Beware of confirmation bias
There is a very strong chance that what you perceive as “almost always” is actually “slightly above if not right on what’s expected”.
Don’t trust your impressions!
It’s far more likely that he once had a streak of good luck and that has coloured your impression ever since - every time he rolls an 18-20 this sets the impression deeper in your head but you forget the far more frequent 1-17.
His perspective
I apologize if this section seems unduly harsh but it’s important that you look at this from the other side. I am concerned that you are “pretty sure” when you are almost certainly wrong and very concerned that you have acted on this erroneous surety to inflict harm on some else. Buckle up - this gets rough.
You accused him of cheating. Not only that, you recruited everyone else at the table to gang up on him. And you don’t trust him when he denies your unfounded accusations. Plus you want to screw with his lucky dice and everyone knows that if you do stuff like that, Lady Luck screws back!
First, I’m almost sure his dice are perfectly fine - not perfect because no dice isare and possibly not good enough that a casino would use them because they pay a premium for high-quality dice but good enough for a casual RPG. If there is something wrong with them, I’m almost sure he doesn’t know it. Put it this way - if I was on a jury, I would have no doubt he was innocent not just reasonable doubt.
What to do
Best option - Forget about it
His dice are almost certainly fine and even if they arearen't, he’s probably not cheating.
If he is cheating in a game of make-believe elves with nothing at stake, and he’s over 12 years old, then he deserves your kindness because he’s a sad, sad individual.
Remember, it’s far, far more likely that you owe him an apology.
If you really, really can't let it go - Write down his results
If you really can’t just take a deep breath and let it go and absolutely have to satisfy yourself that he’s not using unfair dice, jot down all his results on, say, his d20 from the next few sessions.
If you wait long enough you can get enough data to do a full chi-squared test on all 20 numbers but you can do it a lot quicker if you use say 4 “buckets” (1-5, 6-10, 11-15, 16-20) - you need a lot less data for 3 degrees of freedom than 19 - about 25 rolls or so. However, you probably won’t need to do a full test, just looking at the histogram will probably disabuse you of your bias.