Multi-sensor based ontology monitoring destructive test data labeling and modeling method

Abstract

The application discloses a kind of based on multi-sensor ontology monitoring destructive test data labeling and modeling method, comprising the following steps: establishing explosion-proof special experiment cabin, and testing test material;The explosion-proof special experiment cabin includes particle concentration sensor, temperature sensor, program-controlled heating table, electric arc generator, stewing furnace;Using cloud platform to collect experimental data, respectively label normal data and destructive experimental data, model and analyze data by k-means clustering, mean statistics, LSTM algorithm.The application solves the problem of low efficiency in the conventional technology of fire warning algorithm abnormal data collection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fire warning, and in particular to a method for labeling and modeling destructive test data of ontology monitoring based on multiple sensors. BACKGROUND

[0002] Existing warning systems all need to collect fire characteristics for analysis and prediction, but often do not give effective solutions on how to collect relevant feature data. In actual operation, the extraction and labeling of a large amount of effective feature data is the key to realizing accurate warning capability. Existing technical solutions often lack effective solutions. For example, the Chinese invention patent with the publication number CN104766433A and the invention name "electrical fire alarm system based on data fusion" does not give a method for how to collect the key point of each sensor feature. The Chinese invention patent with the publication number CN113538838A and the invention name "electrical fire monitoring method for identifying relics building pyrolysis particles" does not explicitly state how to collect parameters in the core "risk assessment function".

[0003] In order to achieve the preventive effect of preparing for the worst, the energy storage fire prevention warning system needs to analyze a large amount of risks. In the specific algorithm training work, a large amount of abnormal data needs to be labeled. Not only is the collection workload huge, but also some abnormal states rarely occur, and real working condition data cannot cover comprehensively.

[0004] Therefore, the technical personnel in the art are committed to developing a method for labeling and modeling destructive test data of ontology monitoring based on multiple sensors. A specially designed experimental cabin is used in combination with a collection platform to realize automatic data collection and labeling for material ontology with hidden fire hazards, helping us accumulate a large amount of abnormal data in a short period of time and quickly and effectively train risk identification algorithms. SUMMARY

[0005] In view of the above defects of the prior art, the technical problem to be solved by the present application is how to provide a method for labeling and modeling destructive test data of ontology monitoring based on multiple sensors to solve the problem of difficult collection of abnormal data and low efficiency in conventional fire warning algorithm technology.

[0006] To achieve the above-mentioned purpose, the present application provides a method for labeling and modeling destructive test data of ontology monitoring based on multiple sensors, comprising the following steps:

[0007] Step 1, establish an explosion-proof special experimental cabin to test the test material; the explosion-proof special experimental cabin includes a particle concentration sensor, a temperature sensor, a program-controlled heating table, an electric arc generator, and a muffle furnace;

[0008] Step 2, collect experimental data using a cloud platform, label normal data and destructive experimental data respectively, and model and analyze the data through k-means clustering, mean statistics and LSTM algorithm.

[0009] Further, the testing of the test material in step 1 includes: simulating various different heating processes through a program-controlled heating table to finally achieve an overheating result; simulating an electric arc generator to simulate an electric leakage and sparking process; simulating a smoldering process through a smoldering furnace; collecting particle concentration data and temperature data through a sensor.

[0010] Further, step 2 specifically includes the following steps:

[0011] Step 2.1, data collection and cleaning;

[0012] Step 2.2, normal data labeling and destructive experimental data labeling;

[0013] Step 2.3, k-means classification of normal data, and threshold value of clustered data by box plot;

[0014] Step 2.4, prediction training of data;

[0015] Step 2.5, box plot analysis of prediction trend.

[0016] Further, step 2.1 specifically includes the following steps:

[0017] Step 2.1.1, set different heating rates, read pyrolysis data every certain time interval, the pyrolysis data including a plurality of particle concentrations and an ambient temperature;

[0018] Step 2.1.2, record data, including serial number, reporting time and pyrolysis data in step 2.1.1;

[0019] Step 2.1.3, process the collected data set, clean up abnormal jumping single point values and null values;

[0020] Step 2.1.4, data normalization processing.

[0021] Further, step 2.2 specifically includes the following steps:

[0022] Step 2.2.1, label normal data, including environmental stability and environmental disturbance;

[0023] Step 2.2.2, label destructive experimental data.

[0024] Further, step 2.3 specifically includes the following steps:

[0025] Step 2.3.1, extract normal data set, cut according to year, quarter, month, and make 5-class k-means clustering;

[0026] Step 2.3.2, extract normal data, cluster according to ambient temperature, cluster ambient temperature in a year, and make a classification every 5 degrees Celsius to cluster concentration data;

[0027] Step 2.3.3, box plot analysis is performed on the concentration clusters obtained in step 2.3.2 to remove abnormal values and obtain normal upper threshold.

[0028] Further, the data prediction training in step 2.4 includes: predicting the annual, quarterly and monthly data, and adopting three cycle loops.

[0029] Further, the step 2.4 specifically includes the following steps:

[0030] Step 2.4.1, using LSTM algorithm to model normal data set and abnormal data set, including annual, monthly, weekly, daily, every 6 hours, every hour, 30 minutes and 15 minutes trend prediction model;

[0031] Step 2.4.2, convert time series into supervised learning, change the observation angle, and arrange the feature family according to the time series.

[0032] Further, the box plot analysis of the predicted trend in step 2.5 includes: box plot analysis of the predicted trend of each time period model, abnormal judgment of the upper limit of the box plot for the current data, and time window cumulative judgment.

[0033] Further, the specific calculation process of step 2.3.2 is: first, obtain the concentration data set, group the data every 5 degrees Celsius starting from 0 degrees Celsius; cluster each group of data using 9 classifications; and obtain the centroid of each cluster.

[0034] The beneficial effects of the present application are:

[0035] (1) The experiment cabin can accumulate normal data (including external interference) to model the normal environment and identify situations different from the normal environment. From the practical application, it is more comprehensive. For example, scenarios such as construction personnel smoking outside the power distribution cabinet and cutting PVC pipes can be found. These hidden dangers will be ignored by the commonly used threshold judgment technology because they do not reach the threshold.

[0036] (2) Through the simulation of various temperature anomaly scenes, the material anomaly can be clearly labeled, and after training, the algorithm recognition is more comprehensive. For example, the abnormal temperature characteristics of wood, paper and silk in cultural relic protection scenes need to be specially concerned by the algorithm, and in the energy storage scene, we need to focus on testing lithium batteries. These feature data can be quickly collected from the experimental cabin.

[0037] The concept, specific structure and technical effects of the present application will be further described below in combination with the drawings, so as to fully understand the purpose, features and effects of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flowchart of a preferred embodiment of the present application;

[0039] Figure 2 is a normal distribution diagram of particle concentration data of a preferred embodiment of the present application;

[0040] Figure 3 is a schematic diagram of the composition of the experimental cabin of a preferred embodiment of the present application;

[0041] Figure 4 is a schematic diagram of abnormal data modeling of a preferred embodiment of the present application;

[0042] Figure 5 is a schematic diagram of normal prediction of a preferred embodiment of the present application;

[0043] Figure 6 is a system architecture diagram of a preferred embodiment of the present application. DETAILED DESCRIPTION

[0044] The preferred embodiments of the present application are described below with reference to the accompanying drawings, so that the technical content is clearer and easier to understand. The present application can be embodied in many different forms, and the scope of protection of the present application is not limited to the embodiments described herein.

[0045] In the drawings, components of the same structure are denoted by the same reference numerals, and components having similar structures or functions are denoted by similar reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present application is not limited to the size and thickness of each component. In order to make the drawing clearer, the thickness of some components is appropriately exaggerated in some places in the drawing.

[0046] As shown in Figure 1 The present application provides a multi-sensor-based body monitoring destructive test data labeling and modeling method, comprising:

[0047] S1, an explosion-proof special experimental cabin is established, and the test material is heated at various rates, electric arc burning, and smoldering test;

[0048] S2, collect experimental data using a cloud platform, label normal data and destructive experimental data respectively, and model and analyze the data through k-means clustering, mean statistics and LSTM algorithm.

[0049] The specific steps are as follows:

[0050] 1. Establish an explosion-proof special experimental cabin, including but not limited to: particle concentration sensor, temperature sensor, program-controlled heating table, electric arc generator, and muffle furnace. The experimental cabin supports fourth-generation (4G) mobile communication technology and industrial bus communication.

[0051] Simulate various different heating processes through the program-controlled heating table to finally achieve the overheating result.

[0052] Simulate the electric leakage and sparking process through the electric arc generator.

[0053] Simulate the smoldering process through the muffle furnace.

[0054] Collect particle concentration data and temperature data (environmental temperature, cable temperature) through sensors.

[0055] 2. Use the cloud platform to collect experimental data, perform modeling and analysis work.

[0056] Figure 2 is a schematic diagram of the normal distribution of particle concentration data at 28 degrees Celsius.

[0057] As Figure 3 shown is a schematic diagram of the experimental cabin 10 built by the present application, which is installed with: a particle concentration detector 11, a temperature sensor 12, a program-controlled heating table 13, an electric arc generator, and a muffle furnace for testing various rate heating, electric arc burning, and smoldering processes of the test material 14, and the experimental cabin 10 reports the data to the test data processing module of the center cloud platform in real time.

[0058] Figure 4 is a schematic diagram of abnormal data modeling, the vertical axis is the concentration data, the horizontal axis is the time line, and curves 1, 2, and 3 are 10 / 2.5 / 1.0 μm particles. As Figure 4 shown, this is the labeled data set of wood, paper, and wire materials, which shows the comparison relationship between material state, temperature, and particle concentration, which is the data set for modeling the abnormal state.

[0059] As Figure 5 is an example of the system predicting normality, curve 4 is the actual data, and curve 5 is the predicted data

[0060] The working steps of the present application are as follows:

[0061] 1. Data acquisition and cleaning method:

[0062] 1.1, Set different heating rates, read pyrolysis data every 2 seconds, including 3 particle concentrations, and one ambient temperature;

[0063] 1.2, Record data, including serial number, reporting time, and the above concentration and temperature data;

[0064] 1.3, Process the collected data set, and clean up abnormal jumping single point values and null values;

[0065] 1.4, Normalization of data;

[0066] 2, Normal data labeling, destructive experiment data labeling:

[0067] 2.1, Label normal data (data collected in a stable environment), including stable environment (smooth data, no interference), environmental interference (such as construction cutting, worker smoking);

[0068] 2.2, Labeling of destructive experimental data (material state changes, including color, softening, melting, human perception, including touch warm, hot, no smell perception, intermittent slight odor, obvious odor);

[0069] 3, k-means classification of normal data, box plot threshold for clustered data, which can flexibly adapt to local environment.

[0070] 3.1, Extract normal data set, cut by year, season, month, and do 5-class k-means clustering;

[0071] 3.2, Extract normal data, cluster by ambient temperature, cluster ambient temperature throughout the year, and cluster concentration data every 5 degrees Celsius;

[0072] 3.3, The specific calculation process is as follows: first, obtain the concentration data set (0 degrees Celsius, every 5 degrees); cluster each group of data, using 9 categories; obtain the centroid of each cluster;

[0073] 3.4, Box plot analysis of the above concentration clustering, remove abnormal values, and obtain the normal upper threshold;

[0074] 4, Year, season, and month data are predicted and trained in three cycles, t-1 modeling, t data, and t+1 prediction.

[0075] 4.1, using long short-term memory (LSTM) algorithm to model normal data set, abnormal data set, including annual, monthly, weekly, daily, every 6 hours, every hour, 30 minutes and 15 minutes trend prediction model;

[0076] 4.2, the time series is converted into supervised learning, the key is to change the observation angle, and the feature family is arranged according to the time series. For example, when training, if we use 4 columns [concentration 1, concentration 2, concentration 3, concentration to be predicted] to train the model, 3 columns [concentration 1, concentration 2, concentration 3] are needed to provide when predicting, and the LSTM model can calculate the concentration to be predicted, so as to compare with the actual data, judge the coincidence degree of the current data, when the coincidence degree is greater than the preset threshold, such as 90%, the trend can be judged as "consistent with the modeling trend". The specific calculation process is as follows: using the historical data of the same period into the LSTM model, calculating the current prediction, such as using the data of last year today, last month today, last week today, last year today, getting the daily trend prediction model, and putting it into the above prediction algorithm to predict the normal trend and abnormal trend of the day, or using the current one hour, 30 minutes, 15 minutes data into the related model, calculating the coincidence degree and predicting whether the future trend is consistent with the model trend.

[0077] 5, box plot analysis is made on the trend prediction of each time period model (such as weekly, daily, 6 hours, hourly, 30 minutes, 15 minutes, etc.), the upper limit of the box plot is used for abnormal judgment of the current data, and the time window is used for cumulative judgment, such as x minutes within y abnormal accumulations.

[0078] As Figure 6 The system architecture diagram of the present application is shown, and the experimental cabin training model is deployed on the sensor side.

[0079] Real-time experimental data is collected and reported through the sensor;

[0080] The training and iteration of the experimental model are completed in the safe cloud;

[0081] The specific algorithm training action includes:

[0082] Data cleaning (eliminating abnormal data);

[0083] Data segmentation (daily, weekly, monthly, annually);

[0084] The data is modeled by k-means clustering, mean statistics (box plot), LSTM and other algorithms.

[0085] The convergence of experimental data and the deployment of algorithm are completed by the sensor edge cloud and the safe cloud.

[0086] The preferred embodiments of the present application have been described above in detail. It should be understood that modifications and variations to the preferred embodiments could be made by those skilled in the art without departing from the spirit and scope of the application. Accordingly, it is intended that there be included within the scope of the application, all such modifications and variations as would be apparent to those skilled in the art upon reading this disclosure. It is intended to obtain for the inventors such patent rights as are available for any patent granted on the present application.

Claims

1. A method for labeling and modeling destructive test data for body monitoring based on multiple sensors, characterized in that, Includes the following steps: Step 1: Establish an explosion-proof test chamber to test the test material; the explosion-proof test chamber includes a particle concentration sensor, a temperature sensor, a programmable heating platform, an electric arc generator, and a simmering furnace; Step 2: Collect experimental data using a cloud platform, label normal data and destructive experimental data respectively, and model and analyze the data using k-means clustering, box plots, and LSTM algorithms; Step 2 specifically includes the following steps: Step 2.1: Data collection and cleaning; Step 2.2: Labeling of normal data and labeling of destructive experimental data; Step 2.3: Use k-means to classify normal data, and use box plots to apply thresholds to the clustered data; Step 2.4: Perform prediction training on the data; Step 2.5: Perform box plot analysis on the predicted trend; Step 2.3 specifically includes the following steps: Step 2.3.1: Extract the normal dataset, divide it by year, quarter, and month, and perform 5-class k-means clustering; Step 2.3.2: Extract normal data and cluster it according to ambient temperature. Cluster the ambient temperature throughout the year, and make a category for every 5 degrees Celsius. Cluster the concentration data. Step 2.3.3: Perform box plot analysis on the concentration clusters obtained in Step 2.3.2, remove outliers, and obtain the normal upper limit threshold. The prediction training of data in step 2.4 includes: prediction training of annual, quarterly, and monthly data, using three cycles; Step 2.4 specifically includes the following steps: Step 2.4.1: Use the LSTM algorithm to model normal and abnormal datasets, including trend prediction models by year, month, week, day, every 6 hours, every hour, 30 minutes and 15 minutes; Step 2.4.2: Transform the time series into supervised learning by changing the observation perspective and arranging the feature families according to the time series. The box plot analysis of the predicted trend in step 2.5 includes: performing box plot analysis on the predicted trend of the model for each time period, applying the upper limit of the box plot to the current data for anomaly judgment, and using the time window for cumulative judgment.

2. The method for labeling and modeling destructive test data based on multi-sensor body monitoring as described in claim 1, characterized in that, The testing of the test material in step 1 includes: simulating various heating processes through a programmable heating stage to achieve overheating; simulating the leakage and ignition process through an arc generator; simulating the smoldering process through a smoldering furnace; and collecting particle concentration data and temperature data through sensors.

3. The method for labeling and modeling destructive test data based on multi-sensor body monitoring as described in claim 2, characterized in that, Step 2.1 specifically includes the following steps: Step 2.1.1 Set different heating rates and read the pyrolysis data at regular intervals. The pyrolysis data includes several particle concentrations and one ambient temperature. Step 2.1.2 Record the data, including the serial number, reporting time, and the pyrolysis data described in Step 2.1.1; Step 2.1.3 Process the collected dataset, cleaning up abnormally fluctuating single-point values ​​and null values; Step 2.1.4 Data normalization.

4. The method for labeling and modeling destructive test data based on multi-sensor body monitoring as described in claim 3, characterized in that, Step 2.2 specifically includes the following steps: Step 2.2.1: Label the normal data, including environmental stability and environmental interference; Step 2.2.2: Label the destructive experimental data.

5. The method for labeling and modeling destructive test data based on multi-sensor body monitoring as described in claim 1, characterized in that, The specific calculation process of step 2.3.2 is as follows: First, obtain the concentration dataset, and group the data into groups of 5 degrees starting from 0 degrees Celsius; perform clustering on each group of data using a 9-category classification; and obtain the centroid of each cluster.

Images