WO2022227240A1 - 一种基于空间域转换独立树的可控中高压光机异常检测方法 - Google Patents

一种基于空间域转换独立树的可控中高压光机异常检测方法 Download PDF

Info

Publication number
WO2022227240A1
WO2022227240A1 PCT/CN2021/100226 CN2021100226W WO2022227240A1 WO 2022227240 A1 WO2022227240 A1 WO 2022227240A1 CN 2021100226 W CN2021100226 W CN 2021100226W WO 2022227240 A1 WO2022227240 A1 WO 2022227240A1
Authority
WO
WIPO (PCT)
Prior art keywords
calender
data
data set
tree
feature
Prior art date
Application number
PCT/CN2021/100226
Other languages
English (en)
French (fr)
Inventor
屈洪春
胡安明
刘光辉
宋传东
曹旨昊
姜振凤
Original Assignee
枣庄学院
山东明源智能装备科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 枣庄学院, 山东明源智能装备科技有限公司 filed Critical 枣庄学院
Publication of WO2022227240A1 publication Critical patent/WO2022227240A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01D—MEASURING NOT SPECIALLY ADAPTED FOR A SPECIFIC VARIABLE; ARRANGEMENTS FOR MEASURING TWO OR MORE VARIABLES NOT COVERED IN A SINGLE OTHER SUBCLASS; TARIFF METERING APPARATUS; MEASURING OR TESTING NOT OTHERWISE PROVIDED FOR
    • G01D21/00—Measuring or testing not otherwise provided for
    • G01D21/02—Measuring two or more variables by means not covered by a single other subclass
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Definitions

  • the invention relates to the field of abnormality detection of calender equipment faults, in particular to a controllable medium and high voltage calender abnormality detection method based on a spatial domain conversion independent tree.
  • Controllable medium and high pressure calender series equipment (including soft roll calender, non-woven fabric hot rolling mill and embossing machine, hereinafter referred to as calender) is not only an important high-end equipment in the paper industry, but also a meltblown cloth (medical mask) Core materials), reverse osmosis membranes (seawater desalination, water purifiers) and other high-value core equipment in current emerging industries, with complex technical structures and high value, are key nodes in the intelligent manufacturing industry chain.
  • calender is not only an important high-end equipment in the paper industry, but also a meltblown cloth (medical mask) Core materials), reverse osmosis membranes (seawater desalination, water purifiers) and other high-value core equipment in current emerging industries, with complex technical structures and high value, are key nodes in the intelligent manufacturing industry chain.
  • the structure of the calender is composed of frame, soft roll, hot roll, scraper, paper feeding system, curved roll, tension roll, end temperature control system of soft roll, soft roll rubber surface temperature control system, transmission system, hydraulic system, It is composed of many subsystems such as hot oil system, stress perception and control system, distance perception and control system, torque balance and control system. Each subsystem has relatively independent or interrelated components, and its operating state monitoring variables are as high as more than 100. , and the structure of variables is different (including discrete, continuous, vibration, audio and video, etc.), the dimensions are not uniform, and the scope is inconsistent. less than 10GB/day), which is a typical representative of industrial big data.
  • the operating state of the calender is related to various factors such as temperature, friction, stress, and material, and involves various components such as temperature control, lubrication, hydraulic pressure, and torque.
  • various components such as temperature control, lubrication, hydraulic pressure, and torque.
  • manpower only relying on manpower to detect a series of mechanical faults that may occur during the operation of the controllable medium and high pressure optical machine has low detection accuracy and low efficiency.
  • there is a relatively complex driving relationship between the operating states of the calender that is, a single fault may lead to a chain effect.
  • controllable medium and high voltage optical machine has many components and high dimension of monitoring variables, and in its entire life cycle, the operating state will change with the load change, material type and component aging. Diffusion diagnostics, predictive maintenance optimization pose serious challenges.
  • an intelligent fault predictive maintenance system for the calender is developed, which can effectively avoid the direct and indirect losses caused by the downtime, and help reduce the cost of equipment operation and maintenance. It is of great significance to improve the value and service life of equipment, promote the long-term improvement of product quality and reduce equipment failure rates, increase the stability of equipment operation and product quality standards, and increase production capacity.
  • the abnormal detection methods for the running state of calender equipment at home and abroad are mainly divided into statistics-based, distance-based, and machine learning-based.
  • Statistics-based anomaly detection methods use the original data to estimate a statistical model to detect anomalies.
  • the model can capture the distribution of the data and evaluate how well the data instance matches the model. If the probability of the data instance generated by the model is very low, then This data instance is defined as an exception.
  • the use of statistical methods to detect anomalies in data is an effective application of standard statistical methods.
  • the advantage of this method is that this detection method can be very effective when the data and inspection type content are sufficient; the disadvantage is that this method is mainly suitable for data of a single attribute, not multi-dimensional attribute data.
  • the distance-based anomaly detection method collects the data of the running state of the calender by deploying sensors at the important mechanical nodes of the calender, and uses the distance similarity (such as Euclidean distance) to measure the data information of each part of the calender at different time nodes. If the data of some time nodes are significantly different from those of other time nodes, it is determined that the mechanical device of this time node is abnormal.
  • the advantage of this method is that it uses distance to measure the similarity between data objects, which is simpler and easier to use; the disadvantage is that the processing efficiency of large-scale high-dimensional data is significantly reduced, because the increase in the size of the data set will lead to the algorithm's inefficiency. The time complexity grows exponentially.
  • Classification-based methods distinguish abnormal data by dividing the data into two categories: normal data and abnormal data.
  • Clustering-based methods usually use a clustering algorithm to divide the dataset into two or more clusters, and then filter out abnormal data points according to the size of each cluster and the data distance within the cluster. The algorithm usually uses nodes and neighborhoods. to detect the spatiotemporal correlation.
  • the methods based on statistics and distance widely used at home and abroad have some shortcomings: the methods based on statistics are based on standard statistical principles. It can detect the abnormality in the data set, and use it to judge the reason of the abnormality. But in fact, in most cases, it is difficult to clarify the distribution law of the data set, and the actual data set often does not fully conform to an ideal mathematical model. Therefore, this method has limitations, especially when the data volume is large and the distribution is complex. In this case, it is extremely difficult to estimate the distribution of the data. In addition, this method relies on the assumption that the normal data of the operating state of the calender falls in the high probability range of the model, and the abnormal data is relatively in the low probability range, and there must be a certain false alarm rate and false alarm rate during detection.
  • the distance-based method is not efficient in processing big data segmentation, and the detection effect is often not as good as other detection methods, so it is usually used as an abnormal point determination strategy and integrated into other detection methods.
  • the above two methods are aimed at local abnormal data detection, and their detection accuracy is low.
  • the technical problem to be solved by the present invention is to provide a controllable medium and high voltage opto-mechanical anomaly detection method based on the spatial domain conversion independent tree, which adopts the kernel function space domain conversion instance algorithm (Kernel Locality-sensitive hashing, KLSH) and the independent tree algorithm ( The method of integrating isolated tree, itree) is applied to the abnormal detection of the running state of the calender, which solves the problem of local abnormal detection of the running state data of the calender, and also avoids the use of statistical and distance-based methods in the running state of the calender.
  • KLSH kernel function space domain conversion instance algorithm
  • KLSH kernel Locality-sensitive hashing
  • the technical solution adopted in the present invention is: a controllable medium and high voltage opto-mechanical anomaly detection method based on a spatial domain conversion independent tree, which is characterized in that: it includes the following steps:
  • kernel function create nt+1 kernel function space domain transformation instances from the calender running state data set X, and arbitrarily select a kernel function space domain transformation instance to perform kernel function on the data set X, and obtain the kernelization pressure
  • kernel function space domain transformation For the optical machine operating state data set X k , use the remaining nt instances of kernel function space domain transformation to perform kernel function on the calender operating state sample subset to obtain nt kernelized calender operating state sample subsets S i , i ⁇ [1,nt];
  • the independent tree algorithm divides the calender running state data set according to the best dividing feature and dividing the sample value, constructs the itree tree, and ends the construction process of the tree in advance according to the early deadline limit height value of the itree tree;
  • step S04 the calender running state data is mapped from the original space to the high-dimensional space using a Gaussian kernel function-based spatial domain transformation instance algorithm, and the formula is:
  • ⁇ (x) is the mapping function, is a random hyperplane, is a centralized kernel matrix, e ⁇ is a vector of S i *1, ⁇ is a random projection, where the value of ⁇ is min(s i /4, 30), k(x, x i ) is the kernel function, the kernel
  • the function uses the Gaussian kernel function formula, and the width of the kernel function is controlled by the adjustable parameter ⁇ .
  • the Gaussian kernel function is expressed by the following formula:
  • step S05 the current height is compared with the early cut-off limit height value, if the conditions are met, the establishment of the tree is completed, otherwise, the characteristic mean value avg i of the cored calender operating state data set X k is calculated, and then calculated.
  • step S06 is specifically:
  • step S07 the specific process of constructing the itree is:
  • step S08 the data is put into the constructed independent tree model, and the average path length and abnormal score of the recorded data points in the tree are:
  • c(n) represents the average path length of data points in the tree
  • H represents the early cutoff height value of the tree
  • n represents the sample size of the sample subset Si
  • H( i ) is a harmonic number
  • H(i) ln (i)+0.5772156649
  • s(x,n) represents the abnormal score of the data point in the tree
  • E(h(x)) represents the average value of the created nt isolated tree sets h(x).
  • the data set X T formed by the sensor monitoring data is:
  • k represents the number of moments in the time interval T
  • q represents the number of state monitoring sensors
  • the data set X T is composed of data collected by the state monitoring sensors at different times
  • the process of deleting invalid data is: selecting the training data set X T
  • the feature representing the temperature of the roll surface calculate its Q1 quartile, and then traverse from the first eigenvalue of the feature until the first value greater than the Q1 quartile is found, stop the traversal, and record the index of the value information index, and finally delete the data of [0, index] in the training data set X T to obtain a new training data set X' T ;
  • ⁇ is the column mean of the data set X T
  • ⁇ is the column mean square error of the data set X T
  • the data set X is obtained after the data set X T is normalized.
  • a torque sensor is installed on the rotary joint at the end of the heat roller to detect the torque of the rotary joint;
  • a vibration sensor is installed on the bearing seats at both ends of the roller body to detect the harmful vibration caused by the aging and damage of the bearing;
  • inside the hydraulic station Install hydraulic station oil temperature sensor, hydraulic station oil pressure sensor and hydraulic station flow sensor to monitor the temperature, pressure and flow of hydraulic oil respectively;
  • a roll surface temperature sensor is installed to monitor the surface temperature of the hot roll and the rubber surface temperature of the soft roll.
  • the present invention provides a controllable medium and high pressure calender anomaly detection system based on a spatial domain conversion independent tree, the system is simple to deploy, low in cost, and solves the problem of local anomaly detection of calender running state data, It also avoids the problem of low detection rate of local anomalies in calender running state data using statistical and distance-based methods. Therefore, the present invention has wide application value. Compared with the prior art, the advantages of the present invention are:
  • the system uses the Gaussian kernel function to map the running state data of the calender into the feature space, and converts local anomalies into global anomalies;
  • the detection accuracy of the local abnormal points in the running state data of the calender is also improved.
  • the invention has broad application prospects.
  • Fig. 1 is the system construction flow schematic diagram of the present invention
  • Fig. 2 is the optimal division characteristic and division sample value selection strategy flow chart
  • Fig. 3 is an architecture diagram of anomaly detection system of controllable calender based on spatial domain transformation independent tree.
  • This embodiment discloses a method for detecting abnormality of a controllable medium and high pressure calender based on an independent tree of spatial domain transformation. As shown in FIG. 1 , the process of the method is: calender running state data collection, historical data modeling, and real-time data detection , Data classification, calender fault maintenance.
  • the method includes the following steps:
  • a torque sensor is installed on the rotary joint at the end of the heat roller to detect the torque of the rotary joint;
  • a vibration sensor is installed on the bearing seats at both ends of the roller body to detect the harmful vibration caused by the aging and damage of the bearing;
  • hydraulic pressure is installed inside the hydraulic station Station oil temperature sensor, hydraulic station oil pressure sensor, hydraulic station flow sensor, respectively monitor the temperature, pressure and flow of hydraulic oil; install embedded line pressure sensor inside the controllable medium and high roll lagging to monitor line pressure;
  • a roller surface temperature sensor is installed to monitor the surface temperature of the hot roller and the rubber surface temperature of the soft roller.
  • the installation of all sensors can increase or decrease the number of sensors and select the type of sensors according to the model of the specific calender equipment and the actual industrial production environment.
  • a data acquisition card is used to process the data collected by the sensor in parallel and transmit it to the computer terminal in a wireless transmission manner.
  • the sampling frequency of the multi-function data acquisition card for each signal is 50kHz, which ensures that the amount of data collected by each sensor in a unit time is equal.
  • select the corresponding data communication mode to directly connect with the multi-function data acquisition card.
  • the multi-function data acquisition card automatically selects the corresponding signal processing channel according to the difference (analog signal/digital signal) of the output signal of each sensor, and finally converts all the signals into digital signals, which is convenient for computer analysis and processing.
  • the given training dataset X T is:
  • k represents the number of moments in the time interval T
  • q represents the number of state monitoring sensors
  • the data set X T is composed of data collected by the state monitoring sensors at different times. That is, x 11 x 12 ... x 1q represents the data collected by the sensor 1 to the sensor q at the first moment, and x k1 x k2 ... x kq represents the data collected by the sensor 1 to the sensor q at the kth moment.
  • the invalid data collected during the operation of the calender mainly comes from the data collected by the roll surface temperature sensor when the calender equipment starts to run.
  • the data collected by the roll surface temperature sensor is not the data information in the normal working environment of the calender machine.
  • the temperature of the roll surface when the equipment starts running is obviously lower.
  • the process of deleting invalid data is: select the feature representing the roll surface temperature in the training data set X T , calculate its Q1 quartile, and then traverse from the first feature value of the feature until the first one greater than Q1 four is found. The value of the quantile, stop traversing, record the index information index of the value, and finally delete the data of [0, index] in the training data set X T to obtain a new training data set X' T .
  • ⁇ is the column mean of the data set X T
  • ⁇ is the column mean square error of the data set X T
  • the data set X is obtained after the data set X T is normalized.
  • sample size of the sample subset Si is the sample size of the dataset X divided by the number of independent trees nt.
  • kernel function create nt+1 kernel function space domain transformation instances from the calender running state data set X, and arbitrarily select a kernel function space domain transformation instance to perform kernel function on the data set X, and obtain the kernelization pressure
  • kernel function space domain transformation For the optical machine operating state data set X k , use the remaining nt instances of kernel function space domain transformation to perform kernel function on the calender operating state sample subset to obtain nt kernelized calender operating state sample subsets S i , i ⁇ [1,nt].
  • an example algorithm of spatial domain transformation based on Gaussian kernel function is used to map the running state data of the calender from the original space to the high-dimensional space, and the formula is:
  • ⁇ (x) is the mapping function, is a random hyperplane, is the centralized kernel matrix, e ⁇ is the vector of S i *1, ⁇ is a random projection, where the value of ⁇ is min(s i /4, 30), k(x, x i ) is the kernel function, the kernel
  • the function uses the Gaussian kernel function formula, and the width of the kernel function is controlled by the adjustable parameter ⁇ .
  • the Gaussian kernel function is expressed by the following formula:
  • the height of the independent tree constructed by the current sample subset is compared with the height value of the early cut-off limit. If the conditions are met, the tree construction is completed, and nt independent trees are constructed from the nt cored sample subsets, and one sample is selected at a time. The subset builds independent trees until nt independent trees are constructed.
  • this step is as follows:
  • the essence of kernelizing the dataset X is to perform high-dimensional space mapping.
  • the mapped dataset X k has d features, and avgi is a d-dimensional feature vector.
  • the independent tree algorithm divides the calender running state data set according to the optimal division feature and the division sample value, constructs an itree tree, and ends the tree construction process in advance according to the early cutoff limit height value of the itree tree.
  • c(n) represents the average path length of data points in the tree
  • H represents the early cutoff height value of the tree
  • n represents the sample size of the sample subset Si
  • H( i ) is a harmonic number
  • H(i) ln (i)+0.5772156649
  • s(x,n) represents the abnormal score of the data point in the tree
  • E(h(x)) represents the average value of the created nt isolated tree sets h(x).
  • the abnormal score is iteratively calculated and compared with the threshold to determine whether the data point is an abnormal point. If the abnormal score is greater than or equal to the threshold, the data point is judged to be abnormal, then the calender equipment should stop working immediately, and the equipment should be inspected and maintained.
  • This method adopts the fusion of Kernel Locality-sensitive hashing (KLSH) and independent tree algorithm (isolated tree, itree), which is applied to the abnormal detection of calender running state, and solves the problem of calender running state.
  • KLSH Kernel Locality-sensitive hashing
  • independent tree algorithm isolated tree, itree
  • the problem of local anomaly detection of state data also avoids the problem of low detection rate of local anomaly data in calender running state data using methods based on statistics and distance. It is of great significance to stabilize the operation of the calender, reduce the loss caused by abnormal shutdown and predictive maintenance.
  • the method uses the Gaussian kernel function space domain transformation method to map the sensory data of each mechanical device during the calender operation process from the original data space to the high-dimensional feature space, and converts the local anomaly problem into a global anomaly problem for processing.
  • the kernelized calender operating state data set is then used for anomaly detection using the independent tree algorithm. Aiming at the problem of how to choose the best dividing feature and dividing the sample value in the independent tree algorithm, the optimal dividing feature and dividing sample value selection strategy are adopted.
  • the method On the basis of maintaining the ability of the calender running state data to detect the global abnormality, the method also improves the detection accuracy of the local abnormal points of the calender running state data set.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Operations Research (AREA)
  • Probability & Statistics with Applications (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Computing Systems (AREA)
  • Testing And Monitoring For Control Systems (AREA)

Abstract

一种基于空间域转换独立树的可控中高压光机异常检测方法，利用高斯核函数将压光机运行状态数据映射至特征空间中，将局部异常转化为全局异常；在此基础上，构建独立树模型对核化后的数据集进行异常检测。对如何选择最佳划分特征和划分样本值的问题，本系统提出一种最佳划分特征和划分样本值选择策略。在保持检测全局异常能力的基础上，也提高了压光机运行状态数据的局部异常点的检测精度。

Description

一种基于空间域转换独立树的可控中高压光机异常检测方法 技术领域
本发明涉及压光机设备故障异常检测领域，具体是一种基于空间域转换独立树的可控中高压光机异常检测方法。
背景技术
可控中高压光机系列装备(包括软辊压光机、非织造布热轧机及压花机，以下简称压光机)不仅是造纸行业的重要高端装备，也是熔喷布(医用口罩的核心材料)、反渗透膜(海水淡化、净水机)等当前新兴产业的高价值核心设备，其技术结构复杂、价值高、是智能制造产业链的关键节点。
压光机的结构由机架、软辊、热辊、刮刀、引纸系统、弧形辊、张力辊、软辊端部温控系统、软辊胶面温控系统、传动系统、液压系统、热油系统、应力感知与控制系统、间距感知与控制系统、扭矩平衡及控制系统等众多子系统构成，每个子系统存在或相对独立或相互关联的部件，其运行状态的监测变量高达百余个，且变量的结构各异(包含离散、连续、振动、音频和视频等)、量纲不统一、范围大小不一致，在24小时不停机的生产过程中将产生海量的监测数据(单台设备不低于10GB/天)，是工业大数据的典型代表。压光机运行状态与温度、摩擦、应力、材料等各种因素有关，涉及温控、润滑、液压、扭矩等各种部件，运维难度高、需要专业的工作人员进行长时间的维护与保养，然而仅依靠人力对可控中高压光机作业过程中可能产生的一系列机械故障进行检测，其检测准确率低、效率低。且压光机运行状态之间存在较为复杂的驱动关系，即单个故障可能导致连锁效应。如果不及时控制部件状态退化，可能导致关键部件损坏，造成设备停机，产生比预测性维护费用高数十倍的维修成本及停机损失，危及整个生产线，对于24小时不停机的现代制造工厂来说，其带来的间接损失更是难以估算。可控中高压光机的部件众多、监测变量维度高，且在其全生命周期中，运行状态会随着负载变化、材料类型和部件老化发生改变，为其运行时异常状态判断，故障发生和扩散诊断、预测性维护优化带来了严峻挑战。
为了能够及时准确地的发现压光机运行时的异常状态，开发针对压光机的智能故障预测维护系统，有效避免故障停机带来的直接和间接损失，并有助于降低设备运维成本，提升装备价值和使用寿命，促进产品质量的长期提升及降低设备故障率，增加设备运行稳定性及产品质量标准稳定，增加产能，具有极大意义。目前，国内外针对压光机设备运行状态的异常检测方法主要分为基于统计、基于距离、基于机器学习等。
基于统计的异常检测方法利用原有数据估计一个统计模型来检测异常，该模型能够捕捉数据的分布，并评估数据实例与模型的匹配程度，如果通过该模型生成的数据实例的概率非常低，则该数据实例被定义为异常。利用统计学方法对数据做异常检测是标准统计学方法的一种有效应用。该方法的优点是，当数据和检验类型内容足够时，此检测方法会非常有效；缺点是，这种方法主要仅适用于单个属性的数据，不适用于多维属性数据。
基于距离的异常检测方法，通过在压光机重要机械节点部署传感器，采集压光机运行状态的数据，利用距离相似性(例如欧式距离)度量不同时间节点压光机各个部位的数据信息，如果某些时间节点数据与其它时间节点相比，具有明显差异，那么判断该时间节点机械装置出现异常。该方法的优点是使用距离来测量数据对象之间的相近度，该方法更简单，更易于使用；缺点是对大规模高维数据的 处理效率明显下降，因为数据集规模的变大会导致算法的时间复杂度呈指数级增长。
基于机器学习的方法大致分为分类、聚类两种。基于分类的方法通过将数据分为正常数据和异常数据两类来区分出异常数据。基于聚类的方法通常采用聚类算法将数据集分为两个或多个聚类簇，然后依据每个簇的大小和簇内的数据距离筛选异常数据点，该算法通常利用节点与邻域的时空相关性进行检测。
但是，目前国内外广泛使用的基于统计和距离的方法存在一些缺点：基于统计的方法依据标准统计学原理，当构建的数学模型符合压光机运行状态数据集的真实变化规律时，就能快速地检测出数据集中存在的异常，并以此来判断出异常出现的原因。但是实际上，大多数情况下都难以明确数据集的分布规律，实际数据集往往也不完全符合某种理想状态的数学模型，因此这种方法存在局限性，尤其当数据量大且分布复杂的情况下，估计数据的分布状况是极其困难的。另外，这种方法依赖于压光机运行状态的正常数据落在模型的高概率区间，异常数据相对处于低概率区间的假设，检测时也必定存在一定的误报率和漏报率。基于距离的方法处理大数据分段时的效率不高，检测效果往往不如其他检测方法，故通常被用作异常点的判定策略融入到其他检测方法中。以上两种方法针对局部异常数据检测，其检测准确率偏低。
发明内容
本发明要解决的技术问题是提供一种基于空间域转换独立树的可控中高压光机异常检测方法，采用核函数空间域转换实例算法(Kernel Locality-sensitive hashing，KLSH)和独立树算法(isolated tree，itree)相融合的方式应用于压光机运行状态异常检测中，解决了压光机运行状态数据的局部异常检测问题，也避免了使用基于统计和距离的方法在压光机运行状态数据中局部异常数据检测率低的问题。
为了解决所述技术问题，本发明采用的技术方案是：一种基于空间域转换独立树的可控中高压光机异常检测方法，其特征在于：包括以下步骤：
S01)、在压光机的重要、易发生故障的部位部署状态监测传感器，传感器监测压光机运行状态下各重要或易发生故障的部位的状态信息，并将采集的状态信息传输至计算机终端，进行数据存储，形成时间序列数据集；
S02)、无效数据删除、数据归一化，将压光机运行状态数据集X T通过第一四分位数法删除无效数据，然后利用z-score方法进行归一化处理，得到新数据集X；
S03)、随机不重复子采样，使用随机不重复子采样技术，对经过预处理后的压光机运行状态数据集X进行子采样，得到nt个压光机运行状态样本子集S i，i∈[1，nt]，nt为构建独立树的数量；
S04)、核函数化，由压光机运行状态数据集X创建nt+1个核函数空间域转换实例，任意选择一个核函数空间域转换实例对数据集X进行核函数化，得到核化压光机运行状态数据集X k，使用剩余的nt个核函数空间域转换实例对压光机运行状态样本子集进行核函数化，得到nt个核化压光机运行状态样本子集S i，i∈[1，nt]；
S05)、计算特征均值，分别计算核化压光机运行状态数据集X k和每一个核化压 光机运行状态样本子集S i的特征均值；
S06)、选择最佳划分特征和划分样本值，比较核化压光机运行状态样本子集S i的特征均值与核化压光机运行状态数据集X k的特征均值，选出均值差最大的特征作为构建独立树算法模型的最佳划分特征，最佳划分特征所对应的子样本数据集S i的特征均值作为划分样本值；
S07)、独立树算法根据最佳划分特征和划分样本值来划分压光机运行状态数据集，构建itree树，并根据itree树的提前截止限制高度值来提前结束树的构建过程；
S08)、将压光机运行状态实时感知数据经归一化、核函数空间域转换实例核函数化后，依次放入构建的itree树中，计算每个数据点的路径长度、异常得分；
S09)、将异常得分与阈值进行比较，决定数据点是否为异常点，如果出现数据异常，压光机停止工作，由设备维护人员进行相应的故障排查和维护工作。
进一步的，步骤S04中，将压光机运行状态数据由原始空间映射到高维空间采用了基于高斯核函数的空间域转换实例算法，公式为：
Figure PCTCN2021100226-appb-000001
其中φ(x)是映射函数，
Figure PCTCN2021100226-appb-000002
是随机超平面，
Figure PCTCN2021100226-appb-000003
是中心化核矩阵，e ζ是S i*1的向量，ζ是随机投影，其中的ζ的取值为min(s i/4，30)，k(x，x i)是核函数，核函数使用高斯核函数公式，由可调参数γ控制核函数的宽度，高斯核函数通过以下公式表示：
K(x i，x j)＝exp(-γ||x i-x j|| 2)     (3)。
进一步的，步骤S05中，将当前高度与提前截止限制高度值比较，若满足条件，则建树完成，否则，计算核化后的压光机运行状态数据集X k的特征均值avg i，再计算nt个核化后的压光机运行状态子样本数据集S i的特征均值sub_avg i；其中，计算提前截止限制高度值的公式如下：H＝2log 2(len(S i))+0.8327    (4)，式中：S i是第i个子样本集。
进一步的，步骤S06具体为：
S61)、计算压光机运行状态数据集X k的特征均值，记为avg i，1≤i≤d；
S62)、计算每个子样本集S i的的特征均值，记为sub_avg i；
S63)、将数据集X k的特征均值与子样本的特征均值S i作差：diff i＝|avg i-sub_avg i|；选出diff i最大值的特征i作为最佳划分特征，sub_avg i作为划分样本值。
进一步的，步骤S07中，构建itree的具体过程为：
S71)、将对应子样本中的第i个特征的所有元素值x ji与sub_avg i比较，j∈len(S i)，S i∈X，当x ji大于sub_avg i时，将第j个样本向量放入右子树，否则，放入左子树；当分完了这个子样本S i后，构建完成了一棵itree，依次循环nt个子样本子集，得到nt棵itree；
S72)、根据树的提前截止高度值来提前结束树的构建过程，循环构建nt棵itree树。
进一步的，步骤S08中，将数据放入构建的独立树模型中，记录数据点在树的平均路径长度和异常得分为：
Figure PCTCN2021100226-appb-000004
Figure PCTCN2021100226-appb-000005
其中c(n)表述数据点在树的平均路径长度，H表示树的提前截止高度值，n表示样本子集S i的样本容量，H(i)是一个调和数，H(i)＝ln(i)+0.5772156649，s(x，n)表示数据点在树的异常得分，E(h(x))表示创建的nt个隔离树集合h(x)的平均值。
进一步的，在时间间隔T内，传感器监测数据构成的数据集X T为：
Figure PCTCN2021100226-appb-000006
其中，k表示时间间隔T内的时刻数，q表示状态监测传感器的数量，数据集X T由状态监测传感器在不同时刻采集的数据构成；删除无效数据的过程为：选择训练数据集X T内代表辊面温度的特征，计算其Q1四分位数，然后从该特征的第一个特征值开始遍历，直到找到第一个大于Q1四分位数的值，停止遍历，记录该值的索引信息index，最后删除训练数据集X T中[0，index]的数据，得到新的训练数据集X′ T；
对训练数据集X′ T中的每个元素经公式(1)归一化：
Figure PCTCN2021100226-appb-000007
其中μ是数据集X T的列均值，σ是数据集X T的列均方差，数据集X T归一化后得到数据集X。
进一步的，步骤S01中，在热辊端部旋转接头安装扭矩传感器，检测旋转接头扭矩；在辊体两端轴承座安装振动传感器，检测轴承的老化与损伤带来的有害震动；在液压站内部安装液压站油温传感器、液压站油压传感器、液压站流量传感器，分别监测液压油的温度、压力和流量；在可控中高辊包胶内部安装嵌入式 线压力传感器，监测线压力；在压光机机架上，安装辊面温度传感器，监测热辊表面温度和软辊胶面温度。
本发明的有益效果：本发明提供一种基于空间域转换独立树的可控中高压光机异常检测系统，该系统部署简单，成本低，解决了压光机运行状态数据的局部异常检测问题，也避免了使用基于统计和距离的方法在压光机运行状态数据中局部异常检测率低的问题。因此本发明具有广泛的应用价值。与现有的技术相比，本发明的优点在于：
该系统利用高斯核函数将压光机运行状态数据映射至特征空间中，将局部异常转化为全局异常；
在此基础上，构建独立树模型对核化后的数据集进行异常检测。对如何选择最佳划分特征和划分样本值的问题，本系统提出一种最佳划分特征和划分样本值选择策略策略。
在保持检测全局异常能力的基础上，也提高了压光机运行状态数据的局部异常点的检测精度。本发明具有广阔的应用前景。
附图说明
图1为本发明的系统构建流程示意图；
图2为最佳划分特征和划分样本值选择策略流程图；
图3为基于空间域转换独立树的可控压光机异常检测系统架构图。
具体实施方式
下面结合附图，对本发明的优选实施例进行详细的描述。
实施例1
本实施例公开一种基于空间域转换独立树的可控中高压光机异常检测方法，如图1所示，本方法流程为：压光机运行状态数据采集、历史数据建模、实时数据检测、数据分类、压光机故障维护。
具体的，本方法包括以下步骤：
S01)、在压光机的重要、易发生故障的部位部署状态监测传感器，传感器监测压光机运行状态下各重要或易发生故障的部位的状态信息，并将采集的状态信息传输至计算机终端，进行数据存储，形成时间序列数据集。
本实施例中，在热辊端部旋转接头安装扭矩传感器，检测旋转接头扭矩；在辊体两端轴承座安装振动传感器，检测轴承的老化与损伤带来的有害震动；在液压站内部安装液压站油温传感器、液压站油压传感器、液压站流量传感器，分别监测液压油的温度、压力和流量；在可控中高辊包胶内部安装嵌入式线压力传感器，监测线压力；在压光机机架上，安装辊面温度传感器，监测热辊表面温度和软辊胶面温度。
所有传感器的安装可根据具体压光机设备的型号以及实际工业生产环境，增加或减少传感器的数量以及选择传感器的类型。
本实施例中，采用数据采集卡对传感器采集的数据进行并行处理并以无线传输的方式传输到计算机终端。多功能数据采集卡对各信号的采样频率均为为50kHz，保证单位时间内各传感器采集的数据量相等。根据各传感器节点的数据传输方式(有线/无线)，选择相应的数据通信方式直接与多功能数据采集卡相连。多功能数据采集卡根据各传感器的输出信号的差异(模拟信号/数字信号)，自动选择对应的信号处理通道，最后将所有信号都转换为数字信号，便于计算机进行分析和处理。
多功能数据采集卡将来自各个传感器节点的数据进行处理并汇总，某一时刻 的数据v t＝(d 1，d 2，...，d q)，其中q表示传感器的数量，d q是一个p维向量，记录某一类型传感器监测的数据信息。经过一个时间间隔T，计算机将收到n组感知数据V＝{v 1，v 2，...，v n}。定义时间段[0，2T/3]内收到的检测特征集为训练数据，其表示为矩阵X T＝{x 1，x 2，...，x k}。
S02)、无效数据删除、数据归一化，将压光机运行状态数据集X T通过第一四分位数法删除无效数据，然后利用z-score方法进行归一化处理，得到新数据集X。
给定训练数据集X T为：
Figure PCTCN2021100226-appb-000008
其中k表示时间间隔T内的时刻数，q表示状态监测传感器的数量，数据集X T由状态监测传感器在不同时刻采集的数据构成。即x 11 x 12 … x 1q表示第一时刻传感器1至传感器q采集的数据，x k1 x k2 … x kq表示第k时刻传感器1至传感器q采集的数据。
压光机作业过程采集的无效数据主要来源于压光机设备开始运行时，辊面温度传感器所采集的数据并不是压光机常规工作环境下的数据信息，设备开始运行的辊面温度明显低于平时压光机辊面工作的温度。由于数据采样频率较高，在时间间隔T采样的数据总量达几十万条，因此可以采取直接删除无效数据的方法，对训练数据进行数据预处理。
删除无效数据的过程为：选择训练数据集X T内代表辊面温度的特征，计算其Q1四分位数，然后从该特征的第一个特征值开始遍历，直到找到第一个大于Q1四分位数的值，停止遍历，记录该值的索引信息index，最后删除训练数据集X T中[0，index]的数据，得到新的训练数据集X′ T。
对训练数据集X′ T中的每个元素经公式(1)归一化：
Figure PCTCN2021100226-appb-000009
其中μ是数据集X T的列均值，σ是数据集X T的列均方差，数据集X T归一化后得到数据集X。
S03)、随机不重复子采样，使用随机不重复子采样技术，对经过预处理后的压光机运行状态数据集X进行子采样，得到nt个压光机运行状态样本子集S i，i∈[1，nt]，nt为构建独立树的数量。
其中样本子集S i的样本容量为数据集X的样本大小除以独立树的数量nt。
S04)、核函数化，由压光机运行状态数据集X创建nt+1个核函数空间域转换实例，任意选择一个核函数空间域转换实例对数据集X进行核函数化，得到核化压光机运行状态数据集X k，使用剩余的nt个核函数空间域转换实例对压光机运行状态样本子集进行核函数化，得到nt个核化压光机运行状态样本子集S i，i∈[1，nt]。
本实施例中，将压光机运行状态数据由原始空间映射到高维空间采用了基于高斯核函数的空间域转换实例算法，公式为：
Figure PCTCN2021100226-appb-000010
其中φ(x)是映射函数，
Figure PCTCN2021100226-appb-000011
是随机超平面，
Figure PCTCN2021100226-appb-000012
是中心化核矩阵，e ζ是S i*1的向量，ζ是随机投影，其中的ζ的取值为min(s i/4，30)，k(x，x i)是核函数，核函数使用高斯核函数公式，由可调参数γ控制核函数的宽度，高斯核函数通过以下公式表示：
K(x i，x j)＝exp(-γ||x i-x j|| 2)      (3)，其中x i表示压光机运行状态数据集X，x j表示压光机运行状态子样本数据集S i。
S05)、计算特征均值，分别计算核化压光机运行状态数据集X k和每一个核化压光机运行状态样本子集S i的特征均值。
本步骤中，将当前样本子集构建的独立树的高度高度与提前截止限制高度值比较，若满足条件，则建树完成，nt个核化样本子集构建nt棵独立树，每一次选择一个样本子集构建独立树，直到nt棵独立树构建完毕。
否则，计算核化后的压光机运行状态数据集X k的特征均值avg i，再计算nt个核化后的压光机运行状态子样本数据集S i的特征均值sub_avg i；
其中，计算提前截止限制高度值的公式如下：H＝2log 2(len(S i))+0.8327    (4)，式中：S i是第i个子样本集，Len(S i)表示的是求第i个子样本集的样本容量。
S06)、选择最佳划分特征和划分样本值，比较核化压光机运行状态样本子集S i的特征均值与核化压光机运行状态数据集X k的特征均值，选出均值差最大的特征作为构建独立树算法模型的最佳划分特征，最佳划分特征所对应的子样本数据集S i的特征均值作为划分样本值。
如图2所示，本步骤具体为：
S61)、计算压光机运行状态数据集X k的特征均值，记为：avg i，(1≤i≤d)；
对数据集X进行核化的本质是进行高维空间映射，映射后的数据集X k有d个特征，avg i是一个d维特征向量。
S62)、计算每个子样本集S i的特征均值，记为：sub_avg i；
S63)、将数据集X k的特征均值与子样本的特征均值S i作差，记为：diff i＝|avg i-sub_avg i|；选出diff i最大值的特征i作为最佳划分特征，sub_avg i作为划分样本值，特征i为S k中的特征。
S07)、独立树算法根据最佳划分特征和划分样本值来划分压光机运行状态数据集，构建itree树，并根据itree树的提前截止限制高度值来提前结束树的构建过程。
本步骤具体为：
S71)、将对应子样本中的第i个特征的所有元素值x ji与sub_avg i比较，j∈len(S i)，S i∈X，当x ji大于sub_avg i时，将第j个样本向量放入右子树，否则，放入左子树；当分完了这个子样本S i后，构建完成了一棵itree，依次循环nt个子样本子集，得到nt棵itree；
S72)、根据树的提前截止高度值来提前结束树的构建过程，循环构建nt棵itree树。
S08)、将压光机运行状态实时感知数据经归一化、核函数空间域转换实例核函数化后，依次放入构建的itree树中，计算每个数据点的路径长度、异常得分。
其中计算路径长度和异常得分的公式为：
Figure PCTCN2021100226-appb-000013
Figure PCTCN2021100226-appb-000014
其中c(n)表述数据点在树的平均路径长度，H表示树的提前截止高度值，n表示样本子集S i的样本容量，H(i)是一个调和数，H(i)＝ln(i)+0.5772156649，s(x，n)表示数据点在树的异常得分，E(h(x))表示创建的nt个隔离树集合h(x)的平均值。
S09)、将异常得分与阈值进行比较，决定数据点是否为异常点，如果出现数据异常，压光机停止工作，由设备维护人员进行相应的故障排查和维护工作。
本实施例中，将异常得分与阈值进行迭代计算比较，决定数据点是否为异常点。若异常得分大于或者等于阈值，数据点被判断为异常点，那么压光机设备应该立即停止工作，对设备进行检查和维护。
本方法采用核函数空间域转换实例算法(Kernel Locality-sensitive hashing， KLSH)和独立树算法(isolated tree，itree)相融合的方式应用于压光机运行状态异常检测中，解决了压光机运行状态数据的局部异常检测问题，也避免了使用基于统计和距离的方法在压光机运行状态数据中局部异常数据检测率低的问题。为压光机稳定运行、降低异常停机带来的损失和预测性维护具有极其重要的意义。
该方法利用高斯核函数空间域转换方法将传感器采集的压光机作业过程中各机械装置的感知数据由原始数据空间映射至高维特征空间中，将局部异常问题转化为全局异常问题后进行处理；核化后的压光机运行状态数据集再使用独立树算法来进行异常检测。针对独立树算法如何选择最佳划分特征和划分样本值的问题，采取了最佳划分特征和划分样本值选择策略。在保持压光机运行状态数据检测全局异常能力的基础上，本方法还提高压光机运行状态数据集的局部异常点的检测精度。
以上描述的仅是本发明的基本原理和优选实施例，本领域技术人员根据本发明做出的改进和替换，属于本发明的保护范围。

Claims (8)

  1. 一种基于空间域转换独立树的可控中高压光机异常检测方法，其特征在于：包括以下步骤：
    S01)、在压光机的重要、易发生故障的部位部署状态监测传感器，传感器监测压光机运行状态下各重要或易发生故障的部位的状态信息，并将采集的状态信息传输至计算机终端，进行数据存储，形成时间序列数据集；
    S02)、无效数据删除、数据归一化，将压光机运行状态数据集X T通过第一四分位数法删除无效数据，然后利用z-score方法进行归一化处理，得到新数据集X；
    S03)、随机不重复子采样，使用随机不重复子采样技术，对经过预处理后的压光机运行状态数据集X进行子采样，得到nt个压光机运行状态样本子集S i，i∈[1，nt]，nt为构建独立树的数量；
    S04)、核函数化，由压光机运行状态数据集X创建nt+1个核函数空间域转换实例，任意选择一个核函数空间域转换实例对数据集X进行核函数化，得到核化压光机运行状态数据集X k，使用剩余的nt个核函数空间域转换实例对压光机运行状态样本子集进行核函数化，得到nt个核化压光机运行状态样本子集S i，i∈[1，nt]；
    S05)、计算特征均值，分别计算核化压光机运行状态数据集X k和每一个核化压光机运行状态样本子集S i的特征均值；
    S06)、选择最佳划分特征和划分样本值，比较核化压光机运行状态样本子集S i的特征均值与核化压光机运行状态数据集X k的特征均值，选出均值差最大的特征作为构建独立树算法模型的最佳划分特征，最佳划分特征所对应的子样本数据集S i的特征均值作为划分样本值；
    S07)、独立树算法根据最佳划分特征和划分样本值来划分压光机运行状态数据集，构建itree树，并根据itree树的提前截止限制高度值来提前结束树的构建过程；
    S08)、将压光机运行状态实时感知数据经归一化、核函数空间域转换实例核函数化后，依次放入构建的itree树中，计算每个数据点的路径长度、异常得分；
    S09)、将异常得分与阈值进行比较，决定数据点是否为异常点，如果出现数据异常，压光机停止工作，由设备维护人员进行相应的故障排查和维护工作。
  2. 根据权利要求1所述的基于空间域转换独立树的可控中高压光机异常检测方法，其特征在于：步骤S04中，将压光机运行状态数据由原始空间映射到高维空间采用了基于高斯核函数的空间域转换实例算法，公式为：
    Figure PCTCN2021100226-appb-100001
    其中φ(x)是映射函数，
    Figure PCTCN2021100226-appb-100002
    是随机超平面，
    Figure PCTCN2021100226-appb-100003
    是中心化核矩阵，e ζ是S i*1的向量，ζ是随机投影，其中的ζ的取值为min(s i/4，30)，k(x，x i)是核函数，核函数使用高斯核函数 公式，由可调参数γ控制核函数的宽度，高斯核函数通过以下公式表示：
    K(x i，x j)＝exp(-γ||x i-x j|| 2)  (3)。
  3. 根据权利要求1所述的基于空间域转换独立树的可控中高压光机异常检测方法，其特征在于：步骤S05中，将当前高度与提前截止限制高度值比较，若满足条件，则建树完成，否则，计算核化后的压光机运行状态数据集X k的特征均值avg i，再计算nt个核化后的压光机运行状态子样本数据集S i的特征均值sub_avg i；
    其中，计算提前截止限制高度值的公式如下：H＝2log 2(len(S i))+0.8327  (4)，式中：S i是第i个子样本集。
  4. 根据权利要求1所述的基于空间域转换独立树的可控中高压光机异常检测方法，其特征在于：步骤S06具体为：
    S61)、计算压光机运行状态数据集Xk的特征均值，记为avg i，1≤i≤d；
    S62)、计算每个子样本集S i的的特征均值，记为sub_avg i；
    S63)、将数据集X k的特征均值与子样本的特征均值S i作差：diff i＝|avg i-sub_avg i|；选出diff i最大值的特征i作为最佳划分特征，sub_avg i作为划分样本值。
  5. 根据权利要求4所述的基于空间域转换独立树的可控中高压光机异常检测方法，其特征在于：步骤S07中，构建itree的具体过程为：
    S71)、将对应子样本中的第i个特征的所有元素值x ji与sub_avg i比较，j∈len(S i)，S i∈X，当x ji大于sub_avg i时，将第j个样本向量放入右子树，否则，放入左子树；当分完了这个子样本S i后，构建完成了一棵itree，依次循环nt个子样本子集，得到nt棵itree；
    S72)、根据树的提前截止高度值来提前结束树的构建过程，循环构建nt棵itree树。
  6. [根据细则91更正 27.07.2021]　
    根据权利要求3所述的基于空间域转换独立树的可控中高压光机异常检测方法，其特征在于：步骤S08中，将数据放入构建的独立树模型中，记录数据点在树的平均路径长度和异常得分为：
    Figure PCTCN2021100226-appb-100004

    Figure PCTCN2021100226-appb-100005
    其中c(n)表述数据点在树的平均路径长度，H表示树的提前截止高度值，n表示样本子集S i的样本容量，H(i)是一个调和数，H(i)＝ln(i)+ 0.5772156649，s(x，n)表示数据点在树的异常得分，E(h(x))表示创建的nt个隔离树集合h(x)的平均值。
  7. [根据细则91更正 27.07.2021]　
    根据权利要求1所述的基于空间域转换独立树的可控中高压光机异常检测方法，其特征在于：在时间间隔T内，传感器监测数据构成的数据集X T为：
    Figure PCTCN2021100226-appb-100006
    其中，k表示时间间隔T内的时刻数，q表示状态监测传感器的数量，数据集X T由状态监测传感器在不同时刻采集的数据构成；
    删除无效数据的过程为：选择训练数据集X T内代表辊面温度的特征，计算其Q1四分位数，然后从该特征的第一个特征值开始遍历，直到找到第一个大于Q1四分位数的值，停止遍历，记录该值的索引信息index，最后删除训练数据集X T中[0，index]的数据，得到新的训练数据集X′ T；
    对训练数据集X′ T中的每个元素经公式(1)归一化：
    Figure PCTCN2021100226-appb-100007
    其中μ是数据集X T的列均值，σ是数据集X T的列均方差，数据集X T归一化后得到数据集X。
  8. [根据细则91更正 27.07.2021]　
    根据权利要求1所述的基于空间域转换独立树的可控中高压光机异常检测方法，其特征在于：步骤S01中，在热辊端部旋转接头安装扭矩传感器，检测旋转接头扭矩；在辊体两端轴承座安装振动传感器，检测轴承的老化与损伤带来的有害震动；在液压站内部安装液压站油温传感器、液压站油压传感器、液压站流量传感器，分别监测液压油的温度、压力和流量；在可控中高辊包胶内部安装嵌入式线压力传感器，监测线压力；在压光机机架上，安装辊面温度传感器，监测热辊表面温度和软辊胶面温度。
PCT/CN2021/100226 2021-04-28 2021-06-16 一种基于空间域转换独立树的可控中高压光机异常检测方法 WO2022227240A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110469762.6 2021-04-28
CN202110469762.6A CN113255731A (zh) 2021-04-28 2021-04-28 基于空间域转换独立树的可控中高压光机异常检测方法

Publications (1)

Publication Number Publication Date
WO2022227240A1 true WO2022227240A1 (zh) 2022-11-03

Family

ID=77222151

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/100226 WO2022227240A1 (zh) 2021-04-28 2021-06-16 一种基于空间域转换独立树的可控中高压光机异常检测方法

Country Status (2)

Country Link
CN (1) CN113255731A (zh)
WO (1) WO2022227240A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109506963A (zh) * 2018-11-29 2019-03-22 中南大学 一种智能列车牵引故障大数据异常检测辨识方法
WO2019129977A1 (fr) * 2017-12-28 2019-07-04 Worldline Detection d'anomalies par une approche combinant apprentissage supervise et non-supervise
CN112070125A (zh) * 2020-08-19 2020-12-11 西安理工大学 一种基于孤立森林学习的不平衡数据集的预测方法
CN112631155A (zh) * 2020-11-06 2021-04-09 杭州禾控科技有限公司 一种中高可控压光机控制系统

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107292350A (zh) * 2017-08-04 2017-10-24 电子科技大学 大规模数据的异常检测方法
CN111207938B (zh) * 2020-01-13 2020-12-04 西南交通大学 一种铁道车辆故障检测方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019129977A1 (fr) * 2017-12-28 2019-07-04 Worldline Detection d'anomalies par une approche combinant apprentissage supervise et non-supervise
CN109506963A (zh) * 2018-11-29 2019-03-22 中南大学 一种智能列车牵引故障大数据异常检测辨识方法
CN112070125A (zh) * 2020-08-19 2020-12-11 西安理工大学 一种基于孤立森林学习的不平衡数据集的预测方法
CN112631155A (zh) * 2020-11-06 2021-04-09 杭州禾控科技有限公司 一种中高可控压光机控制系统

Also Published As

Publication number Publication date
CN113255731A (zh) 2021-08-13

Similar Documents

Publication Publication Date Title
CN110109445B (zh) 一种船舶机舱辅机监控系统及监控方法
CN110366031B (zh) 一种数字化车间mes系统基于视觉的异常状态监测与故障诊断方法
CN111555716B (zh) 光伏阵列工作状态的确定方法、装置、设备及存储介质
CN206322012U (zh) 一种用于选煤厂的mes系统
CN110362608A (zh) 基于雨流计数法和局部异常因子的能耗异常检测方法
CN111177276A (zh) 一种基于Spark计算框架的动能数据处理系统及方法
CN116189407B (zh) 一种基于数据监测的智能预警系统
CN113569459A (zh) 基于群体智能的液压支架性能退化定量评估方法及系统
CN111401661B (zh) 一种机械设备的预测性维护方法和维护系统
CN107490769A (zh) 一种燃料电池水分布可视化在线检测系统
WO2022227240A1 (zh) 一种基于空间域转换独立树的可控中高压光机异常检测方法
CN104793270A (zh) 一种综合气象观测网中设备运行状态的综合判定方法
CN116168019B (zh) 基于机器视觉技术的电网故障检测方法及系统
CN116951328B (zh) 基于大数据的排水管线运行智慧监控系统
CN114170138A (zh) 无监督工业图像异常检测模型建立方法、检测方法及系统
CN117113135A (zh) 一种可对异常数据整理分类的碳排放异常监测分析系统
CN111733498A (zh) 一种智能纺纱细纱故障检测系统
CN110794799A (zh) 应用于工业生产的具有故障诊断功能的大数据系统
CN116821811A (zh) 基于多层图卷积神经网络的磨煤机组故障诊断方法及系统
CN116466166A (zh) 一种大屏设备的故障检测方法、系统及装置
CN116365705A (zh) 一种光伏组件故障实时监测方法及系统
CN114265377B (zh) 纸机抄造系统的集成优化系统
CN112269778B (zh) 一种设备故障诊断方法
CN115392663A (zh) 一种基于大数据的数据采集处理方法
KR102360004B1 (ko) 진동 신호 기반의 기계 상태 관리 시스템

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21938709

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21938709

Country of ref document: EP

Kind code of ref document: A1