Preventative maintenance gets an upgrade with Online Predictive Processing

Feb. 8, 2018 – Many manufacturers struggle to allocate inventories properly to simultaneously meet their master production schedule of finished goods to satisfy new orders and meet service level agreements for spare parts for their customers. This plagues large and small companies alike. In fact, one of the top ten reasons that startups fail is due to poor inventory management, according to WASP Barcode.

Often field service and manufacturing departments compete for parts, and companies find themselves rushing to recover from customer outages in the field at the expense of the master production schedule. Conversely, at the end of a quarter when revenue is critical, the master production schedule takes priority because delivering new product to customers takes precedence over servicing existing customers in an effort to improve revenue recognition and growth metrics.

The dreaded status quo: Forecast failures
In an ideal world, manufacturers can buffer their inventory to meet these competing factors, but to do so across all parts, locations, and customers is just not financially practical. However if you had more accurate insight into when service or preventative maintenance was required, you could better plan for those events, ensuring the necessary inventory was available for these events without negatively impacting the planned production orders for that same inventory.

The common approach to solving this problem is to implement an ERP system for material planning, a complex spreadsheet for manufacturing planning and a separate complex spreadsheet for service planning. But these processes are often disjointed, causing endless expedites.

More advanced companies augment their ERP systems with supply-chain planning systems and service planning systems. Supply-chain planning systems typically use lead times and capacity models to plan backwards from desired promise dates. These lead times account for assumptions that suppliers provide as well as production lead times typically imported from the Bills of Material (BoMs) in the ERP system. However, these average metrics don’t account for production mishaps and logistics surprises.

Most service planning systems use the mean-time-between-failure (MTBF) rates associated with parts to plan field replenishment. Suppliers or engineering departments typically provide these average metrics. The MTBF data will therefore trigger replenishment to establish inventory levels that meet expected failures. Further, they recommend work orders for preventive maintenance based on average failure rates.

The problem with these approaches to planning is that they are based on approximations and are almost always inaccurate. As a result, a company has to react to late orders, outages and production plan changes.  

Is there a way to customize service planning and manufacturing planning to each circumstance in real time? The answer is yes — this can be done using predictive machine learning.

Predict Failure conditions with machine learning and proactively plan
In every company, there are those who seem to have a sixth sense for spotting issues. They anticipate when a machine is going to go down or when a customer is going to have an outage. How do they do this? It’s because they are really good pattern matchers. They see shop-floor data, service requests, or sensor data from the field and can predict future events. Now, we finally have the tools to automate this ad hoc manual process and build more accurate predictors in the field for all.

Rather than wait to react to failures, the new state-of-the-art is to build predictive models of failure and then proactively plan around these predictions.

Predictive models use machine learning to learn from experience. When a company has historical data regarding field equipment, predictive models can use classification algorithms to classify the historical data into one of a few categories.

For example, let’s assume we want to predict if a product in the field is likely to fail within 30 days, 14 days, or seven days. The machine learning model is trained on snapshots of the historical data; you can take one reading a day for every piece of equipment and feed it into a model with a label such as 0,1,2,3 representing that a piece of equipment failed within 30 days, 14 days, seven days, or did not fail. The readings provided as input to the model are raw data from the equipment such as temperature, pressure and volume sensors, or digital data such as throughput, latency and exceptions.

Often, the most predictive data is net change data, meaning changes in various raw observations. You can imagine calculating net-changes over windows of time, such as change since yesterday, last week or last month. Net-change window data often aggregates data into averages, maximums, and minimums. Temporal data such as recency and frequency of events are often quite predictive, such as time elapsed since installation, time since go-live, time since last maintenance, frequency of replacement over a year, last service call, last alarm and last software upgrade. Models are also often enriched with exogenous data such as weather, news or market data.

Machine learning models train on all of this data. Then, after training, they are deployed to make predictions on live data. Predictive planning applications can then continuously run and make decisions based on the model’s predictions. If a product is likely to fail in the field, then the company can institute rules on how to proactively handle the anticipated failure. They may allocate inventory to a new service work order that replaces the part. They may have to purchase materials from suppliers, or they may have to escalate allocation decisions when parts are necessary for production or sales orders, and there is not enough lead time to procure or produce more parts before the anticipated failure. In this case, customer service or sales may have to be notified if new customer orders are at risk.

The power of this approach is that it is dynamic — it changes over time. As the field equipment matures and changes state, the model constantly tests its attributes. This is in stark contrast to MTBF approaches. If a part deteriorates in a manner similar to training instances, then the model will “see” that pattern and classify the part as likely to fail.

The other dynamic aspect of this approach is that the model can improve over time as it continuously retrains. As it processes more scenarios, it is highly likely that it will become more accurate. The result is that over time, fewer surprises happen because failures are increasingly anticipated.

The power of online predictive processing
The key to this predictive approach to allocating inventory and planning maintenance service is a new foundational platform called Online Predictive Processing (OLPP). OLPP makes predictive analytics actionable and uniquely handles the disparate workloads required for predictive maintenance. Predictive applications require:

1. Streaming ingestion: the ability to continuously process voluminous amounts of real-time data coming from IoT devices such as sensors, mobile devices, field equipment, servers, etc. at high velocity and concurrency

2. Batch ingestion: the ability to extract large batches of data from external systems such as ERP orders

3. Analytical processing: the ability to analyze voluminous amount of data to transform it into usable features for machine learning

4. Transactional processing: this term is misleading in that many view it as the ability to do financial transactions or be an applications’ system of record, but it also entails being able to look up or update a record in a split second, to be able to rollback a bunch of updates when there is an error to keep the platform consistent and accurate, and able to handle high degrees of concurrency, meaning when many users or systems are simultaneously making requests

5. Machine learning: the ability to train and deploy models using a variety of analytics such as logistic regression, decision trees, Naive-Bayes, SVMs or deep-learning networks.

There are two approaches to building an OLPP platform. One is the “decoupled” approach, where companies integrate multiple compute engines and explicitly send data back and forth between the engines. The second is the “integrated” approach, which uses one smart platform that automatically utilizes multiple engines under the hood, but the user does not have to integrate the compute engines or send data back and forth between them.

An example of a decoupled approach is a team that integrates compute engines such as Apache Hadoop for storage, Apache Kafka for streaming, Apache Spark for analytics and Apache HBase for transactional processing. The benefits of this approach is that you have full control of the architecture and can optimize it to your workloads, but this is done at significant cost of implementation due to its complexity. Competition for engineers with distributed system experience that can maintain a decoupled engine is fierce.

On the contrary, one example of an integrated platform is the OLPP platform provided by Splice Machine, which pre-integrates the engines mentioned above but packages the platform as a SQL relational database management system. The benefit of this approach is that most IT personnel know how to use a SQL RDBMS without re-training and the platform’s optimizer makes the decision of which compute engine to use.

Business benefits of OLPP
Companies that use systems to predict failures, and then allocate inventory and dispatch service orders based on those expectations report many business benefits, such as the ability to:

●    Reduce inventory by 20 to 40 per cent
●    Reduce days of supply by 50 per cent
●    Improve SLAs by 5-40pts
●    Reduce expedites by 10 to 30 per cent
●    Improve meeting of customer request dates by 20 to 40 per cent
●    Reduce unplanned downtime by 70 per cent 

Companies that add a predictive capability to their service planning functions can simultaneously lower inventory costs while increasing customer satisfaction. Predictive approaches dynamically modify themselves over time by retraining with more data and capture the changing state of field equipment by streaming data in continuously and extracting order data from systems of record. These new applications predict which components are likely to fail and then provide a what-if testbed to visualize what would happen if you allocated inventory to the expected failures. This new approach is affordable and accessible to any company because of the availability of OLPP platforms. OLPP combines streaming, analytical processing, transactional processing, and machine learning to power predictive applications. Any company that struggles to balance manufacturing planning with service planning based on competing inventories should consider implementing a predictive application powered by OLPP.

With all of the advancements in machine learning and the systems that power it, the time is right for preventive maintenance to become predictive and move past the complexities of the present.

Monte Zweben is the CEO, Splice Machine