line |
stmt |
bran |
cond |
sub |
pod |
time |
code |
1
|
|
|
|
|
|
|
package TimeSeries::AdaptiveFilter; |
2
|
|
|
|
|
|
|
|
3
|
|
|
|
|
|
|
=head1 NAME |
4
|
|
|
|
|
|
|
|
5
|
|
|
|
|
|
|
TimeSeries::AdaptiveFilter - Adaptive filter for data stream with possible outliers |
6
|
|
|
|
|
|
|
|
7
|
|
|
|
|
|
|
=cut |
8
|
|
|
|
|
|
|
|
9
|
|
|
|
|
|
|
our $VERSION = '0.03'; |
10
|
|
|
|
|
|
|
|
11
|
|
|
|
|
|
|
=head1 VERSION |
12
|
|
|
|
|
|
|
|
13
|
|
|
|
|
|
|
Version 0.03 |
14
|
|
|
|
|
|
|
|
15
|
|
|
|
|
|
|
=head1 STATUS |
16
|
|
|
|
|
|
|
|
17
|
|
|
|
|
|
|
=begin HTML |
18
|
|
|
|
|
|
|
|
19
|
|
|
|
|
|
|
|
20
|
|
|
|
|
|
|
|
21
|
|
|
|
|
|
|
|
22
|
|
|
|
|
|
|
|
23
|
|
|
|
|
|
|
=end HTML |
24
|
|
|
|
|
|
|
|
25
|
|
|
|
|
|
|
=cut |
26
|
|
|
|
|
|
|
|
27
|
1
|
|
|
1
|
|
14010
|
use strict; |
|
1
|
|
|
|
|
1
|
|
|
1
|
|
|
|
|
21
|
|
28
|
1
|
|
|
1
|
|
2
|
use warnings; |
|
1
|
|
|
|
|
1
|
|
|
1
|
|
|
|
|
24
|
|
29
|
|
|
|
|
|
|
|
30
|
1
|
|
|
1
|
|
3
|
use List::Util qw(sum max); |
|
1
|
|
|
|
|
3
|
|
|
1
|
|
|
|
|
77
|
|
31
|
|
|
|
|
|
|
|
32
|
1
|
|
|
1
|
|
3
|
use parent qw/Exporter/; |
|
1
|
|
|
|
|
2
|
|
|
1
|
|
|
|
|
6
|
|
33
|
|
|
|
|
|
|
|
34
|
|
|
|
|
|
|
our @EXPORT_OK = qw/filter/; |
35
|
|
|
|
|
|
|
|
36
|
|
|
|
|
|
|
=head1 SYNOPSYS |
37
|
|
|
|
|
|
|
|
38
|
|
|
|
|
|
|
use TimeSeries::AdaptiveFilter qw/filter/; |
39
|
|
|
|
|
|
|
|
40
|
|
|
|
|
|
|
# creation with defaults |
41
|
|
|
|
|
|
|
my $filter = filter(); |
42
|
|
|
|
|
|
|
|
43
|
|
|
|
|
|
|
# create filter with tuned parameters |
44
|
|
|
|
|
|
|
my $filter = filter({ |
45
|
|
|
|
|
|
|
floor => 6 |
46
|
|
|
|
|
|
|
cap => 0.2, |
47
|
|
|
|
|
|
|
lookback_capacity => 20, |
48
|
|
|
|
|
|
|
lookback_period => 4, |
49
|
|
|
|
|
|
|
decay_speeds => [0.03, 0.01, 0.003], |
50
|
|
|
|
|
|
|
build_up_count => 5, |
51
|
|
|
|
|
|
|
reject_criterium => 4, |
52
|
|
|
|
|
|
|
}); |
53
|
|
|
|
|
|
|
|
54
|
|
|
|
|
|
|
# usage |
55
|
|
|
|
|
|
|
my $now = time; |
56
|
|
|
|
|
|
|
$filter->($now, 100.002); # returns true, i.e. all data is valid on learning period |
57
|
|
|
|
|
|
|
$filter->($now + 1, 100.001); # returns true |
58
|
|
|
|
|
|
|
... # it learns form sample of 60 seconds |
59
|
|
|
|
|
|
|
$filter->($now + 60, 100.005); # returns true |
60
|
|
|
|
|
|
|
$filter->($now + 61, 99.9995); # returns true, as value does not differs much |
61
|
|
|
|
|
|
|
$filter->($now + 62, 10_0000); # returns false, outlier data |
62
|
|
|
|
|
|
|
$filter->($now + 63, 10.0001); # returns false, outlier data |
63
|
|
|
|
|
|
|
$filter->($now + 64, 100.011); # returns true, even if the sample is oulier, because |
64
|
|
|
|
|
|
|
# the filter rejected too much values, and has to |
65
|
|
|
|
|
|
|
# re-adapt to time seria again |
66
|
|
|
|
|
|
|
|
67
|
|
|
|
|
|
|
=head1 DESCRIPTION |
68
|
|
|
|
|
|
|
|
69
|
|
|
|
|
|
|
For the details of underlying mathematical model of the filter, configurable paramters |
70
|
|
|
|
|
|
|
and their usage, please, look at the shipped C folder. |
71
|
|
|
|
|
|
|
|
72
|
|
|
|
|
|
|
=cut |
73
|
|
|
|
|
|
|
|
74
|
|
|
|
|
|
|
my $sqrt2pi = 1 / sqrt(2 * atan2(1, 1)); |
75
|
|
|
|
|
|
|
|
76
|
|
|
|
|
|
|
sub filter { |
77
|
4
|
|
|
4
|
0
|
949
|
my $params = shift; |
78
|
|
|
|
|
|
|
|
79
|
|
|
|
|
|
|
############################## |
80
|
|
|
|
|
|
|
## filter tuning parameters ## |
81
|
|
|
|
|
|
|
############################## |
82
|
|
|
|
|
|
|
|
83
|
|
|
|
|
|
|
# minimum amount of values in lookback to take an decision. |
84
|
|
|
|
|
|
|
# Otherwise, input values will be accepted |
85
|
4
|
|
100
|
|
|
16
|
my $floor = $params->{floor} // 6; |
86
|
|
|
|
|
|
|
# maximum share of rejected input values. Upon hit, input values will be accepted |
87
|
4
|
|
50
|
|
|
11
|
my $cap = $params->{cap} // 0.2; |
88
|
|
|
|
|
|
|
# maximum amount ot input values in lookback |
89
|
4
|
|
50
|
|
|
11
|
my $lookback_capacity = $params->{lookback_capacity} // 20; |
90
|
|
|
|
|
|
|
# the retention period for lookback, i.e. max age for input values |
91
|
4
|
|
50
|
|
|
11
|
my $lookback_period = $params->{lookback_period} // 4; |
92
|
4
|
|
50
|
|
|
14
|
my $decay_speeds = $params->{decay_speeds} // [0.03, 0.01, 0.003]; |
93
|
4
|
|
100
|
|
|
9
|
my $build_up_count = $params->{build_up_count} // 5; |
94
|
4
|
|
50
|
|
|
11
|
my $reject_criterium = $params->{reject_criterium} // 4; |
95
|
|
|
|
|
|
|
|
96
|
|
|
|
|
|
|
######################## |
97
|
|
|
|
|
|
|
## enclosed variables ## |
98
|
|
|
|
|
|
|
######################## |
99
|
|
|
|
|
|
|
|
100
|
4
|
|
|
|
|
4
|
my @lookback; |
101
|
4
|
|
|
|
|
2
|
my $lookback_rejected = 0; |
102
|
4
|
|
|
|
|
3
|
my @minute; |
103
|
|
|
|
|
|
|
my $trust_w; |
104
|
0
|
|
|
|
|
0
|
my $ad; |
105
|
0
|
|
|
|
|
0
|
my @ads; # used on build stage only |
106
|
0
|
|
|
|
|
0
|
my $mad; |
107
|
0
|
|
|
|
|
0
|
my @mads; |
108
|
4
|
|
|
|
|
3
|
my $wsum = 0; |
109
|
4
|
|
|
|
|
2
|
my $csum = 0; |
110
|
4
|
|
|
|
|
4
|
my $vol; |
111
|
|
|
|
|
|
|
my $_accepted; |
112
|
|
|
|
|
|
|
|
113
|
|
|
|
|
|
|
# flag initicates, that we still need to accumulate enough time series |
114
|
|
|
|
|
|
|
# before doing actual filtering. |
115
|
4
|
|
|
|
|
3
|
my $build = 1; |
116
|
|
|
|
|
|
|
|
117
|
|
|
|
|
|
|
# resulting adaptive filter function |
118
|
|
|
|
|
|
|
my $fn = sub { |
119
|
213
|
|
|
213
|
|
35594
|
my ($epoch, $spot) = @_; |
120
|
|
|
|
|
|
|
|
121
|
|
|
|
|
|
|
# operating on nature log |
122
|
213
|
|
|
|
|
281
|
$spot = log $spot; |
123
|
|
|
|
|
|
|
|
124
|
|
|
|
|
|
|
# prevent loopback window overgrow |
125
|
213
|
|
100
|
|
|
808
|
while (@lookback > $lookback_capacity or @lookback > $floor and $lookback[0][0] < $epoch - $lookback_period) { |
|
|
|
66
|
|
|
|
|
126
|
178
|
|
|
|
|
147
|
my $leaving = shift @lookback; |
127
|
178
|
50
|
|
|
|
743
|
--$lookback_rejected unless $leaving->[3]; |
128
|
|
|
|
|
|
|
} |
129
|
213
|
|
|
|
|
137
|
my $accepted; |
130
|
|
|
|
|
|
|
|
131
|
213
|
100
|
|
|
|
240
|
if ($build) { |
132
|
|
|
|
|
|
|
# always accept the incoming value |
133
|
173
|
|
|
|
|
138
|
($trust_w, $accepted) = (1, 1); |
134
|
|
|
|
|
|
|
|
135
|
|
|
|
|
|
|
# gather absolute differences |
136
|
173
|
100
|
|
|
|
224
|
unless (@lookback < $build_up_count) { |
137
|
153
|
|
|
|
|
228
|
$ad = abs($spot - $lookback[-$build_up_count][1]) / sqrt($build_up_count); |
138
|
153
|
|
|
|
|
159
|
push @ads, $ad; |
139
|
|
|
|
|
|
|
} |
140
|
|
|
|
|
|
|
|
141
|
|
|
|
|
|
|
# build condition: the current tick is 60s newer |
142
|
|
|
|
|
|
|
# and we have enough absolute differences |
143
|
173
|
100
|
100
|
|
|
548
|
if ( @minute |
|
|
|
66
|
|
|
|
|
144
|
|
|
|
|
|
|
and $minute[0] < $epoch - 60 |
145
|
|
|
|
|
|
|
and @ads >= $build_up_count) |
146
|
|
|
|
|
|
|
{ |
147
|
1
|
|
|
|
|
1
|
$build = 0; |
148
|
1
|
|
|
|
|
6
|
my @new_ads = sort { $a <=> $b } @ads; |
|
263
|
|
|
|
|
149
|
|
149
|
1
|
|
|
|
|
3
|
my $cut = int(@new_ads / $build_up_count); |
150
|
1
|
|
|
|
|
6
|
@new_ads = @new_ads[$cut .. ($build_up_count - 1) * $cut]; |
151
|
1
|
|
|
|
|
9
|
$mad = sum(@new_ads) / @new_ads; |
152
|
1
|
|
|
|
|
3
|
@mads = ($mad) x scalar(@$decay_speeds); |
153
|
|
|
|
|
|
|
} |
154
|
|
|
|
|
|
|
} else { |
155
|
|
|
|
|
|
|
# prevent @minute overgrow |
156
|
40
|
|
66
|
|
|
118
|
while (@minute and $minute[0] < $epoch - 60) { |
157
|
41
|
|
|
|
|
108
|
shift @minute; |
158
|
|
|
|
|
|
|
} |
159
|
40
|
|
|
|
|
45
|
my $density = 60 / (1 + @minute); # the number "1", beacuse we account the current value too |
160
|
40
|
|
|
|
|
36
|
my $ha = $wsum / $csum; |
161
|
40
|
|
|
|
|
44
|
my $vol = $mad / $sqrt2pi; |
162
|
40
|
|
|
|
|
30
|
my $diff = abs($spot - $ha); |
163
|
|
|
|
|
|
|
|
164
|
|
|
|
|
|
|
# ther there is zero-difference, we accept current piece of data; |
165
|
|
|
|
|
|
|
# otherwise, if it is zero-volatility (flat), we reject it. |
166
|
40
|
50
|
|
|
|
58
|
my $reject = |
|
|
50
|
|
|
|
|
|
167
|
|
|
|
|
|
|
$diff |
168
|
|
|
|
|
|
|
? ($vol ? ($diff / $vol) : $reject_criterium + 1) |
169
|
|
|
|
|
|
|
: 0; |
170
|
|
|
|
|
|
|
|
171
|
40
|
|
|
|
|
35
|
$accepted = !($reject > $reject_criterium); |
172
|
40
|
|
|
|
|
86
|
$trust_w = 1 / (1 + ($reject / $reject_criterium)**8); |
173
|
|
|
|
|
|
|
|
174
|
40
|
50
|
66
|
|
|
59
|
if (not $accepted and $lookback_rejected > $cap * @lookback) { |
175
|
0
|
|
|
|
|
0
|
($accepted, $trust_w) = (1, 1); |
176
|
|
|
|
|
|
|
} |
177
|
40
|
|
|
|
|
62
|
$ad = abs($spot - $_accepted->[-$build_up_count][1]) / sqrt($build_up_count); |
178
|
40
|
50
|
|
|
|
51
|
if ($ad) { |
179
|
40
|
|
|
|
|
59
|
for my $idx (0 .. @$decay_speeds - 1) { |
180
|
|
|
|
|
|
|
# MAD (mean absolute deviation) |
181
|
120
|
|
|
|
|
111
|
my $mu = $trust_w * (1 - exp(-$density * $decay_speeds->[$idx])); |
182
|
120
|
|
|
|
|
127
|
$mads[$idx] = (1 - $mu) * $mads[$idx] + $mu * $ad; |
183
|
|
|
|
|
|
|
} |
184
|
|
|
|
|
|
|
} |
185
|
40
|
|
|
|
|
66
|
$mad = max(@mads); |
186
|
|
|
|
|
|
|
} |
187
|
213
|
|
|
|
|
181
|
push @minute, $epoch; |
188
|
213
|
|
|
|
|
300
|
push @lookback, [$epoch, $spot, $trust_w, $accepted]; |
189
|
213
|
100
|
|
|
|
213
|
if ($accepted) { |
190
|
212
|
|
|
|
|
213
|
push @$_accepted, [$epoch, $spot]; |
191
|
212
|
|
|
|
|
510
|
shift @$_accepted while @$_accepted > $build_up_count; |
192
|
|
|
|
|
|
|
} else { |
193
|
1
|
|
|
|
|
2
|
++$lookback_rejected; |
194
|
|
|
|
|
|
|
} |
195
|
213
|
50
|
|
|
|
259
|
if ($trust_w) { |
196
|
213
|
|
|
|
|
199
|
$wsum = 0.5 * $wsum + $spot * $trust_w; |
197
|
213
|
|
|
|
|
148
|
$csum = 0.5 * $csum + $trust_w; |
198
|
|
|
|
|
|
|
} |
199
|
213
|
|
|
|
|
1027
|
return $accepted; |
200
|
4
|
|
|
|
|
24
|
}; |
201
|
4
|
|
|
|
|
12
|
return $fn; |
202
|
|
|
|
|
|
|
} |
203
|
|
|
|
|
|
|
|
204
|
|
|
|
|
|
|
1; |