Construction site accident analysis using text mining and natural language processing techniques
In this paper author is describing concept to provide safety to
workers at construction site from accidents by analysing past accident data by
using machine learning algorithms and text mining technique such as TF-IDF
(Term Frequency-Inverse Document Frequency) and natural language text
processing to remove special symbols, stop words, stemming etc.
Past accident data contains details of accidents and by building
machine learning algorithms model we can analyse data to identify cause of
accident and can prevent future accident by giving test data of new work to
predict causes of accidents and can avoid such causes. This machine learning
algorithms can help in extracting dangerous objects such as misused tools,
sharp objects nearby, damaged equipment etc.
In this paper to provide safety to workers author covering below
points
1)
Various texting
mining and NLP techniques are explored with respectto construction site
accidents analysis. Using this technique we will remove stop words,
punctuations, special symbols and apply stemming technique to clean past
accident data. After data cleaning we will convert all text data to numeric
vector by using TF-IDF technique. TF-IDF contains frequency weight of each word
in vector and using this vector we will build machine learning train model.
Whenever we give new test data then that test data also convert to TF-IDF and
then apply on train model to search for similar data and give output of similar
data as prediction. Below example describe how to convert text to TF-IDF vector.
Suppose
I have 3 sentences
Sentence
1: An apple a day keep doctor away
Sentence
2: apple good for health
Sentence
3: shipment of gold damage in fire
First
we remove stop words such as ‘an, a, of, in’ from sentences and then take
remaining words and form columns of vector. After forming columns put each word
count as values of that vector. See below vector columns
Apple
day keep doctor away good health
shipment gold damage fire
Sentence1 1
1 1 1 1 0
0 0 0 0 0
Sentence2 1
0 0 0 0 1
1 0 0 0 0
Sentence3 0
0 0 0 0 0
0 1 1 1 1
So
I convert all 3 sentences to TF-IDF vector just by putting count of each word
as vector values, if sentence contains that column word then we will put its
count, if sentence not contains work then we put 0 as that column values. Now
to check similarity we can multiply one row with other and if multiply value
greater than 0 then two sentences contains similarity otherwise not.
In
above matrix if multiply sentence 1 row with sentence 2 row then we get value
greater than 0 and similarity is there as both sentences contains 1 common word
called ‘apple’. Similarly if we multiply sentence1 row with sentence3 row then
we will get value 0 which means similarity not there between sentence 1 and 3
and we can see there is no common words in sentence 1 and 3.
2)
Ensemble
algorithm which has not been well studied in this field isproposed to classify
the causes of accidents and SQP algorithm isutilized to search for optimal
weighs of the ensemble model. In this technique we will use ensemble algorithms
such as random forest and voting classifier with SQP (Sequential Quadratic
Programming) to classify causes of accidents. Using SQP we can assign weight to
the classifier which can help classifier in predicting correct causes of
accident.
3)
A rule based chunker is developed for
dangerous objects extraction.Neither SQP optimization algorithm nor rule based
chunker withregard to this field is found in the state of the art. Rule based
chunker means getting Part Of Speech (POS) of each sentence to find dangerous
object detection. When we apply POS on sentence then all dangerous objects will
come under NOUN POS and by extracting noun phrases from sentences we can
identify what are the dangerous objects which causes accidents.
To implement this project author using OSHA dataset and
effectiveness ofthe proposed approaches is verified by the experiment results.
OSHA dataset contains past accident data and by using this dataset we will
analyse performance of various machine learning algorithms such as SVM,
Decision Tree, Naïve Bayes, Logistic Regression, KNN, Ensemble Random Forest
and Propose Voting Classifier which will build on all 5 base classifiers such
as SVM, Naïve Bayes, Decision Tree, KNN and Logistic Regression. Voting
classifier take all 5 classifier and then vote each classifier and whatever
classifier give better accuracy then voting will choose that classifier for
future data prediction.
Dataset saved inside ‘dataset/OSHA.csv’ folder and you can open
and see the details and below are some data of new work and it has no details
what accident can cause by doing that work but machine learning can predict and
display future accident cause.
cutting down a
large horizontal pipe block
installing roof
decking on a flat roof by carrying and placing decking material
portable storage
tank and a running powered industrial truck Caught in or between
electrical transformer
box distribution line electrocuted electric ladder work onto a 13800 volt power
line work electrical parts
In above bold sentences some work details are there and while
doing such work what accident can happen can be predicted with machine learning
algorithms.
video output
thank you for your comment
pls call me on 8125424511