# CNOSSOS-EU railway source emission

> One page of the phonometry documentation, as Markdown. Index: https://docs.jmrp.io/phonometry/llms.txt

Canonical: https://docs.jmrp.io/phonometry/environment/sources/cnossos-rail-emission/
Language: en
Alternate: https://docs.jmrp.io/phonometry/es/environment/sources/cnossos-rail-emission/index.md
Version: phonometry 4.0.0
Install: pip install "phonometry[full]"
Author: José Manuel Requena Plens (https://orcid.org/0000-0003-1250-6212)
Published: 2026-07-30
Updated: 2026-11-16T21:36:14Z
Standards: OJ L 5, 10.1.2018, p. 35; Directive 2002/49/EC
Verified: 12 conformance checks (Directive (EU) 2015/996; Directive (EU) 2021/1226; Directive 2002/49/EC) https://docs.jmrp.io/phonometry/reference/conformance/
License: MIT (https://github.com/jmrplens/phonometry/blob/main/LICENSE)
Summary: CNOSSOS-EU railway emission (Directive 2002/49/EC Annex II): roughness, contact filter, impact, squeal, traction, aerodynamic noise and bridges.

---

A road is one source line 5 cm above the pavement. A railway track is **two**,
one at the axles and one at the roof, because a train radiates from places that
are four metres apart and shield each other differently. Section 2.3 of
**Annex II of Directive 2002/49/EC**, as replaced by the CNOSSOS-EU methods,
builds both of them from a roughness spectrum, a handful of transfer functions
and the traffic flow.

This guide covers the railway source, section 2.3 of Annex II with the
coefficient database of Appendix G. The road source (2.2, Appendix F) and the
CNOSSOS-EU propagation model (2.5) are separate.

## 1. Which text is implemented

Annex II is a moving target, and on the railway side the layering is worse than
on the road side: the corrigendum replaced the **whole** of Appendix G, and the
2021 amendment then replaced five of its seven tables again and rewrote three
pieces of the method.

| Instrument | What it does to the railway source |
| :--- | :--- |
| Commission Directive (EU) **2015/996** | Replaces the whole of Annex II. Supplies formulae 2.3.1 to 2.3.18 and Tables G-1 to G-7. |
| **Corrigendum** of OJ L 5, 10.1.2018 | Replaces the **whole of Appendix G** (dropping its Min/Max validity columns and adding the descriptor codes to the column headers) and corrects the roughness conversion `f = v/lambda`, where the published text says the speed is in km/h. It is in **metres per second**: read literally, the printed formula is wrong by a factor 3.6. |
| Commission Delegated Directive (EU) **2021/1226** | Replaces Tables **G-1b, G-2, G-3a, G-4 and G-7** in their entirety; substitutes a **new curve-squeal rule**; makes bridge noise a **separate source** built on a transfer function instead of a constant; rewrites the **vertical directivity** of source A; corrects three band labels and one pair of values in Table G-5. |

The library implements the consolidated result, and every shipped table records
in its comment which instrument it came from. Three of the 2021 changes are
behavioural, not cosmetic, so a comparison with a pre-2021 study has to account
for them: the rail roughness of an average network is now much higher at long
wavelengths, the bridge correction has a spectrum instead of a flat 1 dB or
4 dB, and the vertical directivity is identically zero below the horizon where
the old form gave a positive correction.

## 2. Two source lines, and what sits on each

The method places **two** incoherent source lines at the centre of the track,
referred to the plane tangent to the two upper rail surfaces:

| Height | Index | Physical sources allocated to it |
| :--- | :--- | :--- |
| **0.5 m** (source A) | $h = 1$ | Rolling noise (rail, wheel and freight superstructure), impact noise, curve squeal, bridge noise, and the traction and aerodynamic components that sit low: gears, motors, shrouds and screens. |
| **4.0 m** (source B) | $h = 2$ | The traction and aerodynamic components that sit high: diesel exhausts, roof apparatus, the pantograph recess. |

Nothing in the method decides the traction split for you. Table G-5 tabulates
the traction sound power **separately for the two heights**, so the split is
read from the database rather than assumed; the Directive only adds that where
a real component sits between the two model heights, its energy is distributed
proportionately over them.

![Two panels of the CNOSSOS-EU railway source geometry. Section: a ballasted double-rail track of 1.435 m gauge with the plane tangent to the two upper rail surfaces drawn as the datum, source A on the track centre 0.5 m above it and source B 4.0 m above it, each listing the components allocated to it, a receiver 25 m away and 4 m above ground with a line of sight to each source, and the vertical angle psi dimensioned from the horizontal plane through the source with the half space below it shaded and labelled as the region where the vertical correction of source A is identically zero. Plan: the track axis, the receiver bearing, the horizontal angle phi dimensioned from the track axis and the dipole polar drawn to scale with 0 dB broadside and minus 20 dB along the track](https://docs.jmrp.io/phonometry/media/diagram_cnossos_rail.svg)

### From a real line to the descriptors

The **descriptors** of Tables [2.3.a] and [2.3.b] are the interface between a
national rolling-stock and track register and the shipped tables: each digit of
the code selects one of them. A vehicle carries four digits and a track six.

| Vehicle digit | What it selects | Codes | Library |
| :--- | :--- | :--- | :--- |
| 1 | Vehicle type | `h` high speed (> 200 km/h), `m` self-propelled passenger coach, `p` hauled passenger coach, `c` city tram or light metro, `d` diesel loco, `e` electric loco, `a` generic freight, `o` other | Chooses the `TractionVehicle` row of Table G-5 and whether the freight superstructure term applies |
| 2 | Number of axles per vehicle | the actual number | `RollingStock(n_axles=...)`, the $10\log_{10} N_\mathrm{a}$ of (2.3.8) |
| 3 | Brake type | `c` cast-iron block, `k` composite or sinter metal block, `n` non-tread braked (disc, drum, magnetic) | `BrakeType.CAST_IRON` / `.COMPOSITE` / `.NON_TREAD` → the wheel roughness of G-1a |
| 4 | Wheel measure | `n` none, `d` dampers, `s` screens, `o` other | Not a table selector in Appendix G |

| Track digit | What it selects | Codes | Library |
| :--- | :--- | :--- | :--- |
| 1 | Track base | `B` ballast, `S` slab, `L` ballasted bridge, `N` non-ballasted bridge, `T` embedded, `O` other | With digit 3, the `TrackTransferClass` of G-3a; `L`/`N` also switch on the bridge source of section 8 |
| 2 | Railhead roughness | `E` well maintained and very smooth, `M` normally maintained, `N` not well maintained, `B` not maintained and bad | `RailRoughnessClass.WELL_MAINTAINED` / `.NORMAL` / `.NOT_WELL_MAINTAINED` / `.BAD` → G-1b |
| 3 | Rail pad type | `S` soft (150-250 MN/m), `M` medium (250-800), `H` stiff (800-1 000) | The second half of the `TrackTransferClass` label: `MONOBLOCK_MEDIUM` is "M/M" |
| 4 | Additional measures | `N` none, `D` rail damper, `B` low barrier, `A` absorber plate on slab, `E` embedded rail, `O` other | Outside the emission model |
| 5 | Rail joints | `N` none, `S` a single joint or switch, `D` two per 100 m, `M` more than two per 100 m | The joint density $n_l$ of (2.3.12) |
| 6 | Curvature | `N` straight, `L` 1 000-500 m, `M` below 500 and above 300 m, `H` below 300 m | The radius passed to `curve_squeal_excess` |

Two selectors have no descriptor digit at all and have to come from the vehicle
data sheet: the **wheel diameter** (`WheelDiameter`, which picks the wheel
transfer function of G-3b) and the **contact filter** (`ContactFilter`, an axle
load in kN paired with that diameter, Table G-2). The snippets on this page name
the enum members directly for exactly that reason.

When the register does not answer, the method itself runs out: roughness classes
`N` and `B` carry no spectrum in Appendix G and must be supplied by the Member
State, and the same is true of a track system that matches no G-3a row: there
is no default and no interpolation, so the honest step is to declare the
substituted spectrum alongside the result.

For each 1/3-octave band the flow equation turns a single-vehicle sound power
into a power per metre of line (2.3.2):

$$
L'_{W,\mathrm{eq,line},i}(\psi,\varphi) = L_{W,0,\mathrm{dir},i}(\psi,\varphi)
  + 10\,\log_{10}\!\left(\frac{Q}{1000\,v}\right),
$$

or, for a stationary vehicle, the same directional power referred to the idling
time over the reference period (2.3.4):

$$
L'_{W,\mathrm{eq,line},i}(\psi,\varphi) = L_{W,0,\mathrm{dir},i}(\psi,\varphi)
  + 10\,\log_{10}\!\left(\frac{T_\mathrm{idle}}{T_\mathrm{ref}\,L}\right).
$$

The second replaces the first; it does not add to it. $Q$ is in vehicles per
hour, $v$ in km/h, $L$ the length of the track section in metres, and
$T_\mathrm{idle}$ and $T_\mathrm{ref}$ share whatever unit of time they are
given in. Note the unit: the flow term uses km/h, while the roughness conversion
two sections down uses m/s. Both are correct, and mixing them up is the single
most expensive mistake available in this model.

![Octave-band source-line power of a conventional main line: grey bars for the two source heights combined, a blue marker line for source A at 0.5 m and a red one for source B at 4.0 m, with source A twenty to thirty decibels above source B in every band](https://docs.jmrp.io/phonometry/media/cnossos_rail_emission.svg)

*At 160 km/h the track is source A almost everywhere: the 4.0 m line only
carries the traction noise of the unit, some twenty-five decibels below. It is
above 200 km/h, when aerodynamic noise appears at the pantograph, that source B
starts to matter.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt
from phonometry import environment

stock = environment.RollingStock(
    n_axles=4,
    wheel_roughness=environment.wheel_roughness(environment.BrakeType.NON_TREAD),
    contact_filter=environment.contact_filter(environment.ContactFilter.LOAD_50_DIAMETER_920),
    wheel_transfer=environment.wheel_transfer(environment.WheelDiameter.MM_920),
    traction=environment.traction_sound_power(environment.TractionVehicle.ELECTRIC_MULTIPLE_UNIT),
    aerodynamic=environment.aerodynamic_sound_power(),
)
track = environment.RailwayTrack(
    rail_roughness=environment.rail_roughness(environment.RailRoughnessClass.NORMAL),
    track_transfer=environment.track_transfer(environment.TrackTransferClass.MONOBLOCK_MEDIUM),
    impact_roughness=environment.impact_roughness_single(),
)
result = environment.railway_source_power(
    environment.RailwayVehicle(stock=stock, flow_rate=96.0, speed_kmh=160.0), track, phi=90.0, psi=10.0,
)
result.plot()
plt.show()
```

![Two panels for a four-axle electric multiple unit on normally maintained monoblock track. Left: the one-third-octave sound power of each component of the 160 km/h scene, rolling noise on source A about twenty decibels above the traction noise, traction noise present on both source heights, and the aerodynamic and bridge components absent because neither is active at this speed. Right: the octave-band total line power of source A at 0.5 m and source B at 4.0 m against speed from 60 to 350 km/h; source A rises smoothly throughout while source B falls slowly to a minimum at 200 km/h and then steps up by nine decibels as the aerodynamic term switches on, narrowing the gap between the two lines from about 30 dB to 19 dB](https://docs.jmrp.io/phonometry/media/cnossos_rail_components.svg)

*Left: the allocation table of this section, drawn. Rolling noise is the whole
of source A at 160 km/h; traction is the only component present on both heights,
and the aerodynamic and bridge entries of `result.components` come back as `nan`
because neither is active here, which is the allocation rule showing itself.
Right: source B is nearly flat and then **steps**, because aerodynamic noise is
only modelled above 200 km/h; source A never stops rising. Source B does not
overtake source A on this vehicle at any speed. The gap closes from 29.9 dB at
200 km/h to 18.7 dB at 350 km/h, and that closing is the sense in which the
pantograph starts to matter.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt
import numpy as np

# `stock`, `track`, `railway_source_power` and `RailwayVehicle` come from the
# figure snippet above; the result carries its own frequency axes.
def total(bands):
    return 10.0 * np.log10(np.sum(10.0 ** (np.asarray(bands) / 10.0)))

fig, (left, right) = plt.subplots(1, 2, figsize=(12, 4.6))
result = environment.railway_source_power(
    environment.RailwayVehicle(stock=stock, flow_rate=96.0, speed_kmh=160.0), track, phi=90.0, psi=10.0)
for name, spectrum in result.components.items():
    for row, height in ((0, "0.5 m"), (1, "4.0 m")):
        values = np.asarray(spectrum)[row]
        if np.any(np.isfinite(values)):
            left.semilogx(result.third_octave_frequencies, values,
                          label=f"{name} @ {height}")
left.set(xlabel="1/3-octave band [Hz]", ylabel="Sound power [dB]")
left.legend(fontsize="small")

speeds = np.linspace(60.0, 350.0, 30)
for row, label in [(0, "Source A (0.5 m)"), (1, "Source B (4.0 m)")]:
    right.plot(speeds, [
        total(environment.railway_source_power(
            environment.RailwayVehicle(stock=stock, flow_rate=96.0, speed_kmh=float(v)),
            track, phi=90.0, psi=10.0).line_power[row]) for v in speeds],
        label=label)
right.axvline(200.0, ls=":", label="Aerodynamic threshold")
right.set(xlabel="Speed v [km/h]", ylabel="Total line power [dB/m]")
right.legend(fontsize="small")
plt.show()
```

## 3. Rolling noise: roughness in, sound power out

Rolling noise is the wheel and the rail exciting each other through the
roughness of the two running surfaces. The method separates the cause from the
radiation:

- **cause**: the rail roughness $L_\mathrm{r,TR}$ (the second table of G-1, called
  G-1b here, selected by digit 2 of
  the track descriptor) and the wheel roughness $L_\mathrm{r,VEH}$ (the first, G-1a,
  selected by the brake type, digit 3 of the vehicle descriptor), energy-summed
  and then filtered by the **contact filter** $A_3$ (Table G-2), which accounts
  for the
  finite size of the contact patch:

$$
L_{\mathrm{R,TOT},i} = 10\,\log_{10}\!\left(10^{L_{\mathrm{r,TR},i}/10} + 10^{L_{\mathrm{r,VEH},i}/10}\right) + A_{3,i};
$$

- **radiation**: three speed-independent transfer functions turn that roughness
  into sound power per axle, one for the track (Table G-3a, selected by the
  sleeper and rail-pad codes), one for the wheel (Table G-3b, selected by the
  wheel diameter) and one for the superstructure of a freight wagon
  (Table G-3c, which is 0.0 dB in every band):

$$
L_{W,0,\mathrm{TR},i} = L_{\mathrm{R,TOT},i} + L_{H,\mathrm{TR},i} + 10\,\log_{10} N_\mathrm{a},
$$

and the same addition for the wheel and the superstructure. The three
components are then energy-summed, and all three sit at source A.

The superstructure term applies to vehicle type `a` (freight) only. A minimum
speed of **50 km/h**, or 30 km/h for trams and light metro, is used to determine
the roughness, which freezes the source spectrum below it; the flow term still
uses the true speed. While a vehicle idles, rolling noise is excluded
altogether.

## 4. The wavelength-to-frequency conversion

Roughness is tabulated against **wavelength**, and read at $\lambda = v/f$, so
the whole roughness spectrum slides along the frequency axis as the train speeds
up. That is the reason a fast train is bright and a slow one is not: the same
millimetre-scale roughness excites a higher band.

![Total effective roughness against 1/3-octave frequency for 60, 160 and 300 km/h, three curves of the same shape displaced upward and to the right as speed rises, crossing zero decibels near 400 Hz, 1 kHz and 2 kHz respectively](https://docs.jmrp.io/phonometry/media/cnossos_rail_roughness_shift.svg)

*The same track and the same wheel at three speeds. The curve does not change
shape, it slides: multiplying the speed by 2.7 moves it by rather more than an
octave and lifts it, because the long wavelengths of an average network are the
rough ones.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt
from phonometry import environment

rail = environment.rail_roughness(environment.RailRoughnessClass.NORMAL)
wheel = environment.wheel_roughness(environment.BrakeType.NON_TREAD)
filt = environment.contact_filter(environment.ContactFilter.LOAD_50_DIAMETER_920)

fig, ax = plt.subplots()
for speed in (60.0, 160.0, 300.0):
    total = environment.total_effective_roughness(
        environment.roughness_to_frequency(
            rail.levels, wavelengths_mm=rail.wavelengths_mm, speed_kmh=speed
        ),
        environment.roughness_to_frequency(
            wheel.levels, wavelengths_mm=wheel.wavelengths_mm, speed_kmh=speed
        ),
        environment.roughness_to_frequency(
            filt.levels, wavelengths_mm=filt.wavelengths_mm, speed_kmh=speed
        ),
    )
    ax.semilogx(environment.RAILWAY_THIRD_OCTAVE_BANDS_HZ, total, label=f"v = {speed:g} km/h")
ax.legend()
plt.show()
```

Two things about this conversion deserve to be stated plainly, because neither
is settled by the Directive.

**The speed is in metres per second.** The published 2015 text says km/h in both
places where the conversion appears, which would put the spectrum 3.6 times too
low in frequency. The corrigendum of 2018 corrected both. This library
implements the corrected form.

**The interpolation rule is prose, not a formula.** All the Directive says is
that "the two corresponding 1/3 octave bands defined in the wavelength domain
shall be averaged energetically and proportionally", with no formula and no
worked example. Two readings are shipped, and the default is the one that
reproduces the Commission's own reference source module:

```python
from phonometry import environment

environment.RoughnessInterpolation.PROPORTIONAL   # levels interpolated linearly in lambda
environment.RoughnessInterpolation.ENERGY         # energies interpolated linearly in lambda
```

They differ by up to about 1 dB on a steep part of a spectrum. This is the
single largest interpretation risk in the railway model, and any implementation
that does not say which rule it uses is not comparable with another one.

A third, smaller ambiguity is handled without a switch. The 2021 amendment put
the tables it replaced on the standard 1/3-octave wavelength series, but its
point (20)(a) replaces "the second table" of G-1 only, so the wheel roughness
of the first is still printed at 120, 12, 3.2 and
1.2 mm where every other table now reads 125, 12.5, 3.15 and 1.25 mm. Rather
than forcing the two onto a common grid, each table is resampled on the grid the
Directive prints for it.

## 5. Impact noise, and curve squeal

**Impact noise** from joints, switches and crossings is added as a fictitious
extra roughness, energy-summed into the total before the transfer functions are
applied (2.3.11). Table G-4 is given for one joint per 100 m, and any other
density scales it (2.3.12):

$$
L_{\mathrm{R,IMPACT},i} = L_{\mathrm{R,IMPACT-SINGLE},i}
  + 10\,\log_{10}\!\left(\frac{n_l}{0.01}\right).
$$

For jointed track a default $n_l = 0.01\ \text{m}^{-1}$ is prescribed. Impact noise is not
modelled below the minimum speed, nor while idling. The increment applies from
50 m before the joint to 50 m after it, which is a **segmentation** instruction:
the model gives the source strength, and it is up to the track segmentation to
place it over the right 100 m.

**Curve squeal** is an excess added to the rolling-noise sound power at all
frequencies. The rule is the one the 2021 amendment substituted, and it is not
the 2015 one:

| Vehicle | Condition | Excess |
| :--- | :--- | :--- |
| Train | $300\ \text{m} < R \le 500\ \text{m}$ over at least 50 m of curve | 5 dB |
| Train | $R \le 300\ \text{m}$ over at least 50 m of curve | 8 dB |
| Train | switch turnout with $R \le 300\ \text{m}$ | 8 dB |
| Tram | curve or switch turnout with $R \le 200\ \text{m}$ | 5 dB |

```python
from phonometry import environment

environment.curve_squeal_excess(280.0)                       # 8.0 dB
environment.curve_squeal_excess(450.0)                       # 5.0 dB
environment.curve_squeal_excess(280.0, tram=True)            # 0.0 dB, trams need R <= 200 m
environment.curve_squeal_excess(280.0, track_length=20.0)    # 0.0 dB, the curve is too short
```

The 2015 text read "8 dB for R < 300 m and 5 dB for 300 m < R < 500 m", which
left the two endpoints with no rule at all even though the same paragraph scoped
the effect to "curves with radius below **or equal to** 500 m". The 2021 wording
closes both intervals.

## 6. Traction and aerodynamic noise

**Traction noise** is tabulated per vehicle and per source height (Table G-5).
The Directive models only two running conditions, constant speed (which it says
is valid as well when the train accelerates or decelerates) and idling, and it
takes the source strength at maximum load in both. Table G-5 therefore tabulates
a **single** value per band and height, and the method reads it for both
conditions: $L_{W,0,\mathrm{const},i} = L_{W,0,\mathrm{idling},i}$ is a
consequence of how the table is printed, not a physical claim that an idling
locomotive is as loud as a running one. Idling is only considered if trains idle for more
than half an hour.

**Aerodynamic noise** is relevant only above **200 km/h** and follows a
logarithmic speed law about $v_0 = 300\ \text{km/h}$ (2.3.13, 2.3.14):

$$
L_{W,0,i} = L_{W,0,h,i}(v_0) + \alpha_{h,i}\,\log_{10}\!\left(\frac{v}{v_0}\right),
$$

with $\alpha_1 = \alpha_2 = 50$ in every band, so a doubling of speed adds
15.05 dB. Table G-6 is the reference spectrum of a 20 m vehicle at $v_0$: at the
lower height it is broadband, at the upper one it is essentially silent except
for a 1.6 kHz to 3.15 kHz plateau at 105 to 110 dB, which is the pantograph.

## 7. Source directivity

Each source is corrected in the two planes before the flow term is added
(2.3.5). The **horizontal** correction is a dipole (2.3.15), 0 dB broadside and
$10\log_{10} 0.01 = -20\ \text{dB}$ along the track:

$$
\Delta L_{W,\mathrm{dir,hor},i} = 10\,\log_{10}\!\left(0.01 + 0.99\,\sin^2\varphi\right).
$$

The **vertical** correction of source A (2.3.16) depends on the band, and it is
one of the pieces the 2021 amendment rewrote. The current text drops the
absolute-value bars of the 2015 form and makes the correction identically zero
for $\psi \le 0$:

$$
\Delta L_{W,\mathrm{dir,ver},i} = \frac{40}{3}
  \left[\tfrac{2}{3}\sin 2\psi - \sin\psi\right]
  \log_{10}\!\left(\frac{f_{\mathrm{c},i} + 600}{200}\right), \quad 0 < \psi < \pi/2,
$$

and zero elsewhere. The two editions disagree over the whole lower half space,
and also above about 41.4 degrees, where the bracket turns negative. Both are
available, because comparing with a study computed before 2021 needs the old
one:

```python
from phonometry import environment

environment.vertical_directivity(-30.0)                                        # zeros
environment.vertical_directivity(-30.0, edition=environment.DirectivityEdition.ORIGINAL_2015)  # positive
```

![Two polar plots of the CNOSSOS-EU railway directivity. Left, the vertical correction of source A at 250 Hz and 4 kHz for both editions: the 2021 form is identically zero everywhere below the horizon and above about 41.4 degrees turns negative, while the 2015 form is symmetric about the horizon and positive on both sides, the two agreeing only in the wedge from 0 to 41.4 degrees. Right, the horizontal dipole in the plane of the track: 0 dB broadside to the track and minus 20 dB along it, with the along-track null marked](https://docs.jmrp.io/phonometry/media/cnossos_rail_directivity.svg)

*Left: the two editions agree only in the first quadrant below 41.4°, which is
exactly where the bracket $\tfrac{2}{3}\sin 2\psi - \sin\psi$ is positive. Below
the horizon the 2021 text returns 0 dB where the 2015 text returned up to
+2.4 dB at 250 Hz and 60° below the horizon; above 41.4° the current form goes negative and the old one
does not. A study computed before 2021 is not comparable band by band unless the
edition is stated. Right: the horizontal dipole is worth 20 dB between a
receiver beside the track and one at the end of it, which is why a track running
towards a dwelling is a different problem from one running past it.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt
import numpy as np

# `vertical_directivity` and `DirectivityEdition` come from the snippet above;
# the horizontal dipole is formula (2.3.15) itself, one line of numpy.
psi = np.linspace(-89.0, 89.0, 179)
bands = {7: "250 Hz", 19: "4 kHz"}       # indices into the 24 third-octaves
fig, (left, right) = plt.subplots(1, 2, subplot_kw={"projection": "polar"},
                                  figsize=(12, 5.0))
for band, name in bands.items():
    for edition, style in [(environment.DirectivityEdition.CURRENT, "-"),
                           (environment.DirectivityEdition.ORIGINAL_2015, "--")]:
        values = [float(environment.vertical_directivity(float(a), edition=edition)[band])
                  for a in psi]
        left.plot(np.radians(psi), values, style,
                  label=f"{name}, {edition.value}")
left.legend(fontsize="small", loc="lower left")

phi = np.radians(np.linspace(0.0, 360.0, 361))
right.plot(phi, 10.0 * np.log10(0.01 + 0.99 * np.sin(phi) ** 2))
right.set_title("Horizontal dipole (2.3.15)", fontsize="small")
plt.show()
```

At source B, only the aerodynamic source is directional, $10\log_{10}(\cos^2\psi)$
for $\psi < 0$ (2.3.17); everything else at 4.0 m is omni-directional, and so is
bridge noise at 0.5 m.

One point of interpretation: 2.3.15 offers the dipole "by default ... for
rolling, impact, squeal, braking, fans and aerodynamic effects". Traction noise
other than the fans is not in that list, and the Directive gives no other
horizontal directivity for it. This library applies the dipole to every source,
which is what the Commission's reference source module does.

## 8. Bridges

On a bridge the deck radiates too. The 2015 text added a constant to the rolling
noise, 1 dB for concrete or masonry and 4 dB for steel with ballasted track. The
2021 amendment replaced that with a **separate source**, built exactly like a
rolling-noise component but with the bridge transfer function of Table G-7
(2.3.18):

$$
L_{W,0,\mathrm{bridge},i} = L_{\mathrm{R,TOT},i} + L_{H,\mathrm{bridge},i}
  + 10\,\log_{10} N_\mathrm{a}.
$$

The rolling noise of the vehicle on the bridge is still computed normally, with
the transfer function of the track system that is actually on the bridge; the
bridge term is added to it. It sits at source A and is omni-directional, which
the amendment states explicitly.

Table G-7 has exactly two columns, and the Directive labels them by the
A-weighted bridge excess they stand for rather than by a construction type:
`BridgeType.PLUS_10_DBA` selects the "+10 dB(A)" column and
`BridgeType.PLUS_15_DBA` the "+15 dB(A)" one. There is no third column and no
interpolation, so choosing between them is a judgement about the deck, made
before the calculation and recorded with it.

```python
from phonometry import environment

environment.RailwayTrack(
    rail_roughness=..., track_transfer=...,
    bridge_transfer=environment.bridge_transfer(environment.BridgeType.PLUS_10_DBA),   # Table G-7, col. 1
)
```

## 9. Handing the source to a propagation model

The railway source is derived in **1/3 octave bands** and then energy-summed
into octave bands for propagation, which is what 2.3.2 requires and what
(EU) 2021/1226 point (1) made explicit. The result object carries both:

The `result` here is the one the section 2 snippet built (open its `<details>`
block for the vehicle and track it describes). **Row 0 is source A at 0.5 m and
row 1 source B at 4.0 m**, the `h = 1` and `h = 2` of the section 2 table:

```python
result.third_octave_line_power   # (2, 24), the two heights in 1/3 octaves
result.line_power                # (2, 8), the same in octave bands
result.total_line_power          # (8,), the two heights energy-summed
result.components["rolling"]     # the per-source breakdown, before directivity
result.frequencies               # the 8 octave centres
result.third_octave_frequencies  # the 24 third-octave centres
```

A component that is not active at the modelled condition comes back as `nan`
rather than as a floor value, so `components["aerodynamic"]` below 200 km/h and
`components["bridge"]` off a bridge are both all-`nan`.

Splitting a source line into equivalent point sources is, in the words of
section 2.5.3, "outside the scope of the current methodology": a point source
standing for a segment of length $dL$ simply carries
$L'_{W,\mathrm{eq,line},i} + 10\log_{10}(dL)$, which is arithmetic.

CNOSSOS-EU has **its own** propagation method in section 2.5 of Annex II, and it
is not ISO 9613-2. Chaining a CNOSSOS emission onto the ISO 9613-2 propagation
of this library is a legitimate engineering estimate, but it is not the
normative CNOSSOS chain and must not be reported as one.

## 10. What is pinned against digits, and what is not

Annex II prints **no worked example and no check value** anywhere in 2.3 or
Appendix G, so it is worth being explicit about what the agreement in the test
suite measures.

**Verified against published digits.** Every Appendix G table, against the
Official Journal text of the instrument that enacted it. And the whole chain of
equations end to end, against the emission test workbook the European Commission
published with its reference source module: feeding the shipped equations the
**2015** coefficient database that workbook was computed with reproduces **all
34 560 published rows of the twenty vehicles whose catalogue entries are well
formed**, that is 17 280 cases at both source heights, to 0.0055 dB, inside the
two decimals it prints. A hundred and twenty-three of those cases are committed
to the repository and run in CI. The three vehicles left out are described in
`tests/data/cnossos/README.md`: their rows record defects of the reference
program, not of the method.

That agreement resolves the loud parts of the database and not the quiet ones:
a coefficient twenty decibels below its neighbour does not move the total, so
the placeholder rows of the reference catalogue and the source-A traction
column, which rolling noise always covers, are carried by the test set without
being constrained by it. `tests/data/cnossos/README.md` reports the measurement.

**Verified as equations times transcribed tables, not end to end.** Everything
the 2021 amendment changed. The workbook predates it, so the new curve-squeal
rule, the bridge transfer function, the current vertical directivity and the
five replaced tables are each pinned individually, by transcription and by
closed-form assertion, but their combination has never been checked against an
independent published result.

**Not anchored on any oracle.** The choice of interpolation rule in section 4;
the proportional distribution of traction sound power when a real component sits
between the two model heights, for which the Directive gives no formula; and the
segmentation of a line into point sources, which the method declares out of
scope.

## What this guide covers

- **Covered.** Section 2.3 of Annex II to Directive 2002/49/EC in the consolidated text: the
  vehicle and track descriptors of Tables [2.3.a] and [2.3.b], the traffic-flow
  line power (2.3.1, 2.3.2, 2.3.4), the directional sound power (2.3.5), rolling
  noise (2.3.6 to 2.3.10), impact noise (2.3.11, 2.3.12), curve squeal, traction
  noise, aerodynamic noise (2.3.13, 2.3.14), the source directivity (2.3.15 to
  2.3.17) and bridge noise (2.3.18), together with the whole Appendix G
  database, through `railway_source_power` and the component functions around
  it.

- **Not covered.** The road source (2.2 and Appendix F), the industrial source (2.4 and Appendix
  H), the aircraft source (2.6 and 2.7) and the CNOSSOS-EU propagation method of
  section 2.5, which is a different model from the ISO 9613-2 one implemented in
  this library. Rail roughness classes `N` (not well maintained) and `B` (bad
  condition) carry no spectrum in Appendix G and have to be supplied by the
  Member State. The other railway sources of 2.3.3, such as depots, stations and
  loudspeakers, are treated by the industrial method.

## See also

- API reference: [`environment.sources.cnossos_rail`](https://docs.jmrp.io/phonometry/reference/api/environment/cnossos-rail/).
- [Outdoor Sound Propagation](https://docs.jmrp.io/phonometry/environment/propagation/outdoor-propagation/): the
  ISO 9613-2 chain that carries a source power to a receiver.
- [Environmental noise levels](https://docs.jmrp.io/phonometry/environment/assessment/environmental-levels/): the
  $L_\mathrm{den}$ and $L_\mathrm{night}$ indicators the resulting maps are drawn for.

## References

- **European Union. (2002).** [*Directive 2002/49/EC of the European Parliament and of the Council of 25 June 2002 relating to the assessment and management of environmental noise*](https://eur-lex.europa.eu/eli/dir/2002/49/oj) (Directive 2002/49/EC).
  The directive whose Annex II the two Commission directives above write: section 2.3 and Appendix G, which this guide implements, are its annex as they amend it.
- **European Union. (2015).** [*Commission Directive (EU) 2015/996 establishing common noise assessment methods according to Directive 2002/49/EC of the European Parliament and of the Council*](https://eur-lex.europa.eu/eli/dir/2015/996/oj) (Directive (EU) 2015/996).
  Replaces Annex II of Directive 2002/49/EC with the CNOSSOS-EU methods. The implemented railway source is section 2.3 (formulae 2.3.1 to 2.3.18) with the coefficient database of Appendix G.
- **European Union. (2018).** [*Corrigendum to Commission Directive (EU) 2015/996*](<https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32015L0996R(01)>) (OJ L 5, 10.1.2018, p. 35).
  Replaces the whole of Appendix G and corrects the railway roughness conversion, where the published text read km/h for a speed that has to be in metres per second.
- **European Union. (2021).** [*Commission Delegated Directive (EU) 2021/1226 amending, for the purposes of adapting to scientific and technical progress, Annex II to Directive 2002/49/EC as regards common assessment methods for noise*](https://eur-lex.europa.eu/eli/dir_del/2021/1226/oj) (Directive (EU) 2021/1226).
  The amendment in force since 31 December 2021. Replaces Tables G-1b, G-2, G-3a, G-4 and G-7, substitutes a new curve-squeal rule and a new bridge model, and rewrites the vertical directivity of source A.
- **International Organization for Standardization. (1996).** [*Acoustics — Attenuation of sound during propagation outdoors — Part 2: General method of calculation*](https://www.iso.org/standard/20649.html) (ISO 9613-2:1996).
  The propagation model used here to carry the emission to a receiver. CNOSSOS-EU has its own propagation method in section 2.5 of Annex II, which is not ISO 9613-2.
- **Kephalopoulos, S., Paviotti, M., & Anfosso-Lédée, F. (2012).** [*Common noise assessment methods in Europe (CNOSSOS-EU)*](https://publications.jrc.ec.europa.eu/repository/handle/JRC72550) (EUR 25379 EN). Publications Office of the European Union. https://doi.org/10.2788/31776
  The Joint Research Centre reference report the Directive text descends from, with the IMAGINE measurement campaigns behind the Appendix G coefficients.
