WO2022116325A1 - 一种边缘计算下神经网络推理的节能方法 - Google Patents

一种边缘计算下神经网络推理的节能方法 Download PDF

Info

Publication number
WO2022116325A1
WO2022116325A1 PCT/CN2020/140017 CN2020140017W WO2022116325A1 WO 2022116325 A1 WO2022116325 A1 WO 2022116325A1 CN 2020140017 W CN2020140017 W CN 2020140017W WO 2022116325 A1 WO2022116325 A1 WO 2022116325A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
edge device
energy consumption
edge
subnet
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2020/140017
Other languages
English (en)
French (fr)
Inventor
代浩
王洋
须成忠
叶可江
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Institute of Advanced Technology of CAS
Original Assignee
Shenzhen Institute of Advanced Technology of CAS
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen Institute of Advanced Technology of CAS filed Critical Shenzhen Institute of Advanced Technology of CAS
Publication of WO2022116325A1 publication Critical patent/WO2022116325A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/06—Testing, supervising or monitoring using simulated traffic
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W52/00—Power management, e.g. Transmission Power Control [TPC] or power classes
    • H04W52/02—Power saving arrangements
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • the invention relates to the technical field of edge computing networks, in particular to an energy-saving method for neural network reasoning under edge computing.
  • edge systems have generated a large amount of data.
  • Such a large amount of data causes the traditional cloud architecture to face considerable challenges in terms of communication, storage, etc.
  • many new applications (such as collaborative autonomous driving) have strict latency requirements, and computing needs to be deployed close to the user equipment to provide real-time response. Therefore, researchers have proposed a mobile edge computing (MEC) method to Pushing artificial intelligence to the edge of the network can effectively solve communication and storage problems and give full play to the potential of edge computing.
  • MEC mobile edge computing
  • the main disadvantage of using deep learning in MEC is the limited resources, which include computing, storage, bandwidth, etc.
  • Another resource that MEC needs to consider is energy consumption, because most end devices are Battery powered, therefore, it is not feasible to put the entire neural network (DNN) on the edge network, and the usual approach is to combine the advantages of the cloud and the edge to synergistically complete the task of neural network inference.
  • DNN neural network
  • the mainstream deployment methods of DNN in MEC are edge collaborative reasoning and cloud center training, that is, the DNN model is trained in the cloud center, and the edge device infers the DNN model.
  • This mode has some specific characteristics: 1) The cloud center will The received data continuously trains the model and updates the parameters; 2) The inference edge needs to be consistent with the model parameters of the cloud center to ensure that the inference result of each request is correct. The request usually continues to arrive at the edge device randomly, and the inference also needs to continue to provide services. , however, because edge devices typically have limited energy, keeping inference programs online for long periods of time is a waste of energy.
  • the main technical problem solved by the present invention is to provide an energy-saving method for neural network reasoning under edge computing, to perform energy consumption comparison and calculation, so as to select the optimal path with the lowest energy consumption, thereby ensuring that all random accesses on the device are completed.
  • it is efficient and energy-saving without wasting energy consumption.
  • a technical solution adopted by the present invention is to provide an energy-saving method for neural network reasoning under edge computing, which includes:
  • the edge device downloading the model from the cloud center compares the energy consumption of the edge device downloading the model from the cloud center with the energy consumption of the edge device transferring the model from other edge devices in the subnet. If the energy consumption of the edge device downloading the model from the cloud center is less than or equal to the energy consumption of the edge device transferring the model from other devices in the subnet, the edge device directly downloads the model from the cloud center; if the energy consumption of the edge device downloading the model from the cloud center is greater than the energy consumption of the edge device transferring the model from other devices in the subnet consumption, the edge device transfers the model from other devices in the subnet; wherein, the subnet consists of an access point that communicates with the cloud center and several edge devices that communicate with the access point;
  • the energy consumption of the continuous operation of the edge device is the energy consumption of the edge device from the last request until the edge device continues to run. Energy consumption for each request.
  • the edge device transfers the model from other edge devices in the subnet
  • the energy consumption required by the edge device to transfer the model from each other edge device in the subnet is compared, Select the transfer model from the edge device in the subnet that requires the least energy consumption.
  • the edge device transfers the model from other edge devices in the subnet
  • the edge device makes a request
  • the model is first transferred to the subnet to continuously run with the least energy consumption
  • the model is then transferred from the edge device with the least energy consumption for continuous operation to the edge device that makes the request.
  • the edge device directly downloads the model from the cloud center.
  • the edge device downloading the model from the cloud center is less than or equal to the energy consumption of the edge device transferring the model from other devices in the subnet, the edge device will directly download the model from the cloud center. Center to download the model.
  • the edge device transfers the model from other devices in the subnet.
  • the edge device downloads the model from the cloud center.
  • the edge device in the case of unpredictable request sequence, if the energy consumption of the edge device for downloading the model from the cloud center is more than twice the energy consumption of the edge device for transferring the model from other devices in the subnet, the edge device will Devices transfer models from other devices in the subnet.
  • the energy consumption of the edge device for downloading the model from the cloud center is greater than twice the energy consumption of the edge device for transferring the model from other devices in the subnet
  • the The model is transferred to the edge device with the least energy consumption for continuous operation in the subnet to run continuously.
  • the model is transferred from the edge device with the least energy consumption for continuous operation to the edge device that sends the request.
  • the beneficial effects of the present invention are: compared with the prior art, the present invention performs energy consumption comparison calculation, so as to select the optimal path with the lowest energy consumption, thereby ensuring the premise of completing the reasoning of all random access requests on the device Under the high-efficiency energy saving, no energy consumption is wasted.
  • Figure 1 is a schematic diagram of an edge computing architecture
  • Figure 2 is a time-space diagram when an edge device sends a request sequence under the condition of a predictable request sequence
  • Figure 3 is a spatiotemporal diagram of continuous operation of edge devices under the condition of predictable request sequences
  • FIG. 4 is a spatiotemporal diagram of an embodiment of an edge device transitioning a model from other edge devices in a subnet under the condition of a predictable request sequence
  • FIG. 5 is a spatiotemporal diagram of another embodiment of an edge device transferring a model from other edge devices in a subnet under the condition of a predictable request sequence;
  • Figure 6 is a spatiotemporal diagram of the edge device only synchronizing with the cloud center in the case of an unpredictable request sequence
  • FIG. 7 is a spatiotemporal diagram of an edge device transferring a model from other edge devices in the subnet under the condition of an unpredictable request sequence
  • Figure 8 is a spatiotemporal diagram of transferring a model to an edge device with the least energy consumption in the case of an unpredictable request sequence
  • Fig. 9 is the spatiotemporal diagram of the active update model of the cloud center
  • Figure 10 is a comparison diagram of experimental results in the case of unpredictable request sequences.
  • the existing cloud center combined with edge computing artificial intelligence services are usually implemented using the architecture of "cloud center training - edge inference". Since the training is carried out in the cloud center, the model in the cloud center is constantly updated, while the reasoning It is done on the user device, each device receives multiple requests in succession and then infers based on the input data of these requests, but for some practical reasons (e.g. privacy concerns) it is usually not possible to route requests to other The device or cloud center performs inference, so it can only request inference on the current local device.
  • cloud center training - edge inference Since the training is carried out in the cloud center, the model in the cloud center is constantly updated, while the reasoning It is done on the user device, each device receives multiple requests in succession and then infers based on the input data of these requests, but for some practical reasons (e.g. privacy concerns) it is usually not possible to route requests to other The device or cloud center performs inference, so it can only request inference on the current local device.
  • the existing technology has the following shortcomings:
  • the model on each device needs to be synchronized with the cloud center in real time, so a large amount of synchronized data transmission is required, resulting in a waste of network bandwidth and device energy consumption;
  • the purpose of the present invention is to study in such a scenario, how to reduce unnecessary energy loss, achieve efficient model synchronization, and provide an energy-saving method for neural network inference under edge computing, including:
  • the edge device downloading the model from the cloud center compares the energy consumption of the edge device downloading the model from the cloud center with the energy consumption of the edge device transferring the model from other edge devices in the subnet. If the energy consumption of the edge device downloading the model from the cloud center is less than or equal to the energy consumption of the edge device transferring the model from other devices in the subnet, the edge device directly downloads the model from the cloud center; if the energy consumption of the edge device downloading the model from the cloud center is greater than the energy consumption of the edge device transferring the model from other devices in the subnet consumption, the edge device transfers the model from other devices in the subnet; wherein, the subnet consists of an access point that communicates with the cloud center and several edge devices that communicate with the access point;
  • the present invention performs energy consumption comparison and calculation, thereby selecting the optimal path with the lowest energy consumption, thereby ensuring high efficiency and energy saving without wasting energy consumption on the premise of completing the reasoning of all random access requests on the device.
  • the edge computing architecture is a three-layer architecture, as shown in Figure 1, including a cloud center (represented by "Internet”), an access point (represented by “access point”), and an edge device (represented by "device”)
  • a cloud center represented by "Internet”
  • an access point represented by "access point”
  • an edge device represented by "device”
  • One access point and multiple edge devices form a subnet, and the cloud center often needs to communicate with the devices through a complex network. It can be seen that the cost of communication between devices under the same subnet is much less than The communication cost between the device and the cloud center.
  • the request for inference occurs on each device, and the device needs to infer the requested data according to the latest model and return the result. Therefore, it is not only necessary to ensure that the latest model exists on the device, but also that the device is running the inference service when the request arrives.
  • the energy consumption on all devices can be calculated using the following formula:
  • CI is the energy consumption required to infer all requests.
  • This part can be regarded as a constant value. Once the request sequence is determined, the constant is determined; Ru represents the number of model updates, and ⁇ i represents the communication between each device and the cloud center.
  • the unit energy consumption consumed, this part represents the energy consumption caused by all devices synchronizing the model with the cloud center; Tmax represents the time when the last request occurred, ⁇ i represents the unit energy consumption consumed by each device running the inference service continuously, this part represents all the The energy consumption caused by the device running the inference service all the time.
  • the model (b) in addition to obtaining the latest model (c) from the cloud center, the model (b) can also be obtained from other devices under the same subnet.
  • algorithms can also be used to ensure continuous operation on the device.
  • the model is up-to-date (a), so it is only necessary to keep the model on a device in a subnet continuously updated, and the model on all devices can be updated by transferring between devices.
  • three ways are defined to serve each request: continuous operation (holding), transfer from other devices (transferring) and download from the cloud center (pulling), and respectively define the energy consumption that the three ways need to consume:
  • optimizing the energy-saving problem of device reasoning in edge computing is transformed into how to combine the above three methods to provide services for all requests; define different devices as si, and the request ri can be expressed as a space-time point (si, ti), ti represents the time when the request occurred, so the request sequence can be transformed into a space-time diagram as shown in Figure 2, where s0 represents the cloud center, that is, the vertical line from s0 is the request satisfied by downloading (pulling) from the cloud center, Vertical lines from s(i) other than s0 represent requests that are satisfied by transferring from other devices, and horizontal lines from s(i) other than s0 represent inference requests being served by holding.
  • the energy consumption of the continuous operation of the edge device is the energy consumption of the edge device from the last request to the current request of the edge device. energy consumption.
  • the edge device transfers the model from other edge devices in the subnet, compare the required energy consumption of the edge device from the transfer model of each other edge device in the subnet, and select the energy consumption from the required energy consumption.
  • the edge device in the least subnet performs the transfer model.
  • the edge device transfers the model from other edge devices in the subnet
  • the edge device makes a request
  • it first transfers the model to the edge device with the least energy consumption for continuous operation in the subnet, and then Shift the model from the edge device that consumes the least energy to run continuously to the edge device that makes the request.
  • the edge device directly downloads the model from the cloud center.
  • the problem can be decomposed into sub-problems for consideration, assuming that the first i-1 requests are optimal, denoted by C(i-1), and For the optimal solution C(i) of the i-th request, it can be divided into the following situations:
  • a minimum cost b(i) can be defined:
  • 1 ⁇ i ⁇ n 1 ⁇ i ⁇ n; it means that when request i has an active inference service in other edge devices, the inference service can be generated on the required edge device with a minimum cost, and Bi is used to represent the requirement of bi and, that is, all points in an interval can be satisfied with the minimum cost, when there are active inference programs of other edge devices in the subnet, so the recurrence of D(i) can be expressed as:
  • the optimal migration mode can be obtained under the condition of known request sequence, so as to save energy consumption to the greatest extent.
  • the present invention provides a It is an energy-saving method that saves the energy consumption of each device as much as possible without knowing when the request occurs.
  • the main idea of this method is to maintain a period of running time after each edge device request occurs to wait for it to happen again. The next request, so how long each edge device needs to maintain the inference program, is divided into the following:
  • the edge device directly downloads the model from the cloud center.
  • the edge device downloading the model from the cloud center is greater than the energy consumption of the edge device transferring the model from other devices in the subnet and less than the energy consumption of the edge device transferring the model from other devices in the subnet
  • the edge device transfers the model from other devices in the subnet; further, if other devices in the subnet do not have the required model, the edge device downloads the model from the cloud center;
  • the model will be transferred to the subnet first.
  • the model runs continuously on the edge device with the least energy consumption for continuous operation.
  • the model is transferred from the edge device with the least energy consumption for continuous operation to the edge device that sends the request.
  • the energy consumption ratio of the sequence can be lower than 2.5 times, that is to say, in the face of an unknown request sequence, it is also guaranteed to serve all inference requests with an energy consumption that is not 2.5 times worse than that of the algorithm whose optimal sequence is known in advance.
  • the edge reasoning architecture often requires the edge device to keep synchronized with the cloud center in real time to ensure the update of the model, and also requires the device to continuously run the reasoning service to meet random requests, which undoubtedly sacrifices a lot of device energy consumption.
  • the invention formalizes the problem into the node connection on the space-time graph, and uses the communication between devices under the same subnet to ensure the update of the model. Compared with the traditional method, we can greatly save the times and bandwidth of the communication between the device and the cloud center. Reduce equipment power consumption.
  • the number of inference requests that can be served by the device which increases the standby time of the device.
  • the synchronization between edge devices is used to replace the synchronization between edge devices and the cloud center to reduce communication energy consumption

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Neurology (AREA)
  • Mobile Radio Communication Systems (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)

Abstract

本发明涉及边缘计算网络技术领域，特别涉及一种边缘计算下神经网络推理的节能方法；本发明进行能耗比较计算，从而选择出最优的能耗最低的路径来，从而保证完成所有随机在设备上访问的请求的推理的前提下，进行高效节能，不浪费能耗，从而提高了在同样的能耗下设备端能够服务的推理请求数量，提升设备的待机时长。

Description

一种边缘计算下神经网络推理的节能方法 技术领域
本发明涉及边缘计算网络技术领域，特别涉及一种边缘计算下神经网络推理的节能方法。
背景技术
近年来，随着智能设备和无处不在的传感器的快速增长，边缘系统产生了大量的数据，如此之大量的数据导致传统的云架构在通信、存储等方面面临着相当大的挑战。此外，许多新型应用(如协同自主驾驶)都有严格的延迟要求，需要在接近用户设备的地方部署计算来提供实时响应，因此研究者提出了移动边缘计算(mobile edge computing,MEC)方式，将人工智能推向网络的边缘，可以有效解决通信和存储问题，充分发挥边缘计算的潜力。
MEC中使用深度学习的主要缺点是资源受限，这些资源包括计算、存储、带宽等，此外，与传统的云中心不同，MEC需要考虑的另一个资源是能源消耗，因为大多数终端设备都是电池供电，因此，将整个神经网络(DNN)放在边缘网络上是不可行的，通常的方法是结合云和边缘的优势，协同完成神经网络推理的任务。
目前，在MEC中DNN的部署方式主流为边缘协同推理和云中心训练，即在云中心训练DNN模型，边缘设备进行推理DNN模型，这种模式有一些具体的特点:1)云中心会根据接收到的数据不断训练模型并更新参数；2)推理边缘需要与云中心的模型参数一致，以保证每个请求的推理结果正确，请求通常会随机地 继续到达边缘设备，而且推理也需要继续提供服务，然而，因为边缘设备的能源通常有限，保持推理程序长期在线是一种能源浪费。
现有的技术较少在进行推理服务时考虑到边缘端设备的能耗问题，通常的方式是实时保持设备上的模型和云中心的设备进行同步，并且一直保持推理服务的持续运行，这种方式在请求比较稀疏的情况下十分浪费网络带宽和设备能耗，如一个预训练的YOLO模型约为200MB，一个预训练的VGG模型约为500MB，在多个设备需要同步时，每次需要上GB的流量来从云中心下载最新的模型到每个设备，造成了不必要的带宽和能耗浪费。
发明内容
本发明主要解决的技术问题是提供一种边缘计算下神经网络推理的节能方法，进行能耗比较计算，从而选择出最优的能耗最低的路径来，从而保证完成所有随机在设备上访问的请求的推理的前提下，进行高效节能，不浪费能耗。
为解决上述技术问题，本发明采用的一个技术方案是：提供一种边缘计算下神经网络推理的节能方法，其中，包括：
在可预知请求序列情况下，将边缘设备从云中心下载模型的能耗与该边缘设备从子网内其它边缘设备转移模型的能耗进行比较，如果边缘设备从云中心下载模型的能耗小于或等于边缘设备从子网内其它设备转移模型的能耗，则边缘设备直接从云中心下载模型；如果边缘设备从云中心下载模型的能耗大于边缘设备从子网内其它设备转移模型的能耗，则边缘设备从子网内其它设备转移模型；其中，子网由与云中心进行通信的接入点和若干个与该接入点通信的边缘设备组成；
在不可预知请求序列情况下，让边缘设备请求发生，维持一段时间，再将边 缘设备从云中心下载模型的能耗与该边缘设备从子网内其它边缘设备转移模型的能耗进行比较。
作为本发明的一种改进，在可预知请求序列情况下，该边缘设备进行请求时，该边缘设备的持续运行的能耗为该边缘设备从上一次进行请求一直持续运行到该边缘设备的这次请求的能耗。
作为本发明的进一步改进，在可预知请求序列情况下，该边缘设备从子网内其它边缘设备转移模型时，比较该边缘设备从子网内每个其它边缘设备的转移模型的所要能耗，选择从所要能耗最少的子网内那个边缘设备进行转移模型。
作为本发明的更进一步改进，在可预知请求序列情况下，该边缘设备从子网内其它边缘设备转移模型时，该边缘设备进行请求时，先将模型转移至子网内持续运行能耗最少的边缘设备上，再将该模型从持续运行能耗最少的边缘设备上转移至发出请求的边缘设备上。
作为本发明的更进一步改进，在可预知请求序列情况下，当上一个请求不来时，该边缘设备直接从云中心下载模型。
作为本发明的更进一步改进，在不可预知请求序列情况下，如果边缘设备从云中心下载模型的能耗小于或等于边缘设备从子网内其它设备转移模型的能耗，则边缘设备直接从云中心下载模型。
作为本发明的更进一步改进，在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗且小于该边缘设备从子网内其它设备转移模型的能耗的两倍时，则该边缘设备从子网内其它设备转移模型。
作为本发明的更进一步改进，在不可预知请求序列情况下，如果该边缘设备 从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗且小于该边缘设备从子网内其它设备转移模型的能耗的两倍时，而且子网内其它设备没有所需的模型，则该边缘设备从云中心下载模型。
作为本发明的更进一步改进，在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗的两倍时，该边缘设备从子网内其它设备转移模型。
作为本发明的更进一步改进，在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗的两倍时，先将模型转移至子网内持续运行能耗最少的边缘设备上持续运行，当有边缘设备发出请求后，再将该模型从持续运行能耗最少的边缘设备上转移至发出请求的边缘设备上。
本发明的有益效果是：与现有技术相比，本发明进行能耗比较计算，从而选择出最优的能耗最低的路径来，从而保证完成所有随机在设备上访问的请求的推理的前提下，进行高效节能，不浪费能耗。
附图说明
图1为边缘计算架构的示意图；
图2为在可预知请求序列情况下，边缘设备发出请求序列时的时空图；
图3为在可预知请求序列情况下，边缘设备进行持续运行的时空图；
图4为在可预知请求序列情况下，边缘设备从子网其它边缘设备转移模型的一种实施方式的时空图；
图5为在可预知请求序列情况下，边缘设备从子网其它边缘设备转移模型的 另一种实施方式的时空图；
图6为在不可预知请求序列情况下，边缘设备只与云中心同步的时空图；
图7为在不可预知请求序列情况下，边缘设备从子网其它边缘设备转移模型的时空图；
图8为在不可预知请求序列情况下，转移模型至能耗最少的边缘设备上的时空图；
图9为云中心主动更新模型的时空图；
图10为在不可预知请求序列情况下，实验结果对比图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
已有的云中心结合边缘计算的人工智能服务通常是采用“云中心训练-边缘端推理”的架构来实现，由于训练是在云中心进行的，因此云中心的模型是不断更新的，而推理则是在用户设备上进行的，每个设备会连续接收多个请求，然后根据这些请求的输入数据进行推理，但由于一些实际的原因(例如，隐私问题)，通常不可能将请求路由到其他设备或者云中心进行推理，因此它只能在当前本地设备上进行请求推理。
目前已有的技术有以下几点缺陷：
1、需要分别实时保持每个设备上的模型和云中心进行同步，因此需要大量的同步数据传输，造成网络带宽和设备能耗的浪费；
2、需要保持设备上的推理服务持续的运行，造成设备产生不必要的能耗；
3、针对随机发生的请求，已有的一些方法并没有理论分析能保证取得较好的在线竞争比。
因此，本发明的目的是研究在这样一个场景下，如何减少不必要的能耗损失，实现高效的模型同步，提供一种边缘计算下神经网络推理的节能方法，包括：
在可预知请求序列情况下，将边缘设备从云中心下载模型的能耗与该边缘设备从子网内其它边缘设备转移模型的能耗进行比较，如果边缘设备从云中心下载模型的能耗小于或等于边缘设备从子网内其它设备转移模型的能耗，则边缘设备直接从云中心下载模型；如果边缘设备从云中心下载模型的能耗大于边缘设备从子网内其它设备转移模型的能耗，则边缘设备从子网内其它设备转移模型；其中，子网由与云中心进行通信的接入点和若干个与该接入点通信的边缘设备组成；
在不可预知请求序列情况下，让边缘设备请求发生，维持一段时间，再将边缘设备从云中心下载模型的能耗与该边缘设备从子网内其它边缘设备转移模型的能耗进行比较。
本发明进行能耗比较计算，从而选择出最优的能耗最低的路径来，从而保证完成所有随机在设备上访问的请求的推理的前提下，进行高效节能，不浪费能耗。
在本发明内，边缘计算架构为三层架构，如图1所示，包含了云中心(“Internet”表示)，接入点(“access point”表示)，以及边缘设备(“device”表示)组成，其中一个接入点和多个边缘设备组成了一个子网，而云中心往往需要通过复杂的网络与设备间联通，可以看出在同一个子网下的设备间通信的代价是远远小于设备与云中心的通信代价的。推理的请求发生在每个设备上，设备需要根据最新的模型来推理请求数据并返回结果，因此不光要保证设备上存在最新 的模型，还需要在请求到达时设备正在运行推理服务。
在本发明内，可以用以下公式来计算所有设备上的能耗：
Figure PCTCN2020140017-appb-000001
其中CI是推理所有请求需要的能耗，这部分可以看作是一个常值，一旦请求序列确定了，这个常数也就确定了；Ru表示模型更新的次数，βi表示每台设备与云中心通信消耗的单位能耗，这一部分表示所有设备都跟云中心同步模型造成的能耗；Tmax表示最后一个请求发生的时间，μi表示每台设备持续运行推理服务消耗的单位能耗，这一部分表示所有设备都一直运行推理服务所带来的能耗。
从图1可以看出，除了从云中心获取最新的模型(c)外，还可以从同一子网下的其它设备上获取模型(b)，当然也可以通过算法来保证在设备上持续运行的模型是最新的(a)，因此只需要保持一个子网内的某台设备上的模型持续更新，即可通过设备间的转移来让所有设备上的模型都更新。
在本发明内，定义三种方式来服务每个请求：持续运行(holding)，从其它设备转移(transferring)以及从云中心下载(pulling)，并分别定义三种方式需要消耗的能耗：
1、持续运行(holding)：μi*δ(p(i),i)，其中，μi是边缘设备si上持续运行服务的单位消耗，而p(i)表示该边缘设备上一次请求发生的时间，i表示当前请求的时间，因此，持续运行(holding)的成本即是从上次请求持续运行到当前请求的能耗。
2、从其它设备转移(transferring)：由于各个边缘设备都处于同一个子网，因此边缘设备之间的传输消耗可以看作同构的，定义为λ。
3、从云中心下载(pulling)：同样由于边缘设备都处于同一个子网，因此所有边缘设备与云中心的传输消耗也是同样的，定义为β。
因此，优化边缘计算中设备推理的节能问题，转化为了如何组合上述三种方式，来为所有的请求提供服务；定义不同的设备为si，请求ri可以表示为一个时空点(si,ti)，ti表示请求发生的时间，因此请求序列可以转化为一个如图2的时空图表示，其中，s0表示云中心，即是从s0下来的竖线是通过从云中心下载(pulling)满足的请求，从除s0之外的s(i)下来的竖线表示从其它设备转移(transferring)满足的请求，除s0之外的s(i)的横线表示通过持续运行(holding)来服务推理请求。
对于请求序列预先可知的情况下，即是一个离线问题，可以看出目标是在这样的图2中时空图上找出最小的斯坦纳树来连接所有的请求点，这样的树即是在设备之间最节省能耗的迁移方案。
在本发明内，在可预知请求序列情况下，该边缘设备进行请求时，该边缘设备的持续运行的能耗为该边缘设备从上一次进行请求一直持续运行到该边缘设备的这次请求的能耗。
在本发明内，分为下面几种情况：
1、在可预知请求序列情况下，该边缘设备从子网内其它边缘设备转移模型时，比较该边缘设备从子网内每个其它边缘设备的转移模型的所要能耗，选择从所要能耗最少的子网内那个边缘设备进行转移模型。
2、在可预知请求序列情况下，该边缘设备从子网内其它边缘设备转移模型时，该边缘设备进行请求时，先将模型转移至子网内持续运行能耗最少的边缘设备上，再将该模型从持续运行能耗最少的边缘设备上转移至发出请求的边缘设备 上。
3、在可预知请求序列情况下，当上一个请求不来时，该边缘设备直接从云中心下载模型。
具体地讲，在本发明内，考虑通过递推公式来分析，可以将问题分解为子问题来考虑，假设前i-1个请求是最优的，用C(i-1)来表示，而对于第i个请求的最优解C(i)，可以分为以下几种情况：
1)通过持续运行(holding)来服务，则这种情况下的i是通过在si上一次请求一直持续运行到这次请求的模型来满足的，如图3所示，我们记为D(i)，可以看出当i到p(i)之间存在连线后，相应中间的请求将会变为通过从其它设备转移(transferring)或者比从其它设备转移(transferring)小的持续运行(holding)来满足，可以定义一个最小成本b(i)：
bi＝min{α*λ+(1-α)*β,μj*δ(p(i),i)},
其中，1≤i≤n；表示请求i在其它边缘设备有活跃的推理服务时，只需要最小的成本就可将推理服务要发生在所需边缘设备上的请求，用Bi来表示bi的求和，即是在一个区间上所有点都可以用最小的成本来满足，当在子网中存在其它边缘设备活跃的推理程序，因此D(i)的递推式可以表示为：
Figure PCTCN2020140017-appb-000002
2)通过从其它设备转移(transferring)来满足，此时其它边缘设备上应该存在着活跃的推理服务，当前边缘设备通过从其它边缘设备上获取最新的模型来完成推理，如图4所示，这时候的能耗成本记为T(i)，需要考虑一个转移成本，用以下递推式表示：
Figure PCTCN2020140017-appb-000003
3)从其它设备转移(transferring)还有一种特殊的形式，由于边缘计算下所有设备的成本i不一致，所以还有一种解决方案，即是将最新的模型转移到成本最低的边缘设备上持续运行，然后需要时再转移到对应的服务器上，如图5所示，这种情况下，需要考虑有可能需要两次从其它设备转移(transferring)才能到达需要服务的设备(先转移到最便宜的设备上，再从该设备转移到目标设备)，用以下递推式表达：
Figure PCTCN2020140017-appb-000004
4)通过从云中心下载(pulling)来满足，当请求迟迟不来时，没有必要在任何设备上维护最新的模型，因此下一个请求需要从云中心下载，这时的成本为β。
因此，最终最优解的递推式可以表示如下：
Figure PCTCN2020140017-appb-000005
通过上述递推式，从而可以在已知请求序列的情况下，得到最优的迁移方式， 来最大程度节省能耗。
在上述中，可预先知道请求序列的情况下可以有最优解，但实际情况中，还有是不可能预知请求什么时候发生的情况，所以在这种情况下，本发明提供了一种在不知道请求什么时候发生的情况下，尽可能的节省每个设备的能耗的节能方法，该方法的主要思想是在每个边缘设备请求发生后，多维持一段运行时间来等待是否还会发生下一个请求，因此每个边缘设备需要维持多长时间的推理程序，具体分为如下：
1、在不可预知请求序列情况下，如果边缘设备从云中心下载模型的能耗小于或等于边缘设备从子网内其它设备转移模型的能耗，则边缘设备直接从云中心下载模型。
2、在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗且小于该边缘设备从子网内其它设备转移模型的能耗的两倍时，则该边缘设备从子网内其它设备转移模型；进一步，如果子网内其它设备没有所需的模型，则该边缘设备从云中心下载模型；
3、在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗的两倍时，该边缘设备从子网内其它设备转移模型；
4、在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗的两倍时，先将模型转移至子网内持续运行能耗最少的边缘设备上持续运行，当有边缘设备发出请求后，再将该模型从持续运行能耗最少的边缘设备上转移至发出请求的边缘设备上。
具体地讲，分三种情况分别讨论：
1)β≤λ；这种情况下，跟云中心同步模型的能耗小于和其它边缘设备同步的能耗，因此只与云中心进行同步才是最优的选择，如图6所示，只需要让每个边缘设备持续运行的代价和与云中心通信的代价一样即可；这种情况下，很容易证明这个算法的代价与离线最优算法最差的情况下也不超过2倍，即是它的竞争比为2，如下公式：
Figure PCTCN2020140017-appb-000006
Figure PCTCN2020140017-appb-000007
2)λ＜β≤2λ；这种情况下，当有其它活跃的推理服务在运行时，可以通过转移来获取最新的模型，而没有时则需要与云中心同步，如图7所示，如(a)，当有其它边缘设备在运行时，当前边缘设备只需要维持成本与设备间转移一样的时间即可，当没有的情况下(如c)，则需要维持到与云中心同步的成本一致的时间，从而可以看出的竞争比是2：
Figure PCTCN2020140017-appb-000008
Figure PCTCN2020140017-appb-000009
Figure PCTCN2020140017-appb-000010
3)β＞2λ；这种情况下，与云中心同步的通信成本远高于设备间同步，尽可能的考虑在设备上让最新的模型持续运行，因此在每个请求结束后，可以将模型转移到代价最低的设备上持续运行，直到其它设备上有请求到来，通过转移来服务，如图8所示，只有当前设备运行时，可以持续运行两倍的转移成本时间，然后将其转移到smin上，运行β-2λ时间，可以到达2.5的竞争比：
Figure PCTCN2020140017-appb-000011
在本发明内，当云中心需要主动更新模型时，如图9所示，只需要选择其中一台边缘设备与云中心进行同步即可，其它边缘设备将立即删除该设备上运行的模型，以保证一致性，之后的请求同样可以通过上述的方法来满足。
如图10所示，在本发明内，经过实验数据的对比，在调整边缘设备的单位能耗成本、与云中心同步能耗以及设备间能耗的不同组合的情况下，面对未知的请求序列时可以低于2.5倍的能耗比，也就是说，在面对未知的请求序列时，也可以保证以不比预先知道序列最优的算法差2.5倍的能耗来服务所有的推理请求。
在本发明内，边缘推理架构往往需要边缘设备实时与云中心保持同步来保证模型的更新，同时还需要设备持续运行推理服务来满足随机发生的请求，这无疑牺牲了大量的设备能耗，本发明则将该问题形式化成了时空图上的节点连接，使用同一子网下的设备间通信来保证模型的更新，比起传统方法，我们可以大大节省设备与云中心通信的次数和带宽，并降低设备的能耗。
除此之外，根据这个新模型，我们可以在减少通信次数的同时，动态关闭和启用推理服务，针对随机产生的请求序列，可以达到最差情况下与最优解至多2.5的能耗比，大大减少了能耗损失，延长了边缘设备的使用时间，从而减少了边缘计算中不必要的能耗和带宽，节约如智能制造、智能工厂等的运营成本；从而提高了在同样的能耗下设备端能够服务的推理请求数量，提升设备的待机时长。
本发明具备如下优点：
1、在模型更新时，使用边缘设备间的同步来代替边缘设备和云中心同步，减少通信能耗；
2、针对随机发生的请求，让模型持续运行的动态时长，来减少不必要的持 续运行能耗；
3、使用迁移来在边缘设备间调度模型来完成请求的推理。
以上所述仅为本发明的实施方式，并非因此限制本发明的专利范围，凡是利用本发明说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本发明的专利保护范围内。

Claims (10)

  1. 一种边缘计算下神经网络推理的节能方法，其特征在于，包括：
    在可预知请求序列情况下，将边缘设备从云中心下载模型的能耗与该边缘设备从子网内其它边缘设备转移模型的能耗进行比较，如果边缘设备从云中心下载模型的能耗小于或等于边缘设备从子网内其它设备转移模型的能耗，则边缘设备直接从云中心下载模型；如果边缘设备从云中心下载模型的能耗大于边缘设备从子网内其它设备转移模型的能耗，则边缘设备从子网内其它设备转移模型；其中，子网由与云中心进行通信的接入点和若干个与该接入点通信的边缘设备组成；
    在不可预知请求序列情况下，让边缘设备请求发生，维持一段时间，再将边缘设备从云中心下载模型的能耗与该边缘设备从子网内其它边缘设备转移模型的能耗进行比较。
  2. 根据权利要求1所述的一种边缘计算下神经网络推理的节能方法，其特征在于，在可预知请求序列情况下，该边缘设备进行请求时，该边缘设备的持续运行的能耗为该边缘设备从上一次进行请求一直持续运行到该边缘设备的这次请求的能耗。
  3. 根据权利要求2所述的一种边缘计算下神经网络推理的节能方法，其特征在于，在可预知请求序列情况下，该边缘设备从子网内其它边缘设备转移模型时，比较该边缘设备从子网内每个其它边缘设备的转移模型的所要能耗，选择从所要能耗最少的子网内那个边缘设备进行转移模型。
  4. 根据权利要求2所述的一种边缘计算下神经网络推理的节能方法，其特征在于，在可预知请求序列情况下，该边缘设备从子网内其它边缘设备转移模型时，该边缘设备进行请求时，先将模型转移至子网内持续运行能耗最少的边缘设备上，再将该模型从持续运行能耗最少的边缘设备上转移至发出请求的边缘设备 上。
  5. 根据权利要求2所述的一种边缘计算下神经网络推理的节能方法，其特征在于，在可预知请求序列情况下，当上一个请求不来时，该边缘设备直接从云中心下载模型。
  6. 根据权利要求1所述的一种边缘计算下神经网络推理的节能方法，其特征在于，在不可预知请求序列情况下，如果边缘设备从云中心下载模型的能耗小于或等于边缘设备从子网内其它设备转移模型的能耗，则边缘设备直接从云中心下载模型。
  7. 根据权利要求1所述的一种边缘计算下神经网络推理的节能方法，其特征在于，在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗且小于该边缘设备从子网内其它设备转移模型的能耗的两倍时，则该边缘设备从子网内其它设备转移模型。
  8. 根据权利要求7所述的一种边缘计算下神经网络推理的节能方法，其特征在于，在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗且小于该边缘设备从子网内其它设备转移模型的能耗的两倍时，而且子网内其它设备没有所需的模型，则该边缘设备从云中心下载模型。
  9. 根据权利要求1所述的一种边缘计算下神经网络推理的节能方法，其特征在于，在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗的两倍时，该边缘设备从子网内其它设备转移模型。
  10. 根据权利要求1所述的一种边缘计算下神经网络推理的节能方法，其特 征在于，在不可预知请求序列情况下，如果该边缘设备从云中心下载模型的能耗大于该边缘设备从子网内其它设备转移模型的能耗的两倍时，先将模型转移至子网内持续运行能耗最少的边缘设备上持续运行，当有边缘设备发出请求后，再将该模型从持续运行能耗最少的边缘设备上转移至发出请求的边缘设备上。
PCT/CN2020/140017 2020-12-04 2020-12-28 一种边缘计算下神经网络推理的节能方法 Ceased WO2022116325A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011400815.0A CN114599055B (zh) 2020-12-04 2020-12-04 一种边缘计算下神经网络推理的节能方法
CN202011400815.0 2020-12-04

Publications (1)

Publication Number Publication Date
WO2022116325A1 true WO2022116325A1 (zh) 2022-06-09

Family

ID=81802379

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/140017 Ceased WO2022116325A1 (zh) 2020-12-04 2020-12-28 一种边缘计算下神经网络推理的节能方法

Country Status (2)

Country Link
CN (1) CN114599055B (zh)
WO (1) WO2022116325A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118869510A (zh) * 2024-07-02 2024-10-29 东南大学 一种边缘计算场景下确定性延迟的协同推理方法及系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107766889A (zh) * 2017-10-26 2018-03-06 济南浪潮高新科技投资发展有限公司 一种云端边缘计算融合的深度学习计算系统和方法
CN109961097A (zh) * 2019-03-20 2019-07-02 西北大学 一种嵌入式场景下基于边缘计算的图像分类调度方法
CN110443298A (zh) * 2019-07-31 2019-11-12 华中科技大学 一种基于云-边缘协同计算的ddnn及其构建方法和应用
CN111031102A (zh) * 2019-11-25 2020-04-17 哈尔滨工业大学 一种多用户、多任务的移动边缘计算系统中可缓存的任务迁移方法
CN111445026A (zh) * 2020-03-16 2020-07-24 东南大学 面向边缘智能应用的深度神经网络多路径推理加速方法
CN111459505A (zh) * 2020-05-22 2020-07-28 南京大学 边缘计算环境下多版本推断模型部署方法、装置和系统

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3155758A4 (en) * 2014-06-10 2018-04-11 Sightline Innovation Inc. System and method for network based application development and implementation
EP3457664B1 (en) * 2017-09-14 2019-11-06 Deutsche Telekom AG Method and system for finding a next edge cloud for a mobile user
CN107682443A (zh) * 2017-10-19 2018-02-09 北京工业大学 联合考虑延迟和能量消耗的移动边缘计算系统计算任务的高效卸载方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107766889A (zh) * 2017-10-26 2018-03-06 济南浪潮高新科技投资发展有限公司 一种云端边缘计算融合的深度学习计算系统和方法
CN109961097A (zh) * 2019-03-20 2019-07-02 西北大学 一种嵌入式场景下基于边缘计算的图像分类调度方法
CN110443298A (zh) * 2019-07-31 2019-11-12 华中科技大学 一种基于云-边缘协同计算的ddnn及其构建方法和应用
CN111031102A (zh) * 2019-11-25 2020-04-17 哈尔滨工业大学 一种多用户、多任务的移动边缘计算系统中可缓存的任务迁移方法
CN111445026A (zh) * 2020-03-16 2020-07-24 东南大学 面向边缘智能应用的深度神经网络多路径推理加速方法
CN111459505A (zh) * 2020-05-22 2020-07-28 南京大学 边缘计算环境下多版本推断模型部署方法、装置和系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
XU ZICHUAN; ZHAO LIQIAN; LIANG WEIFA; RANA OMER F.; ZHOU PAN; XIA QIUFEN; XU WENZHENG; WU GUOWEI: "Energy-Aware Inference Offloading for DNN-Driven Applications in Mobile Edge Clouds", IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, vol. 32, no. 4, 20 October 2020 (2020-10-20), USA, pages 799 - 814, XP011820240, ISSN: 1045-9219, DOI: 10.1109/TPDS.2020.3032443 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118869510A (zh) * 2024-07-02 2024-10-29 东南大学 一种边缘计算场景下确定性延迟的协同推理方法及系统

Also Published As

Publication number Publication date
CN114599055A (zh) 2022-06-07
CN114599055B (zh) 2023-05-19

Similar Documents

Publication Publication Date Title
Xiao et al. Reinforcement learning based energy-efficient collaborative inference for mobile edge computing
Sun et al. Cooperative computation offloading for multi-access edge computing in 6G mobile networks via soft actor critic
Li et al. Task offloading mechanism based on federated reinforcement learning in mobile edge computing
WO2021012584A1 (zh) 一种移动边缘计算场景下制定单任务迁移策略的方法
CN113286329B (zh) 基于移动边缘计算的通信和计算资源联合优化方法
WO2023040022A1 (zh) 一种在随机网络中基于算网协同的分布式计算卸载方法
Li et al. DQN-enabled content caching and quantum ant colony-based computation offloading in MEC
CN114860337B (zh) 一种基于元强化学习算法的计算卸载方法
CN116233926A (zh) 一种基于移动边缘计算的任务卸载及服务缓存联合优化方法
CN114390057A (zh) Mec环境下基于强化学习的多接口自适应数据卸载方法
CN113766540B (zh) 低时延的网络内容传输方法、装置、电子设备及介质
Chatterjee et al. Dynamic and adaptive data caching mechanism for virtualization within sensor-cloud
CN116112488A (zh) 一种面向mec网络的细粒度任务卸载及资源分配方法
CN115665869A (zh) 基于边缘计算和有向无环图的多用户协作平台及其方法
CN113835894A (zh) 一种基于双延迟深度确定性策略梯度的智能计算迁移方法
CN110278611A (zh) 一种无线供电移动边缘计算系统中的资源分配方法
Tang et al. Federated learning-assisted task offloading based on feature matching and caching in collaborative device-edge-cloud networks
CN112162789A (zh) 一种基于软件定义的边缘计算随机卸载决策方法及系统
CN118102395A (zh) 一种区块链驱动的边缘计算系统联合优化方法
CN114281718A (zh) 一种工业互联网边缘服务缓存决策方法及系统
CN104519106B (zh) 一种任务迁移方法及网络控制器
CN114599055B (zh) 一种边缘计算下神经网络推理的节能方法
CN118317365A (zh) 基于深度学习的边缘云计算任务迁移与资源分配联合优化方法
CN118283708A (zh) 一种基于深度强化学习的端边云协作式卸载方法
Karobi et al. Ecoedgetwin: Enhanced 6g network via mobile edge computing and digital twin integration

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20964170

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20964170

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20964170

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 22.08.2024)

122 Ep: pct application non-entry in european phase

Ref document number: 20964170

Country of ref document: EP

Kind code of ref document: A1