Performance Evaluation of an Earthquake Early Warning System in the 2019–2020 M6.0 Changning, Sichuan, China, Seismic Sequence

China is currently building a nationwide earthquake early warning system (EEWS) which will be completed in June 2023. Several regions have been selected as pilot areas for instrumentation, software system and dissemination verification. For these regions, their construction tasks will be completed in advance with trial runs being carried out in June 2021. Before the trial operation, we need to understand the actual processing capabilities of different EEWSs. In this work, we focus on the system deployed in Sichuan province and evaluate its real-time performance during the 2019–2020 M6.0 Changning seismic sequence. This period was divided into two stages. The first stage was the time from the occurrence of the M6.0 (M w5.7) mainshock (June 17, 2019) to the end of October 2019 with no MEMS-based stations around the Changning seismic sequence deployed and most of the broadband and short period seismic stations not upgraded to low latency streaming, and the second one was from the beginning of November 2019 to March 2021 with deployments of more than 700 MEMS-based stations and low latency upgrades of ∼30 seismic stations. Median errors for the epicentral locations, depths and magnitude estimations were almost the same for both stages, 1.5 ± 6.0 km, 0.0 ± 3.6 km and −0.1 ± 0.46 for the first stage and 2.3 ± 3.0 km, −3.0 ± 3.6 km and −0.2 ± 0.32 for the second one. However, an obvious underestimation of the magnitude for earthquakes with M 5.0 + occurring in the first stage was observed, which would be caused by the clipped waveforms, sensors deployed in short period seismic stations and MEMS-based stations, the adopted magnitude estimation method, and the method used to computer the network magnitude. The median reporting time was significantly improved from 10.5 ± 3.0 s after origin time for the first stage to 6.3 ± 3.5 s for the second stage because of introduction of newly deployed MEMS-based stations. The results obtained from the second stage indicate that the system has entered a stable operating stage and we can officially launch the trial operation in the pilot areas for public early warning services.


INTRODUCTION
As an effective way to mitigate seismic hazards, earthquake early warning (EEW) systems (EEWSs) have shown their great potential by providing alarms immediately after a destructive earthquake occurs and before damaging seismic waves reach the target areas. They can enable people and automated systems to take action for preventing the potential damage, such as "drop, cover and hold on," slowing high-speed trains, shutting down gas pipelines, and stopping elevators at the nearest floor (Allen and Melgar, 2019;Cremen and Galasso, 2020). Nowadays, EEWSs are under testing or operational in many earthquake-prone countries and regions, like Japan (Hoshiba et al., 2008), Mexico (Cuéllar et al., 2017), the West Coast of the United States , China (Peng et al., 2011;Zhang et al., 2016;Peng et al., 2019;Peng et al., 2020), Turkey (Alick et al., 2009), Taiwan (Wu, 2014;Wu et al., 2016;Hsu et al., 2018), South Korea (Sheen et al., 2017), Italy (Zollo et al., 2009;Satriano et al., 2011), and India .
In June 2018, China started to build a nationwide EEWS. This project, namely the National System for Fast Seismic Intensity Report and Earthquake Early Warning project (hereafter called the National System project), was led by the China Earthquake Administration (CEA) and its implementation cycle is five years. Detailed information can be found in Peng et al. (2020). To make the project effective as soon as possible, several regions have been selected as pilot areas for instrumentation, software system and dissemination verification, including Sichuan province, Yunnan province, and the Beijing-Tianjin-Hebei region. For these pilot areas, all construction tasks will be completed in June 2021, and the trial run will then be carried out in advance for providing the second-level EEW information service to the public.
Before the trial operation, we need to understand the actual processing capabilities of different EEWSs. We have evaluated the real-time performance of the demonstration EEWS deployed in the Sichuan-Yunnan border region during the test period (2017)(2018). Although the results showed excellent magnitude estimation and epicentral location, the average report time for the first alert was relatively long, approximately 13.4 ± 5.1 s and 26.3 ± 13.5 s after the origin time respectively for the earthquakes occurring inside and outside the seismic network (Peng et al., 2020). This means that the average radius of blind zone was more than 45 km considering an S-wave velocity of 3.5 km/s and a shallow seismicity depth of 10 km, and it was unable to meet the key time goal of the National System project (4-6 s after the first trigger in the four key seismic zones). Here using the time after the first trigger as one of the system goal is that it would not be influenced by the varied focal depth (from several kilometers to more than 30 km) and whether an earthquake occurs inside or outside the seismic network.
Within the past 2 years (2019-2020), the EEWS has been upgraded by introducing more newly constructed stations and optimizing EEW algorithms. And for improving the 3G/4G network environment, all the micro-electro-mechanical system (MEMS) accelerometer based stations belong to the National System project were directly built in the houses of the cellular towers maintained by the China Tower company (https://ir. china-tower.com/en/business/macro.php), and about 50 MEMS-based stations deployed in the Sichuan-Yunnan border region (Peng et al., 2020) were moved to the same environment. In addition, an M6.0 (M w 5.7) Changning earthquake with many aftershocks of M3.0 + occurred in this period, which was the largest event occurring in the four key seismic zones since the build-out of the first prototype EEWS in 2009 (Peng et al., 2011). This seismic sequence provides us an outstanding opportunity to reevaluate the EEWS performance in the current situation.
In this study, we focus on the system deployed in Sichuan province and evaluate its real-time processing capabilities during the 2019-2020 M6.0 Changning, Sichuan, China, seismic sequence. We start by briefly summarizing the Changning seismic sequence. In the Sichuan Seismic Network, we introduce the Sichuan seismic network and present the network latency for different types of seismic stations. Then, we discuss the EEWS performances for the M6.0 Changning mainshock and all the selected aftershocks. The performance analysis is based on comparing the location, origin time, and magnitude estimated by the EEWS with the reference catalogs. In addition, we also analyze the report time of the first alert for the selected aftershocks.

CHANGNING SEISMIC SEQUENCE
The Changning area is located on the southern margin of Sichuan Basin and the western margin of the Yangtze Plate ( Figure 1). It is one of the main salt (well and rock) and shale-gas production regions, with fluid injection wells drilled to ∼3 km depths, and was considered as one of the sites of induced seismic events (Lei et al., 2019;Meng et al., 2019;Hu et al., 2021). The enhanced seismicity has been observed in the last ∼10 years and the occurrence of earthquakes has recently increased significantly. According to the China Earthquake Networks Center (CENC), there were no earthquakes of M 5 or higher in this area before December 2018. However, since the occurrence of December 16, 2018 Xingwen M5.7 earthquake, a devastating and long-lasting earthquake sequence with many M 5 + events started to shake this region, resulting in 13 people dead, more than 200 injured and thousands of buildings damaged (Yi et al., 2019;Li et al., 2020). Here, the generic symbol M denotes either local magnitude M L or the surface-wave magnitude transferred from an empirical estimating equation M S 1.13M L -1.08, for earthquakes below or above M L 5, respectively (Chen et al., 2014).
Until now, the seismic sequence generated more than 100 events with M larger than 3.0 ( Figure 1). The largest magnitude event (M6.0 Changning earthquake) occurred on June 17, 2019, at 22:55 Beijing Time (14:55 UTC), locating in the southwestern margin of Sichuan Basin with a very shallow focal depth of ∼5 km ( Table 1). It was a doublet with two faulting styles (Liu and Zahradník, 2020). The initial thrust faulting could be associated with previously known reverse faults, which were favorably oriented to the regional stress field, while the following strikeslip faulting was probably facilitated by increased pore pressure because of previous injections. The event caused Modified Mercalli Intensity VIII near the epicenter, with the affected area larger than VI being about 2,538 km 2 . The shock was widely felt in many places of Sichuan, Yunnan, Chongqing, and Guizhou provinces, including Chengdu city approximately 260 km from the epicenter.

SICHUAN SEISMIC NETWORK
The Sichuan seismic network is operated by the Sichuan Earthquake Administration. It was constructed by several projects, including the National System project, the Sichuan-Yunnan Demonstration project (Peng et al., 2019;Peng et al., 2020), the Western Sichuan Earthquake Monitoring Capability Improvement project, the Jiuzhaigou Earthquake Recovery and Reconstruction project, and the Kangding Earthquake Recovery and Reconstruction project. The stations are classified into three types: seismic stations, strong-motion stations, and low-cost MEMS-based stations. A seismic station is equipped with a three-component broadband seismometer and a force-balanced accelerometer, and a strong-motion station is only equipped with a force-balanced accelerometer. Detailed information is shown in Table 2. In addition, some stations built by other agencies are also included into the seismic network, such as reservoir stations  (equipped with short period seismometers) and other provincial stations in the border regions. For a newly constructed station, to ensure its stable operation, we need it to go through a trial run of 3 months before being included into the EEWS. Until now, although construction of ∼80% of the seismic stations and strong-motion stations belong to the National System project and the Jiuzhaigou Earthquake Recovery and Reconstruction project was completed, these stations are at the stage of data quality verification and have not been included into the EEWS. The total number of stations currently used in the EEWS is approximately 2000, containing 363 seismic stations and 1631 MEMS-based stations ( Figure 1A). All stations are operated at a sampling rate of 100 Hz and the real-time continuous data are transferred to the Sichuan Earthquake Data Processing and Alert Issuing Center and processed by the Java-based EEWS (JEEWS; Peng et al., 2020).
Most MEMS-based stations were installed in the houses of the cellular towers and used their field supervision unit (FSU) for data transmission, which could significantly improve the network environment and lower maintenance costs. Except for some not upgraded broadband and short period seismic stations, most of the data loggers at the stations integrate a lowlatency data packetizing function which is designed specifically to support the EEWS (Peng et al., 2015;2017a). Theoretically, the data packetizing length can be reduced to 0.1 s. Here, for reducing network load, we chose either an interval of 0.5 or 1.0 s to packetize the real-time data depending on the telemetry type. For those not upgraded seismic stations, they still used the high-latency data transmission protocol with 512 bytes packet.
In JEEWS, the P d scaling from Kuyuk and Allen (2013) is adopted for EEW magnitude estimation. When the system detects S-wave for stations close to an earthquake epicenter (<10 km), the local M L scaling is directly used to calculate the magnitude for avoiding S-wave contamination (Colombelli et al., 2012). The reason is that the peak ground motion will reach these stations soon after the S-wave arrival, generally in 1 or 2 s (Peng et al., 2020). If we still use S-wave data to estimate the P d magnitude, a larger deviation will be introduced. Immediately after the magnitude estimated, a point-source algorithm based on the traditional ground motion prediction equation (GMPE) is used to compute seismic intensities at different target sites.
Using the method similar to Stubailo et al., 2021, we measured data latencies for all stations in March 2021 ( Figure 2). Here data latency is defined as the time difference between the time when the last sample of the data packet was recorded and the time when the same packet is received by the data processing center. Because data processing by the station and waveform-receiving server at the center are very rapid, the measured latency can be interpreted as the delay of data transmission in the telemetry system. Therefore, we ignored the delay awaited between subsequent packets, approximately the packet length. The median values for broadband seismic stations, short period seismic stations, MEMS-based stations and all stations were 2.9 ± 1.61 s, 0.87 ± 1.54 s, 0.5 ± 0.42 s and 0.59 ± 1.1 s, respectively. The relative large standard deviations for broadband and short period seismic stations were due to the high data latencies introduced by those not upgraded stations. FIGURE 2 | Normalized data latency distribution (%) for the Sichuan Seismic Network. Here distribution in each class of instrument is normalized to its total number of sample. All MEMS-based stations sent 0.5 or 1.0 s data packets; for the upgraded broadband and short period seismic stations, they used the same low-latency data packetizing function as the MEMSbased stations and acquired data with data packetizing of 0.5 or 1.0 s; for the broadband and short period seismic stations not upgraded, they sent data after the data packet is full and compressed.

PERFORMANCE OF JEEWS
System Performance for the Changning Seismic Sequence  Frontiers in Earth Science | www.frontiersin.org July 2021 | Volume 9 | Article 699941 5 November 2019 to March 2021. In the first stage, the MEMSbased stations around the Changning seismic sequence were not deployed, and most of the broadband and short period seismic stations were not upgraded ( Figure 1B). The nearest MEMSbased station belong to the Sichuan-Yunnan Demonstration project (Peng et al., 2019;Peng et al., 2020) was 50 km away from the sequence. Additionally, most of the short period seismic stations were temporary stations and transferred real-time data with the old high-latency protocol. Therefore the report time for earthquakes occurring in this stage would be significantly influenced and we needed to evaluate the performance of the two stages separately.
For the 101 earthquakes, 62 occurred in the first stage and 39 happened during the second one. All earthquakes with M 5.0 + occurred in the first stage. Here the first alert was selected to derive all statistics because of its importance for EEW. Although other EEW projects have moved towards more ground motion based metrics and warning time calculations (Cochran et al., 2018;Chung et al., 2020;Meier et al., 2020), at this stage of the National System project, we are mainly concerned about the correctness of results of the first alert. The result of each earthquake identified by the EEWS was compared with those in the China Earthquake Network Center (CENC) catalogs. In addition, three alert levels were defined according to the catalog magnitude of an earthquake. They are successful alert (SA), false alert and missed alert. Based on these definitions, as for the M 3.0 + earthquakes, JEEWS issued 55 SA events and missed seven events for the first stage and 38 SA events and missed one event for the second stage. No false alert was issued during the both periods. Detailed results for the SA events are shown in Figure 3. Figure 3A represents the epicentral error of SA events. The median deviations in epicenter location for the first and second stages were 1.5 ± 6.0 and 2.3 ± 3.0 km respectively. About 84% of the SA events had an error of less than 5 km. Among the earthquakes with epicentral errors of more than 10 km, there were 5 (9.1%) in the first stage and 2 (5.3%) in the second stage.
For an EEWS, depth is usually difficult to be determined correctly, especially for earthquakes occurred outside the seismic network (Carranza et al., 2017). Many studies have ignored this parameter and not analyzed it (Zhang et al., 2016;Sheen et al., 2017;Peng et al., 2020). From Figure 3B, one can observe that depth errors for most earthquakes lie between −5 and 5 km, with median errors of 0.0 ± 3.6 and −3.0 ± 3.6 km for the first and second stages, respectively. This means that the depth value of the first alert is excellent, although the depth error for the second stage is relatively large, which was caused by the MEMS-based stations and the events themselves being of a relatively smaller magnitude. In the second stage, most of the triggered stations contributing to the events were the newly constructed MEMSbased stations. Because of their high noise levels relative to the seismic stations, it is difficult to accurately identify P-wave arrivals from their recorded data for small earthquakes with magnitude less than 4.0, leading to a relatively large depth error.
Generally events with a large deviation in epicenter estimate also had a large deviation in the OT (Figures 3A,C). For the first and second stages, the median differences in OT were 0.4 ± 0.7 s and 0.3 ± 0.4 s, respectively. And ∼69 and ∼89% of the events in both stages had errors of less than 0.5 and 1.0 s. There were eight events (14.5%) with an OT error larger than 1.0 s for the first stage and 4 events (10.5%) for the second stage. Figure 3D shows the deviation in M. About 83% of the events had an error of no more than 0.5 and 61% of the events had a deviation of less than 0.4. The median magnitude differences for the first and second stages were −0.1 ± 0.46 and −0.2 ± 0.32 respectively. For the larger events with M ≥ 4.0 and M ≥ 5.0, the deviations were −0.5 ± 0.55 and −0.9 ± 0.71 respectively, indicating that JEEWS significantly underestimated the magnitude.
We further compared the magnitudes obtained from events in both stages according to JEEWS with those in the CENC catalogs ( Figure 4). The magnitudes estimated at broadband seismic stations, short period seismic stations, and MEMS-based stations are indicated by empty circles, crosses, and triangles, respectively, while the network magnitudes, the average value of all station magnitudes, are presented as blue circles. Here, both P d magnitude and M L based station contributions are shown in the graph. The same trend of magnitude underestimation for larger events could also be found in Figure 4. It is alarming that an extrapolation of Figures 4A,C,E suggests that a M7.0 earthquake would get a M EEW 5.5. However, the broadband seismic station magnitudes are better correlated with the CENC magnitudes relative to the short period station and MEMSbased station magnitudes, especially for the larger earthquakes. In addition, we also considered each type of station's performance in magnitude bins [3, 4), [4, 5) and [5, 6], shown in Table 3. From Figure 4 and Table 3, one can observe that: • Measurements for the broadband seismic stations in general slightly overestimate M3-4 events and slightly underestimate earthquakes with higher magnitude. Scatter is generally lower for this station type. • Measurements for the short period seismic stations in general more strongly overestimate M3-4 events, are good for M4-5 and slightly underestimate earthquakes with higher magnitude. However, scatter is very large (up to two magnitude units) for this station type. This may be a result of the deployment sites for this sensor type. Most of the short period seismic stations were temporary stations directly deployed on the hard soil layer. This would lead to amplification effects on the recorded ground motion. • For the MEMS-based stations, measurements in general slightly overestimate M3-4 events, are good for M4-5 and more strongly underestimate earthquakes with higher magnitude. Scatter is similar to the broadband seismic stations, except for the magnitude bin [5, 6].
Reporting time, as another important feature, directly influences the effectiveness of an EEWS. From Figure 3E we can find that there was a huge difference in the reporting time for earthquakes occurring between the first and second stages. The median reporting times for the first and second stages were 10.5 ± 3.0 s and 6.3 ± 3.5 s, respectively.

Real-Time Results for the M 5 + Earthquakes
From the occurrence of the M6.0 mainshock, five earthquakes with magnitudes larger than 5.0 struck the Changning area. The list of events is presented in Table 1. All these earthquakes occurred in the first stage with no MEMS-based stations introduced and no broadband and short period seismic  stations upgraded ( Figure 1B). Therefore the first reporting time for these earthquakes was long ( Figure 5), with an average of 11.6 ± 1.2 s after origin time. For the two Gongxian events, they are faster due to closer to the non-MEMS sub-networks. Additionally, although the location and depth errors were small with slight variations, the magnitudes for these earthquakes were significantly underestimated with large fluctuations and did not seem to converge toward the catalog values, except the Changning M5.1 event. Considering an S-wave velocity of 3.5 km/s, these earthquakes had a blind zone with radius of more than 30 km.

DISCUSSION AND CONCLUSION
After several years of construction, the number of stations of the Sichuan seismic network had reached ∼2000, and the average distance between stations was about 12 km, which is comparable to the Italian Strong Motion Network used for PRESTo (Festa et al., 2018) and the California Integrated Seismic Network used for ShakeAlert . In some regions, such as the Xianshuihe, Longmenshan and Xiaojiang fault zones ( Figure 1A), the average interstation distance was less than 10 km. In addition, most stations integrated a low-latency data packetizing function for real-time continuous data transmission (Peng et al., 2015;2017a). These factors have greatly improved the EEW capabilities, which could be observed from the EEW results of all earthquakes with M3.0 or more in the Changning seismic sequence.
However, during the first stage, the MEMS-based stations around the Changning seismic sequence were still under construction and most data loggers installed in the broadband and short period seismic stations were not changed to those specially designed for EEW purposes, such as EDAS-24GN (http://www.geolight.com.cn/p.aspx?id 49&&type 2) and HG-D (http://www.szadpr.cn/a/yewufangxiang/xianjindizhenyiqishebei/ Frontiers in Earth Science | www.frontiersin.org July 2021 | Volume 9 | Article 699941 20200716/442.html), leading to reporting time more than 10 s for the first alerts of most earthquakes. In the second stage, with the introduction of more than 1,000 MEMS-based stations, the reporting time has been significantly improved to 6.3 s, meaning that radii of the blind zones for most events have been reduced to ∼20 km under an S-wave velocity of 3.5 km/s and a shallow seismicity depth of 7 km. This is an impressive result and is comparable to those obtained by other excellent EEWSs, such as ShakeAlert , NEEWS (Hsu et al., 2018), on-site P-Alert system , and PRESTo in Italy (Festa et al., 2018). In addition, this result can meet the key time goal requirement of the National System project.
As to the seismic stations not upgraded, they adopted a fixed-size packet of 512 bytes for data transmission. Generally, during an earthquake, fixed-size data packet will be filled out quicker with large amplitudes, leading to shorter duration. However, the delay for transmitting and receiving the packet might become problematic during a large earthquake as stations sending high data rates will be a significantly additional stress for the telemetry infrastructure (Stubailo et al., 2021). One example is shown in Figure 6. For those not upgraded seismic stations transferring data with cellular modems ( Figure 6A), huge latency could be observed after shaking onset. However, for the seismic stations delivering data over fiber lines ( Figure 6B), a clear downward trend was presented because the packet was filled out quicker during the earthquake. Only a slight fluctuation was observed for the MEMS-based stations because of their low resolution (14-15 bits). The sensitivity of all MEMS-based accelerometers has been corrected to 500 counts/cm/s 2 . This means that 1 count is equal to 0.002 cm/s 2 . Usually, in the quiet time, the recorded data for an MEMS-based station vary between ±0.1 gals (±50 counts). During a large event, if the PGA value recorded can reach ±100 cm/s 2 , the counts will fluctuate between ±50,000, significantly lower than the seismic stations with waveforms clipped and approximately ±8,000,000 counts (23 bits resolution). Therefore, to lower the large latency introduced by the seismic stations, we can upgrade the data loggers by integrating a low-latency data packetizing function and increase the telemetry bandwidth by substituting the cellular modem links with fiber lines.
In terms of earthquake location and magnitude estimation, most of the deviations were within 5 km and 0.5 magnitude units, a significant improvement compared to the results reported in Peng et al. (2020). In addition, most of the depth errors lied in ±5 km. The reason is mainly related to the increase in station density and the improvement of the network environment. However, eight earthquakes with M3.0 or more were missed by JEEWS, all of which occurred within a short time after earthquakes at the same location. Currently, a time threshold of 49 s, an empirical value, is used for an event detection to avoid declaring a secondary trigger at the same station as a new earthquake. This is an inherent limitation of JEEWS, especially in case of a foreshock and the mainshock occurring in a short time, which cannot be avoided because reducing the time threshold for new event detection would be controversial and could lead to many false alerts (Carranza et al., 2017;Peng et al., 2020). To reduce the impacts of this limitation on JEEWS, some effective methods like the integrated particle filter (IPF; Tamaribuchi et al., 2014), propagation of local undamped motion (PLUM; Kodera et al., 2016) algorithms or using an  intensity threshold can be introduced to process multiple simultaneous events within a short time and distance. In JEEWS, one of the alert filters was the event being triggered by at least two stations. It can be seen from the results in Table 4 that this alert filter would lead to larger location and magnitude errors. For earthquakes occurring in the second stage, with seismic network density increased and seismic stations upgraded, most of the first alerts were triggered by three or more stations and the reporting time was not significantly influenced. Therefore, when in the trial operation, the alerting strategy should be adjusted for the pilot areas, requiring at least three stations to take part in the EEW processing for declaring an earthquake. However, for other regions with different station densities, the number of stations would probably varies from place to place.
From the results of earthquakes with magnitude larger than 5.0, there was an obvious underestimation of the magnitude, which persisted even as time increased with more triggered stations. This may be caused by the following four reasons: One is the seismic stations with clipped waveforms. Due to these larger earthquakes occurring in the first stage, except for some MEMS-based stations 50 km away from the epicenter, most stations used to estimate magnitude were seismic stations. Waveforms from the seismic stations close to the epicenter clipped, and some even clipped in the first second immediately after the P-wave arrival. These clipped data were not excluded and still used for EEW processing. To avoid this problem, we need to introduce a strategy to eliminate the clipped seismic stations, such as with a threshold on the number of counts.
The second reason comes from the sensors deployed in short period seismic stations and MEMS-based stations. It is well-known that the short period sensors have saturation problems when using to estimate the magnitude of a large earthquake (Xu et al., 2008), because they will miss the long period vibration of large events. However, the magnitude contributions are surprisingly good for these stations ( Figure 4C). The reason may be related to the energy release of the large earthquakes themselves, which is out of the scope of this paper. As to the MEMS-based stations, except for the GL-P2B device (Peng et al., 2017a(Peng et al., , 2019(Peng et al., , 2020, most of the deployed sensors, including Palert (Wu, 2014;Wu et al., 2019;Peng et al., 2020), with a low signal resolution (14-15 bits) and dynamic range (less than 90 dB), had a relative high noise level, making them difficult to clearly record p wave arrivals of small earthquakes and more suitable for earthquakes with M4.0+ ( Figure 4E).
The third one is the adopted magnitude estimation method itself (Kuyuk and Allen, 2013). The data selected to deriving the P d magnitude relationship were records from broadband seismic stations and strong motion stations. In JEEWS, we also used short period seismic stations and MEMS-based stations. For the sensors deployed in these stations, they had different frequency bands and noise levels. From Figures 4C-F, we could find that this magnitude estimation method was not working well for these two types of stations. Therefore we need to calibrate new relationships with data directly recorded by the stations.
The final one lies in the method used to compute the network magnitude. In the current system, the network magnitude outputted is directly obtained by averaging the magnitude of each triggered station. This would lead to magnitude underestimation because of the magnitudes calculated from the newly triggered stations (Melgar and Hayes, 2019;Trugman et al., 2019;Chung et al., 2020) using a short P-wave time window (less than 2 s), especially for the MEMS-based stations ∼50 km away from the epicenter (Table 3). Usually, for a large earthquake with rupture duration longer than 4 s, the magnitude calculated from such a short window will be significantly underestimated (Meier et al., 2016(Meier et al., , 2021Trugman et al., 2019). To avoid this risk, we can use the time window length of each triggered station as a weight for computing the network magnitude. The effectiveness has been proved in Colombelli et al. (2012) and Peng et al. (2017b).
For the M 3.0 + earthquakes that occurred in the second stage, JEEWS obtained excellent results, indicating that the system has entered a stable operating state. However there is a high risk that is lack of M 7.0 + earthquake verification. So far, the largest intranetwork event handled by JEEWS was the 2019 Changning M6.0 earthquake. In JEEWS we adopted a simple point-source algorithm to calculate seismic intensities at different target areas. This would lead to significant ground-motion underprediction for devastating (M 7.0+) earthquakes because fault finiteness was not considered . This limitation can be efficiently solved by introducing FinDer (Böse et al., 2012(Böse et al., , 2018 or other algorithms (Yamada et al., 2007;Crowell et al., 2016) into JEEWS.
Recently we are developing a decision module which is used for combining results from different EEW algorithms. In this module, some rules are currently adopted, such as 1) the first alert is the fastest received result; 2) if multiple results are received in a time window of less than 1 s, the origin time, epicenter and depth from the one with the largest number of triggered stations are selected as the source parameters, and the magnitude is set as the largest one. Based on the results after decision, peak ground motion estimates at different target sites are calculated by using GMPEs. As the system runs, more suitable rules will be integrated into this module for further improving accuracy of the system.
Besides, in the near future we need to consider providing customized services from the perspectives of an end user, such as triggering interpretable alerts according to probabilistic riskbased estimation optimized for the preferences of a given stakeholder (Cremen and Galasso, 2020), or providing an open public (event or ground motion-based) service to users for building their own customized applications. This will further improve the efficiency of an EEWS and transform it into a more useful tool.

DATA AVAILABILITY STATEMENT
The datasets for this study, including seismic waveforms recorded by broadband and short period seismic stations, the microelectromechanical system (MEMS) based recordings, processing results of the earthquake early warning system (EEWS) and data latency for each station, can be obtained from the Sichuan Earthquake Administration (http://www. scdzj.gov.cn). Please contact jiang_0057@163.com for information about how to access the data for research purposes. The earthquake catalogs used for compared with results of the EEWS are acquired from the China Earthquake