Machine Learning may seem like a distant technology reserved for scientists and tech companies. The reality is that there are a plethora of applications of Machine Learning throughout a wide span of industries. One of these industries is Supply Chain Management where Machine Learning has numerous use cases.

**How can Machine Learning improve supply chain efficiency?

The Swish Machine Learning team highlights the use cases and how they can benefit Supply Chain Management and Logistics businesses.**

What should I know about Machine Learning?

Machine learning describes the application of an algorithm (mathematical formulas) to a set of data to extract insights, predictions, conclusions. More broadly, Machine Learning is a field of computer science with methods that aim to equip machines with human-like reasoning. From the genesis of this field in the 1950s to today, Machine Learning has grown tremendously. Step by step, humans taught machines how to perceive the external environment, interact with it, and drive insight and decisions autonomously. Today’s state of the technology allows a wide array of use cases serving the purpose of helping managers and decision-makers process data more easily and take the best business decision based on the insights provided by the Machine Learning model.

For the Supply Management industry and teams, Machine Learning is a powerful tool to improve business operations and optimize the whole workflow.

1. Predictive machine learning for optimal inventory management.

Inventory (or Warehouse) Management starts with predicting demand. Once a demand level is determined, it is a matter of reverse engineering to predict the number of parts required.

Predicting demand for a product is a difficult task across all industries and can be very costly when done wrong. Predicting too little demand results in missed revenue opportunities and unsatisfied customers. Predicting too much demand creates product waste and extra storage cost.

Warehouse managers have several inputs they must keep on hand and regularly restock. Choosing how many and when to rebuy them is an ever growing in complexity optimization problem.

Predictive machine learning techniques can help automate an optimal solution to stock and replenishing prediction and planning.

The key to all successful machine learning models is quality data and fortunately, warehouses have a lot of it. Time to delivery, previous years sales, storage constraints, margins of error and even social media posts around a product are just a few data points that help predict demand in a machine learning model.

Using the large amounts of data, machine learning models are able to discover patterns/correlations not visible to the human eye.

One common approach to solve this problem is to use LSTM Network which is a type of recurrent neural network. It is widely used and successful because it takes into account not only a current static picture but all past data points (hence “recurrent”) as well to help predict demand. Time series analysis is useful because it often discovers patterns like seasonal correlation or time-lagged correlation (past affect future) between input variables.

2. Prediction and planning maintenance.

Whether it is delivery trucks or manufacturing robotics, machines require maintenance that can be quite costly. Generally most companies believe preventative maintenance is less costly than reparative maintenance and for the most part, they are correct.

However, companies are often not optimizing the amount and type of preventative maintenance. The correct type of preventative maintenance is often performed too much, causing unnecessary downtime & additional preventative cost. Or, worse, cases when the wrong type of preventative maintenance is performed causing even more downtime with additional reparative maintenance.

With proper data tracking, machine learning models can help predict when reparative maintenance will and won’t occur.

Given the type of preventative & reparative maintenance, respective costs & downtimes, general use of machinery and other relevant data points, a classification machine learning model can help predict when reparative maintenance will occur.

This model will show what factors affect the machinery most, and with this, it is possible to find the minimum amount of preventative maintenance required to successfully prevent a repair. There are many different types of classification model types that can be used. Artificial neural networks, a deep learning model, has proven superior to most in recent years.

These predictive maintenance models allow operators to better assess the effects of preventative maintenance and how to optimize it.

3. Route & Delivery Optimization.

Two conditions for goods to be delivered are a route and a transportation means.

1. Delays in transportation can prove detrimental, adding unforeseen costs or missing critical deadlines.

2. Lowering the cost/time taken to deliver is very often a strategic advantage.

Both of these areas can benefit from machine learning.

Transport delays are an inevitable costly aspect of supply chain management. Predicting these delays allow for prevention or proactive management, reducing its negative impact. Using historical data from transport providers, weather patterns, current transport sensor data, and other factors like departure time, machine learning models are able to predict the likelihood of a delay. This is another classification problem that can be solved by Artificial Neural Networks or other similar algorithms. Once equipped with this knowledge of delays, managers are able to mitigate its impact.

Depending on delivery requirements (shared cargo, door to door, same day) the optimal route to take is not always obvious and is increasing in complexity. Fortunately, a relatively new branch of machine learning called Reinforcement learning is perfect for this problem. Different from other types in the fact it doesn’t need large sums of data to start, it requires “only” the rules and landscape to the problem. In this context, a simulation of possible delivery routes would be mapped out, and the agent would be given an end goal to complete along with actions to take. Through vast amounts of simulated cases, the model is able to find optimal patterns not visible to humans due to sheer size.

Note: This is the type of machine learning that is beating pros at Go, Chess, and Poker.

4. Fraud Detection

With so many different third parties and vendors, all collaborating on one supply chain, detecting fraud can be difficult.

Checking for fraud is often a very repetitive task with similar invoices coming in routinely from the same provider. The more moving parts there are the more that must be monitored, increasing the decision cycle time. This risk management practice is also unnecessary in more than 95% of the cases as everything is usually correct.

The task could be described as looking for a “needle in a haystack”… which is perfect for a computer.

Anomaly detection is a branch of machine learning that is designed to find the needle/outlier in vasts amounts of data.

It can be done both in a supervised training, providing “normal” labeled train data sets, or in an unsupervised that clusters normal data & outlier data. With the help of natural language processing, invoices can be automatically scanned and inputted into the anomaly detection model. Then an operator will be alerted if new items are flagged for concern and what % of confidence the alert has, drastically reducing the time it takes to monitor and prevent fraud.

5. Supply Chain Digital Assistants

Many of the above applications solve problems to a level where a person is still required to take action on the analyzed data. Although this will still often be the case, and companies want it that way, it is becoming less so with the rise of digital assistants.

Digital assistants are being applied in multiple fields and their abilities are growing, including those in supply chain management.

There are multiple uses for which a digital assistant can be applied throughout a supply chain.

Inventory orders can be automated completely by having the digital assistant send an email to place the order. It can accept orders by conversing with the supplier, reviewing the order for completeness and alerting a human for any anomalies found. In general digital assistants can be linked with all machine learning models so it is continuously scanning the outcome and can alert respective parties if action is required.

Digital assistants have benefited from the advances made in Deep Learning. They are no longer just a large web of predefined rules and responses. A conversation is a sequential set of steps which makes this a problem for recurrent neural networks.

A popular methodology is a sequence-to-sequence network, which has two recurrent neural networks, where the first sequence is the encoder (or the message) and the second sequence is the decoder (or the reply). Both of these sequences often use LSTM networks as they can “remember” past sequences that can provide greater context to the current response.

To create a well functioning assistant you will need to train it on data that is relevant to the tasks you want to perform. For example: for a customer service assistant, feed it historical customer service conversations. To prevent errors and have it perform additional tasks, rule-based systems can be added. A combination of well thought out rules and a large training dataset makes for a high functioning digital assistant.

Machine learning’s impact on supply-chain management is profound if players in the industry use the tremendous potential the technology holds. The many moving parts of supply-chain management, procurement, manufacturing, and delivery can all be enhanced by machine learning.

Demand forecasting, operational maintenance, accepting and placing orders and risk management are just a few of the applications that are being impacted by Machine Learning today.

The organizations who adopt machine learning early will be at a competitive advantage among their peers, with many cost savings and increased operational efficiencies.

This analysis was brought to you by the Machine Learning team at Swish.

We build Machine Learning models for Supply Chain and Logistics businesses and teams. Do you have an operational efficiency challenge? A bottleneck in your supply workflow? A difficulty that can be solved with Machine Learning? We’ll be happy to help. Let’s chat!