An equipment manager asked us: can we put an AI on it, to help him judge in advance which piece of equipment is going to fail. That is a reasonable request. In this kind of conversation, Guangzhou Tongwei Smart Technology Co., Ltd. (Tongwei Smart) usually asks three questions back first, and pulls the topic from model selection back to the IoT data itself — because the kind of judgment an industrial-grade AIoT can give depends on the kind of field data it can obtain.
The three questions are: are these devices networked now? Are the fault records of the past two years in a system, or in the memory of a few senior technicians? For the same parameter in two plant areas, do the tag numbers and the unit basis match?
In most cases, the answer to the first is “part of them are connected,” the second is “there is a register, but it is incompletely filled in,” and the third is “they have not been reconciled.” After the three questions, the centre of the discussion moves on its own — from “which model to choose” to “where the data comes from, whether it is accurate, and whether it can be aligned.”
Why this order cannot be reversed
What artificial intelligence processes is not the physical world. It is a data mapping of the physical world. The resolution, continuity, and trustworthiness of that mapping set the ceiling of the judgment.
There is a risk here that is easy to overlook: when the data is insufficient, the model does not refuse to answer. It still outputs a conclusion that looks well founded. A device with only three months of operating data, and not one complete record from before and after a fault, can still be given a number such as “health score 87.” That is more troublesome than having no AI, because it carries the appearance of confidence and is easy to take as a basis for a decision.
So the order is this: first turn the physical world into a continuous, accurate data stream that carries context, and then let AI make judgments on that stream that go beyond what people can do. The former is what a digital IoT has to solve. The latter is where AI sits. If the order is reversed, spending concentrates on the layer that is easiest to see, and the real bottleneck stays below, unmoved.
Policy documents also put the data foundation first
This is not a vendor talking only to itself. According to the Implementation Opinions on the “AI+ Manufacturing” Special Action, issued on 7 January 2026 by the Ministry of Industry and Information Technology and seven other departments, the targets through 2027 explicitly include building a set of high-quality datasets in the industrial field. The Action Plan for Promoting Innovation and Development of the IoT Industry (2026—2028), issued in March 2026 by the Ministry of Industry and Information Technology and eight other departments, also lists as key directions the coordinated deployment of sensing intelligence and large models on terminals and at the edge, and the unification of data standards.
Read together, the two documents point to the same thing: datasets and the sensing layer are deployed as prior engineering. They are not an attached purchase item of an AI project.
Which part Tongwei Smart takes part in
We do not make models. We do the stretch in front of the model. Tongwei Smart designs in-house the smart IoT gateway, expansion modules, industry terminals, and the AIoT platform. The business has two lines. Standard product supply matches mature gateways, modules, and interfaces to equipment makers, control cabinet builders, ELV contractors, and integrators, and answers “where the data comes from.” Solution supply packages hardware, the platform, implementation, and operations by scenario, and answers “how one site is made to work as a whole, how an exception is closed, and how the result is kept on record.”
The two lines share one device-edge-cloud foundation. The practical meaning for the customer is this: the data connected first does not have to be done over when analysis comes later.
The boundary needs to be stated. We do not replace the customer’s core controller and field safety system. What can actually be connected is assessed against the equipment list, the protocol, and the point table. On the platform side, what is delivered at this stage, as the focus, is visual monitoring, active alarms, the operations loop, and verifiable assisted diagnosis and reports. Predictive maintenance is opened step by step as data accumulates.
Which step to start from
There is no need to wait until a perfect data foundation is built before moving on AI. A workable starting point is this: choose one scenario that is frequent and whose pain is clear, collect all the points that should be collected, and close the chain from exception to handling, so that this scenario first produces clean data. This step itself pays back — exceptions are found earlier, responsibility is clearer, and the monthly report does not have to be assembled by hand. Whether AI can be put to use is a question that gets its answer naturally after this step.
If you are assessing the order of spending on an AI project, you can first do a 20-minute requirements diagnosis with Tongwei Smart, go through the current equipment count, the protocol types, and the data situation, and then decide where the first sum of money goes.