A method for fault diagnosis and prediction of molecular pumps in fusion devices based on machine learning

Abstract

The present invention will provide a method for diagnosing and predicting molecular pump faults in a fusion device based on machine learning. The method diagnoses and predicts molecular pump faults in the fusion device in real time, and replaces and repairs molecular pumps that are about to fail according to the diagnosis and prediction results, thereby effectively reducing the possibility of major accidents of molecular pumps during the operation of the fusion device, thereby improving the safety of the vacuum system and the entire device in the fusion device, and solving the problems of lack of pertinence in the unified maintenance of conventional molecular pumps and the maintenance of molecular pumps after accidents. This is of great significance for the automated operation of future fusion devices, and effectively solves the problems faced by future fusion devices in pre-decision-making and pre-maintenance. The present invention adopts a fault diagnosis method based on a support vector machine and a prediction method based on an improved bidirectional long short-term memory network model, which can effectively improve the fault diagnosis performance of molecular pumps and the accuracy of fault prediction.

Description

Technical Field

[0001] The present invention relates to the field of vacuum technology of fusion devices, and in particular to a method for diagnosing and predicting molecular pump faults of fusion devices based on machine learning. Background Art

[0002] The vacuum pumping system is one of the core subsystems of the fusion device. The molecular pump is a precision machine used in a high vacuum environment. In order to improve the vacuum degree, pumping speed and other performance indicators, the molecular pump usually runs at an extremely high speed. The quench caused by the failure of the vacuum system will cause the local temperature of the magnet of the fusion device to rise rapidly, thereby destroying the internal structure of the superconductor. Therefore, the molecular pump is an extremely important part of the entire vacuum system and affects the automatic operation of the fusion device.

[0003] As society develops, mechanical equipment becomes more and more sophisticated, and the risk of equipment failure is increasing. At the same time, the cost of equipment maintenance is also increasing. In this context, traditional maintenance methods, including "after-the-fact maintenance" for repairs after a failure and "planned maintenance" for regular equipment inspection and maintenance, cannot meet the needs.

[0004] Traditional fault diagnosis methods are based on workers' personal experience, which is time-consuming, labor-intensive, inaccurate, and unable to predict the occurrence of faults, resulting in the failure to perform predictive maintenance on instruments and equipment.

[0005] In recent years, with the rapid development of artificial intelligence, fault diagnosis and prediction technology based on machine learning has achieved efficient and accurate results. In its most basic form, it uses algorithms to parse data, learn data, and then make predictions or judgments about certain things in the real world.

[0006] By using advanced sensor technology to collect system fault status data, the current status monitoring and fault diagnosis of the equipment can be realized. However, the current technical research in this field lacks a method to predict the life of the molecular pump in the fusion device. Therefore, the life of the molecular pump in the fusion device cannot be effectively managed, affecting the vacuum system and even the safe operation of the fusion device. Summary of the invention

[0007] In order to solve the above-mentioned existing problems, the present invention provides a method for diagnosing and predicting molecular pump faults of a fusion device based on machine learning, studies the fault diagnosis and predictability of molecular pumps of fusion devices, develops a fault diagnosis method based on SVM and a prediction method of an improved Bi-LSTM model, which can effectively improve the fault diagnosis performance of molecular pumps and the safety of early fault prediction. The method predicts the life and faults of molecular pumps in fusion devices in advance, and replaces and repairs molecular pumps that are about to fail according to the prediction results, thereby effectively reducing the possibility of major accidents of molecular pumps during the operation of fusion devices, thereby improving the safety of the vacuum system and the entire device in the device, and solving the problem of lack of pertinence in the unified maintenance of conventional molecular pumps and the maintenance of molecular pumps occurring after accidents. This is of great significance for the automated operation of future fusion devices, and can solve the problems faced by future fusion device pre-decision-making and pre-maintenance.

[0008] In view of the problems that traditional fault diagnosis methods are time-consuming and labor-intensive, and cannot make timely predictions of equipment failures, the present invention proposes SVM fault diagnosis based on time-frequency domain analysis and an improved Bi-LSTM fault prediction model.

[0009] A method for diagnosing and predicting molecular pump faults in a fusion device based on machine learning comprises the following steps:

[0010] Step 1, collecting and storing real-time status information data of the molecular pump;

[0011] Step 2, analyzing and preprocessing the real-time status information data of the molecular pump to extract the characteristic information of the molecular pump;

[0012] Step 3: Based on the molecular pump fault diagnosis model of the support vector machine algorithm, multiple fault diagnoses are realized; based on the improved bidirectional long short-term memory network molecular pump fault prediction model, molecular pump fault prediction is realized.

[0013] Furthermore, in step 1, the collected molecular pump real-time status information data includes data collected by the molecular pump speed sensor, acceleration sensor, current sensor, voltage sensor, and temperature sensor, and the data collection is realized on the industrial computer through a signal acquisition card.

[0014] Furthermore, the step 2 includes: constructing a matrix X for the data after preprocessing and feature information extraction Lm :

[0015]

[0016] Wherein, n represents the number of samples each time, r represents the molecular pump speed signal after preprocessing and feature information extraction, a represents the molecular pump acceleration signal after preprocessing and feature information extraction, s represents the molecular pump vibration signal after preprocessing and feature information extraction, t represents the molecular pump temperature signal after preprocessing and feature information extraction; L represents the physical state of the molecular pump, and the physical state includes r, a, s, and t; m represents the number of sampling times;

[0017] The above matrix X Lm Vectorize as input for model training;

[0018] Matrix X Lm Vectorized to:

[0019] vec(X Lm )=[X r1 , …, X rn , X a1 , …, X an , X s1 , …, X sn , …, X t1 , …X tn ] T

[0020] Among them, the characteristic information includes time domain characteristics and frequency domain characteristics. The time domain characteristics include peak-to-peak value, mean value, variance, standard deviation, root mean square, root amplitude, skewness, kurtosis, peak factor, pulse factor, and margin factor; the frequency domain characteristics include spectrum mean, spectrum root mean square value, frequency center of gravity, root mean square frequency, and standard deviation frequency.

[0021] Furthermore, in step 3, the molecular pump fault diagnosis model based on the support vector machine algorithm realizes multiple fault diagnoses, including: firstly collecting molecular pump fault type data, respectively extracting time domain features and frequency domain features to form a feature vector set, and then standardizing it as the input of the molecular pump fault diagnosis model based on the support vector machine algorithm;

[0022] The molecular pump fault prediction model based on the improved bidirectional long short-term memory network realizes molecular pump fault prediction, including: firstly collecting the complete life cycle time series information of the molecular pump, grouping the collected full life cycle time series data at equal intervals, extracting time domain features and frequency domain features to form a feature vector set, and finally standardizing it as an input of the molecular pump fault prediction model based on the improved bidirectional long short-term memory network to realize molecular pump fault prediction.

[0023] The Bi-LSTM (bidirectional long short-term memory network) introduces back propagation on the basis of the unidirectional LSTM (long short-term memory network) to realize the bidirectional propagation of the model, and improves the prediction accuracy of the model by learning the forward and reverse historical information of the time series;

[0024] The calculation of Bi-LSTM is based on unidirectional LSTM:

[0025]

[0026]

[0027]

[0028] Where: is the forward vector; is the reverse vector; is the forward vector at the previous moment; is the reverse vector at the next moment; y t is the forward and reverse vector action output; W is the corresponding weight coefficient; b y is the bias term; f is the activation function.

[0029] Furthermore, the improved bidirectional long short-term memory network model introduces average pooling on the output layer of the bidirectional long short-term memory network to suppress noise, reduce information redundancy, reduce model calculation amount, prevent overfitting, and further improve the overall prediction accuracy of the model.

[0030] In summary, the advantages and positive effects of the present invention are as follows: the present invention uses SVM fault diagnosis based on machine learning and improved Bi-LSTM fault prediction. By introducing the method of artificial intelligence autonomous learning, the inherent connection hidden in the data is analyzed, which solves the problems of low efficiency of traditional fault diagnosis and prediction of molecular pump late faults, and provides guarantee for the later maintenance of precision instruments.

[0031] Compared with the maintenance of traditional molecular pumps, the present invention has the following advantages: firstly, based on the improved Bi-LSTM prediction model, the accuracy of molecular pump maintenance can be greatly improved from the original post-fault maintenance to the maintenance before the failure is about to occur, thereby avoiding the secondary harm of the chain reaction of subsequent devices caused by the molecular pump failure; secondly, compared with the original post-fault maintenance and repair, which requires higher costs, this system can realize fault diagnosis and prediction, replace damaged components in advance to avoid the wear of multiple components after the molecular pump has a real failure, and has better economy. In addition, accurate positioning of the fault point can reduce maintenance costs and improve maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1It is an overall block diagram of a method for diagnosing and predicting molecular pump faults of a fusion device based on machine learning of the present invention;

[0033] Figure 2 It is a fault diagnosis flow chart of a molecular pump of a fusion device based on machine learning of the present invention;

[0034] Figure 3 It is a flow chart of the fault prediction of the molecular pump of the fusion device based on machine learning of the present invention;

[0035] Figure 4 It is the Bi-LSTM principle diagram in the fault prediction method of the present invention;

[0036] Figure 5 This is the principle diagram of the improved Bi-LSTM for fault prediction of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with 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. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0038] The present invention is a study on the diagnosis and prediction of molecular pumps for fusion devices, which has important social engineering application significance. Based on a molecular pump experimental platform, the present invention applies radial or axial force to the bearings of the molecular pump to accelerate the degradation rate of the bearings of the molecular pump and realize the degradation experiment. Through a data acquisition card, Labview is used to collect speed, acceleration, current, voltage, and temperature fault data from the stable period to the end of degradation from the sensor, which is a critical step in the degradation experiment. By preprocessing the data, the characteristic information of the data is extracted as the input of the entire diagnosis and prediction model.

[0039] The molecular pump status information collection of the present invention collects the pump body speed during the operation of the molecular pump through a speed sensor, collects the vibration acceleration during the operation of the molecular pump through a vibration acceleration sensor, collects the real-time current during the operation of the molecular pump through a current sensor, collects the real-time voltage during the operation of the molecular pump through a voltage sensor, and collects the real-time temperature during the operation of the molecular pump through a temperature sensor. The molecular pump data information storage stores the real-time status operation data of multiple molecular pumps in an industrial computer. Finally, training is performed through a fault diagnosis and prediction model, and a diagnosis prediction output is made.

[0040] like Figure 1 As shown, a method for diagnosing and predicting molecular pump faults of a fusion device based on machine learning of the present invention comprises the following steps:

[0041] Step 1: Through manual intervention on the molecular pump, some accelerated degradation experiments are performed, and the fault data of the stable operation stage, the initial stage of degradation, the middle stage of degradation, and the final stage of degradation are collected through Labview software;

[0042] Step 2: Do some preprocessing on the collected fault data, including the peak-to-peak value, mean value, variance, standard deviation, root mean square, root amplitude, skewness, kurtosis, peak factor, pulse factor, and margin factor in the time domain characteristic parameters; including the spectrum mean, spectrum root mean square value, frequency center of gravity, root mean square frequency, and standard deviation frequency in the frequency domain characteristic parameters.

[0043] Step 3: The data collected at different degradation stages are segmented into each sample, processed in the time-frequency domain, and the characteristic information of the molecular pump is extracted to obtain characteristic samples. The characteristic vector sets constructed from different characteristic samples are used as the input of the SVM (support vector machine) and the improved Bi-LSTM model, such as Figure 2 ,like Figure 3 shown.

[0044] Furthermore, the SVM classification model is: in a given series of data samples, each sample has a corresponding label. Facing a two-dimensional plane, there are two different types of samples. We need to find a straight line that can completely separate the two types of samples, and this line has the largest interval to the sample points on both sides. At this time, this line is the ideal separator. Similarly, facing a high-dimensional space, we need to find a plane with the largest interval to each type of sample point. This plane is the defined maximum interval hyperplane. Any hyperplane can be described by the following linear equation:

[0045] W T X+b=0

[0046] Among them, X is the input feature variable; W T is the normal vector, which determines the direction of the hyperplane; b is the displacement term, which determines the distance between the hyperplane and the origin.

[0047] Furthermore, the LSTM model is a variation of the traditional recurrent neural network RNN, in which a forget gate, an input gate, and an output gate are introduced into each memory unit of the network. Each LSTM memory unit has four inputs, and the three door switches use the Sigmoid function.

[0048] The input gate is calculated as:

[0049] i t =σ(w i ·[h t-1 , x t ]+b i )

[0050] Among them, i trepresents the output result of the input gate; σ is the Sigmoid function; w i and b i is the weight coefficient and bias corresponding to the input gate; h t-1 is the output of the previous moment; x t The input at the current moment.

[0051] The forget gate is calculated as:

[0052] f t =σ(w f ·[h t-1 , x t ]+b f )

[0053] Among them, f t is the output result of the forget gate; σ is the Sigmoid function; w f and b f are the weight coefficient and bias term corresponding to the forget gate respectively; h t-1 is the output of the previous moment; x t The input at the current moment.

[0054] New value input calculation:

[0055] C tx =tanh(w c ·[h t-1 , x t ]+b c )

[0056] Among them, C tx is the calculated value after the information is input but before the input gate; tanh is the tanh activation function; w c and b c are the corresponding weight coefficients and bias terms respectively; h t-1 is the output of the previous moment; x t The input at the current moment.

[0057] The final result of Cell (network unit) is:

[0058] C t =f t ·C t-1 +i t ·C tx

[0059] Among them, C t It is the final output result of the cell after adding the current input information and the information in the cell at the previous moment after the forget gate.

[0060] Output gate calculation:

[0061] O t =σ(w o ·[h t-1 , x t ]+b o )

[0062] Among them, O t is the output result of the output gate; σ is the Sigmoid function; w o and b o are the weight coefficients and bias terms corresponding to the input gate and output gate respectively; h t-1 is the output of the previous moment; x t The input at the current moment.

[0063] The final output of the LSTM model is:

[0064] h t =O t tanh(C t )

[0065] The Bi-LSTM (bidirectional long short-term memory network) introduces back propagation on the basis of the unidirectional LSTM (long short-term memory network) to realize the bidirectional propagation of the model. By learning the forward and reverse historical information of the time series, the prediction accuracy of the model is better improved. Figure 4 shown.

[0066] Among them, the Bi-LSTM calculation principle is based on unidirectional LSTM.

[0067]

[0068]

[0069]

[0070] Where: is the forward vector; is the reverse vector; is the forward vector at the previous moment; is the reverse vector at the next moment; y t is the forward and reverse vector action output; W is the corresponding weight coefficient; b y is the bias term; f is the activation function.

[0071] Improved Bi-LSTM model: In the unidirectional LSTM, the reverse LSTM is introduced to construct the Bi-LSTM model. The forward LSTM is used to capture the feature information of the previous text, and the reverse LSTM is used to capture the information of the following text. At the same time, Mean pooling is added to the output layer of the Bi-LSTM neural network to equally weight all feature extracted information to suppress noise, reduce information redundancy, reduce model calculations, and prevent overfitting. Compared with the Bi-LSTM model, the improved model can capture more feature information for model training, thereby improving the prediction accuracy of the model, such as Figure 5 shown.

[0072] Furthermore, by extracting the core data that affects the molecular pump's operating failures and service life from the molecular pump's status information data, and assigning weight values ​​to the training model based on the effects of different parameters on the molecular pump in the historical database, the weight matching is automatically updated through the model's forward and backward propagation.

[0073] Furthermore, the process of the molecular pump fault diagnosis and prediction method of the fusion device based on machine learning is as follows:

[0074] The molecular pump state information collected above, including the molecular pump speed signal, the molecular pump acceleration signal, the molecular pump current signal, the molecular pump voltage signal and the molecular pump temperature signal, is preprocessed and feature extracted, and a matrix is ​​constructed using the data after preprocessing and feature extraction, wherein n in the matrix represents the number of samples each time, r represents the molecular pump speed signal after preprocessing and feature extraction, a represents the molecular pump acceleration signal after preprocessing and feature extraction, s represents the molecular pump vibration signal after preprocessing and feature extraction, and t represents the molecular pump temperature signal after preprocessing and feature extraction;

[0075]

[0076] In the above matrix, L represents the different physical states of the molecular pump, which refers to r, a, s, and t; m represents the number of sampling times, which refers to n;

[0077] Vectorize the above matrix model as input for model training;

[0078] The matrix model is vectorized as:

[0079] vec(X Lm )=[X r1 , …, X rn , X a1 , …, X an , X s1 , …, X sn , …, X t1 , …X tn ] T

[0080] The vector constructed from the above molecular pump operation data is input into the SVM and improved Bi-LSTM models for training. The weight coefficient is continuously adjusted through the autonomous learning of the model to achieve the optimal effect of model training.

[0081] The above method process steps are to use Labview to collect the data from the fault information, perform preprocessing analysis on the fault data through feature information preprocessing, use it as the input of the SVM classification model and the improved Bi-LSTM model, and output the results through training and testing. According to the results, formulate corresponding maintenance strategies to reduce maintenance and operation costs, and avoid damage to the entire fusion device caused by damage to the molecular pump.

[0082] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for diagnosing and predicting molecular pump faults in a fusion device based on machine learning, characterized in that: The steps include: Step 1, collecting and storing real-time status information data of the molecular pump; Step 2, analyzing and preprocessing the real-time status information data of the molecular pump to extract the characteristic information of the molecular pump; Step 3: Based on the molecular pump fault diagnosis model of the support vector machine algorithm, multiple fault diagnoses are realized; based on the improved bidirectional long short-term memory network molecular pump fault prediction model, molecular pump fault prediction is realized; The molecular pump fault diagnosis model based on the support vector machine algorithm realizes multiple fault diagnoses, including: firstly, collecting data of multiple fault types of molecular pumps, extracting time domain features and frequency domain features respectively, forming a feature vector set, and then standardizing it as the input of the molecular pump fault diagnosis model based on the support vector machine algorithm; The molecular pump fault prediction model based on the improved bidirectional long short-term memory network realizes molecular pump fault prediction, including: firstly collecting the complete life cycle time series information of the molecular pump, grouping the collected full life cycle time series data at equal intervals, extracting time domain features and frequency domain features to form a feature vector set, and finally standardizing it as the input of the improved bidirectional long short-term memory network molecular pump fault prediction model to realize molecular pump fault prediction; The bidirectional long short-term memory network, namely Bi-LSTM, introduces back propagation on the basis of unidirectional LSTM, namely long short-term memory network, realizes bidirectional propagation of the model, and improves the prediction accuracy of the model by learning forward and reverse historical information of time series; The calculation of Bi-LSTM is based on unidirectional LSTM: ； ； ； Where: is the forward vector; is the reverse vector; is the forward vector at the previous moment; is the reverse vector at the next moment; is the forward and reverse vector action output; W is the corresponding weight coefficient; is the bias term; is the activation function; x t The input at the current moment.

2. The method for diagnosing and predicting molecular pump faults of a fusion device based on machine learning according to claim 1, characterized in that: In the step 1, the collected molecular pump real-time status information data includes data collected by the molecular pump speed sensor, acceleration sensor, current sensor, voltage sensor, and temperature sensor, and the data collection is realized on the industrial computer through a signal acquisition card.

3. The method for diagnosing and predicting molecular pump faults of a fusion device based on machine learning according to claim 1, characterized in that: The step 2 comprises: constructing a matrix X for the data after preprocessing and feature information extraction Lm : ； Wherein, n represents the number of samples each time, r represents the molecular pump speed signal after preprocessing and feature information extraction, a represents the molecular pump acceleration signal after preprocessing and feature information extraction, s represents the molecular pump vibration signal after preprocessing and feature information extraction, t represents the molecular pump temperature signal after preprocessing and feature information extraction; L represents the physical state of the molecular pump, and the physical state includes r, a, s, and t; m represents the number of sampling times; The above matrix X Lm Vectorize as input for model training; Matrix X Lm Vectorized to: ； Among them, the characteristic information includes time domain characteristics and frequency domain characteristics. The time domain characteristics include peak-to-peak value, mean value, variance, standard deviation, root mean square, root amplitude, skewness, kurtosis, peak factor, pulse factor, and margin factor; the frequency domain characteristics include spectrum mean, spectrum root mean square value, frequency center of gravity, root mean square frequency, and standard deviation frequency.

4. The method for diagnosing and predicting molecular pump faults of a fusion device based on machine learning according to claim 1, characterized in that: The improved bidirectional long short-term memory network model introduces average pooling on the output layer of the bidirectional long short-term memory network to suppress noise, reduce information redundancy, reduce model calculation amount, prevent overfitting, and further improve the overall prediction accuracy of the model.

Images