How a Water Treatment Plant Connected 1,200 Sensors in 8 Weeks

Water treatment is one of those industries where the phrase "can't afford downtime" is literal, not rhetorical. The plant serves 340,000 residents. Treatment processes run continuously, 24 hours a day. Any unplanned outage has regulatory consequences and public health implications. The operations team doesn't have patience for multi-year technology rollouts.

When the plant's director of operations contacted us in late 2025, they had already run a two-sensor pilot on their primary chlorination system. The results were solid — they caught two dosing anomalies that could have led to compliance failures. The question was how to scale from 2 sensors to 1,200 without disrupting operations.

What They Had to Work With

The plant was built in two phases: the original intake and primary treatment facilities from the early 1990s, and a membrane filtration expansion added in 2011. This meant two generations of control systems with no common integration layer.

The older section ran a legacy SCADA system communicating with field devices over Modbus RTU on RS-485 at 9600 baud. Approximately 280 instruments were on this network — flow meters, turbidity sensors, pressure transmitters, and chemical dosing controls. The newer membrane filtration section had a more modern DCS with OPC UA connectivity, managing another 340 instruments.

Outside the control system boundary, they had 580 sensors that needed to be added as net-new instrumentation: pump motor temperature sensors, electrical panel current transformers for energy monitoring, ambient temperature/humidity sensors in electrical rooms, and vibration sensors on the three large centrifugal pumps that feed the distribution system. These would be MQTT-native wireless sensors communicating over a dedicated 900MHz ISM band network.

The Network Architecture

The core constraint was the IT/OT boundary. The plant's OT network was physically isolated from corporate IT and from the internet. Any data leaving the OT environment had to go through a demilitarized zone with firewall inspection and explicit allow rules.

We deployed three edge gateway nodes — one in each major building — running the SensorVault edge agent. Each gateway sat on the OT network and handled local protocol translation: Modbus RTU polling on one interface, OPC UA client subscriptions on another, and the 900MHz radio bridge for the wireless sensors. Local buffering was set to 72 hours to handle any network interruptions.

Data from the gateways transited the DMZ over TLS 1.3 on port 8883 (MQTT/TLS) to the plant's on-premises analytics server. The IT security team specifically requested that no data leave the facility's physical network, so this was a fully on-premises deployment — nothing in the cloud.

The firewall allow rule took eleven days to get approved through the change control process. We built the entire integration and tested it on a lab network in parallel while waiting. By the time the firewall rule was in place, we had one day of field work left rather than three weeks.

The Sensor Deployment Itself

The 580 net-new wireless sensors were the fastest part. The pump vibration sensors — triaxial MEMS accelerometers rated to IP67, -40°C to +85°C operating range — went on the bearing housings of each pump and took about 45 minutes per pump to install and calibrate. The current transformers clipped onto the motor leads without any power interruption. Temperature/humidity sensors in electrical rooms required only a small mounting screw and a battery change every 18 months.

Connecting the existing Modbus instruments was more involved. Each Modbus segment required a register map audit: we had to verify that the register addresses in the SCADA configuration actually matched the current firmware on each field device. On two flow meters, the register maps had been changed in a firmware update that nobody had documented. That discovery cost us two days and pushed us to add register map verification as a standard step in our deployment checklist.

The OPC UA integration with the membrane filtration DCS was straightforward once we had certificates provisioned. The DCS vendor provided a security policy endpoint that required mutual certificate authentication and 256-bit encryption. We provisioned application certificates through the plant's internal PKI, verified certificate revocation list distribution, and subscribed to 340 monitored items with a 1-second sampling interval. Total data volume from the OPC UA feed: approximately 28,000 readings per minute.

What Eight Weeks Actually Looks Like

Week 1-2: Site survey, register map audit, network architecture review, edge gateway procurement. The hardware lead time for the gateways was 6 business days; we ordered immediately after the site survey.

Week 3: Edge gateway installation, physical sensor mounting (current transformers, temperature sensors). Parallel: waiting on firewall change control.

Week 4-5: Modbus integration testing, OPC UA certificate provisioning, wireless sensor commissioning. Parallel: firewall change control still in review.

Week 6: Firewall approved. Connected gateways to on-premises analytics server. Data flowing from all three sources. Immediate finding: the centrifugal pump on Feed Line 2 was showing vibration signatures at 2.3× the baseline on its outboard bearing — a classic sign of developing bearing fault.

Week 7: Dashboard configuration, alert rule tuning, staff training. The bearing on Feed Line 2 was replaced during a scheduled maintenance window; visual inspection confirmed moderate wear on the outer race.

Week 8: Parallel operation (both old and new monitoring), final handoff, documentation.

The Numbers

Total sensor count: 1,247 (including 47 spares commissioned for future expansion). Data throughput at steady state: approximately 68,000 readings per minute across all three protocol sources. Alert delivery latency from sensor trigger to operator notification: average 1.4 seconds, 99th percentile 2.8 seconds. On-premises storage footprint: 14 months of full-resolution data in approximately 380GB on a standard NAS.

Within the first 90 days of operation, the team identified four maintenance items from sensor data — the bearing fault described above, an abnormal pressure drop across a sand filter indicating media fouling, an HVAC unit in the main electrical room running 12°C above design temperature, and a chlorine dosing pump with inconsistent flow rates traced to a faulty check valve.

The plant estimated the four findings, caught early, avoided roughly $180,000 in emergency repair costs and a regulatory compliance incident that would have triggered a consent order review.