Espresso quality attracts the interest of many stakeholders: from consumers to local business activities, from cofee-machine vendors to international cofee industries. So far, it has been mostly addressed by means of human experts, electronic noses, and chemical approaches. The current work, instead, proposes a datadriven analysis exploiting time-series feature engineering. We analyze a real-world dataset of espresso brewing by professional cofee-making machines. The novelty of the proposed work is provided by the focus on the brewing time series, from which we propose to engineer features able to improve previous data-driven metrics determining the quality of the espresso. Thanks to the exploitation of the proposed features, better quality-evaluation predictions are achieved with respect to previous data-driven approaches that relied solely on metrics describing each brewing as a whole (e.g., average flow, total amount of water). Yet, the engineered features are simple to compute and add a very limited workload to the cofee-machine sensor-data collection device, hence being suitable for large-scale IoT installations on-board of professional cofee machines, such as those typically installed in consumer-oriented business activities, shops, and workplaces. To the best of the authors' knowledge, this is the first attempt to perform a data-driven analysis of real-world espresso-brewing time series. Presented results yield to three-fold improvements in classification accuracy of high-quality espresso cofees with respect to current data-driven approaches (from 30% to 100%), exploiting simple threshold-based quality evaluations, defined in the newly proposed feature space.
Espresso is an almost syrupy beverage generated by a machine, typically using a motor-driven pump, forcing pressurized hot water through finely ground cofee. Each espresso shot in a bar can generate one or two cups of cofee, being called, respectively, single or double, and requiring proportional amounts of ground cofee.
Drinking espresso cofee is a ritual rooted in the pleasure of
its taste. In some countries, such as Italy, where 97% of adults
drink espresso daily [
In 2018, each Italian had 2.2 daily espresso cups on average,
i.e., 6 kg yearly, in one of the 150 thousand bars, with each bar
using 1.2 kg of ground cofee daily to serve almost 200 cofees on
average, and most of them were espresso, representing
approximately one third of a medium bar turnover [
According to common knowledge and online sources [
In the current work, among the diferent quality-influencing
variables, we focus on (i) cofee ground size, (ii) ground amount,
and (iii) water pressure. Regarding the quality-evaluation
variables, we exploit the following common metrics as selected by
domain experts and related works: (i) total extraction time, (ii)
the total volume of cofee in cup, and (iii) the derived average
lfow of the extraction [
The ideal portion [
The cofee ground derives from the process of cofee grinding from cofee beans. Small changes in the grind size can drastically afect the taste and the quality of the brewed espresso. In general, if the cofee is ground too coarse, the espresso can be underextracted and less flavorful. On the other hand, too fine ground may result in an over-extracted and bitter cofee. The amount of ground itself impacts on quality, resulting in a too watery or bitter cofee. Water pressure must be set to brew the right cofee amount in a proper time, thus leading to the right flow rate determining an intense flavour.
The novelty of the proposed work is provided by the exploitation of the brewing time series, from which we propose to engineer features able to improve the standard data-driven metrics determining the quality of the espresso, i.e., extraction time, volume, and flow (as the ratio of volume and time). The proposed features are applied on a real-world dataset where we show that they can provide better quality-evaluation predictions, by allowing to reduce the false positives, i.e., apparently good cofees, without any loss in true positives.
Since the engineered features are simple to compute and add a very limited workload to the cofee-machine sensor-data collection device, they are also suitable for large-scale IoT installations on-board of professional cofee machines, such as those typically installed in consumer-oriented business activities, shops, and workplaces.
Presented results uncover insights into the espresso quality evaluation, its relationships with the main quality variables, leading to positive impacts on both cofee consumers and cofeemaking industries, respectively enjoying and providing more pleasure in drinking higher-quality espresso cofee.
The rest of the paper is structured as follows. Section 2 discusses related works, Section 3 describes the dataset and the experimental design, Section 4 introduces the time-series feature engineering algorithm, and Section 5 presents experimental results. Finally, Section 6 draws conclusions and outlines future works.
Espresso quality assessment is traditionally performed with sensory analysis, the scientific discipline that statistically and experimentally analyze reactions to stimuli perceived through the human senses (sight, smell, taste, touch and hearing). Sensory evaluation is however time-consuming and afected by subjectiveness and low-reproducibility due to the human component.
Considering these limitations, objective analysis as chemical
techniques, electronic noises and data-driven approaces are
commonly exploited for cofee quality control. Diferent chemical
techniques adopt Gas Chromatography (GC) and Mass
Spectroscopy (MS) analysis. Several works study the efect of external
variables (e.g. water pressure, water temperature) or of cofee
characteristics on the final espresso quality. Some works are
focused on the influence of water, as its composition, pressure [
However, GC and MS analysis often require a significant
amount of time and human intervention. Many studies exploit
Electronic Nose (EN) systems to overcome the complexity and
cost of GS/MS techniques. An electronic nose is a device intended
to mimic human olfaction. It consists of an array of chemical
sensors for chemical detection and a pattern recognition system
capable of identifying the specific components of an odor [
Finally, data-driven approaches can be applied for large-scale
and real-time espresso quality assessment, exploiting Internet of
Things (IoT) sensors in place of the more sensitive and unstable
EN devices. Recently, a data-driven approach that exploits
association rule mining has been proposed to analyze the correlation
of cofee-making machine parameters and espresso quality [
Time series analysis is a popular and well-known approach in
many application fields [
With respect to the state of the art, the current work contributes by cleverly transferring known and simple time-series feature engineering techniques into the espresso quality evaluation domain, leading to significant improvement in classification performance with respect to the state of the art. To the best of the authors’ knowledge, this is the first attempt to perform a datadriven analysis of real-world espresso-brewing time series, as until now the focus has been limited to whole-extraction metrics. 3
The dataset under analysis consists of real-world espresso brewing data. Since the dataset is provided by a leading cofee company, we cannot disclose exact details of the real-world settings (e.g., the cofee-machine maker and model, the precise location and name of the involved business activities). Each espresso extraction has been performed on professional cofee-making machines and the values of the quality-evaluation variables have been collected every 300 ms. In particular, our time series consist of the values of the amount of water at each time interval, as provided by flow-meter pulse counter, then deriving the instant lfow rate (i.e., the ratio of the amount of water and the time).
Each extraction has been performed with specific values of the quality-influencing variables, hence allowing us to know the ground-truth labels of high-quality espresso cofees, i.e., those having all optimal settings for (i) cofee ground size, (ii) ground amount, and (iii) water pressure. An exhaustive set of cofees has been produced to observe the efect of non-optimal values on the espresso quality. For each quality-influencing variable, diferent values are considered: ground size can be coarse, optimal, or fine; ground amount can be high, optimal, or low; brewing water pressure can be high, optimal, or low. All possible combinations of the three external-variable values (e.g., optimal, high, low) have been included in the dataset, hence generating 33 = 27 possible input configurations. For each configuration among the 27 combinations of external variables (for instance: coarse ground size, optimal ground amount, and high water pressure), 20 espresso extractions have been performed. Experiments have been repeated on a professional cofee-making machine, generating a datasets consisting of 540 espresso extractions.
The domain-expert quality thresholds used in our experiments
are as follows: espresso volume from 20–30 ml, extraction time
from 20–30 s. The values have been selected according to public
literature, e.g., those published by the Specialty Cofee
Association of Europe [
Given such thresholds, espresso extractions can be labelled with their quality assessment. Quality labels are optimal, too low or too high for each of the quality variables: volume, time, and flow. Table 1 recaps the domain-based threshold values and corresponding labels.
The problem tackled by this work stems from the fact that
analyzing the standard quality-evaluation variables without the
additional time-series novel features, many false positives are
provided: some espresso extractions are characterized by
highquality values in terms of water amount, flow rate and extraction
time, however, their ground size, ground amount or water
pressure were not optimal (compensation efect [
Feature engineering refers to the process of extracting features from raw data. It is typically executed to improve the performance of predictive or classification models. In the current work, we exploit feature engineering to leverage the cofee-brewing time series with the aim of improving the espresso quality assessment.
For each cofee extraction, the time series of the flow-meter pulses is stored, with sampling time equal to 300 ms. Flow-meter pulses are firstly converted to quantity of brewed water q, as follows: (1) q = nump ∗ pulseq
numc where nump is the number of pulses of the flow-meter, pulseq represents the quantity of brewed water per pulse of the flowmeter and numc represents the number of brewed cofees. In the experimental data under analysis, pulseq =0.5 ml, as given by the cofee-machine datasheet, and numc =2, since two espresso coffees are brewed for each extraction. The time series captures the water quantity over time, hence the instant flow rate is known.
Figure 1 shows an example of a real time series from the dataset. We notice a clear two-segment trend that is observable for any arbitrary extraction: a first steeper phase is followed by a second part having a lower flow rate. This phenomenon is known by domain experts. In the first, transient, phase of cofee brewing, water is forced in the cofee panel inside the filter holder, and cofee grounds do not slow the water flow yet. On the contrary, in the second phase, water penetrate and dampen cofee grounds yielding the actual cofee extraction.
We propose to extract the following new features to capture the two-fold behavior of the extraction. We firstly determine the point where a significant flow variation is observed. We refer to this point as trend point. The trend point is used to approximate the water quantity time series as a polygonal chain. The approximate polygonal chain is constituted by two line segments that represent the two phases of the water flow and its vertex of intersection is the trend point. The trend point is estimated by considering the maximum variation of the slope average of the points in two consecutive not-overlapping sliding windows of size W . The slope si (or gradient) of two consecutive points pi = (ti , qi ) and pj = (tj , qj ) is computed as follows. si = qj − qi tj − ti In Equation 2, t is the time reference and q is the water quantity, and they represent the axes of Figure 1. The slope s describes the steepness of the water flow.
The procedure for the trend point estimation is reported in Algorithm 1.
The maximum variation of the slope and the corresponding points are initialized in Lines 1 and 2. In Lines 4 and 5, two consecutive not-overlapping sliding windows of size W are defined.
Let wk be a time window of size W . The slope average wkmean of all consecutive points of the time window is computed as follows
1 wkmean = W − 1 j=1 tj − tj−1
WÕ−1 qj − qj−1 (2) (3) where pj = (tj , qj ) and pj−1 = (tj−1, qj−1) are consecutive points of the time window.
The slope average is estimated for the two sliding windows, as reported in Lines 6 and 7. The two terms capture the average flow rate in the corresponding time window. The diference of the two slope averages is computed in Line 8. The maximum slope variation and the corresponding point are updated in Line 9.
The point of maximum variation corresponds to the intersection point of the two considered sliding windows. The process is repeated until all N points of the time series are considered. Finally, the trend point is returned (Line 12).
The trend point ptp = (ttp , qtp ) represents the intersect vertex of an approximate polygonal chain of the water quantity time series. It is exploited to compute two features that capture the two phases of the espresso extraction. Let be p0 = (t0, q0) and pN = (tN , wqN ) the first and last points of the time series, respectively. We define s1 and s2 as follows. l)125 m ( r e ta100 w f o ty 75 i t n a uQ50 25 0 0 5 10 Time (s) 15 20 25
In Figure 2, the approximate polygonal chain of a cofee extraction time series is reported. The dashed line indicates the average water flow. The slope s1 represents the average flow of the first phase of the espresso brewing while slope s2 the average lfow of the second phase. These two features are exploited in the analysis to better characterize the cofee extraction, providing additional information with respect to the overall average flow. The extracted features will also be exploited to compute new ranges for the optimal quality parameters, hence improving the recognition of high-quality cofees. 5
This section provides a description of the data cleaning procedures applied to the dataset (Section 5.1), a discussion of the data characterization of the extracted features (Section 5.2), and their contribution to the espresso quality assessment improvement (Section 5.3). 5.1
The dataset has been pre-processed by applying the data
cleaning steps described in [
We firstly analyze the relationship between the extracted features and the quality-evaluation variables (i.e., total extraction time, average flow rate, total water amount). The trend point and the consequent slope values have been computed with a window size W set to 10.
The correlation analysis shows that slope s2 is highly correlated with the average flow rate (over the whole extraction), with a Pearson correlation coeficient equal to 0.95, and the total brewing time, with a correlation coeficient of -0.94. As expected, lower flow rates lead to longer extraction times, since the total amount of cofee is an almost constant goal of the cofee machine.
We then investigate the relationship between the two average flows (i.e. s1 and s2) and the three external quality-influencing variables: water pressure, cofee ground amount and cofee ground size, also known as grinding setting).
Figure 3 shows the pressure behavior with respect to s1 and s2. The pressure values (low, optimal, and high) are represented by the label in the scatter plot. We can observe that cofee extractions in the (s1, s2) space are clearly divided in three macro-areas, determined by s1 value. The central partition is characterized by an optimal pressure, while the first and last areas by low and high values of pressure respectively. Hence, the value of the external variable highly influence the first phase of cofee extractions, when water is forced into the cofee panel. To a low pressure corresponds a low water flow in the initial phase and vice versa for the high pressure. The flow in the second phase is instead almost independent from the pressure value.
Regarding the total amount of water, we report in Figure 4 the cofee extractions as a function of s1 and s2. Diferently from the pressure-labeled scatter plot, it is not observable a sharp distinction. We can however identify a relationship with s2. Higher amounts of cofee ground lead to lower values of the flow s2. In this case, the average flow in the second phase of the extraction is hindered by the higher amount of cofee ground. Hence, the water flow is reduced. Likewise, the lower quantity of cofee ground facilitates the flow of water, with a consequent increase in flow s2. The cofee ground amount, instead, do not influence s1, since it captures the average flow of the water when it is forced in the cofee panel and before the cofee ground tampering.
Finally, we observe a similar behavior when considering the cofee ground size (i.e., grinding settings), hence we do not report the plot. A coarser grinding generally corresponds to a higher lfow. The finer cofee grinding instead hinders the water flow. This results in a lower flow s2 in the second phase of the cofee extraction. 5.3
In this section, we evaluate the extracted feature ability to
characterize espresso quality and to improve the detection of
highquality espresso cofees. All the three external variables are under
the barista control. However, brew pressure is set at first in the
espresso machine calibration phase and it is periodically checked
and configured, typically with the support of technicians. On
the other hand, the grinding settings and the amount of cofee
ground are determined by the barista at each espresso brewing.
Hence, it is particularly relevant to control that these two
external variables are set properly by the barista. In existing works,
domain-experts and data-driven thresholds on quality indexes,
such as espresso volume, extraction time and brewing flow rate,
have been applied to evaluate cofee quality. The analysis in [
6.5 5.0
Slope 1 5.5 5.0
Slope 1
5.5
explored the phenomena of compensating sub-optimal values
of diferent external variables. A compensation efect is
observable when configurations of values of external variables allow
to achieve apparently high-quality cofees, in terms of quality
indexes, despite one or more values are, in fact, not optimal.
Interpretable exploration techniques highlighted that high amounts
of cofee ground, that generally hinder the water flow and lead
to long percolation times, could be compensated by a coarser
grinding that, on the other hand, facilitates the flow [
As a reference, we consider domain-driven thresholds on
coffee quality indexes. In Figure 5 the espresso extractions with
optimal values of quality indexes are reported in the s1 and s2
space. They can be grouped as follows. (i) True high-quality
extractions present optimal values for both the quality-evaluation
indexes and, in particular, for all external variables. (ii) False
high-quality extractions present optimal quality-index values
with respect to domain-expert thresholds, but at least an external
variable has a sub-optimal value [
We refer to true high-quality extractions as optimal, and we characterize them as a function of the proposed time-series features s1 and s2. Let O be the set of optimal extractions {o1, o2, ..., oN }, where each point oi ∈ O is defined in terms of s1 and s2, i.e., oi = (oi_s1, oi_s2 ). We define novel quality thresholds for optimal extractions To_min and To_max in the (s1, s2) space as follows: To_min = (min(oi_s1 ), min(oi_s2 )) (6)
To_max = (max (oi_s1 ), max (oi_s2 )) (7) Among the whole set of espresso extractions E = {e1, e2, ..., eM }, a generic sample ej = (ej_s1, ej_s2 ) ∈ E is labeled as optimal e ∈ O, with O ⊆ E, if its values of flow rate ( ej_s1 , ej_s2 ) are within the thresholds To_min and To_max .
In Figure 5 two rectangular areas are shown. The green area contains the optimal extractions. Its boundaries are defined by the thresholds To_min and To_max . The orange dashed area contains the false high-quality extractions, which current state-of-theart solutions would (incorrectly) classify as high-quality cofees. Exploiting the proposed thresholds in the new feature space, we can detect many false positives (orange squared points in the plot). Specifically, instead of assigning an optimal label to the overall 67 extractions (green and orange ones), we can correctly detect the 20 true optimal extractions (green ones), and we can discard 31 out of 47 false positives (orange ones). State of the art thresholds would lead to the same true positive detection (20 out of 67), while the proposed approach leads to a drastically better accuracy (76% instead of 30%) and precision of high-quality extractions (56% instead of 30%).
To drill down the analysis, we further distinguished two types of false positives, stemming from diferent compensation efects: (i) low amount of cofee ground with fine grinding and (ii) high amount of cofee ground with coarse grinding. The former is less common, since very few baristas intentionally use higher amounts of cofee ground, being a cost for them. On the contrary, the latter is much more frequent, because it brings savings on cofee ground costs. For this reason, extractions afected by the latter are of greater interest.
In Figure 6 three areas are shown. The green one still contains the true optimal extractions, the blue one contains the extractions belonging to the first type of compensation and the orange one now contains only the extractions belonging to the second type of compensation. Again, exploiting thresholds in the new feature space, the target extractions can be correctly classified and the compensation efect can be detected. Results show that all 23 extractions from type-(ii) compensation can be correctly detected, besides 8 extractions out of 24 from type-(i) compensation, which means improving from 30% accuracy of data-driven state of the False high-quality extractions
True optimal extractions 4.0 4.5 5.0
Slope 1 5.5 6.0 6.5 2.0 1.5 4.0 4.5 6.0
6.5 5.0
Slope 1 5.5 art to 100% accuracy considering only true optimal and type(ii) compensation extractions. To this aim, in our dataset, the new feature thresholds have been set as 5.19 < s1 < 5.48 and 2.64 < s2 < 3.73. 6
This work presented a data-driven analysis of a real-world timeseries dataset of espresso brewing by professional cofee-making machines. The proposed feature space, despite being simple and easy to compute, brought a large improvement in the classification accuracy of high-quality espresso with respect to current state-of-the-art data-driven approaches: results yielded to threefold improvements in accuracy, from 30% to 100%, with specific focus on currently misclassified extractions due to common compensation efects. The proposed methodology can be applied in similar contexts to improve current data-driven analyses of espresso quality.
Future works aim to widen the scope of the analysis including additional quality variables, definitely diferent models of professional cofee-making machines, diverse cofee blends, and environmental variables. Furthermore, we plan to apply clustering techniques for determining the quality-index thresholds.
This work is partially funded by the SmartData@PoliTO center.