Datathons Solutions

Datathon Kaufland Solution – LSTM and EDM Models for Predictive Maintenance

In this paper we propose the use of a combination of LSTM and EDM models to address the issue of anomaly classification and prediction in time series data. Working with sensor data for automated storage and retrieval systems for a German hypermarket chain, we show that predictors based on variance and median methods show sufficient promise in the handling of anomalies.

16
votes

1. Business Understanding

As part of the Kaufland challenge, we are tasked to build a predictive maintenance model to avert machine failure. Internet of Things (IoT) sensors allow us to monitor the machine’s performance in real time. The model can detect when the machine performs anomalously and suggest the need for action on specific parts of the machine that is about to fail. A predictive system that can detect anomalies in the machine’s behaviour just before it breaks down will help minimise maintenance costs. Compared to a preventive maintenance system, a predictive system is able to reduce costs associated with preventive maintenance done even when maintenance is not required. The business success criteria would therefore be to lower maintenance costs relative to a preventive maintenance system.

Predictive maintenance is a strategy to reduce cost of:

  • Labour for skilled repairman
  • Preventative part replacement
  • Post damage repair which might be higher than preventive maintenance
  • Overcome the availability of experts to repair machines

The predictive system will be a success if it detects anomalies accurately.

Our project plan is summarised in Figure 1. The first step consists of exploratory data analysis and data preparation. Based on the first step, we decide on our modelling approach and then construct a strategy for deployment.

Figure 1: Project plan

In this project, we have access to a dataset of sensor data (velocity and acceleration) attached to seven machines over a period of 5 months to two years.

The tools that we consider for this purpose are Python and R, which are popular languages for statistical and data analysis. Given that our team has mixed proficiency in both languages, we conduct our analysis in a mix of both. The models that we consider for this purpose at the preliminary stage are Long-Short-Term Memory (LSTM) models.

The goal is not to determine how much vibration a machine will withstand before failure. The aim is to obtain a trend in vibration characteristics that can warn of trouble. For example a sudden short peak or increase does not necessarily means the machine is going to fail, it might be caused by sensor error. Instead, look for odd occurrence that persists.

 

2. Data Understanding

The data is loaded in R and Python. The data we use is provided by Kaufland, a large hypermarket chain based in Germany. We are provided with data on 7 separate automated storage and retrieval systems machines. Each machine has 6 different sensors (Figure 2). Data from these machines are collected for different time ranges:

  • 2 years for Machine 1 and 7 (RBG1 and 7).
  • 5 months for the other machines.

RBG 1 and 7 are the pilot study to test the efficiency of the sensors, this explains why they have long range of data. Sensors are embedded to the other machines after the sensors are proved to be working.

 Figure 2: Sensors on machine. Image from Data Science Society.

The sensors attached to the machine are as follows:

1. Lifting mechanism:

  • Sensor 1: Lifting motor
  • Sensor 2: Lifting gear

2. Drive mechanism:

  • Sensor 3: Drive motor
  • Sensor 4: Drive gear
  • Sensor 5: Drive wheel

3. Sensor 6: Idle (non-powered) wheel

These six sensors measure the effective velocity of vibrations and the maximum acceleration of shocks, collected under controlled conditions thrice a day.

Figure 3 shows a sample of data from the drive gear sensor on Machine 1. All these sensors measure both the effective velocity of vibrations and the maximum acceleration of shocks. Each sensor collects data under controlled conditions three times a day.  Two of the machines have collected data for almost two years, the other five for five months.

machine_name sensor_type date_measurement realvalue unit start_time end_time day_of_week sensor_group
0 RBG1 drive_gear_V_eff 2/9/2016 0.395 mm/s 26:42.8 26:42.8 Friday driving_motor
1 RBG1 drive_gear_V_eff 2/9/2016 0.577 mm/s 26:45.7 26:45.7 Friday driving_motor
2 RBG1 drive_gear_V_eff 2/9/2016 0.717 mm/s 26:48.5 26:48.5 Friday driving_motor
3 RBG1 drive_gear_V_eff 2/9/2016 0.832 mm/s 26:51.3 26:51.3 Friday driving_motor
4 RBG1 drive_gear_V_eff 2/9/2016 0.941 mm/s 26:54.1 26:54.1 Friday driving_motor
5 RBG1 drive_gear_V_eff 2/9/2016 1.042 mm/s 26:56.9 26:56.9 Friday driving_motor

Figure 3: Data from drive gear sensor of Machine 1.

3. Data Exploration

Data exploration is performed with purpose of finding relationships among factors to motivate a modelling approach.

The values of vibrations of sensors for each sensor is plotted using scatter plot (Figure 4 and 5). The values of six sensors are included. Null values are not shown in the figures.

Figure 4: Values of sensors for Machine 1 (RBG1). This machine shows a significant decrease in mean vibration read after a documented maintenance event on 8 Feb 2018.

 

Figure 5: Values of sensors for Machine 7 (RBG7)

Correlation matrix is plotted in Figure 6. Matrix is plotted against each sensor. The sensors are grouped together based on location of the sensors. Group 1 consists of sensor 1 and 2; group 2 sensor 3, 4 and 5; group 3 sensor 6. Sensors are highly correlated within groups.

Figure 6: Correlation matrix of sensors  from Machine 1. Grey to black indicates high correlation score.

The time series is plotted for each sensors of each machine. No seasonality/repeating patterns/cycles is found. Spectral analysis on Figure 7 confirms the absence of periodicity within the data. This provides a strong case to abandon ARIMA method. At this stage, we consider not using classical time series approaches in favour of Long-Short-Term Memory (LSTM) unit for recurrent neural networks (RNN).

Figure 7: Spectral analysis of Sensor 1 from Machine 1.

Figure 8 shows the plot according to sensor for 7 machines.  The patterns in the series suggests that spread may be an indication of the health of the machine. This forms the basis of our study.

Figure 8: Values of Driver Gear Sensor from 7 machines are plotted at the same axis.

 

Extreme values in the data are filtered using Tukey’s method. Figure 9 illustrates the occurrence rate of outliers that are present in the sensor data aggregated by machine and sensor type. Visual inspection indicates that Machine 7 shows a significantly higher occurrence of outliers, followed closely by Machines 2 and 3. Interpreting the occurrence of outliers as an indicator of machine reliability, we hypothesize that machine health for Machines 1, 4, 5, and 6 were superior to that of the remaining machines during the period of data recorded. We thus propose that a controlled experiment be designed to test said hypothesis and identify supplementary factors that may contribute towards overall machine health.

Figure 9: Distribution of outliers

3. Data Preparation

Operations below are performed on the data:

  1. Datetime variables are converted into machine readable variables.
  2. Data aggregation is performed to reduce number of rows as well as fill in missing values. Mean is calculated for time interval of 1 minute.
  3. Some columns are grouped together for exploratory data analysis purpose.
  4. Outlier removal using standard deviation (Tukey’s Hinges). Any departure from standard operating characteristics are removed.
  5. Moving variance is calculated with window length of 10 minutes.

4. Modelling and Evaluation

A. AR-LSTM approach

i. Modeling

To detect anomalies in a sensor, we first identify periods for each machine which appear to be well-behaved, fit an AR(10) autoregressive model (10 lags of the variable of interest) after pooling well-behaved series, then forecast the values. If the difference between the forecasted and actual value is ‘high’, the observation will be highlighted as an anomaly.

We begin by modelling the velocity measure, a common parameter for analysing vibration (Scheffer & Girdhar 2004), of the Idle Wheel sensor. This is done by training on data pooled time series data between March 2018 and July 2018, where data is available across all machines. Pooling time series data is not a new idea in the time series literature, although scarcely explored (see Maharaj and Inder 1999). While the functioning of the machine is unknown, the maintenance events (Figure 10) are the next-best option for labelling the data. These models are evaluated by back-casting the series and cross-checking against maintenance events to identify false positives and false negatives.

We consider the use of neural network models for sequence data, particularly recurrent neural network (RNN) models. Traditional RNN units are unable to remember long-term dependencies and susceptible to the vanishing gradient problem, and for this purpose LSTM units may be more suitable.

 

machine_name date repair_description
RBG 1 13.11.2017 Correction of skewed movement of the machine on the rail
08.02.2018 Change of idle wheel
12.02.2018 – 14.02.2018 Grinding of the rail
RBG 2 02.04.2018 – 08.04.2018 Grinding of the rail
RBG 3 02.04.2018 – 08.04.2018 Grinding of the rail
05.04.2018 Change of bearings of drive motor
RBG 7 02.04. 2018  - 08.04.2018 Grinding of the rail

 Figure 10: Repair description for maintenance that has occurred.

As each of these series have different centers and spread, we assume that the number of standard deviations away from the mean is similarly distributed and can be ‘learned’ in the same way (Figure 11).

Figure 11: Distribution of Idle Wheel velocity (standardised) across all machines and Time series of each of the machines’ Idle Wheel velocity (standardised)

ii. Evaluation

Evaluation of forecast model  

The AR-LSTM model provides forecasts that correspond to the level and spread of data. The spread of the AR-LSTM forecasts (Figure 12) are rather conservative, and the AR-LSTM model performs similarly well out-of-sample.

Figure 12: In-sample performance

 

Evaluation of classification criteria  

Figure 13 plots the AR-LSTM forecast errors. We set 3 standard deviations above and below the mean as the threshold for anomalous behaviour.  This standard is convenient to interpret as our data is already scaled. This corresponds to approximately 95% of the data, assuming the data is normally distributed. This cut-off rule has a relatively low false positive rate of roughly 18/7000 (0.2%).

 

Figure 13: Absolute value of forecast error (standard deviations)

B. Moving variance approach

i. Model

Motivation

We depart from our initial approach of fitting an autoregressive model using LSTM because we cannot easily evaluate our model as the data is unlabelled. Even though a maintenance record is provided, the functioning of the machine remains an unobserved variable. Specifically, we don’t have an idea of machine health, so we can’t evaluate AR-LSTM. Since our EDA suggests we can approximate machine health using the variation in the series, we assume variation to be a general indication of machine health and use LSTM to detect anomalous patterns in the past hour of data.

At a high level, the differences between these two approaches are as follows:

AR-LSTM Moving variance approach
Predictors Lagged measure (velocity) Moving variance of lagged measures
Outlier threshold Standard deviation Median which is less sensitive to outlier is used

Figure 15: Major differences in approach

Our use of the spread as a measure of machine health is justified as follows. Figure 16 and 17 show the time series of idle wheel velocity and the moving average respectively, pooled across all machines. It suggests that there are distinct intervals in the data with different spread and centers. This implies that spread may be a good way to indicate machine health.

Figure 16: Minute-aggregated time series of idle wheel velocity (machine 1)

Figure 17: Moving Variance of Idle Wheel velocity (machine 1)

We begin by modelling the velocity measure, a common parameter for analysing vibration (Scheffer & Girdhar 2004), of the Idle Wheel sensor. We generate a series j, a pooled time series of velocity for a single sensor j across all machines. Pooling time series data is not a new idea in the time series literature, although scarcely explored (see Maharaj and Inder 1999).  

The training set for v_j is generated as a concatenation of velocity data from machines 2 to 7. Data for machine 1 will be used as our test set. Machine 1 is selected because repairs on the idle wheel were only conducted for machine 1.

To use supervised learning methods, we define an anomaly flagbeta_{j,t} to equal True if an observation lies outside of Tukey’s hinges across v_j. The anomaly flag is used to flag abnormal behaviour in the sensors. Predictions from the model would indicate that repairs are necessary. In addition, we evaluate the model by back-casting the series. Using the two maintenance events recorded against the idle wheel (Table 1), we can obtain an idea of false positives and false negatives in the data. We emphasise again that we cannot obtain the true rate of false positives and false negatives as machine condition remains latent.  

Our exploratory analysis suggests that the variation in the series may indicate the healthy functioning of the machine. We may want to use the instantaneous variation of the series; however the acceleration data does not suggest a relationship between shocks and repair events. As a next-best solution, we use the moving variance to approximate the variation in the series at a point in time.

The moving variance for sensor j across 60 periods, which is equivalent to an hour, is defined as
s_{j,t}^2=1/60 sumlimits_{i=t-59}^t(v_{j,i}-bar v_{j,t})^2
where

bar v_{j,t}=1/60 sumlimits_{i=t-59}^t v_{j,i}

We fit a classification model to classify beta_{j,t} using s_{j, t-1}, s_{j, t-2},...,s_{j, t-10} as predictors. Given that the predictors are sequence data, we consider the use of recurrent neural network (RNN) models for classifying anomalies. Traditional RNN units are unable to remember long-term dependencies and susceptible to the vanishing gradient problem, and for this purpose LSTM units may be more suitable.

We fit two LSTM models for this purpose, with different train/test set pairs (Figure 18). The model summary is shown in Figure 19.

Train Test
Model 1 Velocity measure from idle wheel sensors, machines 2 to 7 Velocity measure from idle wheel sensors, machine 1
Model 2 Velocity measure from idle wheel sensors, machines 1 to 7 Velocity measure from lifting motor and drive wheel, machines 1 to 7

Figure 18: Train and test set for Models 1 and 2

Figure 19: Model summary

ii. Evaluation

Model 1

By visually inspecting Figure 20, observations 300-1300 should be classified as anomalies because the spread of the series is abnormally high. Figure x shows the anomaly flag for Machine 1’s velocity on the idle wheel sensor as classified by Tukey’s method. The anomaly flag correctly highlights these observations as anomalous. The drop-off point at approximately observation 1700 is likely due to a maintenance event. This event is also flagged as anomalous according to Tukey’s Method. Model 1’s predictions also closely follow the anomaly labels.

Figure 20: Velocity series of idle wheel for machine 1

Model 2

It is reasonable to assume that the distribution of the normal condition of the sensors is similar across sensors as well as machines. To provide some evidence for our hypothesis, a LSTM model is fitted on the idle wheel data across all machines and tested against lifting motor and drive wheel data across all machines. The results of the predictions on the drive wheel sensors are illustrated in Figure 21.

Figure 21: Drive wheel velocity measure for machines 1-7

Again, the predicted data closely trails the actual labels in all drive wheel sensors except for Machine 5, where a false negative was detected. However, the series does not appear to have an abnormal distribution in that region. This suggests that the LSTM approach is able to generalise anomalous patterns better than a naive approach using Tukey’s method to flag anomalies.

As for the lifting motor (Figure 22), the model detects anomalies in Machine 1, 2, 6 and 7. Both the anomaly labels and LSTM predictions detect anomalies similarly. The abnormalities in Machine 1’s idle wheel are reflected in the predictions and anomaly labels for the drive wheel and lifting motor. This suggests that this modelling approach may find it hard to distinguish between different sections of the machine when predicting faults.

Figure 22: Velocity measures for lifting motor for machines 1-7

Assuming the anomaly labels to be the ‘ground truth’, we can calculate false positive and false negative rates. False negatives are lower than false positives in all sensor-machine combinations evaluated except for Machine 5’s drive wheel sensor. Even though the LSTM model was not optimsed to minimise false negatives, Model 2 has a low share of false negatives.

 

C. E-Divisive with Medians (EDM)

E-Divisive with Medians (EDM) is a model for breakpoint (changepoint) detection. Breakpoint detection describes the process of detecting change points in time series. We decided to analyse the breakpoints with an unsupervised approach as we noticed that there are visually obvious changes of distribution resulting in distinct segments in the sensor’s time series data, due to different possible reasons (could be anomalies, repairs or maintenances, in some cases it’s missing data). The EDM algorithm (James et al. 2014) is reported to have several advantages over traditional breakout detection techniques. They include (1) robustness against anomalies, (2) able to detect divergence and change in distributions, and (3) faster computation process. In Figure 23, we segment out series into 3 sections by visually examining the data. We expect the EDM algorithm to automate this process.

Figure 23: Visual inspection of series

Outcome

We tried to detect the breakouts using different metrics, namely mean (Figure 24), distributions (Figure 25), and median (Figure 26). With no labeled data available, we visually evaluated the graphs and observed that median-based method draws a cleaner boundary between different segments in the time series. We concluded that median is the best option to detect breakouts in our case as it is robust against outliers. One downside of this approach is that computing the median is computationally expensive, with a computational cost of </b>O(n!). Due to time constraint, we did not generate breakout detection reports for all the sensors.

Figure 24: Mean-based EDM

 

Figure 25: Distribution-based method

Figure 26: Median-based EDM

Breakpoint detection can indicate gradual shifts in the time series that may not be visible to the human eye. As a complement to actual anomaly detection, it can be deployed to provide alerts before an abrupt anomalous event occurs. This in theory can be used as an indicator for scheduling preventative maintenance.

 

6. Deployment

A. AR-LSTM

Figure 27 shows the deployment of the AR-LSTM model. Firstly, 11 minutes of data is collected. 10 minutes of data would be sent to LSTM model to make a ‘normal’ prediction, which is the 11th minute. The prediction is then compared to the data collected from the sensor, at the 11th minute since the start.

The difference between the prediction and the real data is calculated. If the difference is greater than the threshold, anomaly is detected since the data deviates from the predicted ‘normal’ value. Otherwise it means there is no anomaly, the machine is operating as normal. The loop continues.

Figure 27: AR-LSTM deployment plan

B. Moving Variance

Figure 28 shows the deployment of the moving variance model. Data is collected for 70 minutes. Data is transformed by computing moving variances for 60 minute windows. Then, predictions are made based on the LSTM model. Based on the prediction, the model suggests whether preventive maintenance is required.

 

 

Figure 28: Moving variances-LSTM deployment plan

Conclusion

To sum up our analysis, we have achieved two main actionable outcomes (1) a classification model for outlier detection, and (2) an implementation of breakout detection. With that, we suggest that Kaufland can (1) implement the breakout detection method to alert operators that the machines need to be inspected, (2) use the trained classification model to improve performance of outlier and breakout detection, and finally, (3) do outlier detection in conjunction with detailed repair logs to correlate repair parts with anomalous patterns. In addition, we also recommend several steps that can be taken to improve the overall analyses and modeling: collection of temperature data, machine utilization data, machine movement data and maintaining updated detailed repair logs.

References

  1. James, N., Kejariwal A.m, and Matteson, D. (2014). Leveraging Cloud Data to Mitigate User Experience from ‘Breaking Bad’.Available at: https://arxiv.org/pdf/1411.7955.pdf
  2. Girdhar,Paresh dan C. Scheffer., 2004, Practical Machinery Vibration Analysis and Predictive Maintenance., Elsevier, Burlington. Gates corp., 2015.
  3. Forecasting time series from clusters. Maharaj, E. A. & Inder, B. A. 1999 Melbourne Vic Australia: Monash University Publishing. 27 p.

Code

The code for this submission is available here.

Share this

30 thoughts on “Datathon Kaufland Solution – LSTM and EDM Models for Predictive Maintenance

  1. 4
    votes

    Positive Aspects
    —————————-
    – very good introduction
    – excellent descriptions & explanations
    – good flow & project plan
    – decent data description
    – comparisons on machine and sensor level
    – reasonable modeling approaches
    – evaluation of different approaches
    – suggestion of combination of methods

    Negative Aspects
    —————————-
    – no word on hyperparameter tuning

    1. 0
      votes

      GET RICH WITH BLANK ATM CARD … Whatsapp: +18033921735

      I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich. {[email protected]} I email them also and they sent me the blank atm card. I have use it to get 500,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

      You can also contact them for the service below

      * Western Union/MoneyGram Transfer

      * Bank Transfer

      * PayPal / Skrill Transfer

      * Crypto Mining

      * CashApp Transfer

      * Bitcoin Loans

      * Recover Stolen/Missing Crypto/Funds/Assets

      Email: [email protected]

      Text & Call or WhatsApp: +18033921735

      Website: https://darkwebonlinehackers.com

    2. 0
      votes

      URGENT AND EFFECTIVE LOVE SPELL CASTER TO HELP YOU GET BACK YOUR EX LOVER VERY FAST WHATSAPP +2348124644470

      My Name is Micheline Peric from Ireland, I want to say thank you to PRIEST WISDOM for the good thing he has done for me,Though am not sure if this is the best forum to show my joy and happiness for what he has done for me but i can’t hide happiness and my Joy so i have to share it with people, my marriage got crashed about two years ago and i tried all i could within my power but to no avail. I saw a post and testimonial about the good things PRIEST WISDOM has been doing so I decided to give it a try. though he is always a busy man but when he responded back to my email, he gave me 24hours for my marriage to be restored really just like he said my marriage was restored since then I am happy and i am living happily i am so grateful, PRIEST WISDOM you can always email him here: Supernaturalspell@ yahoo. com or WHATSAPP / cell phone +2348124644470 or view his blog his web page supernaturalspell21.blogspot.com/ or go to his page facebook.com/PRIESTWISDOM11

  2. 0
    votes

    Very interesting article, thx for sharing this!

    Have you tested it on more than 1 minute interval? I mean in real world you need to predict within factory lets say in range of 10-30 minutes, if supporting remote sites even hours. Did you try to predict lets say 1 hour into the future? What could be bottleneck and possible tuning such approach. On the other hand in my case I have dataset for past 4 years with very detailed markings on different actions within devices (maintainances, runtime, warning, error, etc.)

  3. 0
    votes

    INTERNET SCAM ALERT‼️

    The internet today is full of SCAM ADS, mostly in comments of various sites and blogs. A large number of individuals have been victims of scam and lost a lot of money to SCAMMERS. Most of the common scam you can see
    ❌BANK LOAN SCAM. ❌BINARY OPTIONS SCAM.
    ❌LOTTERY SCAM. ❌HACKING SCAM. and lost more……

    ✳️The big Question is “Can someone Recover their money lost to Internet Scam⁉️
    I will say yes, and will tell you how.

    The only way to Recovery your is to hire a Hacker who will help you take back your money from this Scammers and today is your lucky day, you just met the guys perfect for the job.

    WHO ARE WE❔
    We are PYTHONAX, A group of skilled Hackers and have dedicated our time to help individuals get back thier money from INTERNET SCAMMERS. There is a research was carried one to calculate the amount of money lost through Scam, and it was confirmed that more than USD $3billion annually.

    HOW DO WE OPERATE❔
    We first of all study the scammer brought to us by hacking the person device(phone or computer) to get information of How, Where, this person keeps money he/she as defrauded people of ( so many of this scammers don’t actually save the money in banks, they mostly stack the money in a Bitcoin wallet, that way it is safe and untraceable to authorities) and we work on a strategy to get back the money and give it back to whom they have defrauded.

    Contacting us is simple, just give us a message through the email below.
    Email-: [email protected]

    If you a victim of internet scam or you know someone who is, make contact to us immediately. You are 💯 % safe to contact us, our email is very secure.

    We also provide Legit Hacking Services such as-:
    🔸Phone Hacking/Cloning
    🔸Email Hacking & Password Recovery
    🔸Social Media Hacking & Passowrd Recovery
    🔸Deleted Files Recovery 🔸Mobil Tracking
    🔸Virus detection & Elimination. e.t.c

    Email-: [email protected]

    Pythonax.
    2020 © All Right Reserved.

  4. 1
    votes

    Hello everyone,

    I want to testify about Luis Carlos who help me invest my bitcoin and made me who i am today, i never believe in bitcoin investment till i met Luis Carlos, I saw so many testimony about Luis Carlos helping people to invest their bitcoin, So i decide to contact him and invested $500 and after 7 days i receive my profit of $5,000 on my bitcoin wallet, Ever since i have been investing with them and i always receive my profit without any delay, so if you want to invest your bitcoin, Luis Carlos is the best trader you can invest with and getting your profit is guarantee, So if you want to invest just contact them and they will guide you on how start your investment . Email: [email protected] or WhatsApp : +14234516435

  5. 0
    votes

    Are you looking for a profitable investment where you can start with a little amount and earn a reasonable profit within a short period of time?. I never believed in any online investment because I was scared and never wanted to be cheated, until I saw a review about Mr Pablo Martinez. He’s a Forex/Crypto trading account manager who can help you manage your trading account with his trading strategies and winning signals. I started with an investment of $500 and earned a profit of $6,650 within 7 days. I now earn quite a lot on a weekly basis and I owe everything to Mr Pablo Martinez. Thank you Mr Pablo Martinez for turning my financial life around, and I will keep recommending your good works. If you want to invest in Stock, Binary options and Forex/Crypto trading, kindly contact Mr Pablo Martinez and you’ll be glad you did. There are no hidden charges.

    Contact Mr Pablo Martinez through

    E-mail: [email protected]

    WhatsApp: +44 7520 636249

    Facebook: https://www.facebook.com/PM-Fast-Trade-104912912155335/

  6. 0
    votes

    BITCOIN RECOVERY IS REAL!!!

    ( MorrisGray830 At gmail Dot Com, is the man for the job )
    This man is dedicated to his work and you can trust him more than yourself. I contacted him a year and a half Ago and he didn’t succeed. when i got ripped of $491,000 worth of bitcoins by scammers, I tried several recovery programs with no success too. I kept on. And now after so much time Mr Morris Gray contacted me with a success, and the reward he took was small because obviously he is doing this because he wants to help idiots like me who fell for crypto scam, and love his job. Of course he could have taken all the coins and not tell me , I was not syncing this wallet for a year, but he didn’t.
    He is the MAN guys , He is! If you have been a victim of crypto scam before you can trust Morris Gray 10000000%. I thought there were no such good genuine guys anymore on earth, but Mr Morris Gray brought my trust to humanity again. GOD bless you sir…you can reach him via ( MORRIS GRAY 830 at Gmaill dot com ) or Whatsapp +1 (607)698-0239..

  7. 0
    votes

    GET RICH WITH BLANK ATM CARD … Whatsapp: +18033921735

    I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich.([email protected]) I email them also and they sent me the blank atm card. I have use it to get 250,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

    You can also contact them for the service below

    * Western Union/MoneyGram Transfer

    * Bank Transfer

    * PayPal / Skrill Transfer

    * Crypto Mining

    * CashApp Transfer

    Email: [email protected]

    Call or WhatsApp: +18033921735

  8. 0
    votes

    GET RICH WITH BLANK ATM CARD … Whatsapp: +18033921735

    I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich.([email protected]) I email them also and they sent me the blank atm card. I have use it to get 250,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

    You can also contact them for the service below

    * Western Union/MoneyGram Transfer

    * Bank Transfer

    * PayPal / Skrill Transfer

    * Crypto Mining

    * CashApp Transfer

    Email: [email protected]

    Text & Call or WhatsApp: +18033921735

  9. 0
    votes

    GET RICH WITH BLANK ATM CARD … Whatsapp: +18033921735

    I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich. {[email protected]} I email them also and they sent me the blank atm card. I have use it to get 500,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

    You can also contact them for the service below

    * Western Union/MoneyGram Transfer

    * Bank Transfer

    * PayPal / Skrill Transfer

    * Crypto Mining

    * CashApp Transfer

    * Bitcoin Loans

    * Recover Stolen/Missing Crypto/Funds/Assets

    Email: [email protected]

    Text & Call or WhatsApp: +18033921735

    Website: https://darkwebonlinehackers.com

  10. 0
    votes

    GET RICH WITH BLANK ATM CARD, Whatsapp: +18033921735

    I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich. {[email protected]} I email them also and they sent me the blank atm card. I have use it to get 500,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

    You can also contact them for the service below

    * Western Union/MoneyGram Transfer

    * Bank Transfer

    * PayPal / Skrill Transfer

    * Crypto Mining

    * CashApp Transfer

    * Bitcoin Loans

    * Recover Stolen/Missing Crypto/Funds/Assets

    Email: [email protected]

    Text & Call or WhatsApp: +18033921735

    WEBSITE:: https://darkwebonlinehackers.com

  11. 0
    votes

    GET RICH WITH BLANK ATM CARD, Whatsapp: +18033921735

    I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich. {[email protected]} I email them also and they sent me the blank atm card. I have use it to get 500,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

    You can also contact them for the service below

    * Western Union/MoneyGram Transfer

    * Bank Transfer

    * PayPal / Skrill Transfer

    * Crypto Mining

    * CashApp Transfer

    * Bitcoin Loans

    * Recover Stolen/Missing Crypto/Funds/Assets

    Email:>>>>[email protected]

    Text & Call or WhatsApp: +18033921735

    Website: https://darkwebonlinehackers.com

  12. 0
    votes

    My ex-husband and I had always managed to stay friendly after our divorce in February 2017. But I always wanted to get back together with him, All it took was a visit to this spell casters website last December, because my dream was to start a new year with my husband, and live happily with him.. This spell caster requested a specific love spell for me and my husband, and I accepted it. And this powerful spell caster began to work his magic. And 48 hours after this spell caster worked for me, my husband called me back for us to be together again, and he was remorseful for all his wrong deeds. My spell is working because guess what: My “husband” is back and we are making preparations on how to go to court and withdraw our divorce papers ASAP. This is nothing short of a miracle. Thank you Dr Emu for your powerful spells. Words are not enough.
    Email [email protected]
    Phone/WhatsApp +2347012841542.

  13. 0
    votes

    GET RICH WITH BLANK ATM CARD, Whatsapp: +18033921735

    I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich. {[email protected]} I email them also and they sent me the blank atm card. I have use it to get 500,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

    You can also contact them for the service below

    * Western Union/MoneyGram Transfer

    * Bank Transfer

    * PayPal / Skrill Transfer

    * Crypto Mining

    * CashApp Transfer

    * Bitcoin Loans

    * Recover Stolen/Missing Crypto/Funds/Assets

    Email: [email protected]

    Text & Call or WhatsApp: +18033921735

    WEBSITE: https://darkwebonlinehackers.com

  14. 0
    votes

    I am very happy, I am here to share my testimonies with the general public about what this man called Dr Adodo has just done for me , this man has just brought back my lost Ex husband to me with his great spell, i was married to this man called Steven we were together for a long time and we loved our self’s but when i was unable to give he a child for 2 years he left me and told me he can’t continue anymore then i was now looking for ways to get him back until a friend of mine told me about this man and gave his contact email ([email protected]) then you won’t believe this when i contacted this man on my problems he prepared this spell cast and bring my lost husband back, and after a month i miss my month and go for a test and the result stated am pregnant am happy today am a mother of a baby girl, thank you once again the great Dr Adodo for what you have done for me, if you are out there passing through this same kind of problems you can contact he today on his mail ( [email protected]) on his Whatspp +18144484537 and he will also help you as well,,,,,,,,,,,,,,,,,

  15. 0
    votes

    i want to thank Dr Adodo for the herbal HIV medicine he gave to me and my daughter, i was suffering from HIV when i gave birth to my daughter and that was how my daughter got the sickness indirect from me, but to God be the glory that i am heal with the herbal medicine that Dr Adodo gave to me when i contacted him. i want to use this medium to tell everyone that the solution to our sickness has come, so i will like you to contact this great healer on his Whatspp +18144484537 with him all your pains will be gone, i am really happy today that i and my daughter are cured of HIV, we are now negative after the use of his medicine,my doctor confirm it. once more i say a big thank to you Dr Adodo for healing hands upon my life and my daughter, i say may God continue to bless you abundantly and give you more power to keep helping those that want your help in their lives. email him now he is waiting to receive you. [email protected] Whatspp +18144484537

    Dr Adodo CAN AS WELL CURE THE FOLLOWING DISEASE:-

    1. HIV/AIDS
    2. HERPES
    3. CANCER
    4. ALS
    5. Love spell
    6. Lottery Spell

  16. 0
    votes

    GET RICH WITH BLANK ATM CARD, Whatsapp: +18033921735

    I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich {[email protected]} I email them also and they sent me the blank atm card. I have use it to get 500,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

    You can also contact them for the service below

    * Western Union/MoneyGram Transfer

    * Bank Transfer

    * PayPal / Skrill Transfer

    * Crypto Mining

    * CashApp Transfer

    * Bitcoin Loans

    * Recover Stolen/Missing Crypto/Funds/Assets

    Email: [email protected]

    Text & Call or WhatsApp: +18033921735

    Website: https://darkwebonlinehackers.com

  17. 0
    votes

    URGENT AND EFFECTIVE LOVE SPELL CASTER TO HELP YOU GET BACK YOUR EX LOVER VERY FAST WHATSAPP +2348124644470

    My Name is Micheline Peric from Ireland, I want to say thank you to PRIEST WISDOM for the good thing he has done for me,Though am not sure if this is the best forum to show my joy and happiness for what he has done for me but i can’t hide happiness and my Joy so i have to share it with people, my marriage got crashed about two years ago and i tried all i could within my power but to no avail. I saw a post and testimonial about the good things PRIEST WISDOM has been doing so I decided to give it a try. though he is always a busy man but when he responded back to my email, he gave me 24hours for my marriage to be restored really just like he said my marriage was restored since then I am happy and i am living happily i am so grateful, PRIEST WISDOM you can always email him here: Supernaturalspell@ yahoo. com or WHATSAPP / cell phone +2348124644470 or view his blog his web page supernaturalspell21.blogspot.com/ or go to his page facebook.com/PRIESTWISDOM11

  18. 0
    votes

    GET RICH WITH BLANK ATM CARD, Whatsapp: +18033921735

    I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich {[email protected]} I email them also and they sent me the blank atm card. I have use it to get 500,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

    You can also contact them for the service below

    * Western Union/MoneyGram Transfer

    * Bank Transfer

    * PayPal / Skrill Transfer

    * Crypto Mining

    * CashApp Transfer

    * Bitcoin Loans

    * Recover Stolen/Missing Crypto/Funds/Assets

    Text & Call or WhatsApp: +18033921735

    Website:: https://darkwebonlinehackers.com

  19. 0
    votes

    GET RICH WITH BLANK ATM CARD, Whatsapp: +18033921735

    I want to testify about Dark Web blank atm cards which can withdraw money from any atm machines around the world. I was very poor before and have no job. I saw so many testimony about how Dark Web Cyber hackers send them the atm blank card and use it to collect money in any atm machine and become rich {[email protected]} I email them also and they sent me the blank atm card. I have use it to get 500,000 dollars. withdraw the maximum of 5,000 USD daily. Dark Web is giving out the card just to help the poor. Hack and take money directly from any atm machine vault with the use of atm programmed card which runs in automatic mode.

    You can also contact them for the service below

    * Western Union/MoneyGram Transfer

    * Bank Transfer

    * PayPal / Skrill Transfer

    * Crypto Mining

    * CashApp Transfer

    * Bitcoin Loans

    * Recover Stolen/Missing Crypto/Funds/Assets

    Email: [email protected]

    Text & Call or WhatsApp: +18033921735

    Website:: https://darkwebonlinehackers.com

  20. 0
    votes

    I have made several unfortunate decisions based on the lifestyle I choose. I took several debts to gamble with, and when I got the chance to make instant money with little risk and lots of profits. I jumped into it, I thought I would pay off all my debts and still have plenty. I wasn’t stupid, I was just too greedy to see that I was being scammed that I kept investing hundreds of thousands in these crypto racketeers. Writing this review is the only good thing I have done in my life for years as I will be saving tons of people from falling prey to these Bitcoin investment racketeers. I got in touch with this programmer online, I was very open and honest with him, I just wanted to know if they can really help me retrieve my hijacked funds before dragging myself into another mess. He was very honest, and dealing with this programmer was enlightening. We achieved success together as he always says. My hijacked tokens were successfully traced and retrieved. “Cybergenie @ cyberservices . com–WA (+1)2525120391”. You can never go wrong with CYBER GENIE..

Leave a Reply