Electronic – Using Ethernet Bus

In 10BASE5 and 10BASE2, a link relation (between whom?) cannot be resolved at the physical layer because there is no physical (i.e. "inside" the corresponding IEEE 802.3 Physical sublayer) link ("logical connection/association") at all here, despite the presence of a link ("hardwired connection"), typically 1:n as it is a bus, at the medium (i.e. by the coax itself). Therefore, a special protocol operating over the MAC sublayer was designed for, including but not limited to, resolve such a link. It was the Logical Link Control (LLC) protocol forgotten later as TCP/IP became popular.

In other words, in 10BASE5 and 10BASE2 you can state that a local station is linked with a distant station only by a higher layer association activated/supported, e.g. a TCP session established between the stations. At the physical layer itself, a station can observe only packet activity/collision: this is CDMA/CD in its pure "reincarnation".

It seems logical because on a point-to-point Ethernet, a link shows only that a station is already connected to a distant, compatible physical layer, knowing nothing about what is it: a switch, a terminal device, or a test loop, i.e. the (physical) link itself does not make you informed about whose MAC or higher (IP, socket) entities are your counterparts in this topology (virtual bus) too and to resolve are the selected stations linked or not, a higher layer association (identified by at least pair of MAC/IP addresses) is also needed.

You are correct in that surges are a concern for cascade failures; it is also true that USB is not galvanically isolated by default, and in fact is a pain in the rump to isolate due to its half-duplex, bidirectional, differential-but-not-always (blast the SE0!) nature. Thankfully, the good folks at Analog Devices stepped up and put together the ADuM4160 to take care of all of the truly hard parts of USB isolation -- while it only supports low or full speed operation, and cannot pass low/full speed negotiation signals, meaning it can't be built into a general purpose isolated USB "lump" without help, it is still the closest thing to a generic USB isolator available.

Assuming that a reliable 5V supply to the hubs is available, and preferably one 5V supply per hub (not hard!), I would split the hubs into 4 ports each and isolate the upstream port for each hub -- pick your favorite hub chipset here. Along with good per-port surge suppression, for which I'd specify a pair of Bourns TBU devices in series with D+/D- and a shunt low-capacitance, high-energy suppressor in addition to a MOV and TVSS clamp + current limiter or PPTC for Vbus and the obligatory TVSS/clamp network for D+/D-, this should limit the propagation of surges through the USB network.