CN114238854A - Anomaly detection method in mining scene based on graph regular incremental non-negative matrix factorization - Google Patents

Anomaly detection method in mining scene based on graph regular incremental non-negative matrix factorization Download PDF

Info

Publication number
CN114238854A
CN114238854A CN202111509423.2A CN202111509423A CN114238854A CN 114238854 A CN114238854 A CN 114238854A CN 202111509423 A CN202111509423 A CN 202111509423A CN 114238854 A CN114238854 A CN 114238854A
Authority
CN
China
Prior art keywords
matrix
data
sample
sets
decomposition
Prior art date
Legal status (The legal status 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 status listed.)
Granted
Application number
CN202111509423.2A
Other languages
Chinese (zh)
Other versions
CN114238854B (en
Inventor
陈自刚
肖琪
陈龙
张镇江
潘鼎
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University of Post and Telecommunications
Original Assignee
Chongqing University of Post and Telecommunications
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 Chongqing University of Post and Telecommunications filed Critical Chongqing University of Post and Telecommunications
Priority to CN202111509423.2A priority Critical patent/CN114238854B/en
Publication of CN114238854A publication Critical patent/CN114238854A/en
Application granted granted Critical
Publication of CN114238854B publication Critical patent/CN114238854B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • 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
    • 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
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G08—SIGNALLING
    • G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
    • G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
    • G08B21/18—Status alarms
    • G08B21/182—Level alarms, e.g. alarms responsive to variables exceeding a threshold

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Algebra (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Computing Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Emergency Management (AREA)
  • Business, Economics & Management (AREA)
  • Operations Research (AREA)
  • Probability & Statistics with Applications (AREA)
  • Testing Or Calibration Of Command Recording Devices (AREA)

Abstract

The invention provides a mining scene abnormity detection method based on graph regular increment nonnegative matrix decomposition, and belongs to the technical field of abnormity detection and diagnosis and the field of intelligent safety. The method comprises the steps that firstly, mining environment information under a normal state is collected by two sets of equipment, and data obtained by the two sets of equipment are processed as follows; preprocessing data to obtain a training set X'; then obtaining the optimal base matrix W through regular incremental nonnegative matrix decomposition of the graphnewSum coefficient matrix Hnew(ii) a Thereby establishing a monitoring statistic N2And SPE, calculating the control of the training sets of the two sets of equipmentLimiting; then, data (a test set X ") is collected again for detection, the statistic of the test set X" is calculated, and finally the statistic is compared with two sets of control limits, so that whether the mining scene is abnormal or not is judged; when the scene is abnormal, the maximum or larger contribution values are uploaded to the control interface as the abnormal reasons to be displayed. The method solves the problems that the traditional mining scene abnormity detection is not timely and accurate, and the like, and creates digital mining industry.

Description

Mining scene abnormity detection method based on graph regular increment nonnegative matrix decomposition
Technical Field
The invention relates to the technical field of anomaly detection and diagnosis and the field of intelligent safety, in particular to a mining scene anomaly detection method based on graph regular increment non-negative matrix decomposition.
Background
Mining safety issues have been a concern. At present, most mines, especially metal mines, adopt an underground mining mode, however, in the underground construction process, toxic gas, mine collapse and the like pose great threats to the safety of constructors; therefore, the method has important significance in risk detection and early warning of the underground construction environment. At present, a plurality of anomaly detection methods are provided, but the method is challenging work for ensuring timeliness and effective utilization of network resources, comprehensively analyzing multi-sensor (multi-factor) data to ensure accuracy and determining the root cause of scene anomaly.
A mining scene anomaly detection method based on graph regular increment nonnegative matrix decomposition is provided. The graph regular incremental nonnegative matrix decomposition can overcome the difficulty of the traditional nonnegative matrix decomposition in the online processing of a large data set, can online represent data content and maintain the geometric structure of data, and can also fully utilize the decomposition result of the previous step by combining incremental learning to avoid repeated calculation, thereby reducing the operation time; meanwhile, the dimension is obviously reduced, and the clustering precision is better; based on the operation of the point elements by the matrix points, the decomposition result of the graph regular increment non-negative matrix method is more likely to represent the local characteristics of the data; in addition, the nonnegative basis vectors obtained by the regular incremental nonnegative matrix decomposition of the graph have certain linear independence and sparsity, and large-scale high-dimensional data generated in an industrial process can be better described; finally, the regular incremental non-negative matrix method of the graph does not make a special ideal assumption on data distribution in the decomposition process, so that non-Gaussian data can be processed, and only a proper monitoring statistic needs to be designed by combining a density estimation method.
Disclosure of Invention
In view of the above, the invention provides a mining scene abnormity detection method based on graph regular increment non-negative matrix decomposition, which realizes dynamic data training and multi-sensor data comprehensive analysis, saves time and space cost, and can analyze the reason of abnormity if abnormity occurs. The method comprises the following steps:
step 1: collecting data; sampling environments at a plurality of normal working moments by using two sets of equipment; the data obtained by the two sets of equipment are processed in the same way as follows, and the operation performed by one set of equipment data will be described below, and will not be described repeatedly hereinafter.
Step 2: preprocessing data; the measured data in the industrial process do not necessarily satisfy the non-negative condition, for example, the readings of the sensors such as temperature, pressure, etc. may be negative, and the unit can be adjusted to make the values non-negative; then preprocessing the data collected in the step 1 such as graying, vectorization, normalization and the like to obtain a training set X', and initially training a sample matrix
Figure BDA0003404694610000021
The matrix (k samples in the matrix) is composed of data samples in a certain time period in the training set X ', and the data samples at the later moment in the training set X' are sequentially used as the next new sample.
Initial training matrix
Figure BDA0003404694610000022
Wherein t is0≤ti≤…≤ti+k-1≤t1Is an initial period t0～t1Collected data;
Figure BDA0003404694610000023
represents data acquired by the ith device in the jth sample at tiCollected all the time; for example, i-1 indicates that the device is a camera, i-2 indicates that the device is a gas sensor,
Figure BDA0003404694610000024
data representing the gas sensor in a third sample at ti+2Time samplingCollected;
Figure BDA0003404694610000025
represents the first data sample, is at tiCollected all the time;
Figure BDA0003404694610000026
represents the (k + 1) th data sample, is at t2The samples are collected at the moment and are also used as the next newly added sample, and the rest is analogized by the newly added sample.
And step 3: a training stage; the initial training sample matrix obtained in the step 2
Figure BDA0003404694610000027
Carrying out graph regular increment non-negative matrix decomposition on the newly added samples to obtain an optimal base matrix WnewSum coefficient matrix Hnew(ii) a Therefore, the geometrical structure information of the sample can be kept in a low-dimensional space, the decomposition result of the previous step can be fully utilized by combining incremental learning, repeated calculation is avoided, and the operation time is reduced. The method comprises the following specific steps:
step 3.1: firstly, an initial training sample matrix is
Figure BDA0003404694610000028
SVD is carried out to obtain a singular value matrix sigma and a singular vector matrix U, V, and the singular value matrix and the singular vector matrix are respectively used for pairing
Figure BDA0003404694610000029
Initializing a base matrix and a coefficient matrix in the regular non-negative matrix decomposition of the graph, and updating and iterating the initialized base matrix and coefficient matrix until a target function tends to be stable to obtain Wk,Hk(ii) a Therefore, a better global optimal solution effect can be obtained, the input matrix does not need to be changed in any data structure, the data structure of the original data cannot be damaged, more detailed information can be reserved, and the decomposition effect is improved.
SVD decomposition formula:
Figure BDA0003404694610000031
Wk,Hkinitialization:
Figure BDA0003404694610000032
where | U | represents taking the absolute value of the matrix U, VΤRepresenting a transpose of the matrix V.
An objective function:
Figure BDA0003404694610000033
an iteration rule:
Figure BDA0003404694610000034
wherein R represents a weight matrix and D is a diagonal matrix
Figure BDA0003404694610000035
LkIs a Laplace matrix (L)kD-R), λ is a regularization parameter.
Step 3.2: when a new sample is added at the next moment, the optimal base matrix W is obtained by adopting the regular incremental nonnegative matrix decomposition of the graphnewSum coefficient matrix Hnew；
For example when
Figure BDA0003404694610000036
The objective function at the time of addition was:
Figure BDA0003404694610000037
iterative formula
Figure BDA0003404694610000038
Step 3.3: repeating the above operations on all newly added samples, and obtaining the optimal base matrix W after the updating is finishednewSum coefficient matrix Hnew。
And 4, step 4: calculating a control limit; w from step 3new、HnewCalculating a monitoring statistic N2And SPE, N2In order to monitor the change of the characteristic space, SPE is used for monitoring the change of the residual error space; then, probability density estimation is carried out on the process data by adopting a kernel density estimation (KED) method, and actual distribution information of the data is extracted, so that statistic control limits corresponding to two sets of equipment training samples are determined
Figure BDA0003404694610000041
SPE1'；
Figure BDA0003404694610000042
SPE'2。
And 5: a testing stage; the data is collected again (as a test set X ') for detection, the same processing is carried out on the test set X' at S2 and S3, corresponding statistic is obtained, and the monitoring statistic is compared with two sets of control limits, wherein the following three conditions can be adopted:
the first condition is as follows: when the two statistics are within the two sets of control limits, the scene is normal, and mining can be carried out.
Case two: when any one or two statistics are outside the two sets of control limits, the situation is necessarily abnormal, a first-level alarm is immediately carried out, and mining cannot be carried out; and calculating and sequencing the contribution values, and uploading the largest or larger contribution values to a control interface as the abnormality reasons for displaying.
Case three: and when any one or two statistics are only outside one set of control limit, further confirming whether the equipment is in failure or not, if not, immediately carrying out secondary alarm, calculating contribution values and sequencing, and uploading the largest or larger contribution values to a control interface as abnormal reasons for display.
The invention has the beneficial effects that: the invention avoids the defects of the traditional detection, can carry out dynamic training, real-time tracking and prediction and ensures the timeliness; various factors such as toxic gas, water burst, mine collapse and the like can be comprehensively considered, and the data of the plurality of sensors are comprehensively analyzed to ensure the accuracy; if the scene is abnormal, the source of the scene abnormality can be found; and effective utilization of network resources can be ensured, and efficiency is improved.
The invention has the beneficial effects that: the method for carrying out abnormity detection by using the graph regular increment non-negative matrix decomposition can overcome the difficulty of the traditional non-negative matrix decomposition in the online processing of a large data set, can online represent the data content and maintain the geometric structure of the data, and simultaneously obviously reduces the dimension, reduces the operation time and has better clustering precision; based on the operation of the point elements by the matrix points, the decomposition result of the graph regular increment non-negative matrix method is more likely to represent the local characteristics of the data; meanwhile, the nonnegative basis vectors obtained by the regular incremental nonnegative matrix decomposition of the graph have certain linear independence and sparsity, and large-scale high-dimensional data generated in an industrial process can be better described; in addition, the graph regular increment non-negative matrix method does not make a special ideal assumption on data distribution in the decomposition process, so that non-Gaussian data can be processed, and only a proper monitoring statistic needs to be designed by combining a density estimation method.
Drawings
Fig. 1 is a general flowchart of a mining scene anomaly detection method based on graph regular incremental nonnegative matrix factorization according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of mining environment data collection provided by an embodiment of the present invention;
FIG. 3 is a schematic diagram of data structure provided by an embodiment of the present invention;
FIG. 4 is a specific flowchart of regular incremental non-negative matrix decomposition of a graph according to an embodiment of the present invention;
Detailed Description
The invention will be further described with reference to the accompanying drawings and specific embodiments.
The invention aims to provide a mining scene abnormity detection method based on graph regular increment nonnegative matrix decomposition, which can track and predict the safety of a mining scene in real time, achieve the effect of reducing the safety risk of mining personnel, feed back the reason of abnormity when the environment is abnormal, ensure the effective utilization of network resources and improve the efficiency.
As shown in fig. 1, a general flowchart of a mining scene anomaly detection method based on graph regular incremental non-negative matrix factorization is provided in the embodiment of the present invention. The method comprises the following steps:
step 1: collecting data; two sets of equipment are used (each set of equipment comprises a camera and a plurality of sensors (a gas sensor and a CO sensor)2Sensors, etc.) in a single quantity) samples the environment at multiple times of normal operation.
Fig. 2 is a schematic diagram of mining environment data acquisition. And data acquired by the two sets of equipment are transmitted to the nodes and then uploaded to the cloud server for processing. The equipment is not limited to the above three types and may be tailored to the particular mining environment.
The data obtained by the two sets of equipment are respectively processed in the following way, and the operation performed by the data collected by the set of equipment will be described below, and is respectively described if necessary, and the repeated description is not repeated hereinafter.
Step 2: preprocessing data; carrying out graying, vectorization, normalization and other processing on the data acquired in the step 1 to obtain a training set X', and obtaining an initial training sample matrix
Figure BDA0003404694610000051
The matrix (k samples in the matrix) is composed of data samples in a certain time period in the training set X ', and the data samples at the later moment in the training set X' are sequentially used as the next new sample.
Fig. 3 is a schematic diagram of data composition. Preprocessing the acquired data, graying the image data acquired by the camera to obtain an image matrix V belonging to a multiplied by b.
Vectorizing the obtained image matrix V, first extracting each column and recombining into a column vector as shown in FIG. 3
Figure BDA0003404694610000061
Then, the data collected by other sensors in a corresponding set of equipment are combined into a column vector in sequence as shown in FIG. 3
Figure BDA0003404694610000062
Finally, the set of equipment is put at tiThe data collected at the moment are normalized to be between 0 and 1 and combined into a column vector
Figure BDA0003404694610000063
Indicating the first sample data, is at tiAcquired at a moment of time, wherein
Figure BDA0003404694610000064
Data representing the second plant gas sensor in the first sample, also at tiAnd (4) acquiring at any moment.
Continuously collecting data at multiple moments, performing repeated processing according to the above manner, and finally obtaining an initial training sample matrix
Figure BDA0003404694610000065
The data collected at different later times are sequentially used as the added samples.
In this embodiment, 200 groups of normal data samples are collected within 5 hours as samples in an initial training sample matrix, 40 groups are collected every hour, 10 groups are collected every hour after 5 hours as added samples, and 100 groups are added as newly added samples.
And step 3: a training stage; the initial training sample matrix obtained in the step 2
Figure BDA0003404694610000066
Carrying out graph regular increment non-negative matrix decomposition on the newly added samples to obtain an optimal base matrix WnewSum coefficient matrix Hnew。
As shown in fig. 4, a specific process of regular incremental non-negative matrix factorization is provided for the graph according to the embodiment of the present invention. Firstly, an initial training sample matrix is
Figure BDA0003404694610000067
SVD is performed to obtain a singular value matrix sigma and a singular vector matrix U, V.
SVD decomposition:
Figure BDA0003404694610000068
using singular value matrix and singular vector matrix respectively
Figure BDA0003404694610000069
And initializing a base matrix and a coefficient matrix in the regular non-negative matrix decomposition of the graph.
Wk,HkInitialization:
Figure BDA00034046946100000610
where | U | represents taking the absolute value of the matrix U, VΤIs the transpose of matrix V.
Updating and iterating the initialized base matrix and coefficient matrix until the objective function tends to be stable to obtain Wk,Hk。
An objective function:
Figure BDA0003404694610000071
an iteration rule:
Figure BDA0003404694610000072
wherein R represents a weight matrix and D is a diagonal matrix
Figure BDA0003404694610000073
LkIs a Laplace matrix (L)kD-R), λ is a regularization parameter.
Sequentially increasing samples
Figure BDA0003404694610000074
Performing graph regular increment non-negative matrix decomposition once every adding one sample, and iteratively updating until the target function meets the convergence condition to obtain the optimal base matrix and coefficient matrix Wnew,Hnew。
For example when
Figure BDA0003404694610000075
The objective function at the time of addition was:
Figure BDA0003404694610000076
Wk+1and Hk+1Respectively representing sample sets Xk+1Base matrix and coefficient matrix L obtained by performing regular nonnegative matrix decomposition on graphk+1Is a sample set Xk+1The laplacian matrix of. When the number of training samples is large enough, the influence of adding a new training sample on the base matrix and the coefficient matrix is small, so that it is assumed that the coefficient matrix H is added when a new sample is addedk+1Is approximately equal to HkColumn vector of (i.e. H)k+1＝[Hk,hk+1]At this time, the objective function Fk+1Rewritable in the form:
Figure BDA0003404694610000077
obtain an objective function Fk+1After the incremental expression is obtained, a corresponding iterative updating formula can be deduced by using a gradient descent method:
Figure BDA0003404694610000081
repeating the above operations on all newly added samples, and obtaining the optimal base matrix W after the updating is finishednewSum coefficient matrix Hnew。
And 4, step 4: calculating a control limit; obtaining corresponding W from two sets of training setsnew、HnewCalculating a corresponding monitoring statistic N2And SPE.
Statistic N for monitoring feature spatial variation2：N2(i)＝XΤ(i)WWΤX(i)。
Statistic SPE for degree of deviation of reaction data:
Figure BDA0003404694610000082
wherein
Figure BDA0003404694610000083
The reconstructed value representing the ith sample vector is calculated as:
Figure BDA0003404694610000084
the control limits for two statistics are calculated: probability density estimation is carried out on the two statistics by using a nuclear density estimation method, the actual distribution condition of the statistics is extracted, and the control limits of the statistics of the two sets of equipment training samples are respectively calculated by setting the significance level alpha
Figure BDA0003404694610000085
SPE1'；
Figure BDA0003404694610000086
SPE2'。
And 5: a testing stage; the data is collected again (as a test set X ') for detection, the same processing as the step 2 and the step 3 is carried out on the test set X', and the corresponding statistic N is obtained2And SPE, comparing the monitoring statistic with two sets of control limits.
When in use
Figure BDA0003404694610000087
When the two statistics are within the control limit trained by the two equipment training sets, the scene is normal, and mining can be performed.
When in use
Figure BDA0003404694610000088
When any one or two statistics are beyond the control limit trained by the two sets of equipment training sets, the situation is necessarily abnormal, a first-level alarm is immediately carried out, and mining cannot be carried out; calculating and sequencing the contribution values, uploading several maximum or larger contribution values serving as abnormal reasons to a control interface to be displayedShown in the figure.
Calculating the contribution value
Figure BDA0003404694610000091
The subscript j represents the label of the variable, abs indicates the absolute value; and deltajIs the jth column of the n × n identity matrix; suppose there are four devices, δ for the 2 nd device (variable) gas sensor2＝[0 1 0 0]Τ。
And when any one or two statistics are only larger than the control limit trained by one set of equipment training set, further confirming whether the equipment is in failure or not, if not, immediately carrying out secondary alarm, calculating contribution values and sequencing, and uploading the largest or larger contribution values serving as abnormal reasons to a control interface for display.

Claims (4)

1.基于图正则增量非负矩阵分解的采矿场景异常检测方法,其特征在于，包括以下步骤：1. A mining scene anomaly detection method based on graph regular incremental non-negative matrix decomposition, is characterized in that, comprises the following steps: S1：数据采集；使用两套设备(每套设备均包括摄像头、多种传感器(瓦斯传感器、CO2传感器等)，数量均为一个)对多个正常工作时刻下的环境进行采样；两套设备获得的数据均分别做以下相同处理，下面将描述一套设备数据所进行的操作，必要时分别描述，此后不做重复说明。S1: Data collection; use two sets of equipment (each set of equipment includes a camera, a variety of sensors (gas sensor, CO 2 sensor, etc.), the number is one) to sample the environment at multiple normal working hours; two sets of equipment The obtained data are processed in the same manner as follows. The operations performed by a set of device data will be described below, and described separately when necessary, and the description will not be repeated hereafter. S2：数据预处理；将S1采集的数据进行灰度化、向量化、归一化等预处理后得到训练集X'，初始训练样本矩阵
Figure FDA0003404694600000011
(该矩阵中有k个样本)由训练集X'中某个时段的数据样本组成，将训练集X'中此后时刻的数据样本依次作为下一个新增样本。S2: Data preprocessing; after preprocessing the data collected by S1, such as grayscale, vectorization, and normalization, the training set X' is obtained, and the initial training sample matrix
Figure FDA0003404694600000011
(There are k samples in the matrix) It is composed of data samples in a certain period of time in the training set X', and the data samples in the training set X' at subsequent times are sequentially used as the next new sample. S3：训练阶段；将S2得到的初始训练样本矩阵
Figure FDA0003404694600000012
和新增样本进行图正则增量非负矩阵分解得到最优基矩阵Wnew和系数矩阵Hnew；其中使用
Figure FDA0003404694600000013
进行SVD分解得到的奇异值矩阵、奇异向量矩阵分别对
Figure FDA0003404694600000014
进行图正则非负矩阵分解中的基矩阵和系数矩阵初始化；这样不会破坏原始数据的数据结构，能在低维空间保持样本的几何结构信息，还能结合增量学习充分利用上一步的分解结果，避免重复计算，从而降低了运算时间。S3: training phase; the initial training sample matrix obtained by S2
Figure FDA0003404694600000012
Perform graph-regular incremental non-negative matrix decomposition with new samples to obtain the optimal basis matrix W new and coefficient matrix H new ; use
Figure FDA0003404694600000013
The singular value matrix and singular vector matrix obtained by SVD decomposition are respectively
Figure FDA0003404694600000014
Initialize the basis matrix and coefficient matrix in the normalized non-negative matrix decomposition of the graph; in this way, the data structure of the original data will not be destroyed, the geometric structure information of the sample can be maintained in the low-dimensional space, and the decomposition of the previous step can be fully utilized in combination with incremental learning. As a result, repeated calculation is avoided, thereby reducing the operation time. S4：计算控制限；由S3得到的Wnew、Hnew计算监控统计量N2和SPE，N2为监控特征空间的变化，SPE为监控残差空间的变化；再采用核密度估计(KED)方法对过程数据进行概率密度估计，提取出数据的实际分布信息，从而确定每套设备训练样本对应的统计量控制限
Figure FDA0003404694600000015
SPE′1；
Figure FDA0003404694600000016
SPE′2。S4: Calculate the control limit; calculate the monitoring statistics N 2 and SPE from the W new and H new obtained in S3, where N 2 is the change in the monitoring feature space, and SPE is the change in the monitoring residual space; then use the kernel density estimation (KED) The method estimates the probability density of the process data, extracts the actual distribution information of the data, and determines the statistical control limit corresponding to each set of equipment training samples.
Figure FDA0003404694600000015
SPE′ 1 ;
Figure FDA0003404694600000016
SPE' 2 . S5：测试阶段；重新采集数据(作为测试集X”)进行检测，对测试集X”进行S2、S3相同处理，求出对应的统计量，将该监控统计量与两套控制限进行对比；如果统计量均在两套设备训练集的控制限之内则表示正常状态；如果任意一个或两个统计量均在两套设备训练集的控制限之外则表示异常状态，立即进行一级警报；如果任意一个或两个统计量仅在其中一套设备训练集的控制限之外则进一步排查是否为设备异常，若不是则为环境异常，进行二级警报；当场景异常时，计算贡献值并进行排序，将最大的或较大的几个贡献值作为异常原因上传至控制接口以展示。S5: testing stage; re-collect data (as the test set X") for detection, perform the same processing of S2 and S3 on the test set X", obtain the corresponding statistics, and compare the monitoring statistics with the two sets of control limits; If the statistics are within the control limits of the two sets of equipment training sets, it indicates a normal state; if any one or two statistics are outside the control limits of the two sets of equipment training sets, it indicates an abnormal state, and a first-level alarm will be issued immediately. ;If any one or two statistics are only outside the control limit of one of the equipment training sets, further check whether the equipment is abnormal, if not, the environment is abnormal, and a secondary alarm will be issued; when the scene is abnormal, calculate the contribution value And sort, upload the largest or larger contribution values to the control interface as abnormal causes for display. 2.根据权利要求书1所述的基于图正则增量非负矩阵分解的采矿场景异常检测方法，其特征在于：步骤S2中数据样本具体参数如下：2. the mining scene anomaly detection method based on graph regular incremental non-negative matrix decomposition according to claim 1, is characterized in that: in step S2, the specific parameters of the data sample are as follows: 初始训练样本矩阵
Figure FDA0003404694600000021
其中t0≤ti≤…≤ti+k-1≤t1，表示是初始时段t0～t1采集的数据；
Figure FDA0003404694600000022
表示第j个样本中第i个设备所采集到的数据，该样本是在ti时刻采集到的；例如i＝1表示设备是摄像头，i＝2表示设备是瓦斯传感器，
Figure FDA0003404694600000023
表示第三个样本中瓦斯传感器的数据，该样本是在ti+2时刻采集到的；
Figure FDA0003404694600000024
表示第一个数据样本，是在ti时刻采集的；
Figure FDA0003404694600000025
表示第k+1个数据样本，是在t2时刻采集的，也是作为下一次的新增样本，此后的新增样本以此类推。Initial training sample matrix
Figure FDA0003404694600000021
where t 0 ≤t i ≤...≤t i+k-1 ≤t 1 , indicating that it is the data collected in the initial period t 0 to t 1 ;
Figure FDA0003404694600000022
Represents the data collected by the i-th device in the j-th sample, which is collected at time t i ; for example, i=1 means the device is a camera, i=2 means the device is a gas sensor,
Figure FDA0003404694600000023
Represents the data of the gas sensor in the third sample, which is collected at time t i+2 ;
Figure FDA0003404694600000024
represents the first data sample, which was collected at time t i ;
Figure FDA0003404694600000025
Indicates that the k+1th data sample is collected at time t 2 and is also used as the next new sample, and so on for the subsequent new samples. 3.根据权利要求书1所述的基于图正则增量非负矩阵分解的采矿场景异常检测方法，其特征在于：步骤S3中训练阶段具体步骤如下：3. the mining scene anomaly detection method based on graph regular incremental non-negative matrix decomposition according to claim 1, is characterized in that: in step S3, the concrete steps of training stage are as follows: S3.1：先将初始训练样本矩阵
Figure FDA0003404694600000026
进行SVD分解，得到奇异值矩阵∑和奇异向量矩阵U、V,使用奇异值矩阵、奇异向量矩阵分别对
Figure FDA0003404694600000027
进行图正则非负矩阵分解中的基矩阵和系数矩阵进行初始化，对初始化后的基矩阵、系数矩阵再进行更新迭代，直至目标函数趋于稳定，得到Wk,Hk；这样可获得更优的全局最优解的效果，且对于输入矩阵无需进行任何数据结构的改变处理，不会破坏原始数据的数据结构，可以保留更多的细节信息，从而提高分解效果。S3.1: First set the initial training sample matrix
Figure FDA0003404694600000026
Perform SVD decomposition to obtain singular value matrix ∑ and singular vector matrix U, V, use singular value matrix and singular vector matrix to respectively
Figure FDA0003404694600000027
Initialize the basis matrix and coefficient matrix in the decomposition of the regular non-negative matrix of the graph, and then update and iterate the initialized basis matrix and coefficient matrix until the objective function tends to be stable, and obtain W k , H k ; The effect of the global optimal solution, and the input matrix does not need to be changed in any data structure, the data structure of the original data will not be destroyed, and more detailed information can be retained, thereby improving the decomposition effect. SVD分解公式：
Figure FDA0003404694600000028
SVD decomposition formula:
Figure FDA0003404694600000028
 Wk,Hk初始化：
Figure FDA0003404694600000031
W k ,H k initialization:
Figure FDA0003404694600000031
 目标函数：
Figure FDA0003404694600000032
Objective function:
Figure FDA0003404694600000032
 迭代规则：
Figure FDA0003404694600000033
Iteration rules:
Figure FDA0003404694600000033
 注：其中R表示权重矩阵，D是对角矩阵
Figure FDA0003404694600000034
Lk是拉普拉斯矩阵(Lk＝D-R)，λ是正则化参数。Note: where R represents the weight matrix and D is the diagonal matrix
Figure FDA0003404694600000034
L k is the Laplacian matrix (L k =DR) and λ is the regularization parameter. S3.2：当下一个时刻的新增样本加入时，采用图正则增量非负矩阵分解获得最优基矩阵和系数矩阵。S3.2: When a new sample is added at the next moment, the optimal basis matrix and coefficient matrix are obtained by using graph-regular incremental non-negative matrix decomposition. 例如当
Figure FDA0003404694600000035
加入时目标函数为：For example when
Figure FDA0003404694600000035
The objective function when joining is:
Figure FDA0003404694600000036
Figure FDA0003404694600000036
 迭代规则
Figure FDA0003404694600000037
Iteration rules
Figure FDA0003404694600000037
 S3.3：将所有新增样本重复上述操作，更新结束后得到最优基矩阵Wnew和系数矩阵Hnew。S3.3: Repeat the above operation for all newly added samples, and obtain the optimal basis matrix W new and coefficient matrix H new after the update is completed. 4.一种计算机可读存储介质，其特征在于：所述计算机可读存储介质存储有计算机指令，该指令用于使计算机执行如权利要求1-3任意一项所述的采矿场景异常检测方法。4. A computer-readable storage medium, characterized in that: the computer-readable storage medium stores a computer instruction, the instruction is used to make the computer execute the mining scene abnormality detection method according to any one of claims 1-3 .
CN202111509423.2A 2021-12-10 2021-12-10 Anomaly detection method for mining scenarios based on graph-regularized incremental non-negative matrix factorization Active CN114238854B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111509423.2A CN114238854B (en) 2021-12-10 2021-12-10 Anomaly detection method for mining scenarios based on graph-regularized incremental non-negative matrix factorization

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111509423.2A CN114238854B (en) 2021-12-10 2021-12-10 Anomaly detection method for mining scenarios based on graph-regularized incremental non-negative matrix factorization

Publications (2)

Publication Number Publication Date
CN114238854A true CN114238854A (en) 2022-03-25
CN114238854B CN114238854B (en) 2025-05-16

Family

ID=80754729

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111509423.2A Active CN114238854B (en) 2021-12-10 2021-12-10 Anomaly detection method for mining scenarios based on graph-regularized incremental non-negative matrix factorization

Country Status (1)

Country Link
CN (1) CN114238854B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117576568A (en) * 2023-12-08 2024-02-20 成都理工大学 Depth robust non-negative matrix factorization method based on incremental learning
CN119559768A (en) * 2025-01-09 2025-03-04 四川省华地建设工程有限责任公司 A ground collapse early warning method and system influenced by karst groundwater dynamics

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108985356A (en) * 2018-06-28 2018-12-11 中国人民解放军国防科技大学 Image Decomposition Method Based on NMF
CN110334761A (en) * 2019-07-03 2019-10-15 北京林业大学 Supervised Image Recognition Method Based on Incremental Nonnegative Matrix Factorization Constrained by Orthogonality
CN110516026A (en) * 2019-07-15 2019-11-29 西安电子科技大学 An online single-modal hash retrieval method based on graph regularized non-negative matrix factorization

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108985356A (en) * 2018-06-28 2018-12-11 中国人民解放军国防科技大学 Image Decomposition Method Based on NMF
CN110334761A (en) * 2019-07-03 2019-10-15 北京林业大学 Supervised Image Recognition Method Based on Incremental Nonnegative Matrix Factorization Constrained by Orthogonality
CN110516026A (en) * 2019-07-15 2019-11-29 西安电子科技大学 An online single-modal hash retrieval method based on graph regularized non-negative matrix factorization

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
汪金涛;曹玉东;孙福明;: "稀疏约束图正则非负矩阵分解的增量学习算法", 计算机应用, vol. 37, no. 04, 10 April 2017 (2017-04-10) *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117576568A (en) * 2023-12-08 2024-02-20 成都理工大学 Depth robust non-negative matrix factorization method based on incremental learning
CN117576568B (en) * 2023-12-08 2024-07-05 成都理工大学 Depth robust non-negative matrix factorization method based on incremental learning
CN119559768A (en) * 2025-01-09 2025-03-04 四川省华地建设工程有限责任公司 A ground collapse early warning method and system influenced by karst groundwater dynamics
CN119559768B (en) * 2025-01-09 2025-06-03 四川省华地建设工程有限责任公司 Ground subsidence early warning method and system for karst groundwater dynamic influence

Also Published As

Publication number Publication date
CN114238854B (en) 2025-05-16

Similar Documents

Publication Publication Date Title
CN109308522B (en) A GIS Fault Prediction Method Based on Recurrent Neural Network
CN111401471B (en) Spacecraft attitude anomaly detection method and system
CN112508105B (en) Fault detection and retrieval method for oil extraction machine
CN113391622B (en) Spacecraft attitude system anomaly detection method using multivariate multistep prediction technology
CN108549908B (en) Chemical process fault detection method based on multi-sampling probability kernel principal component model
CN112464567B (en) Intelligent data assimilation method based on variational assimilation framework
CN114238854B (en) Anomaly detection method for mining scenarios based on graph-regularized incremental non-negative matrix factorization
CN115438726A (en) Device life and fault type prediction method and system based on digital twin technology
CN113988210B (en) Method, device and storage medium for repairing distorted data of structural monitoring sensor network
CN118196662B (en) Coal mine digital twin model building method based on Bayesian network algorithm
CN118378900B (en) Coal mine risk intelligent monitoring method and computer equipment based on machine learning model
CN105938524A (en) Microorganism association network prediction method and apparatus
Zhao et al. A neural architecture search method based on gradient descent for remaining useful life estimation
CN115099286A (en) Attention-enhanced harmonic reducer tiny fault diagnosis method and system
Yang et al. A semi-supervised feature contrast convolutional neural network for processes fault diagnosis
CN112434750B (en) Dam monitoring data development mode identification method based on convolutional neural network
CN114037012A (en) Flight data anomaly detection method based on correlation analysis and deep learning
CN114383648A (en) Temperature instrument fault diagnosis method and device
CN119049559A (en) Deep learning-based fermentation process soft measurement method and device
Ashraf et al. DESIGN AND IMPLEMENTATION OF ERROR ISOLATION IN TECHNO METER
CN117831351A (en) A real-time detection method for aviation flight data anomaly
CN112417709A (en) Dynamic modal analysis method based on schlieren image
Wang et al. Aero-engine remaining useful life prediction with missing data using diffusion model
Politi et al. Comparing data assimilation filters for parameter estimation in a neuron model
CN119147563B (en) A material life prediction method based on microstructure analysis

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant