Remaining useful life prediction method for laboratory equipment based on automatic machine learning

Abstract

The present invention relates to a method for predicting the remaining useful life of laboratory equipment based on automatic machine learning, comprising obtaining original time series signal data of the equipment; extracting the time series features of the original time series signal using the TSFRESH algorithm; screening the time series features according to the weight coefficients of the features to obtain a feature subset; matching the feature subset with a machine learning pipeline using a hierarchical task planning network; inputting the feature subset into the machine learning pipeline to obtain the remaining useful life of the corresponding equipment. Compared with the prior art, the present invention has the advantages of more accurate prediction results and the like.

Description

Technical Field

[0001] The present invention relates to the field of fault prediction and health management, and in particular to a method for predicting the remaining useful life of laboratory equipment based on automatic machine learning. Background Art

[0002] With the continuous progress of science and technology and the continuous improvement of industrial level, the government, universities and other enterprises and institutions pay more and more attention to laboratory construction. Advanced instruments and research methods have greatly improved the efficiency of experiments and research fields. Automation and informatization have become important development trends of laboratories in the future. In response to the development needs of education, universities have invested more special funds in the planning and construction of laboratories. The replacement of laboratory instruments and equipment has also put forward higher requirements for laboratory management. How to achieve efficient and intelligent management of laboratories and equipment to maximize their role is an important topic in laboratory management. At the same time, in order to ensure the orderly and smooth progress of experimental work and to protect personal and property safety, laboratory safety management is of vital importance. Due to the large number of equipment, long operation time, large load, dense personnel and complex environment in university laboratories, once an accident occurs, it will have a greater impact and harm. Therefore, operation maintenance and daily maintenance are important tasks of laboratory management. Safety management should be based on prevention, reduce or eliminate the probability of accidents from the source; improve the reliability of equipment, extend the trouble-free operation time of equipment, and thus improve the overall utilization rate of the laboratory.

[0003] In the prior art, timely maintenance of laboratory equipment can be performed by predicting the remaining service life of the equipment. However, the existing prediction methods cannot achieve accurate prediction of the remaining service life, and thus cannot guarantee the reliable operation of the equipment, and are therefore of poor practicality. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for predicting the remaining useful life of laboratory equipment based on automatic machine learning.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for predicting the remaining useful life of laboratory equipment based on automatic machine learning (AutoML) comprises the following steps:

[0007] S1. Obtain the original timing signal data of the device;

[0008] S2, extracting the time series features of the original time series signal using the time series feature extraction (TSFRESH) algorithm;

[0009] S3, screening the time series features according to the weight coefficient of the features to obtain a feature subset;

[0010] S4, using Hierarchical Task Planning Network (HTN) to match machine learning pipelines for feature subsets;

[0011] S5. Input the feature subset into the machine learning pipeline to obtain the remaining useful life (RUL) of the corresponding equipment.

[0012] Furthermore, the machine learning pipeline includes a preprocessor and a classifier, wherein the preprocessor is used to perform data preprocessing on the feature subset, and the classifier maps the preprocessed feature subset to the prediction algorithm.

[0013] Furthermore, after obtaining the machine learning pipeline, the parameters of the machine learning pipeline are tuned using a hyperparameter optimization method.

[0014] Furthermore, the hyperparameter optimization method includes random search, grid search or Bayesian optimization.

[0015] Furthermore, the step S3 is specifically as follows:

[0016] S31. Use hypothesis testing method to obtain evaluation results of time series characteristics;

[0017] S32, obtaining a weight coefficient of the feature according to the evaluation result and the fault extraction rate;

[0018] S33. Sort and filter the features according to the weight coefficients.

[0019] Furthermore, in step S31, different hypothesis testing methods are used for different time series features:

[0020] If both features and labels are binary, use the Fisher test;

[0021] If the feature is binary and the label is continuous, the KS test is used;

[0022] If the feature is continuous and the label is binary, use the KS test or the Mann-Whitney U test;

[0023] If both features and labels are continuous values, the Kendall correlation coefficient test is used.

[0024] Furthermore, an asymmetric loss function is used to evaluate the prediction results.

[0025] Furthermore, the asymmetric loss function L(D test ) is as follows:

[0026]

[0027] In the formula, N represents the number of devices, a k is the remaining lifetime estimation accuracy of device k.

[0028] Furthermore, the method for acquiring the original timing signal data of the device is as follows:

[0029] The raw data from the sensors on the acquisition device are extracted, transformed and loaded, and combined into a unified raw time series signal data source that is easy to analyze.

[0030] Furthermore, after obtaining the original time series signal, the data is preprocessed using data cleaning, data integration, data transformation and data reduction.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] 1. The present invention obtains the timing signal data of the equipment, uses the TSFRESH algorithm to quickly extract features, screens the features, and uses the hierarchical task planning network to automatically match the corresponding machine learning pipeline, processes the features and inputs them into the machine learning algorithm to obtain the prediction results of the remaining service life of the equipment, thereby achieving effective maintenance of the equipment. The present invention combines machine learning and hierarchical task planning networks, has high prediction accuracy and low computational complexity, and is suitable for real-time maintenance needs.

[0033] 2. The present invention selects different hypothesis testing methods for different time series features and labels, so that the correlation calculation of time series and its features is more accurate, further improving the prediction accuracy.

[0034] 3. Since overestimating the remaining service life has more serious consequences than underestimating the remaining service life, the present invention uses an asymmetric loss function to evaluate the prediction results, so that the optimization direction of the results is more accurate.

[0035] 4. The present invention performs data preprocessing and integration on both the original data and the obtained features, simplifies the calculation and improves the prediction efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the process of the present invention.

[0037] Figure 2 Schematic diagram of the data preprocessing process of the present invention.

[0038] Figure 3 A schematic diagram of the application principle of the hierarchical task planning network of the present invention.

[0039] Figure 4 It is a flowchart of the overall execution of the present invention.

[0040] Figure 5A flowchart of obtaining prediction results and subsequent diagnosis process for the present invention. DETAILED DESCRIPTION

[0041] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0042] This embodiment provides a method for predicting the remaining useful life of laboratory equipment based on automatic machine learning. Figure 1 As shown, the specific steps include:

[0043] Step S1, acquiring and integrating the original timing signal data of the device through the sensor;

[0044] Step S2, using the TSFRESH algorithm to adopt a variety of time series characterization methods to extract the time series features of the original time series signal;

[0045] Step S3: obtaining a weight coefficient of a feature based on a hypothesis test and a filter, and screening the time series features according to the weight coefficient of the feature to obtain a feature subset;

[0046] Step S4, using a hierarchical task planning network to match a machine learning pipeline for a feature subset;

[0047] Step S5: Input the feature subset into the machine learning pipeline to obtain the remaining service life of the corresponding equipment.

[0048] Among them, step S1 is specifically expanded as follows:

[0049] First, the data transmitted by the sensor is written into the corresponding data table in the database using an asynchronous queue. The same type of sensor on the same device is written into the same table, with an automatically increasing ID as the primary identifier, and a timestamp is automatically written. Then, multiple one-to-one, one-to-many, or many-to-many relationship tables are operated to integrate data from different sources. The data is aggregated into the required analysis level through various levels of identification, i.e., the ETL process, which makes the data easier to calculate and analyze later.

[0050] Step S2 is specifically expanded as follows:

[0051] First, the data needs to be preprocessed to simplify the subsequent calculation of feature extraction. The main contents include data cleaning, data integration, data transformation and data reduction. Data cleaning mainly deletes irrelevant data, duplicate data, abnormal data and noise data in the original data set, and uses Lagrange interpolation method to interpolate missing data, such as Figure 2As shown; data integration mainly deals with different forms of data expressions from multiple data sources, converts, refines and integrates the source data from the bottom layer, and reduces the redundancy and inconsistency of data expression through integration, thereby improving the speed and quality of data mining; data transformation includes normalization, function transformation, etc., which mainly standardizes the data and converts the data into an appropriate form to meet the needs of the algorithm; data reduction is used to obtain a reduced representation of the data set from the source data. The main methods include feature reduction, sample reduction and eigenvalue reduction. The space occupied by the reduced data becomes smaller, but the amount of information remains basically unchanged.

[0052] After preprocessing, the data is related to sensors and devices and is expressed as multivariate data of different lengths. Therefore, the TSFRESH algorithm is used to automatically extract the characteristics of time series signals. The device is represented as s i , relative to sensor k, whose measurement value is a time series continuous variable s relative to t i,k (t). Assume that the sampling information is a sequence of length q, expressed as: s i,k (t1),s i,k (t2),...,s i,k (t q ). The number of devices is i=1,...,m, and the average number of sensors per device is k=1,...,n, so the number of time series is m·n. In order to reduce the amount of data, the time series is extracted, and the feature matrix and target vector formed by it are the basis of the supervised classification algorithm.

[0053] Step S3 includes the following steps:

[0054] Step S31, using a hypothesis testing method to obtain an evaluation result of the time series feature;

[0055] Step S32, obtaining a weight coefficient of the feature according to the evaluation result and the fault extraction rate;

[0056] Step S33: sort and filter the features according to the weight coefficients.

[0057] The details can be expanded as follows:

[0058] After all the feature sets are obtained, the redundant features in the feature set are filtered out through the correlation analysis filter. First, different hypothesis testing methods are adopted according to different time series: if the features and labels are both binary, the Fisher test is used; if the features are binary and the labels are continuous values, the KS test is used; if the features are continuous values ​​and the labels are binary, the KS test or the Mann-Whitney U test is used; if the features and labels are both continuous values, the Kendall correlation coefficient test is used. The Benjamini-Yekutieli process is used as a filter. This process is a global inference error measure used to control the reference fault extraction rate (false extraction rate, FER), and its calculation formula is as follows:

[0059]

[0060] The weight coefficient w of the extracted feature can be determined by the calculated FER value, and the features are sorted according to w. Whether to retain this feature is determined based on the sorting result. The unretained features are deleted from the feature set to obtain a feature subset.

[0061] Step S4 is specifically expanded as follows:

[0062] Hierarchical Task Planning Network is an intelligent planning technology that breaks through previous planning technologies. Hierarchical Task Network (HTN) is a partially ordered task set T. Task t(v0,...,v n ) is a name with a list of parameters, which are variables or constants from L. Tasks named by operators are called atomic tasks, otherwise they are complex tasks. Tasks with constant parameters are basic. The goal of planning is to achieve a set of tasks. The input of the planning system contains a set of actions to achieve the goal and a set of methods to decompose the tasks. The planning process is to recursively decompose non-atomic tasks into atomic tasks that can be completed by directly executing the planned actions. The principle is as follows Figure 3 shown.

[0063] In this embodiment, the planning goal of the hierarchical task planning network is to find a machine learning pipeline, which is used to obtain the prediction results of the life span, and is composed of a parameterized preprocessor and a learner, which associates the input data with the adopted processing method. In this case, two types of algorithms can be applied. The preprocessor processing objects include preprocessing methods such as dimensionality reduction methods, feature selection, interpolation, discretization, and normalization. The learner maps the data set to the prediction algorithm. The planning goal is usually solved by simplifying it to a standard graph search problem, such as depth-first search, best-first search and other algorithms, and there is a one-to-one correspondence between the search space elements (machine learning pipelines) and the target nodes.

[0064] This embodiment uses a best-first search algorithm to identify good pipelines. The best-first search algorithm explores an implicitly given graph by assigning a number to each node and selecting the node with the current best (usually the lowest) known expansion value in each iteration. Expansion means calculating all successor nodes of a node. The graph description consists of a root node, a successor calculation function for expansion, and a predicate on the node that indicates whether the node is a target node. The task is to find a path from the root to the target node with the smallest score, and the best-first algorithm attempts to find such a path by expanding the intermediate nodes with the smallest scores.

[0065] After obtaining the machine learning pipeline, you can use the hyperparameter optimization tool Ray.tune to optimize the parameters of the machine learning pipeline. It integrates a variety of hyperparameter optimization methods, including random search, grid search, or Bayesian optimization, and evaluates it based on a set of hyperparameter configurations taken from the hyperparameter space; each set of hyperparameter configurations is called a Trail, and Tune supports multiple Trails running in parallel. The configuration can be generated from Tune or obtained from a user-specified search algorithm.

[0066] Given a set of labeled feature data D, the task consists of combining the above algorithms into a pipeline C that takes D as input and generates an optimal predictor ψ = C(D) as output. The objective function of this predictor is as follows:

[0067]

[0068]

[0069] In the formula, A pp and A learn They represent parameterized preprocessing and learning algorithm space respectively, loss(y,ψ(x)) represents the penalty term, x represents historical data, y represents the prediction result, X and Y are the corresponding data sets, and P represents the joint probability distribution.

[0070] Finally, in step S5, the optimal predictor is applied to predict the remaining service life of the equipment, and the predicted data can be measured by the loss function. Generally speaking, underestimating the remaining service life has less negative impact than overestimating the remaining service life, because the latter will lead to failures, and the loss caused by failures is greater than replacing the device prematurely. Therefore, an asymmetric loss function is used to describe the problem. Asymmetric loss function L(D test ) is as follows:

[0071]

[0072] In the formula, N represents the number of devices, a k is the remaining lifetime estimation accuracy of device k.

[0073] After obtaining the prediction results, the mean absolute percentage error can be used as an indicator to evaluate the performance, and a model for predicting the remaining service life of the equipment can be generated. In the automatic modeling process, the training loss curve can be printed for users to monitor its changing rules. The final overall process flowchart is as follows: Figure 4 shown.

[0074] After obtaining the remaining service life prediction results of the equipment, the equipment can be diagnosed, prognosed and health managed based on the results. The specific process flow chart is as follows: Figure 5 shown.

[0075] This embodiment also provides a device for predicting the remaining useful life of laboratory equipment based on automatic machine learning, including a memory and a processor; the memory is used to store computer programs; the processor executes the above-mentioned method for predicting the remaining useful life of laboratory equipment based on automatic machine learning.

[0076] This embodiment further provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by the processor, the remaining useful life prediction method of laboratory equipment based on automatic machine learning as mentioned in this embodiment can be implemented, and any combination of one or more computer-readable media can be used. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, - but not limited to - an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0077] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for predicting the remaining useful life of laboratory equipment based on automatic machine learning, characterized in that: The following steps are involved: S1. Obtain the original timing signal data of the device; S2, extract the timing features of the original timing signal using the TSFRESH algorithm; S3, screening the time series features according to the weight coefficient of the features to obtain a feature subset; S4, using a hierarchical task planning network to match machine learning pipelines for feature subsets; S5. Input the feature subset into the machine learning pipeline to obtain the remaining service life of the corresponding equipment; The S3 is as follows: S31. Use hypothesis testing method to obtain evaluation results of time series characteristics; S32, obtaining a weight coefficient of the feature according to the evaluation result and the fault extraction rate; S33, sorting and screening the features according to the weight coefficients; In step S31, different hypothesis testing methods are used for different time series features: If both features and labels are binary, use the Fisher test; If the feature is binary and the label is continuous, the KS test is used; If the feature is continuous and the label is binary, use the KS test or the Mann-Whitney U test; If both the features and labels are continuous values, the Kendall correlation coefficient test is used; The weight coefficient w of the extracted feature is determined by the calculated reference fault extraction rate FER, and the features are sorted according to w. Whether to retain the feature is determined according to the sorting result. The unretained feature is deleted from the feature set to obtain a feature subset; The planning goal of the hierarchical task planning network is to find a machine learning pipeline that is used to obtain the prediction results of life span. The machine learning pipeline includes a preprocessor and a classifier, wherein the preprocessor is used to perform data preprocessing on a feature subset, and the classifier maps the preprocessed feature subset to a prediction algorithm; After obtaining the machine learning pipeline, use the hyperparameter optimization method to tune the parameters of the machine learning pipeline; The hyperparameter optimization method includes random search, grid search or Bayesian optimization; Use an asymmetric loss function to evaluate the prediction results; The asymmetric loss function L(D test ) is as follows: In the formula, N represents the number of devices, a k is the remaining lifetime estimation accuracy of device k.

2. The method for predicting the remaining useful life of laboratory equipment based on automatic machine learning according to claim 1, characterized in that: The method for obtaining the original timing signal data of the device is as follows: The raw data from the sensors on the acquisition device are extracted, transformed and loaded, and combined into a unified raw time series signal data source that is easy to analyze.

3. The method for predicting the remaining useful life of laboratory equipment based on automatic machine learning according to claim 1, characterized in that: After obtaining the original time series signal, the data is preprocessed using data cleaning, data integration, data transformation and data reduction.

Images