Predictive Maintenance for Connected Products

Most equipment failures are not sudden. A compressor, a motor, a heating element: degradation is a gradual process, and its signals are visible in the telemetry well before the device stops working. Predictive maintenance means using those signals to detect problems before they become failures.

If your product runs continuously, has measurable performance indicators, and your customers care about uptime, predictive maintenance is a realistic and commercially attractive capability to offer.

The challenge with predictive maintenance​

Traditional threshold-based alerting catches obvious problems: if temperature > 90°C, send an alert. But most real failures don't appear as a single threshold crossing. They appear as subtle pattern changes: a metric that drifts slowly over weeks, a correlation between two variables that normally move together starting to diverge, a seasonal anomaly that simple rules miss.

Building predictive models requires:

• Time-series data at sufficient resolution and history: anomaly detection needs baseline behavior, which means weeks to months of consistent telemetry from the device under normal operating conditions.
• Relevant sensors: not every product has the right instrumentation. Useful signals for wear detection include temperature, vibration, current draw, pressure, cycle counts, and response-time metrics. Products that only report binary state (on/off) have limited predictive surface.
• Feature engineering: extracting meaningful signals from raw sensor data is a specialist skill.
• Model training and validation: training an ML model, evaluating its performance, and avoiding false positives requires data science expertise.
• Model serving and real-time inference: running models at scale on streaming data requires infrastructure.
• Feedback loops: models need to improve as they encounter new failure patterns.

Most device manufacturers don't have a data science team, and hiring one to solve this problem is hard to justify before the revenue model is proven.

What it takes to get it right​

A practical predictive maintenance system for hardware manufacturers needs:

1. Automatic baseline learning: the system learns what "normal" looks like for each device type, without manual configuration.
2. Anomaly detection: deviations from the baseline trigger alerts, not just threshold crossings.
3. Forecasting: extrapolating trends to predict when a metric will cross a critical threshold, so maintenance can be scheduled before the failure.
4. Context-aware alerting: alerts are routed to the right people (the installer, the manufacturer's support team, the end user) based on the severity and type.
5. Integration with maintenance workflows: an alert that doesn't trigger a concrete action is not useful.

How Connhex solves it​

Connhex AI provides out-of-the-box anomaly detection and forecasting for time-series telemetry collected from devices. There is no model training step: Connhex AI automatically learns the baseline behavior of each device type from historical data and flags deviations.

This means:

• No data science expertise required, the platform handles the modeling; your team configures which metrics to monitor and what severity to assign to anomalies.
• Per-device baselines, each device has its own learned baseline; an anomaly for device A is assessed against A's own history, not a fleet-wide average that may not apply.
• Forecasting, for metrics with predictable trends (e.g. component wear indicators), Connhex AI projects the trend forward and triggers alerts before the metric becomes critical.

Connhex Rules Engine translates anomaly events into operational actions: create a maintenance ticket, notify an installer, send a push notification to the end user, trigger an OTA update.

Connhex Notifications delivers the alert to the right person through the right channel, with per-recipient preferences and delivery tracking.

A realistic workflow​

A commercial HVAC unit has been reporting compressor current draw and outlet temperature for six months. Here is how a failure is caught before it happens:

• Connhex AI learns the unit's baseline behavior from historical telemetry.
• One morning, the current draw pattern deviates from its learned range: not enough to trigger a threshold alert, but statistically anomalous.
• Connhex AI flags the anomaly. The Rules Engine matches it against a configured rule and notifies the assigned installer via push notification.
• The installer visits the site and finds a failing capacitor before the compressor seizes. The customer never experienced a failure.

The same workflow runs on every monitored device, continuously.

See it in practice​

Industrial and commercial equipment manufacturers use Connhex's anomaly detection to monitor deployed devices and proactively schedule maintenance:

Connhex for HVAC · Connhex for Vending Machines · Connhex for Industrial Washing Machines

Read the technical docs​

• Connhex AI, Overview
• Edge Addons, System Monitor
• Core, Aggregation and Decimation