Effectiveness of Hamming Single Error Correction

Codes under Harsh Electromagnetic Disturbances

Jonas Van Waes∗ , Jonas Lannoo† , Jens Vankeirsbilck∗ , Andy Degraeve† , Joan Peuteman† , Dries Vanoost† ,
Davy Pissoort† and Jeroen Boydens∗
∗ Dept.
of Computer Science
† Dept.
of Electrical Engineering
KU Leuven, Bruges Campus
Spoorwegstraat 12, 8200 Brugge
[email protected]

Abstract—Modern hi-tech systems rely heavily on communi- In previous work [10], the effectiveness of Cyclic Redun-
cation networks while operating in increasingly harsher electro- dancy Checks (CRCs) was investigated. Residing as a detection
magnetic conditions. To protect transmitted data from corruption, mechanism, it would generate up to 50% false positives under
Error Correcting Codes are widely used. In this paper, the
effectiveness of the Hamming code is evaluated under harsh the right conditions. False positives occur when the error
electromagnetic disturbances. Our simulations show that, under detection scheme assumes the data to be error free, while the
certain conditions, the impact of the introduced overhead cannot data differs from the originally sent data. This work considers
be compensated by the single error correcting capabilities of equally harsh environments, but studies error correction instead
the Hamming codes. Moreover, for specific bit and disturbance of detection. Furthermore, a transition is made from plane-wave
frequencies and for larger data sets, the use of a Hamming code
provides limited to no advantage. to reverberation room conditions.
Index Terms—Hamming Code, Electromagnetic Interference, Harsh and continuous electromagnetic disturbances can cause
Error Correction Code, Embedded Systems, Fault Tolerance, multiple upsets in the data. These disturbances are replicated
Resilience in an in-house built framework for simulating reverberation
room conditions. Different Hamming codes are considered,
I. I NTRODUCTION depending on the amount of data bits: 4, 11 or 26 bits. By
More and more electric, electronic and programmable elec- comparison to an unprotected baseline, whereby the data bits
tronic (E/E/PE) devices are being used in our everyday lives, are sent over the channel without any ECC, the effectiveness of
ranging from smartphones and laptops to systems fulfilling the different Hamming codes is investigated. The reverberation
mission- or safety-critical tasks. Applications such as au- chamber simulation framework is based on the Plane Wave
tonomous systems are being developed and, very often, robust Integral Representation for Reverberation Chambers [6].
communication channels are crucial. Communication channels The organization of this paper is as follows. The Hamming
in general are increasingly affected by Electromagnetic Inter- code is presented in Section II. The reverberation simulation
ference (EMI), most notably due to three trends: framework is described in Section III with the results given in
1) Internal voltages of the E/E/PE systems are constantly Section IV. Section V covers the conclusions of this work and
lowered to reduce power consumption and heat dissipa- the options for future work are presented in Section VI.
tion;
II. H AMMING C ODE
2) The decrease of the minimum feature sizes in order to in-
crease the transistor density, resulting in more processing The Hamming code is a linear, Single Error Correcting (SEC)
power per area and reduced heat generation; Code named after its inventor R.W. Hamming [5]. In mathemat-
3) Harsher ElectroMagnetic (EM) environments due to ev- ical terms, the code parameters are described by Equation (1),
ermore powerful and/or transmitting devices. where r is the number of redundancy (also called parity) bits,
The combination of these trends leads to a lowered intrinsic k the message (data bits) length and n the block (code word)
immunity to EMI, which is often perceived as voltages induced length. For further notation, these codes are abbreviated as
on the channel, possibly leading to bit errors. To revert these H[n, k, r].
corruptions, Error Detection Codes (EDCs) and Error Correc- 
tion Codes (ECCs) have been used since the fifties [5]. By r > 2

adding redundant information, error detection and/or recovery k = 2r − r − 1 (1)
 r
n=2 −1=k+r

is performed at receiver’s side. This methodology is known as
Forward Error Correction (FEC). A FEC code example is the In this work, three message length cases are considered:
Hamming Code [5]. 4, 11 and 26 bits of data, resulting in H[7,4,3], H[15,11,4]
and H[31,26,5], respectively. These are the smallest Hamming
codes for r ≥ 3. For r = 2 and k = 1, simple triplication is DATA ENCODER + DECODER DATA
obtained. As running example for the code generation, H[7,4,3]
is used, but each code is constructed in the same way. The codes EMI

have the layout as displayed in Equation (2). Dx are the data

bits and P x are the parity bits. SENDER SIDE CHANNEL RECEIVER SIDE
(Experiment Specific) (Reverberation Core) (Experiment Specific)
[D3, D2, D1, P 2, D0, P 1, P 0] (2) Figure 1: Framework Outline
The Hamming code is constructed using the following three
Read Configuration:
steps. E_Inc_Max
Loop_Max
1) Each position in the code which is a power of two, is the
position of a parity bit. Therefore, the parity bits are only E_Inc = 0

dependent on the data bits. The data bits themselves fill Encode Data
the other positions.
2) Initialize the parity bits to zero. Determine Word Error
F E_inc < E_Inc_Max
3) The parity bits themselves are calculated by a specific Rate per E_inc

algorithm which starts from the back (at P 0). For the T

P i parity bit: skip 2i bits, check 2i bits, skip 2i bits, Loops = 0 E_Inc+5

check 2i and so on, until the end of the code word is

reached. With this algorithm, the generated parity will Generate EMI-sine T Loops < Loop_Max F Save WER_Per_E_Inc

automatically be even. Note: for the first skip operation,

count one position outside of the coding array. In other
Data += EMI-sine Loops++
words, for the three parity bits of H[7, 4, 3], Equation (3)
T
can be used. P x denotes a parity bit and Dx are data
bits. 
Decode data
RX_data is equal to
original data ? F WER_Per_E_Inc++

P 0 = P 0 ⊕ D0 ⊕ D1 ⊕ D3

P 1 = P 1 ⊕ D0 ⊕ D2 ⊕ D3 (3) Figure 2: Framework Work Flow

P 2 = P 2 ⊕ D1 ⊕ D2 ⊕ D3


Since Hamming codes are linear, matrices can be used to 1


1 1 1 0 0 0

encode and decode Hamming codes. The generator matrix H= 1 1 0 0 1 1 0  (5)
G(k×n) is used for encoding. The data vector is multiplied by 1 0 1 0 1 0 1
G to generate the code word. In the correction and decoding  
process, three matrices are needed: the parity check matrix 1 0 0 0 0 0 0
H(r×n) , the syndrome matrix S(r×1) and the decoding matrix  0 1 0 0 0 0 0 
R=  (6)
R(k×n) . Multiplying H with the received code word yields S.  0 0 1 0 0 0 0 
When a correct code word was received, the syndrome S is 0 0 0 0 1 0 0
an all-zero matrix. If an error is present in the code word,
III. R EVERBERATION ROOM S IMULATION
the syndrome S indicates the bit position of the error. After
correction, the code word is decoded. This section covers the in-house built reverberation room
Note that if two or more errors reside in the code word, a non- simulation framework used to test the effectiveness of Ham-
zero syndrome is obtained as in previous paragraph. However, ming code protection. Its global outline is depicted in Fig. 1.
only one bit-error can be corrected. A successive multiplication The core of the framework is described first: the channel.
between the code word and the parity check matrix, will yield Next, both sender and receiver sides are presented. These form
an S = 0, indicating an error free code word. At this point, the additions to the core. Finally, the specific parameters of the
the code word is regarded as error free, but there are still experiments are denoted.
uncorrected errors.
For all Hamming codes it is mandatory that H × GT yields A. Reverberation Core
an all-zero matrix. The following matrices (Equations (4) to (6)) The core reverberation simulation uses a numerical approach
are used to yield the code presented in Equation (2). to simulate the reverberation room conditions while avoiding
many full-wave simulations. The results within this simulation
framework are based on the superposition of a sufficiently large
 
1 0 0 1 0 1 1
 0 1 0 1 0 1 0  set of randomly chosen plane waves to statistically represent
G=
 0
 (4) the behavior a reverberation chamber [7]. The generic process
0 1 1 0 0 1 
0 0 0 0 1 1 1 of the reverberation chamber conditions is visualized in Fig. 2.
 Three specific parameters should be known at the start: B. Coding Modules
the maximum strength of the electromagnetic disturbance In essence, the core in Section III-A only introduces noise
(EincM ax ), the amount of repetitions to be made (Loopmax ) (EMI) to the considered transmission line. The core was sur-
and the considered code word (output of the encoder). The start rounded with data encoder and decoder modules. Two different
value for Einc was chosen as 0 V/m, simulating an EMI-free encoder modules are used; one simply converting logic ’0’ and
environment. For each of the considered strengths, Loopmax ’1’ to voltages (thus without any data protection schemes) and
repetitions are performed. In each repetition, an undisturbed one with the Hamming code.
code word is considered; there is no accumulation of EMI. After The voltages themselves (and thresholds) depend on the
the generation of the disturbance, it is added to the encoded initial configuration of the framework. The encoding voltage Vs
data. The disturbed data is then decoded and checked for errors. (at the source) is 1 V, i.e. Non-Return to Zero Level (NRZ-L)
If an error is present, the error counter (W ERP erEinc ) is coding. Due to the assumption of matched transmission lines
incremented. In the end, this counter is divided by Loopmax (the characteristic impedance of the transmission line equals
to yield the Word Error Rate (WER) per strength of Einc as both the source and load impedance), the voltage incurred at
denoted in Equation (7). When all repetitions are completed, the end of the transmission line is half the original voltage. The
the next iteration is started with an increased Einc . When Einc voltage seen by the receiver is therefore Vrec = Vs /2. The core
equals EincM ax , the simulation is completed. adds the reverberation EMI disturbance before the code word
is decoded (see Fig 1).
T otal W rong Data W ords Similarly, the thresholds for decoding are a function of Vs as
W ERP erEinc = (7)
Loopmax well. Since a NRZ-L is used, the receiver assumes that anything
above half the maximum theoretical receivable voltage is a ’1’,
In order to calculate the induced voltages on the ports of a below it is considered ’0’. The threshold lies thus at Vth =
Device Under Test, many full-wave simulations are required: Vs /4.
one simulation for each incident plane wave with a specific All other parameters needed in this framework, are specific
angle of incidence and polarization. As this would require a sig- to the conducted experiments. These parameters are described
nificant amount of simulating time, an improved algorithm was in the following Section.
developed [9]. The improved algorithm only requires one full-
wave simulation, and the induced voltages on the ports of the C. Experiment Specific Parameters
DUT are calculated via a reciprocity-based methodology. Since Six experiment specific parameters are needed: the data, the
this paper considers reverberant conditions, a set of random bit frequency Fbit , the disturbance frequency Femi , the maxi-
plane waves is required to create this in accordance to the Plane mum number of repetitions Loopmax , the maximum considered
Wave Integral Representation [6]. From that representation, we EMI strength EincM ax and the difference between considered
know that the resulting induced voltages and currents follow a strengths ∆Einc .
Rayleigh distribution.. Since our DUT satisfies all requirements For the initial experiments presented here, a random 4-bit
from [6], the parameter σ can be calculated, characterizing the data sequence was generated. For the larger size experiments,
Rayleigh distribution. From this distribution, the amplitude of the bit stream was expanded to the desired length (11 or 26
the induced voltages can be calculated. bits). Equation (9) provides the actual data used in the experi-
The amplitude A itself is based on an EMI strength of 1 V/m, ments. For further experiments, all data sets of the prescribed
so multiplication with the incoming strength Einc must be lengths are to be simulated.
performed. Furthermore, a random phase shift ϕ is added to 
account for the arrival times of the incident waves. The phase


 D(4) = {1, 1, 0, 0}
D(11) = {1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1}

shift ϕ is uniformly distributed between −π and π. Equation (8)


represents the induced voltage U (i) at the ports of the DUT. D(26) = {1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, (9)

1, 1, 1, 1, 1, 1, 0, 1, 0,





 1, 0, 0, 1, 0, 0}
U (i) = Einc × A × sin [2π × Femi × Ti + ϕ] ; The values chosen in these experiments for Fbit , Femi and
i 1 EincM ax are based on previous work, published in [2] and [3].
where Ti = +
Fbit 2 × Fbit The values used are depicted in Equation (10).
and i ∈ [0, k + r − 1] (8)


 Femi ∈ [200 M Hz; 5 GHz] | ∆Femi = 200 M Hz
U (i) also depends on the considered disturbance frequency 
F = {197; 200; 211; 1000} M Hz
bit
Femi and bit frequency Fbit , in order to maintain the time (10)
Einc ∈ [0; 10.000] V /m | ∆Einc = 5 V /m
dependencies between both frequencies. At receiver side, each 

Loopmax = 100.000

bit is sampled according to Fbit in the middle of the bit, and
not in the beginning of the bit. Therefore, the sample time Ti The values for ∆Einc and Loopmax were chosen at the
from Equation (8) considers half a bit period extra. beginning of the experiments.
 IV. R ESULTS susceptible to corruption outweighs the possibility to correct
In total, 300 figures were generated (4 Fbit × 25 Femi × a single bit error. The difference between the baseline and
3 DataSets), therefore not all figures for the WER are pre- Hamming curves on Fig. 4a are regarded as the percentage of
sented in this paper; a selection was made. The results in the single bit errors in the data. The percentage below the Hamming
subsections are grouped according to the data-set length. curve, are multi-bit errors.
Fig. 4b shows again that, when Femi is an integer multiple of
A. 4-bit data Fbit , Hamming produces the same WER as unprotected data.
The first result is shown in Fig. 3a. This is the expected As discussed in Section IV-A, this is due to the generation of
outcome, where Hamming code is able to correct single bit all-zero or all-one code words. The Hamming decoder considers
errors and the actual WER is slightly lower. For the parameters the code words as correct, but they differ from the sent data.
specified in the caption of the Figure, the introduced overhead The mean (d) and maximum deviation (dmax ) between the
of Hamming results in a better WER. The difference in WER two curves can be calculated and yield (Equation 11):
is the percentage of corrected single bit errors in the data. The
addition of the Hamming parity bits, result in an increased  Xp
ability to recover errors.  (W ERN P (i) − W ERH (i))2
d¯ = (11)
In contradiction, Fig. 3c show Hamming performing worse p 2000
dmax = M ax( (W ERN P (i) − W ERH (i))2

than sending the data without protection. This means that in a
number of repetitions, the unprotected data is intact, but there where the i represents the value for each considered EMI
are errors in the Hamming parity bits such that the correction strength at a specific frequency of Fbit and Femi . There
mechanism is unable to correct. When the code word with is a mean deviation of 0.1083% between the curves and a
multiple bit errors is processed by the decoder, the parity check maximum of 0.554%. Given the nature of the simulation, this
matrix will identify one bit as false. That bit will be flipped, is identified as being equal. Note that the Hamming code is
resulting in a correction, but an incorrect one. No double error able to compensate the introduced overhead by correcting the
detection is in place, thus this exposes the disadvantage of received data blocks; the raw correct data throughput is higher.
Hamming.
It should be noted that Fig. 3b includes both situations C. 26-bit data
described in the previous paragraphs. For simulations with
Einc < 2000 V /m, Hamming provides an advantage in terms Now that 26 data bits are encoded with a H[31,26,5], Ham-
of WER. This advantage however, is lost upon consideration ming is rapidly losing its advantages. Since a code word is 31
of larger incident reverberation waves. bits long, the probability of having only one bit-flip in the data
For the specific situation depicted in Fig. 3d, where Femi is has been reduced to almost zero for great strengths of Einc . The
an integer multiple of Fbit , there is no difference in the word WER curves for unprotected and Hamming encoded data are
error rates at all. Depending on the generated wave amplitude almost equal. Using the formula provided in Equation (11), the
of the Rayleigh distribution, the disruption is large enough to mean deviation is 0.223% and the maximum deviation is 2.12%.
generate all-zero or all-one code words (which will result in Note that the mean deviation is about double the percentage
wrong data) or a disruption not large enough to corrupt the that was considered equal in Section IV-B. The overhead for a
data, therefore yielding correct data). Both code words are valid H[31,26,5] is about 19% (5 parity bits for each 26 bits of data)
according to the Hamming decoding logic, but are different and the maximum deviation is 2.12%. Therefore, the actual
from the data originally sent. In essence, the same problem of correct data block rate of Hamming is lower. Performance-wise,
false positives (as in [10] with CRC) arises here: data is seen sending the unprotected data is the best option. However, since
as correct while it is not. only 18.88% of the received data blocks are correct, this is not
Figures for an Fbit of 1000 MHz are not provided. The results acceptable in real-life applications unless there are other error
are similar to the ones provided in Fig. 3. Also note here that detection schemes. The Hamming code considered here cannot
Femi is an integer multiple of Fbit for specific frequencies, distinguish a recoverable single error and an unrecoverable
which yields the same graph as in Fig. 3d. multi-bit error. To recover multi-bit errors, Hamming cannot be
used: other types of code are required, such as Reed-Solomon
B. 11-bit data Codes [4], Bose–Chaudhuri–Hocquenghem Codes [1], etc.
This set of experiments shows a much more stable situation Extended Hamming codes are linear codes based on the orig-
than in Section IV-A: none of the graphs shown in Fig. 4 depicts inal Hamming code, but do not follow the specific requirements
a Hamming code yielding worse WER than the unprotected set in Equation (1). An example is the code developed in [8],
data. The relative overhead of parity bits is lower than for 4 bits which can correct a single error, while detecting double and
data. It is unlikely that the unprotected data is unaffected, while triple adjacent bit errors. To counter harsh EM environments,
only the parity bits are corrupted. Furthermore, the presence these codes might prove better and are to be considered in future
of double (or more) bit errors (non-correctable by Hamming) experiments. It should be noted that these Extended Hamming
are likely to have affected the original data bits as well. At Codes can provide multi-bit error detection, but are always
this point, the disadvantage of the added parity bits being Single Error Correction only.
 102 102
Baseline Baseline
Hamming Hamming
101 101
Word Error Rate (%)

Word Error Rate (%)

100 100

10−1 10−1

10−2 10−2

10−3 10−3
101 102 103 104 101 102 103 104
Einc (V/m) Einc (V/m)

(a) Fbit = 197 MHz, Femi =2200 MHz (b) Fbit = 197 MHz, Femi =4200 MHz

102 102
Baseline Baseline
Hamming Hamming
101 101
Word Error Rate (%)

Word Error Rate (%)

100 100

10−1 10−1

10−2 10−2

10−3 10−3
101 102 103 104 101 102 103 104
Einc (V/m) Einc (V/m)

(c) Fbit = 211 MHz, Femi =1000 MHz (d) Fbit = 200 MHz, Femi =1000 MHz

Figure 3: Results of the 4-bit experiments

102 102
Baseline Baseline
Hamming Hamming
101 101
Word Error Rate (%)

Word Error Rate (%)

100 100

10−1 10−1

10−2 10−2

10−3 10−3
101 102 103 104 101 102 103 104
Einc (V/m) Einc (V/m)

(a) Fbit = 197 MHz, Femi =3400 MHz (b) Fbit = 200 MHz, Femi =3400 MHz

Figure 4: Results of the 11-bit experiments

 102 102
Baseline Baseline
Hamming Hamming
101 101
Word Error Rate (%)

Word Error Rate (%)

100 100

10−1 10−1

10−2 10−2

10−3 10−3
101 102 103 104 101 102 103 104
Einc (V/m) Einc (V/m)

(a) Fbit = 211 MHz, Femi =1800 MHz (b) Fbit = 1000 MHz, Femi =1000 MHz

Figure 5: Results of the 26-bit experiments

Furthermore, for continuous EMI, other strategies such as in- Secondly, more ranges for Femi and Fbit will be evaluated.
terleaving might prove useful. However, when the EMI persists Building on the increased research of Femi and Fbit , WER-
beyond the interleaving time interval, other measure must be curve characterization will be attempted.
taken as well. As a third option, more ECCs can be simulated and verified.
In essence, the used technique must be able to overcome the Migrating beyond Hamming codes, multi-bit ECCs are to be
induced overhead in terms of correction before usefulness can simulated. Line coding is another variable to consider. Besides
be claimed. However, each code has its limits and needs careful NRZ-L, a multitude of encoding schemes exist, for example
consideration. Manchester-encoding.
V. C ONCLUSIONS R EFERENCES
This paper presented the effectiveness of Hamming codes [1] R.C. Bose and D.K. Ray-Chaudhuri. On a class of error correcting binary
group codes. Information and Control, Elsevier, 3(1):68–79, March 1960.
operating in harsh EM environments. Resulting from our in- [2] A. Degraeve and D. Pissoort. Study of the effectiveness of spatially em-
house built framework, multiple conclusions could be drawn. diverse redundant systems under plane-wave illumination. In Proc. Asia-
In terms of WER, it is possible to achieve a slightly increased Pacific Int. Symp. Electromagnetic Compatibility (APEMC), volume 01,
pages 211–213, May 2016.
performance. The results show a difference in WER of the [3] A. Degraeve and D. Pissoort. Study of the effectiveness of spatially em-
Hamming code compared to unprotected transmissions. The diverse redundant systems under reverberation room conditions. In Proc.
results showed that special consideration is in order when IEEE Int. Symp. Electromagnetic Compatibility (EMC), pages 374–378,
July 2016.
the disturbance frequency is an integer multiple of the bit [4] P. M. Ebert and S. Y. Tong. Convolutional reed-solomon codes. The Bell
frequency. In practice this can arise as harmonics of the System Technical Journal, 48(3):729–742, March 1969.
sending frequency of another, similar device. At this point, [5] R. W. Hamming. Error detecting and error correcting codes. The Bell
System Technical Journal, 29(2):147–160, April 1950.
for large disturbances, the code words are easily transformed [6] D. A. Hill. Plane wave integral representation for fields in reverber-
into all-one or all-zero data words. Those code words are ation chambers. IEEE Transactions on Electromagnetic Compatibility,
then regarded as correct by the receiver. The effectiveness 40(3):209–217, August 1998.
[7] Kent Rosengren and Per-Simon Kildal. Study of distributions of modes
of the implemented Hamming code significantly drops under and plane waves in reverberation chambers for the characterization of
these conditions. Moreover, those harsh disturbances introduce antennas in a multipath environment. Microwave and Optical Technology
multiple bit errors, to which a Hamming code is not the best Letters, 30(6):386–391, 2001.
[8] A. Sanchez-Macian, P. Reviriego, and J. A. Maestro. Hamming sec-daed
solution. Harsh electromagnetic environments clearly show the and extended Hamming sec-ded-taed codes through selective shortening
limits (and operational conditions) of the Hamming code. In and bit placement. IEEE Transactions on Device and Materials Reliabil-
operational systems, increased efforts are needed to protect ity, 14(1):574–576, March 2014.
[9] F. Vanhee, D. Pissoort, J. Catrysse, G. A. E. Vandenbosch, and G. G. E.
the transmitted data from corruption. Examples thereof are Gielen. Efficient reciprocity-based algorithm to predict worst case induced
Double or Triple Error Correction (DEC/TEC) Codes, possibly disturbances on multiconductor transmission lines due to incoming plane
supplemented with interleaving of data. waves. IEEE Transactions on Electromagnetic Compatibility, 55(1):208–
216, February 2013.
VI. F UTURE W ORK [10] J. Van Waes, J. Lannoo, A. Degraeve, D. Vanoost, D. Pissoort, and
J. Boydens. Effectiveness of cyclic redundancy checks under harsh
This work only considered one specific data stream as noted electromagnetic disturbances. In Proc. Int. Symp. Electromagnetic Com-
in Equation (9). In future work, the same comparisons will need patibility - EMC EUROPE, pages 1–6, September 2017.
to be made to verify the claims are applicable to all data sets.