural network (LSTM_3) is located. Through several more short-term prediction calculations and state identification, the multimodal neural network model performs segmented relay calculations. Finally, when the state identification result indicates a failed tube, the remaining life value is given. This strategy, based on state localization and matching with a multimodal neural network, can solve the problems of "accelerated depreciation" and "delayed aging" in the remaining life prediction process of ion-tubes. Figure 8 The figure shows the predicted state calculation results of the ion tube based on the segmented relay calculation method described above. The three Lissajous figures in the figure are plotted using the total predicted voltage and measured voltage signals of the three segments, corresponding to the continuous operation of the ion tube for three months. As can be seen from the figure, the discharge capacity of the ion tube gradually weakens over time (corresponding to a gradual decrease in the area of ​​the Lissajous figure). The trend shown by the predicted calculation results is basically consistent with the actual situation, indicating that the designed calculation method is feasible.

Claims

1. A multi-modal neural network computing method for ion tube residual life prediction, characterized in that, The method comprises the following steps: S1. Analyzing ion tube running state characteristic data and extracting data modal features; The ion tube refers to a dielectric barrier discharge device for generating atmospheric pressure low-temperature plasma, which adopts a sealed glass tube structure, a metal sheet as an inner electrode lined in the glass tube, and a stainless steel mesh as an outer electrode wrapped outside, and forms stable corona discharge by applying high voltage between the inner and outer electrodes; the analysis of ion tube running state characteristic data refers to the analysis of data representing the continuous running state of the ion tube, including the measured voltage and total voltage of the ion tube, wherein the measured voltage and total voltage represent the charge amount and working voltage during the running of the ion tube, and the Lissajous figure of the ion tube is drawn by the measured voltage and total voltage, and the running state of the ion tube is represented by the Lissajous figure; the extraction of data modal features refers to the analysis of the time series data of the total voltage and measured voltage during the running of the ion tube, and the extraction of the trend features, frequency features and correlation features of the data to obtain the multi-modal features of the data; S2. Constructing and training a multi-modal neural network; The multi-modal neural network refers to a long short-term memory neural network (LSTM) that uses the total voltage, measured voltage and modal features of the ion tube as input to realize the prediction and calculation of the time series signal of the total voltage and measured voltage; the structure of the constructed long short-term memory neural network is: input layer, LSTM layer, fully connected layer, output layer, wherein the feature matrix of the input layer is 500×6, including 6 features, and the sequence step is 500 points; the LSTM layer includes three layers of LSTM neurons, and the number of neurons in each layer is 64, 32 and 16 respectively, and the dropout rate of each layer of neurons is set to 20%; the number of neurons in the fully connected layer is 32, and the dropout rate is set to 20%; the number of neurons in the output layer is 2, which outputs the total voltage and measured voltage signals respectively; the training process of the multi-modal neural network is: constructing a sample data set from the obtained total voltage and measured voltage and modal features of the ion tube, and the size of the sample data set is 1.08 million data, and the time span of the sample data is 15 days; training and testing the constructed LSTM neural network using the sample data set; S3. Predicting the remaining life of the ion tube using the multi-modal neural network.

2. The multi-modal neural network computing method for ion tube residual life prediction according to claim 1, characterized in that, The step S3 adopts the trained multi-modal LSTM neural network to predict the ion tube remaining life, including a short-term prediction calculation and a long-term prediction calculation, and the remaining life of the ion tube is in units of hours; the effective time length of the short-term prediction calculation is 15 days, the long-term prediction calculation adopts a segmented relay calculation mode and is composed of multiple short-term prediction calculation processes; the specific calculation process is: ①obtaining current running state data of the ion tube, ②identifying the ion tube state based on the data, ③judging whether the ion tube is failed based on the state, if failed, giving the remaining life according to the accumulated short-term prediction calculation time length, if not failed, positioning the multi-modal LSTM neural network model participating in the prediction calculation based on the current state, ④carrying out the short-term prediction calculation based on the multi-modal LSTM neural network model to obtain state prediction data, and finally returning to the ion tube state identification process to realize cyclic calculation.

Images