WO2021164404A1 - 一种巡检方法及装置 - Google Patents

一种巡检方法及装置 Download PDF

Info

Publication number
WO2021164404A1
WO2021164404A1 PCT/CN2020/137411 CN2020137411W WO2021164404A1 WO 2021164404 A1 WO2021164404 A1 WO 2021164404A1 CN 2020137411 W CN2020137411 W CN 2020137411W WO 2021164404 A1 WO2021164404 A1 WO 2021164404A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
training
mobile device
time slice
recognition
Prior art date
Application number
PCT/CN2020/137411
Other languages
English (en)
French (fr)
Inventor
杨洁
何东杰
Original Assignee
中国银联股份有限公司
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 中国银联股份有限公司 filed Critical 中国银联股份有限公司
Publication of WO2021164404A1 publication Critical patent/WO2021164404A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/30—Monitoring
    • G06F11/3058—Monitoring arrangements for monitoring environmental properties or parameters of the computing system or of the computing system component, e.g. monitoring of power, currents, temperature, humidity, position, vibrations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/30—Monitoring
    • G06F11/3065—Monitoring arrangements determined by the means or processing involved in reporting the monitored data
    • G06F11/3072—Monitoring arrangements determined by the means or processing involved in reporting the monitored data where the reporting involves data filtering, e.g. pattern matching, time or event triggered, adaptive or policy-based reporting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G07—CHECKING-DEVICES
    • G07C—TIME OR ATTENDANCE REGISTERS; REGISTERING OR INDICATING THE WORKING OF MACHINES; GENERATING RANDOM NUMBERS; VOTING OR LOTTERY APPARATUS; ARRANGEMENTS, SYSTEMS OR APPARATUS FOR CHECKING NOT PROVIDED FOR ELSEWHERE
    • G07C1/00—Registering, indicating or recording the time of events or elapsed time, e.g. time-recorders for work people
    • G07C1/20—Checking timed patrols, e.g. of watchman
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L47/00—Traffic control in data switching networks
    • H04L47/10—Flow control; Congestion control

Definitions

  • the present invention relates to the technical field of image processing, in particular to a patrol inspection method and device.
  • the Internet data center (IDC) computer room is a standardized computer room environment established on the basis of Internet communication lines and bandwidth resources.
  • the IDC computer room can accommodate multiple types of equipment, such as servers, monitoring equipment, management equipment, or security equipment.
  • equipment such as servers, monitoring equipment, management equipment, or security equipment.
  • it is usually necessary to inspect the IDC computer room, such as collecting the image of the signal lamp on the device to identify the status of the signal lamp on the device, collecting the image of the device to identify the type of the device, collecting odor data and/or temperature data for Analyze the computer room environment.
  • how to effectively identify the patrol data is very important for maintaining the security of the computer room and troubleshooting in time.
  • each mobile device collects training data from the computer room and reports it to the server for centralized training. After the server has trained the recognition model, the recognition model is issued to each mobile device for use by each mobile device. The recognition model recognizes the collected inspection data and completes the inspection of the computer room.
  • this method requires each mobile device to report the training data to the server, and the training data (such as images, smell or temperature) generally corresponds to a larger amount of data, so this inspection method will have a larger communication overhead. As a result, the inspection efficiency is low.
  • the invention provides a patrol inspection method and device, which are used to solve the technical problems of high communication overhead and low patrol inspection efficiency caused by the use of a server to centrally train a recognition model in the prior art.
  • the present invention provides a patrol inspection method, which is applied to a model management device.
  • the method includes: the model management device sends a model training instruction to each mobile device to instruct each mobile device to follow its own patrol route in the computer room.
  • the training data is collected while traveling, and the intermediate model is obtained based on the local model and training data training and reported to the model management device.
  • the model management device receives the model parameters of the intermediate model reported by each mobile device, it can be based on each mobile device
  • the model parameters of the reported intermediate model are trained to obtain a recognition model, and the recognition model is used to determine the health status of the computer room.
  • the training process of the intermediate model is placed on the side of the mobile device for execution, so that the mobile device can only report the model parameters of the intermediate model to the model management device, instead of reporting the full amount of training data, because the model parameters are relative to the training data. It is said that it has a small amount of data, so this kind of inspection method can save communication expenses and effectively improve the efficiency of inspection.
  • this kind of inspection method can save communication expenses and effectively improve the efficiency of inspection.
  • the recognition model to determine the health status of the computer room during the inspection process of the mobile device, it is helpful to realize the joint operation of model training and model recognition, and further improve the inspection efficiency.
  • the recognition model refers to the recognition model corresponding to the current time slice.
  • the model management device can also retrieve the model from all mobile devices in the computer room before sending model training instructions to each mobile device.
  • the mobile device corresponding to the current time slice is selected as each mobile device that performs model training corresponding to the current time slice.
  • the model training instruction is used to instruct each mobile device to collect training data while traveling in the computer room according to its own inspection route, and to instruct each mobile device to obtain an intermediate model based on the local model and training data training. Report to the model management device.
  • each time the model management device receives a set number of model parameters it can construct a comprehensive model parameter based on at least the set number of model parameters, and then send the comprehensive model parameters to each mobile device so that Each mobile device updates the local model based on the comprehensive model parameters.
  • the model management device may send a model training end instruction to each mobile device.
  • the model corresponding to the comprehensive model parameter is the recognition model.
  • the above implementation method executes multiple model synthesis operations in each time slice, and uses a set number of model parameters in each model synthesis operation to obtain the synthesis model parameters, instead of using the model parameters of all mobile devices for synthesis. While using as many model parameters as possible to obtain comprehensive model parameters, it is compatible with the failure of some mobile devices, which helps to ensure the smooth progress of model training.
  • the model management device may refer to a server.
  • the model management device obtains the recognition model based on the model parameters of the intermediate model reported by each mobile device, it can also determine whether the recognition model meets the end conditions of the model training.
  • the mobile device selects the mobile device corresponding to the next time slice, and sends a model training instruction to the mobile device corresponding to the next time slice. If it is satisfied, the recognition model can be used as the target recognition model.
  • the server as a model management device, not only can the model training process of each mobile device be uniformly managed, but also the pressure on the mobile device can be effectively reduced, and the efficiency of training the intermediate model of the mobile device can be improved.
  • the resources of the mobile device can also be effectively allocated, and the differences between different mobile devices can be fully considered to improve the accuracy of the recognition model.
  • the model management device may refer to a mobile device.
  • the model management device obtains the recognition model based on the model parameters of the intermediate model reported by each mobile device, it can also determine whether the recognition model meets the end condition of the model training. If not, you can select a mobile device as the model management device corresponding to the next time slice through communication with other mobile devices, and send a model start instruction to the model management device corresponding to the next time slice to indicate the next time.
  • the model management device corresponding to the slice selects other mobile devices corresponding to the next time slice from other mobile devices and sends a model training instruction to other mobile devices corresponding to the next time slice. If it is satisfied, the recognition model can be used as the target recognition model.
  • a mobile device as a model management device, there is no need to set up an additional server, thereby helping to reduce the cost of model training.
  • the resources of the mobile device can also be effectively allocated, and the differences between different mobile devices can be fully considered to improve the accuracy of the recognition model.
  • the present invention provides a patrol inspection method, which is applied to a mobile device.
  • the method includes: the mobile device receives a model training instruction sent by a model management device, and according to the model training instruction, in the computer room according to the corresponding inspection route Collect training data while traveling, and then train based on the local model and training data to obtain the intermediate model corresponding to the mobile device, and report the model parameters of the intermediate model to the model management device, so that the model management device is based on the model of the intermediate model reported by each mobile device
  • the parameter training obtains the recognition model.
  • the recognition model is used to determine the health status of the computer room.
  • the mobile device can perform the following operations in a loop according to the model training instructions: travel in the computer room according to the inspection route and collect training data, and train based on the local model and training data to obtain an intermediate model corresponding to the mobile device. Report the model parameters of the intermediate model to the model management device.
  • the mobile device can also receive the integrated model parameters sent by the model management device, and then use the integrated model parameters to update the local model.
  • the mobile device after the mobile device uses the comprehensive model parameters to update the local model, it can also travel in the computer room according to the inspection route and collect the data to be identified, and then use the updated local model to identify the data to be identified. Determine the health status of the computer room.
  • the effects of training models, optimization models, and recognition models can be achieved, and the efficiency of training and recognition can be improved.
  • the recognition effect can also be improved.
  • the present invention provides a patrol inspection device, the device includes: a transceiver module for sending model training instructions to each mobile device, and receiving model parameters of the intermediate model reported by each mobile device; wherein the model training instruction is used Collect training data when each mobile device is traveling in the computer room according to its own inspection route, and instruct each mobile device to train based on the local model and training data to obtain an intermediate model; the training module is used for the model based on the intermediate model reported by each mobile device The parameter training obtains the recognition model. Among them, the recognition model is used to determine the health status of the computer room.
  • the recognition model may be a recognition model corresponding to the current time slice.
  • the training module can also select the mobile device corresponding to the current time slice from all the mobile devices in the computer room, as each mobile device that performs training in the current time slice. Mobile devices.
  • the model training instruction is used for each mobile device to cycle in the computer room according to its own inspection route and collect training data, and an intermediate model is obtained and reported based on the local model and training data training.
  • the training module is specifically used to: each time a set number of model parameters are received, the comprehensive model parameters are constructed at least based on the set number of model parameters; the transceiver module is specifically used to: send the comprehensive model parameters to Each mobile device, and when the model corresponding to a certain comprehensive model parameter meets the end condition of the current time slice, sends a model training end instruction to each mobile device to indicate that the model corresponding to the comprehensive model parameter is the recognition model.
  • the comprehensive model parameters are used for each mobile device to update the local model.
  • the inspection device is a server.
  • the training module obtains the recognition model based on the model parameters of the intermediate model reported by each mobile device, it can also determine whether the recognition model meets the end conditions of the model training. If not, it can move from all parts of the computer room.
  • the mobile device corresponding to the next time slice is selected from the device, and a model training instruction is sent to the mobile device corresponding to the next time slice. If it is satisfied, the recognition model can be used as the target recognition model.
  • the inspection device is a mobile device.
  • the training module can also determine whether the recognition model meets the end conditions of the model training. If not, it can communicate with other mobile devices. The communication selects a certain mobile device as the model management device corresponding to the next time slice, and sends a model start instruction to the model management device corresponding to the next time slice to instruct the model management device corresponding to the next time slice from other mobile devices Select other mobile devices corresponding to the next time slice and send a model training instruction to other mobile devices corresponding to the next time slice; if it is satisfied, the recognition model can be used as the target recognition model.
  • the present invention provides a patrol inspection device, which includes: a transceiver module for receiving a model training instruction sent by a model management device; a training module for receiving a model training instruction according to the model training instruction and following the corresponding inspection route.
  • the training data is collected while traveling in the computer room, and the intermediate model corresponding to the mobile device is obtained based on the local model and training data training; the transceiver module is also used to report the model parameters of the intermediate model to the model management device.
  • the model parameters of the intermediate model reported by each mobile device are used to train the model management device to obtain the recognition model, and the recognition model is used to determine the health status of the computer room.
  • the training module may perform the following operations in a loop according to the model training instructions: travel in the computer room according to the inspection route and collect training data, and train based on the local model and the training data to obtain an intermediate model corresponding to the mobile device.
  • the transceiver module can perform the following operations cyclically according to the model training instruction: report the model parameters of the intermediate model to the model management device.
  • the transceiver module can also receive the comprehensive model parameters sent by the model management device, and the training module can also use the comprehensive model parameters to update the local model.
  • the device may further include an identification module.
  • the recognition module can collect the data to be recognized while traveling in the computer room according to the inspection route, and use the updated local model to recognize the data to be recognized to determine the health status of the computer room.
  • the present invention provides a computing device including at least one processor and at least one memory, wherein the memory stores a computer program, and when the program is executed by the processor, the processor executes The inspection method described in any of the above-mentioned first aspect or second aspect.
  • the present invention provides a computer-readable storage medium that stores a computer program executable by a computing device.
  • the program runs on the computing device, the computing device executes the first aspect or Any of the inspection methods described in the second aspect.
  • FIG. 1 is a schematic structural diagram of an IDC computer room provided by an embodiment of the present invention
  • FIG. 2 is a schematic diagram of the system architecture of a patrol inspection system provided by an embodiment of the present invention
  • FIG. 3 is a schematic flowchart of a corresponding inspection method according to an embodiment of the present invention.
  • FIG. 4 is a schematic diagram of the interaction flow corresponding to the inspection method when the server is used as the model management device;
  • FIG. 5 is a schematic diagram of the interaction flow corresponding to the inspection method when the mobile device is used as the model management device;
  • FIG. 6 is a schematic structural diagram of a patrol inspection device provided by an embodiment of the present invention.
  • FIG. 7 is a schematic structural diagram of another inspection device provided by an embodiment of the present invention.
  • FIG. 8 is a schematic structural diagram of a computing device provided by an embodiment of the present invention.
  • Fig. 1 is a schematic structural diagram of an IDC computer room provided by an embodiment of the present invention.
  • at least one row of cabinets may be provided in the IDC computer room, such as cabinets 101 to 106, and cabinets 101 to 104 can be arranged in parallel.
  • the cabinet 105 and the cabinet 106 can be arranged side by side, and each row of cabinets can be provided with multiple devices, such as server equipment, data acquisition equipment, monitoring equipment, or temperature control equipment.
  • the cabinet may have a single-layer structure, and multiple devices are placed in parallel on the single-layer structure.
  • the cabinet may also have a multi-layer structure, and multiple devices can be placed on the multi-layer structure, and one or more devices can be placed side by side on each layer structure, which is not specifically limited.
  • FIG. 2 is a schematic diagram of the system architecture of an inspection system provided by an embodiment of the present invention.
  • the system architecture includes a model management device 110 and at least one mobile device, such as Mobile device 121, mobile device 122, and mobile device 123.
  • the model management device 110 can be connected to any mobile device, for example, it can be connected in a wired way or can be connected in a wireless way, which is not specifically limited.
  • each mobile device can be deployed in the same IDC computer room, or can be deployed in different IDC computer rooms. If a mobile device is deployed in each IDC computer room, the mobile device can be responsible for patrolling the entire IDC computer room. If multiple mobile devices are deployed in each IDC computer room, each mobile device can be responsible for inspecting an area in the IDC computer room, and multiple mobile devices can jointly complete the inspection operation on the entire IDC computer room.
  • the inspection method in the embodiment of the present invention may be used to inspect one IDC computer room, and may also be used to inspect multiple IDC computer rooms, which is not specifically limited.
  • FIG. 3 is a schematic diagram of the process corresponding to an inspection method provided by an embodiment of the present invention.
  • the method can be applied to model management devices and mobile devices, such as the model management device shown in FIG. 110, and one or more of the mobile device 121 to the mobile device 123.
  • the method includes:
  • Step 301 The model management device sends a model training instruction to each mobile device.
  • the target recognition model can be trained through multiple time slices.
  • the model management device can first select the mobile device corresponding to the time slice from all mobile devices in the computer room, and then Send a model training instruction to the mobile device corresponding to the time slice.
  • the mobile device that receives the model training instruction can perform model training in conjunction with the model management device to update the recognition model corresponding to the previous time slice, and obtain the recognition model corresponding to the time slice.
  • the mobile device that has not received the model training instruction may only perform the recognition operation without participating in the model training in the time slice.
  • each mobile device described in step 301 may be a mobile device corresponding to any time slice.
  • a fixed number or a random number of mobile devices can be randomly selected as the mobile device corresponding to any time slice, or a fixed number of mobile devices can be selected in turn.
  • a random number of mobile devices are used as mobile devices corresponding to any time slice, or a fixed or random number of mobile devices with strong processing capabilities can be selected as mobile devices corresponding to any time slice, which is not specifically limited.
  • a mobile device with a set ratio can be selected from all mobile devices as the mobile device corresponding to any time slice.
  • the set ratio can be set by those skilled in the art based on experience, for example, it can be set to 50% to 80%, so as to avoid overfitting caused by too much training data while preserving the characteristics of most of the training data, and improve Identify the accuracy of the model.
  • the training data and the mobile devices can be effectively distributed and evenly used.
  • the uniform distribution of training data can avoid the problem that the parameter iteration fails to converge due to excessive data when training the model.
  • the uniform distribution of mobile devices can improve the ability of the model management device to respond to mobile device failures and improve the availability of training and identifying models.
  • Step 302 According to the model training instruction, the mobile device collects training data while traveling in the computer room according to the inspection route, and trains to obtain an intermediate model corresponding to the mobile device based on the local model and the training data.
  • the model training instruction sent by the model management device is used to instruct the mobile device to perform the following operations in a loop: collect training data while traveling in the computer room according to the inspection route, and train the mobile device based on the local model and training data
  • the corresponding intermediate model reports the intermediate model to the model management device.
  • the mobile device can continuously collect training data. After a fixed amount of training data is collected, the collected training data pairs can be used.
  • the local model is trained to obtain the intermediate model corresponding to the mobile device and report it to the model management device.
  • the fixed number can be 1, or any integer greater than 1, and is not limited.
  • the training data can be set by those skilled in the art as required, for example, it can be any one or more of the image of the equipment room, the image of the signal lamp, the temperature information, or the odor information, which is not limited.
  • the method of optimizing the local model may be as follows: first use the local model to predict each training data to obtain the predicted label of each training data, and then according to the degree of matching between the predicted label of each training data and the real label, from each In the training data, the number of training data with correct prediction labels and the number of training data with incorrect prediction labels are determined, and the loss function is calculated according to these numbers. Finally, the model parameters of the local model are adjusted according to the loss function to obtain the intermediate model.
  • the mobile device when collecting training data, can collect training data according to a fixed frequency, or it can collect continuous data first, and then intercept the training data from the continuous data according to the fixed frequency.
  • the training data of is regarded as one-time training data and is not limited.
  • the training data is a device image and the fixed time period is 1 minute
  • the fixed frequency is 6 milliseconds/each time
  • the mobile device can take a device image every 6 milliseconds, and then change each one minute
  • the 600 pieces of equipment images captured during the time period of the Among the captured device images, 600 continuous device images are selected as a training data.
  • the mobile device can also use the collected training data to train the local model after each collection of training data for a fixed period of time, or the mobile device can also train the local model after each collection of a random number or random period of training data. Use the collected training data to train the local model, which is not specifically limited.
  • the local model of the mobile device may be any one of the initial model, the recognition model corresponding to the previous time slice, and the model corresponding to the comprehensive model parameters.
  • the local model in the mobile device In the initial stage of model training for the first time slice, the local model in the mobile device is the initial model. In the start phase of model training in other time slices, the local model in the mobile device is the recognition model corresponding to the previous time slice. In the execution stage of model training in any time slice, the local model of the mobile device is the model corresponding to the comprehensive model parameters.
  • the initial model can be trained by a model management device or any mobile device. Take the model management equipment training the initial model as an example.
  • the model management equipment can first obtain the initial training data, then use the initial training data to train to obtain the initial model, and then send the model parameters of the initial model to all mobile devices in the computer room .
  • the initial training data can be obtained in a variety of ways, for example, it can be downloaded via the network, it can also be collected from the computer room before the first time slice, or it can be obtained from a third-party model management device, which is not limited.
  • the mobile device may be in a waiting state without participating in the recognition model training corresponding to the time slice.
  • the mobile device In the next time slice, if the mobile device receives the model training instruction, the mobile device can participate in the recognition model training corresponding to the next time slice. If the model training instruction is still not received, the mobile device can continue to be in the waiting state .
  • Step 303 The mobile device reports the model parameters of the intermediate model to the model management device.
  • model parameters of the intermediate model may include gradient and loss value, and may also include other information, which is not limited.
  • the mobile device may first compress the model parameters of the intermediate model, and then report the compressed package to the model management device, so as to reduce communication overhead, reduce data transmission time, and improve inspection efficiency.
  • the mobile device can also perform an encryption operation on the compressed package, and report the encrypted compressed package to the model management device, so as to improve the security of the data during transmission.
  • step 304 the model management device obtains a recognition model based on the model parameter training of the intermediate model corresponding to each mobile device.
  • the model management device may be a server or a mobile device.
  • the model management device can directly use the model parameters of each intermediate model sent by other mobile devices corresponding to the current time slice to calculate the comprehensive model parameters.
  • the model management device is a mobile device, in addition to sending model training instructions to other mobile devices corresponding to the current time slice, the model management device can also use training data to train to obtain an intermediate model corresponding to the model management device, which is based on the current time slice.
  • the model parameters of the corresponding other mobile devices and the model parameters of the intermediate model corresponding to the model management device are calculated to obtain the comprehensive model parameters.
  • the model management device may train in a variety of ways to obtain the recognition model corresponding to the time slice. For example, after receiving all or most of the model parameters sent by the mobile devices corresponding to the time slice, the model management device may use all or most of the model parameters corresponding to the mobile devices to train to obtain the recognition model corresponding to the time slice. Alternatively, the model management device may first instruct the mobile device corresponding to the time slice to perform multiple model training independently to obtain multiple model parameters corresponding to each mobile device, and then use each model parameter training corresponding to each mobile device to obtain the time slice
  • the corresponding recognition model is not limited.
  • the model management device can train the recognition model corresponding to any time slice in the following manner:
  • Step a Each time the model management device receives a set number of model parameters, it can calculate a comprehensive model parameter based on the set number of model parameters.
  • the model management device can continuously receive the model parameters sent by each mobile device .
  • the model management device can calculate a comprehensive model parameter based on the set number of model parameters received this time.
  • the set number can be set by those skilled in the art based on experience. For example, it can be set to be slightly smaller than the total number of mobile devices corresponding to the time slice, so as to use as many model parameters as possible to obtain comprehensive model parameters while being compatible The failure of some mobile devices ensures the smooth progress of model training.
  • the model management equipment can calculate the comprehensive model parameters in many ways.
  • the average parameters of a set number of model parameters (and the model parameters of the model management equipment) can be used as the comprehensive model parameters, or the design parameters can also be set.
  • the weighted average parameter of a certain number of model parameters (and model parameters of the model management equipment) is used as a comprehensive model parameter, or the set number of model parameters (and model parameters of the model management equipment) can be screened out that do not meet the requirements
  • the model parameters, and then the average parameters or weighted average parameters of the model parameters that meet the requirements are used as the comprehensive model parameters, which are not limited.
  • Step b The model management device issues the comprehensive model parameters to all mobile devices in the computer room.
  • any mobile device in the computer room may also collect data to be identified during the travel, and use a local model to identify the data to be identified, so as to complete the identification operation of the computer room during the inspection process.
  • the recognition operation can be triggered by a recognition instruction, or can be executed according to a set cycle, which is not limited.
  • the model management device can also deliver the comprehensive model parameters to all mobile devices in the computer room during the training process of each time slice.
  • any mobile device in the computer room a mobile device that has not received the model training instruction or a mobile device that has received the model training instruction
  • receives the integrated model parameters it can also use the integrated model parameters to update the local model.
  • the mobile device can use the updated model to identify the data to be identified, so that the joint operation of training, optimization, and identification can be performed during the inspection process to realize the side training model, Improve the efficiency of training and recognition by optimizing the model and recognizing the effect of the data.
  • the recognition effect can also be improved.
  • Step c The model management device determines whether the comprehensive model parameter meets the end condition corresponding to the current time slice, if it meets, execute step d 1 , if not, execute step d 2 .
  • the end condition corresponding to any time slice may be any one or more of the model parameters are not received in the set time period, the number of model training times is greater than or equal to the set times, and the model training time is greater than or equal to the set time.
  • Step d 1 the management device determines that the current model model training time slice has ended, a comprehensive model parameters of model parameter identification is the current time slice corresponding to the model, the model parameters corresponding to the integrated model is the current time slice corresponding to the identified model, and therefore
  • the model management device may send a model training end instruction to each mobile device corresponding to the current time slice.
  • Step d 2 the management device determines that the current model model training time slot has not been completed, a comprehensive model parameters other than the current time slice identification parameter corresponding to the model, the model parameters corresponding to the integrated model is not the current time slice corresponding recognition model, so the model management device No special treatment is required. Since each mobile device corresponding to the current time slice repeatedly performs the operations of collecting training data, training the intermediate model, and reporting model parameters, the model management device can repeat step a to step d 1 or step a to step d 2 until the type management device Make sure that the model training for the current time slice has ended.
  • the model management device may also obtain the recognition effect of each mobile device using the recognition model corresponding to the current time slice, and determine whether the recognition effect meets the end condition of the model training. If so, the model management device may use the recognition model corresponding to the current time slice as the target recognition model, and send a model training end instruction to each mobile model corresponding to the current time slice. If not, the model management device can start the model training for the next time slice.
  • the model management device can first establish a communication connection with the mobile device in the new inspection area or inspection room, and then store the information stored in the model management device.
  • the latest model parameters are issued to newly connected mobile devices so that the newly connected mobile devices can use the latest model parameters for identification.
  • the model management device can re-select the mobile device corresponding to the next time slice from all devices (including newly connected mobile devices), and jointly execute the model training for the next time slice, thereby Quickly realize the inspection operation of the new inspection area, and improve the flexibility of the collaborative training process.
  • the model management device may be a server or a mobile device.
  • the method of starting the model training for the next time slice is also different, specifically:
  • the model management device When the model management device is a server, the model management device may be in communication connection with each mobile device in the computer room, and each mobile device in the computer room may not be in communication connection.
  • the model management device determines to start the model training for the next time slice, it can first select some mobile devices from all mobile devices as the mobile devices corresponding to the next time slice, and then combine the mobile devices corresponding to the next time slice. The equipment is trained to obtain the recognition model corresponding to the next time slice.
  • any two mobile devices in the computer room can be communicatively connected. In this way, each mobile device in the computer room can form a decentralized distributed cluster.
  • all mobile devices can first select a mobile device as the model management device corresponding to the next time slice through communication interaction, and then manage the model corresponding to the next time slice The device selects some mobile devices from other mobile devices as other mobile devices corresponding to the next time slice, and then the model management device corresponding to the next time slice is trained in conjunction with other mobile devices corresponding to the next time slice to obtain the corresponding mobile device for the next time slice. Identify the model.
  • the mobile device with the most votes can be used as the model management device corresponding to the next time slice by voting, or the computing power can be the strongest.
  • the model management device corresponding to the next time slice mobile devices that have not been used as model management devices can be selected alternately or randomly as the model management device corresponding to the next time slice, which is not limited.
  • the following describes the specific implementation process of the inspection method when the server is used as the model management device and the mobile device is used as the model management device.
  • FIG. 4 is a schematic diagram of the interaction process corresponding to a patrol inspection method provided by an embodiment of the present invention.
  • the method can be applied to model management devices and mobile devices.
  • the model management device may refer to a server.
  • the method includes:
  • Step 401 The model management device uses the initial training data to train to obtain the initial model.
  • Step 402 The model management device issues the model parameters of the initial model to all mobile devices in the computer room.
  • Step 403 The model management device selects the mobile device corresponding to the current time slice from all the mobile devices in the computer room.
  • Step 404 The model management device sends a model training instruction to the mobile device corresponding to the current time slice.
  • Step 405 For any mobile device corresponding to the current time slice (that is, any mobile device that receives the model training instruction), the mobile device performs the following operations in a loop according to the model training instruction: according to the inspection route corresponding to the mobile device Collect training data while traveling in the computer room, perform model training based on the local model and training data, and obtain an intermediate model corresponding to the mobile device.
  • the local model may be any one or multiple of the initial model, the recognition model corresponding to the previous time slice, and the model corresponding to the comprehensive model parameters. If the current time slice is the first time slice, and this training is the first training in the first time slice, the local model is the initial model. If the current time slice is any time slice after the first time slice, and this training is the first training in any time slice, the local model is the recognition model corresponding to the previous time slice. If the current time slice is not the first training in any time slice, the local model is the model corresponding to the comprehensive model parameters.
  • the mobile device does not participate in the model training of the current time slice, but can perform a recognition operation.
  • a recognition operation For example, when traveling according to the corresponding inspection route in the computer room, you can also collect the data to be identified, and then use the local model to identify the data to be identified, or you can perform model update operations, such as following the corresponding inspection route in the computer room
  • any mobile device corresponding to the current time slice cyclically reports the model parameters of the intermediate model to the model management device according to the model training instruction.
  • Step 407 Whenever the model management device receives a set number of model parameters, it calculates a comprehensive model parameter based on the set number of model parameters.
  • Step 408 The model management device issues the comprehensive model parameters to all mobile devices in the computer room.
  • Step 409 After any mobile device in the computer room receives the comprehensive model parameters sent by the model management device, it uses the comprehensive model parameters to update the local model to use the updated local model to identify the to-be-identified data collected on the inspection route operate.
  • any mobile device in the computer room can be any mobile device corresponding to the current time slice, or any other mobile device, such as any mobile device that has not received the model training instruction.
  • step 410 the model management device judges whether the model corresponding to the comprehensive model parameter meets the end condition of the current time slice, if yes, execute step 411, and if not, execute step 407.
  • the end condition of the current time slice can be that the number of training times is greater than or equal to the preset number of training times, the training duration is greater than or equal to the preset training duration, the model effect corresponding to the comprehensive model parameter meets the preset model effect, and the number of trainings is not received within the set time.
  • Any one or any number of model parameters sent by the mobile device is not limited.
  • Step 411 The model management device sends a model training end instruction to the mobile device corresponding to the current time slice.
  • Step 412 The model management device determines that the model corresponding to the comprehensive model parameters is the recognition model corresponding to the current time slice, and judges whether the recognition model corresponding to the current time slice meets the end condition of the model training, if yes, execute step 413, if not, execute Step 414.
  • the end condition of the model training may be any one or more of the model effects satisfying the preset effects, the model training duration is greater than or equal to the preset duration, and the number of time slices is greater than or equal to the preset number of time slices, which are not limited.
  • the model management device can obtain the identification corresponding to each mobile device (the mobile device corresponding to the current time slice, or other mobile devices) using the current time slice.
  • the model performs the recognition effect of the recognition operation, and then obtains the comprehensive recognition effect according to the recognition effect of each mobile device. If the comprehensive recognition effect does not reach the preset effect, it is determined that the recognition model corresponding to the current time slice does not meet the end condition of the model training. If the comprehensive recognition effect has reached the preset effect, it is determined that the recognition model corresponding to the current time slice meets the end condition of the model training.
  • Step 413 The model management device uses the recognition model corresponding to the current time slice as the target recognition model, and ends the model training.
  • step 414 the model management device determines that the model training of the current time slice is finished, takes the next time slice as the current time slice, and executes step 403 to start the model training of the next time slice.
  • the model training process of each mobile device can be uniformly managed through the server, which not only helps to reduce the pressure on the mobile device, but also improves the efficiency of the mobile device training intermediate model .
  • the mobile devices can be effectively allocated, and the differences between different mobile devices can be fully considered, and the accuracy of the recognition model can be improved.
  • FIG. 5 is a schematic diagram of the interaction flow corresponding to another inspection method provided by an embodiment of the present invention.
  • the method is applicable to model management devices and mobile devices, such as model management device 110 and mobile device 121 to mobile device 123 as shown in FIG. 2 .
  • the model management device in this example may refer to a mobile device.
  • the method includes:
  • any mobile device uses the initial training data to train to obtain an initial model.
  • any mobile device synchronizes the model parameters of the initial model to other mobile devices in the computer room, such as other mobile devices except any mobile device.
  • Step 503 Each mobile device in the computer room performs communication interaction, and a certain mobile device is selected from each mobile device as the model management device corresponding to the current time slice.
  • any two mobile devices can be set to send their respective resource occupancy status to each other, and the mobile device with the lowest resource occupancy rate can be agreed through each mobile device to select the one with the strongest processing capability.
  • the mobile device is used as a model management device; or you can set a number for each mobile device.
  • Each mobile device stores the corresponding relationship between the number of other mobile devices and the Internet Protocol (IP) address.
  • IP Internet Protocol
  • Step 504 The model management device corresponding to the current time slice selects other mobile devices corresponding to the current time slice from other mobile devices (ie, mobile devices other than the model management device corresponding to the current time slice).
  • Step 505 The model management device corresponding to the current time slice sends a model training instruction to other mobile devices corresponding to the current time slice.
  • Step 506 For any other mobile device corresponding to the current time slice, after the mobile device receives the model training instruction, it executes in a loop according to the model training instruction: collect training data while traveling in the computer room according to the corresponding inspection route, and Perform model training based on the local model and training data to obtain an intermediate model corresponding to the mobile device.
  • the model management equipment corresponding to the current time slice is also executed cyclically: training data is collected while traveling in the computer room according to the corresponding inspection route, and model training is performed based on the local model and training data to obtain the model management equipment corresponding to the current time slice The corresponding intermediate model.
  • Step 507 any other mobile device corresponding to the current time slice cyclically executes: synchronizing the model parameters of the intermediate model to the model management device corresponding to the current time slice.
  • Step 508 Each time the model management device corresponding to the current time slice receives a set number of model parameters, a comprehensive model is calculated based on the model parameters of the intermediate model obtained through training of the model management device corresponding to the current time slice and the set number of model parameters. parameter.
  • Step 509 The model management device corresponding to the current time slice issues the comprehensive model parameters to each mobile device in the computer room.
  • any mobile device in the computer room uses the integrated model parameters to update the local model, and uses the updated local model to perform a recognition operation on the to-be-identified data collected on the corresponding patrol route to determine the health status of the computer room.
  • Step 511 The model management device corresponding to the current time slice judges whether the model corresponding to the comprehensive model parameter meets the end condition of the current time slice, if yes, execute step 512, and if not, execute step 508.
  • Step 512 The model management device corresponding to the current time slice sends a model training end instruction to other mobile devices corresponding to the current time slice.
  • Step 513 The model management device corresponding to the current time slice determines that the model corresponding to the comprehensive model parameters is the recognition model corresponding to the current time slice, and judges whether the recognition model corresponding to the current time slice meets the end condition of the model training, and if so, execute step 514, If not, go to step 515.
  • Step 514 The model management device corresponding to the current time slice uses the recognition model corresponding to the current time slice as the target recognition model.
  • step 515 the model management device corresponding to the current time slice determines that the model training of the current time slice is completed, and the next time slice is regarded as the current time slice, and step 503 is executed.
  • the model management device corresponding to the current time slice can send interactive instructions to other mobile devices (mobile devices other than the model management device corresponding to the current time slice), and the interactive instructions are used by any mobile device to send its own to other mobile devices
  • the model management device corresponding to the current time slice may send a start instruction to the IP address of the next numbered mobile device. The start instruction is used to instruct the next numbered mobile device to start the model training for the next time slice, etc., etc. limited.
  • a mobile device as a model management device, there is no need to set up an additional server, thereby reducing the cost of model training.
  • the mobile devices can be effectively allocated, and the differences between different mobile devices can be fully considered, and the accuracy of the recognition model can be improved.
  • step numbers in Figures 4 and 5 are only an example of the execution process, and do not constitute a limitation on the order of execution of each step.
  • step 409 or step 510 can occur in any of the processes. The time is not limited.
  • the model management device sends a model training instruction to each mobile device, so that each mobile device collects training data when traveling in the computer room according to the model training instruction according to its own inspection route, and based on the local model and the location
  • the training data trains to obtain the intermediate model and reports it to the model management device.
  • the model management device can train to obtain a recognition model based on the model parameters of the intermediate model reported by each mobile device.
  • the recognition model is used to determine the health status of the computer room.
  • the mobile device by placing the training process of the intermediate model on the side of the mobile device for execution, the mobile device can only report the model parameters of the intermediate model to the model management device without reporting the full amount of training data. Data has a small amount of data, so this method can save communication overhead and improve inspection efficiency. Moreover, by using the recognition model to determine the health status of the computer room during the inspection process of the mobile device, the joint operation of model training and model recognition is realized, which helps to further improve the inspection efficiency.
  • an embodiment of the present invention also provides a patrol inspection device, and the specific content of the device can be implemented with reference to the foregoing method.
  • Fig. 6 is a schematic structural diagram of an inspection device provided by an embodiment of the present invention. As shown in Fig. 6, the device includes:
  • the transceiver module 601 is configured to send a model training instruction to each mobile device and receive model parameters of the intermediate model reported by each mobile device; wherein the model training instruction is used to: each mobile device according to its own inspection
  • the training data is collected when the route travels in the computer room, and the intermediate model is obtained and reported based on the local model and the training data training;
  • the training module 602 is configured to train to obtain a recognition model based on the model parameters of the intermediate model reported by each mobile device, and the recognition model is used to determine the health status of the computer room.
  • the recognition model is a recognition model corresponding to the current time slice.
  • the training module 602 is also used to: select the mobile device corresponding to the current time slice from all the mobile devices in the computer room as the mobile device. Describe each mobile device.
  • the model training instruction is used to collect training data when each mobile device cyclically travels in the computer room according to their respective inspection routes, and obtain and report an intermediate model based on the local model and the training data training.
  • the training module 602 is specifically configured to: each time a set number of model parameters are received, construct a comprehensive model parameter based on at least the set number of model parameters.
  • the transceiver module 601 is specifically configured to: send the comprehensive model parameters to the respective mobile devices, and when a model corresponding to a certain comprehensive model parameter meets the end condition of the current time slice, move to the respective mobile devices The device sends a model training end instruction, and the model corresponding to the comprehensive model parameter is the recognition model.
  • the comprehensive model parameter is used for each mobile device to update the local model.
  • the device is a server.
  • the training module 602 obtains a recognition model based on the model parameters of the intermediate model reported by each mobile device, it is also used to determine whether the recognition model meets the end condition of the model training; if If not satisfied, select the mobile device corresponding to the next time slice from all mobile devices in the computer room, and send a model training instruction to the mobile device corresponding to the next time slice; if it is satisfied, the recognition model is taken as the target Identify the model.
  • the device is a mobile device.
  • the training module 602 obtains a recognition model based on the model parameters of the intermediate model reported by each mobile device, it is also used to determine whether the recognition model meets the end condition of the model training; if If it is not satisfied, a certain mobile device is selected as the model management device corresponding to the next time slice through communication with other mobile devices, and a model start instruction is sent to the model management device corresponding to the next time slice to instruct the next time slice.
  • the model management device corresponding to the time slice selects other mobile devices corresponding to the next time slice from other mobile devices and sends a model training instruction to the other mobile devices corresponding to the next time slice; if it is satisfied, the recognition model is used as Target recognition model.
  • FIG. 7 is a schematic structural diagram of an inspection device provided by an embodiment of the present invention. As shown in FIG. 7, the device includes:
  • the transceiver module 701 is configured to receive a model training instruction sent by a model management device;
  • the training module 702 is configured to collect training data when traveling in the computer room according to the corresponding patrol route according to the model training instruction, and obtain an intermediate model corresponding to the mobile device based on the local model and the training data;
  • the transceiver module 701 is further configured to report the model parameters of the intermediate model to the model management device.
  • model parameters of the intermediate model reported by each mobile device are used to train the model management device to obtain a recognition model, and the recognition model is used to determine the health status of the computer room.
  • the training module 702 is specifically configured to: perform cyclically according to the model training instruction: collect training data while traveling in the computer room according to the inspection route, and train the mobile device based on the local model and the training data.
  • the transceiver module 701 is specifically configured to: according to the model training instruction, cyclically execute: report the model parameters of the intermediate model to the model management device.
  • the transceiver module 701 is further configured to receive the comprehensive model parameters sent by the model management device
  • the training module 702 is also used to update the local model using the comprehensive model parameters.
  • the device further includes an identification module 703.
  • the identification module 703 is configured to collect data to be identified while traveling in the computer room according to the inspection route, and use the updated local model. The model recognizes the to-be-identified data to determine the health status of the computer room.
  • the model management device sends a model training instruction to each mobile device to instruct each mobile device to collect training data when traveling in the computer room according to the model training instruction according to its own inspection route. , And train based on the local model and the training data to obtain an intermediate model and report it to the model management device.
  • the model management device can be based on the intermediate model reported by each mobile device.
  • the model parameter training obtains the recognition model. Among them, the recognition model is used to determine the health status of the computer room.
  • the mobile device by placing the training process of the intermediate model on the side of the mobile device for execution, the mobile device can only report the model parameters of the intermediate model to the model management device without reporting the full amount of training data. Data has a small amount of data, so this method can save communication overhead and improve inspection efficiency. Moreover, by using the recognition model to determine the health status of the computer room during the inspection process of the mobile device, the joint operation of model training and model recognition is realized, which helps to further improve the inspection efficiency.
  • an embodiment of the present invention provides a computing device. As shown in FIG. 8, it includes at least one processor 801 and a memory 802 connected to the at least one processor.
  • the embodiment of the present invention does not limit the processor.
  • the connection between the processor 801 and the memory 802 in FIG. 8 is taken as an example.
  • the bus can be divided into address bus, data bus, control bus and so on.
  • the memory 802 stores instructions that can be executed by at least one processor 801. By executing the instructions stored in the memory 802, the at least one processor 801 can execute the inspection method described in any of the above steps.
  • the processor 801 is the control center of the computing device. It can use various interfaces and lines to connect various parts of the computing device, and realize data by running or executing instructions stored in the memory 802 and calling data stored in the memory 802. deal with.
  • the processor 801 may include one or more processing units.
  • the processor 801 may integrate an application processor and a modem processor.
  • the application processor mainly processes the operating system, user interface, and application programs.
  • the adjustment processor mainly handles issuing instructions. It can be understood that the foregoing modem processor may not be integrated into the processor 801.
  • the processor 801 and the memory 802 may be implemented on the same chip, and in some embodiments, they may also be implemented on separate chips.
  • the processor 801 may be a general-purpose processor, such as a central processing unit (CPU), a digital signal processor, an application specific integrated circuit (ASIC), a field programmable gate array or other programmable logic devices, discrete gates or transistors Logic devices and discrete hardware components can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention.
  • the general-purpose processor may be a microprocessor or any conventional processor or the like. The steps of the method disclosed in combination with the inspection embodiment can be directly embodied as being executed and completed by a hardware processor, or executed and completed by a combination of hardware and software modules in the processor.
  • the memory 802 as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules.
  • the memory 802 may include at least one type of storage medium, for example, may include flash memory, hard disk, multimedia card, card-type memory, random access memory (Random Access Memory, RAM), static random access memory (Static Random Access Memory, SRAM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), magnetic memory, disk , CD, etc.
  • the memory 802 is any other medium that can be used to carry or store desired program codes in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.
  • the memory 802 in the embodiment of the present invention may also be a circuit or any other device capable of realizing a storage function for storing program instructions and/or data.
  • embodiments of the present invention also provide a computer-readable storage medium that stores a computer program executable by a computing device, and when the program runs on the computing device, the computing device executes Figure 3 to Figure 5 arbitrarily described inspection method.
  • the embodiments of the present invention can be provided as a method or a computer program product. Therefore, the present invention may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
  • a computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device.
  • the device implements the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.
  • These computer program instructions can also be loaded on a computer or other programmable data processing equipment, so that a series of operation steps are executed on the computer or other programmable equipment to produce computer-implemented processing, so as to execute on the computer or other programmable equipment.
  • the instructions provide steps for implementing the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Mobile Radio Communication Systems (AREA)
  • Testing Or Measuring Of Semiconductors Or The Like (AREA)
  • Transplanting Machines (AREA)
  • Indicating And Signalling Devices For Elevators (AREA)

Abstract

一种巡检方法及装置，用以提高巡检效率。其中方法包括：模型管理设备向各个移动设备发送模型训练指示，以指示各个移动设备按照对应的巡检路线在机房中行进时采集训练数据，基于本地模型和训练数据训练得到中间模型并上报给模型管理设备，如此，模型管理设备可以基于各个中间模型的模型参数训练得到识别模型，该识别模型用于确定机房的健康状态。通过将中间模型的训练过程放置在移动设备侧执行，使得移动设备可以仅上报中间模型的模型参数给模型管理设备，而无需上报全量的训练数据，该方式有助于节省通信开销，有效提高巡检效率。

Description

一种巡检方法及装置
相关申请的交叉引用
本申请要求在2020年02月20日提交中国专利局、申请号为202010103868.X、申请名称为“一种巡检方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及图像处理技术领域，尤其涉及一种巡检方法及装置。
背景技术
互联网数据中心(internet data center，IDC)机房是在互联网通信线路和带宽资源的基础上建立的标准化的机房环境。IDC机房可以容纳多种类型的设备，比如服务器、监控设备、管理设备或安全设备等。在实际操作中，通常需要对IDC机房进行巡检，比如采集设备上的信号灯的图像以识别设备上的信号灯的状态、采集设备的图像以识别设备的类型、采集气味数据和/或温度数据以分析机房环境。在对IDC机房中的设备进行巡检时，如何有效地对巡检数据进行识别，对于维护机房安全、及时排查故障是非常重要的。
在一种现有的实现方式中，各个移动设备从机房中采集训练数据并上报给服务器进行集中训练，服务器训练得到识别模型后，将识别模型下发给各个移动设备，以由各个移动设备使用识别模型对采集到的巡检数据进行识别，完成对机房的巡检。然而，该种方式需要各个移动设备将训练数据上报给服务器，而训练数据(比如图像、气味或温度)一般会对应较大的数据量，因此这种巡检方式会存在较大的通信开销，导致巡检效率较低。
综上，目前亟需一种巡检方法，用以解决现有技术采用服务器集中训练识别模型所导致的通信开销大、巡检效率低的技术问题。
发明内容
本发明提供一种巡检方法及装置，用以解决现有技术采用服务器集中训练识别模型所导致的通信开销大、巡检效率低的技术问题。
第一方面，本发明提供一种巡检方法，该方法应用于模型管理设备，该方法包括：模型管理设备向各个移动设备发送模型训练指示，以指示各个移动设备按照各自的巡检路线在机房中行进时采集训练数据、以及基于本地模型和训练数据训练得到中间模型并上报给模型管理设备，如此，模型管理设备在接收到各个移动设备上报的中间模型的模型参数之后，可以基于各个移动设备上报的中间模型的模型参数训练得到识别模型，该识别模型用于确定机房的健康状态。
在上述设计中，通过中间模型的训练过程放置在移动设备侧执行，使得移动设备可以仅上报中间模型的模型参数给模型管理设备，而无需上报全量的训练数据，由于模型参数相对于训练数据来说具有较小的数据量，因此该种巡检方式可以较好地节省通信开销，有效提高巡检效率。且，通过在移动设备的巡检过程中使用识别模型确定机房的健康状态，有助于实现模型训练和模型识别的联合操作，进一步提高巡检效率。
在一种可能的实现方式中，识别模型是指当前时间片对应的识别模型，在这种情况下， 模型管理设备在向各个移动设备发送模型训练指示之前，还可以从机房的全部移动设备中选取出当前时间片所对应的移动设备，以作为执行当前时间片对应的模型训练的各个移动设备。通过将整个时间维度上的模型训练分解为多个时间片分别执行，并在每个时间片中选取部分移动设备作为训练设备，能够有效分配并均匀利用训练数据和移动设备，避免在训练模型的过程中出现由于数据过大而造成参数迭代无法收敛的问题，提高识别模型的准确性。
在一种可能的实现方式中，模型训练指示用于指示各个移动设备循环按照各自的巡检路线在机房中行进时采集训练数据，以及指示各个移动设备基于本地模型和训练数据训练得到中间模型并上报给模型管理设备。在这种情况下，模型管理设备每接收到设定数量的模型参数，则可以至少基于该设定数量的模型参数构建得到综合模型参数，然后将综合模型参数下发给各个移动设备，以使各个移动设备基于综合模型参数更新本地模型。当某一综合模型参数对应的模型满足当前时间片的结束条件时，模型管理设备可以向各个移动设备发送模型训练结束指令。其中，该综合模型参数对应的模型即为识别模型。上述实现方式在每个时间片内执行多次模型综合操作，并在每次模型综合操作中使用设定数量的模型参数得到综合模型参数，而不使用全部移动设备的模型参数进行综合，能在使用尽可能多的模型参数得到综合模型参数的同时，兼容部分移动设备故障的情况，有助于保证模型训练的顺利进行。
在一种可能的实现方式中，模型管理设备可以是指服务器。在这种情况下，模型管理设备在基于各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还可以判断识别模型是否满足模型训练的结束条件，若不满足，则可以从机房的全部移动设备中选取下一时间片对应的移动设备，并向下一时间片对应的移动设备发送模型训练指示，若满足，则可以将识别模型作为目标识别模型。通过将服务器作为模型管理设备，不仅能对各个移动设备的模型训练过程进行统一管理，还能有效降低移动设备的压力，提高移动设备训练中间模型的效率。且，通过在每个时间片启动时重新选取执行训练的移动设备，还能有效分配移动设备的资源，充分考虑到不同移动设备的差异，提高识别模型的准确性。
在一种可能的实现方式中，模型管理设备可以是指移动设备。在这种情况下，模型管理设备在基于各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还可以判断识别模型是否满足模型训练的结束条件。若不满足，则可以通过与其他移动设备的通信选取某一移动设备作为下一时间片对应的模型管理设备，并向下一时间片对应的模型管理设备发送模型启动指示，以指示下一时间片对应的模型管理设备从其它移动设备中选取下一时间片对应的其它移动设备并向下一时间片对应的其它移动设备发送模型训练指示。若满足，则可以将识别模型作为目标识别模型。通过将移动设备作为模型管理设备，可以无需再额外设置服务器，从而有助于降低模型训练的成本。且，通过在每个时间片启动时重新选取执行训练的移动设备，还能有效分配移动设备的资源，充分考虑到不同移动设备的差异，提高识别模型的准确性。
第二方面，本发明提供一种巡检方法，该方法应用于移动设备，该方法包括：移动设备接收模型管理设备发送的模型训练指示，根据模型训练指示，按照对应的巡检路线在机房中行进时采集训练数据，然后基于本地模型和训练数据训练得到移动设备对应的中间模型，并将中间模型的模型参数上报给模型管理设备，以使模型管理设备基于各个移动设备上报的中间模型的模型参数训练得到识别模型。其中，该识别模型用于确定机房的健康状 态。
在一种可能的实现方式中，移动设备根据模型训练指示，可以循环执行如下操作：按照巡检路线在机房中行进并采集训练数据，基于本地模型和训练数据训练得到移动设备对应的中间模型，将中间模型的模型参数上报给模型管理设备。其中，在每次循环中，移动设备在将中间模型的模型参数上报给模型管理设备之后，还可以接收模型管理设备发送的综合模型参数，然后使用综合模型参数更新本地模型。
在一种可能的实现方式中，移动设备使用综合模型参数更新本地模型之后，还可以按照巡检路线在机房中行进并采集待识别数据，然后使用更新后的本地模型对待识别数据进行识别，以确定机房的健康状态。如此，通过在巡检过程中联合执行训练、优化和识别操作，能够实现边训练模型、边优化模型、边识别模型的效果，提高训练和识别的效率。且，通过使用实时优化后的模型进行识别，还能提高识别的效果。
第三方面，本发明提供一种巡检装置，该装置包括：收发模块，用于向各个移动设备发送模型训练指示，以及接收各个移动设备上报的中间模型的模型参数；其中该模型训练指示用于各个移动设备按照各自的巡检路线在机房中行进时采集训练数据，以及指示各个移动设备基于本地模型和训练数据训练得到中间模型；训练模块，用于基于各个移动设备上报的中间模型的模型参数训练得到识别模型。其中，识别模型用于确定机房的健康状态。
在一种可能的实现方式中，识别模型可以为当前时间片对应的识别模型。在这种情况下，在收发模块向各个移动设备发送模型训练指示之前，训练模块还可以从机房的全部移动设备中选取出当前时间片对应的移动设备，作为在当前时间片内执行训练的各个移动设备。
在一种可能的实现方式中，模型训练指示用于各个移动设备循环按照各自的巡检路线在机房中行进并采集训练数据，基于本地模型和训练数据训练得到中间模型并上报。在这种情况下，训练模块具体用于：每接收到设定数量的模型参数，则至少基于设定数量的模型参数构建得到综合模型参数；收发模块具体用于：将综合模型参数下发给各个移动设备，以及当某一综合模型参数对应的模型满足当前时间片的结束条件时，向各个移动设备发送模型训练结束指令，以指示该综合模型参数对应的模型即为识别模型。其中，综合模型参数用于各个移动设备更新本地模型。
在一种可能的实现方式中，该巡检装置为服务器。在这种情况下，训练模块在基于各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还可以判断识别模型是否满足模型训练的结束条件，若不满足，则可以从机房的全部移动设备中选取下一时间片对应的移动设备，并向下一时间片对应的移动设备发送模型训练指示，若满足，则可以将识别模型作为目标识别模型。
在一种可能的实现方式中，该巡检装置为移动设备。在这种情况下，训练模块在基于各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还可以判断识别模型是否满足模型训练的结束条件，若不满足，则可以通过与其他移动设备的通信选取某一移动设备作为下一时间片对应的模型管理设备，并向下一时间片对应的模型管理设备发送模型启动指示，以指示下一时间片对应的模型管理设备从其它移动设备中选取下一时间片对应的其它移动设备并向下一时间片对应的其它移动设备发送模型训练指示；若满足，则可以将识别模型作为目标识别模型。
第四方面，本发明提供的一种巡检装置，该装置包括：收发模块，用于接收模型管理 设备发送的模型训练指示；训练模块，用于根据模型训练指示，按照对应的巡检路线在机房中行进时采集训练数据，基于本地模型和训练数据训练得到移动设备对应的中间模型；收发模块，还用于将中间模型的模型参数上报给模型管理设备。其中，各个移动设备上报的中间模型的模型参数用于模型管理设备训练得到识别模型，识别模型用于确定机房的健康状态。
在一种可能的实现方式中，训练模块可以根据模型训练指示，循环执行如下操作：按照巡检路线在机房中行进并采集训练数据，基于本地模型和训练数据训练得到移动设备对应的中间模型。相应地，收发模块可以根据模型训练指示，循环执行如下操作：将中间模型的模型参数上报给模型管理设备。在每次循环中，在收发模块将中间模型的模型参数上报给模型管理设备之后：收发模块还可以接收模型管理设备发送的综合模型参数，训练模块还可以使用综合模型参数更新本地模型。
在一种可能的实现方式中，该装置还可以包括识别模块。在训练模块使用综合模型参数更新本地模型之后，识别模块可以按照巡检路线在机房中行进时采集待识别数据，使用更新后的本地模型对待识别数据进行识别，以确定机房的健康状态。
第五方面，本发明提供一种计算设备，包括至少一个处理器以及至少一个存储器，其中，所述存储器存储有计算机程序，当所述程序被所述处理器执行时，使得所述处理器执行上述第一方面或第二方面任意所述的巡检方法。
第六方面，本发明提供一种计算机可读存储介质，其存储有可由计算设备执行的计算机程序，当所述程序在所述计算设备上运行时，使得所述计算设备执行上述第一方面或第二方面任意所述的巡检方法。
本发明的这些实现方式或其他实现方式在以下实施例的描述中会更加简明易懂。
附图说明
为了更清楚地说明本发明实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简要介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域的普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为本发明实施例提供的一种IDC机房的结构示意图；
图2为本发明实施例提供的一种巡检系统的系统架构示意图；
图3为本发明实施例提供的一种巡检方法对应的流程示意图；
图4为服务器作为模型管理设备时巡检方法对应的交互流程示意图；
图5为移动设备作为模型管理设备时巡检方法对应的交互流程示意图；
图6为本发明实施例提供的一种巡检装置的结构示意图；
图7为本发明实施例提供的另一种巡检装置的结构示意图；
图8为本发明实施例提供的一种计算设备的结构示意图。
具体实施方式
为了使本发明的目的、技术方案和优点更加清楚，下面将结合附图对本发明作进一步地详细描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有 其它实施例，都属于本发明保护的范围。
图1为本发明实施例提供的一种IDC机房的结构示意图，如图1所示，IDC机房中可以设置有至少一排机柜，比如机柜101～机柜106，机柜101～机柜104可以并列设置，机柜105和机柜106可以并列设置，每排机柜上可以设置有多台设备，比如服务器设备、数据采集设备、监控设备或温控设备等。
本发明实施例中，机柜可以为单层结构，多台设备并列放置在单层结构上。或者，机柜也可以为多层结构，多台设备分别放置在多层结构上，每层结构上可以并列放置一台或多台设备，具体不作限定。
基于图1所示意的IDC机房，图2为本发明实施例提供的一种巡检系统的系统架构示意图，如图2所示，该系统架构中包括模型管理设备110和至少一个移动设备，比如移动设备121、移动设备122和移动设备123。其中，模型管理设备110可以与任一移动设备连接，比如可以通过有线方式连接，也可以通过无线方式连接，具体不作限定。
本发明实施例中，各个移动设备可以部署在同一IDC机房中，也可以部署在不同的IDC机房中。若每个IDC机房中部署一个移动设备，则该移动设备可以负责巡检整个IDC机房。若每个IDC机房中部署多个移动设备，则每个移动设备可以负责巡检IDC机房中的一片区域，多台移动设备共同完成对整个IDC机房的巡检操作。
需要说明的是，本发明实施例中的巡检方法可以用于对一个IDC机房进行巡检，也可以用于对多个IDC机房进行巡检，具体不作限定。
基于图2所示意的系统架构，图3为本发明实施例提供的一种巡检方法对应的流程示意图，该方法可以适用于模型管理设备和移动设备，如图2所示意出的模型管理设备110、以及移动设备121～移动设备123中的一个或多个。如图3所示，该方法包括：
步骤301，模型管理设备向各个移动设备发送模型训练指示。
在一种可能的实现方式中，目标识别模型可以经由多个时间片训练得到，在任一时间片中，模型管理设备可以先从机房的全部移动设备中选取出该时间片对应的移动设备，再向该时间片对应的移动设备发送模型训练指示。相应地，接收到模型训练指示的移动设备可以联合模型管理设备进行模型训练，以更新上一时间片对应的识别模型，得到该时间片对应的识别模型。而未接收到模型训练指示的移动设备可以只执行识别操作，而不参与该时间片内的模型训练。
如此，按照上述实现方式，步骤301所述的各个移动设备可以为任一时间片对应的移动设备。
具体实施中，选取任一时间片对应的移动设备的方式可以有多种，比如可以随机选取固定数量或随机数量的移动设备作为任一时间片对应的移动设备，或者也可以轮流选取固定数量或随机数量的移动设备作为任一时间片对应的移动设备，或者还可以选取处理能力较强的固定数量或随机数量的移动设备作为任一时间片对应的移动设备，具体不作限定。
在一个示例中，可以从全部移动设备中选取设定比例的移动设备作为任一时间片对应的移动设备。其中，设定比例可以由本领域技术人员根据经验进行设置，比如可以设置为50％～80％，以在保留大部分训练数据的特征的同时，避免训练数据过多所造成的过拟合，提高识别模型的精度。
在该实现方式中，通过设置多个时间片，并在每个时间片中选取部分移动设备参与模型训练，能够有效分配并均匀利用训练数据和移动设备。如此，训练数据的均匀分配能够 避免训练模型时出现由于数据过大而造成参数迭代无法收敛的问题。移动设备的均匀分配能够提高模型管理设备应对移动设备故障的能力，提高训练识别模型的可用性。
步骤302，移动设备根据模型训练指示，按照巡检路线在机房中行进时采集训练数据，基于本地模型和训练数据训练得到移动设备对应的中间模型。
在一种可能的实现方式中，模型管理设备发送的模型训练指示用于指示移动设备循环执行如下操作：按照巡检路线在机房中行进时采集训练数据，基于本地模型和训练数据训练得到移动设备对应的中间模型，将中间模型上报给模型管理设备。如此，在任一时间片内，针对于接收到模型训练指示的任一移动设备，该移动设备可以不断采集训练数据，每采集到固定数量的训练数据后，就可以使用采集到的这些训练数据对本地模型进行训练，得到该移动设备对应的中间模型，并上报给模型管理设备。其中，固定数量可以是1，也可以是大于1的任意整数，不作限定。其中，训练数据可以由本领域技术人员根据需要进行设置，比如可以为机房设备图像、信号灯图像、温度信息或气味信息中的任意一项或任意多项，不作限定。
本发明实施例中，优化本地模型的方式可以为：先使用本地模型对各个训练数据进行预测，得到各个训练数据的预测标签，再根据各个训练数据的预测标签与真实标签的匹配程度，从各个训练数据中确定出预测标签正确的训练数据的数量以及预测标签错误的训练数据的数量，根据这些数量计算得到损失函数，最后根据损失函数调整本地模型的模型参数以得到中间模型。
相应地，在采集训练数据时，移动设备可以按照固定频率来采集训练数据，也可以先采集连续数据，再按照固定频率从连续数据中截取训练数据，将截取的训练数据中每个固定时段内的训练数据作为一次训练数据，不作限定。举例来说，当训练数据为设备图像，固定时段为1分钟时，若固定频率为6毫秒/每次，则：移动设备可以每隔6毫秒拍摄得到一张设备图像，然后将每个1分钟的时段内拍摄得到的600张设备图像作为一个训练数据；也可以先录制时长为1分钟(或者更长)的设备视频，再从设备视频中每隔6毫秒截取一张设备图像，每次从截取的设备图像中选取600张连续的设备图像作为一个训练数据。
需要说明的是，上述所述的固定数量仅是一种示例性的说明，并不构成对本方案的限定。在具体实施中，移动设备也可以在每采集到固定时段的训练数据后使用采集到的这些训练数据对本地模型进行训练，或者移动设备也可以在每采集到随机数量或随机时段的训练数据后使用采集到的这些训练数据对本地模型进行训练，具体不作限定。
本发明实施例中，移动设备的本地模型可以为初始模型、上一时间片对应的识别模型和综合模型参数对应的模型中的任意一种。在第一时间片的模型训练的启动阶段，移动设备中的本地模型为初始模型。在其它时间片的模型训练的启动阶段，移动设备中的本地模型为上一时间片对应的识别模型。在任一时间片的模型训练的执行阶段，移动设备的本地模型为综合模型参数对应的模型。
具体实施中，初始模型可以由模型管理设备或任一移动设备训练得到。以模型管理设备训练初始模型为例，具体实施中：模型管理设备可以先获取初始训练数据，然后使用初始训练数据训练得到初始模型，再将初始模型的模型参数下发给机房中的全部移动设备。其中，初始训练数据可以通过多种方式获取，比如可以通过网络下载，也可以在第一时间片之前从机房中采集，还可以从第三方模型管理设备获取，不作限定。
相应地，针对于未接收到模型训练指示的任一移动设备，该移动设备可以处于等待状 态，而不参与该时间片对应的识别模型训练。在下一时间片时，若该移动设备接收到模型训练指示，则该移动设备可以参与下一时间片对应的识别模型训练，若还是未接收到模型训练指示，则该移动设备可以继续处于等待状态。
步骤303，移动设备将中间模型的模型参数上报给模型管理设备。
此处，中间模型的模型参数可以包括梯度和损失值，还可以包括其它信息，不作限定。
具体实施中，移动设备可以先对中间模型的模型参数进行压缩，再将压缩包上报给模型管理设备，以减少通信开销，降低数据传输耗费的时间，提高巡检效率。或者，移动设备还可以对压缩包进行加密操作，并将加密后的压缩包上报给模型管理设备，以提高传输过程中数据的安全性。
步骤304，模型管理设备基于各个移动设备对应的中间模型的模型参数训练得到识别模型。
本发明实施例中，模型管理设备可以为服务器，也可以为移动设备。当模型管理设备为服务器时，模型管理设备可以直接使用当前时间片对应的其它移动设备发送的各个中间模型的模型参数计算得到综合模型参数。当模型管理设备为移动设备时，模型管理设备除了可以向当前时间片对应的其它移动设备发送模型训练指令之外，还可以使用训练数据训练得到模型管理设备对应的中间模型，从而基于当前时间片对应的其它移动设备的模型参数和模型管理设备对应的中间模型的模型参数计算得到综合模型参数。
具体实施中，针对于任一时间片，模型管理设备可以采用多种方式训练得到该时间片对应的识别模型。比如，模型管理设备可以在接收到该时间片对应的全部或大部分移动设备发送的模型参数后，即使用全部或大部分移动设备对应的模型参数训练得到该时间片对应的识别模型。或者，模型管理设备也可以先指示该时间片对应的移动设备独自执行多次模型训练以得到每个移动设备对应的多个模型参数，再使用各个移动设备对应的各个模型参数训练得到该时间片对应的识别模型，不作限定。
在一种可能的实现方式中，模型管理设备可以采用如下方式训练得到任一时间片对应的识别模型：
步骤a，模型管理设备每接收到设定数量的模型参数，即可根据设定数量的模型参数计算得到综合模型参数。
具体实施中，由于模型训练指示用于指示各个移动设备循环执行采集训练数据、训练中间模型以及中间模型的模型参数上报的操作，因此模型管理设备可以持续不断的接收到各个移动设备发送的模型参数。当每接收到设定数量的模型参数时，模型管理设备可以根据本次接收到的设定数量的模型参数计算得到一个综合模型参数。此处，设定数量可以由本领域技术人员根据经验进行设置，比如可以设置为略小于该时间片对应的移动设备的总数量，以在使用尽可能多的模型参数得到综合模型参数的同时，兼容部分移动设备故障的情况，保证模型训练的顺利进行。
具体实施中，模型管理设备计算得到综合模型参数的方式可以有多种，比如可以将设定数量的模型参数(以及模型管理设备的模型参数)的平均参数作为综合模型参数，或者也可以将设定数量的模型参数(以及模型管理设备的模型参数)的加权平均参数作为综合模型参数，或者还可以先从设定数量的模型参数(以及模型管理设备的模型参数)中筛除不满足要求的模型参数，再将满足要求的模型参数的平均参数或加权平均参数作为综合模型参数，不作限定。
步骤b，模型管理设备将综合模型参数下发给机房中的全部移动设备。
在一个示例中，机房中的任一移动设备还可以在行进过程中采集待识别数据，并使用本地模型对待识别数据进行识别，以完成在巡检过程中对机房的识别操作。其中，识别操作可以由识别指令触发，也可以按照设定周期执行，不作限定。
相应地，模型管理设备在每个时间片的训练过程中，还可以将综合模型参数下发给机房中的全部移动设备。而机房中的任一移动设备(未接收到模型训练指示的移动设备或接收到模型训练指示的移动设备)接收到综合模型参数后，还可以使用综合模型参数更新本地模型。如此，针对于后续采集到的任一待识别数据，移动设备可以使用更新后的模型对待识别数据进行识别，从而能够在巡检过程中执行训练、优化和识别的联合操作，实现边训练模型、边优化模型、边识别数据的效果，提高训练和识别的效率。且，通过使用实时优化后的模型进行识别，还能提高识别的效果。
步骤c，模型管理设备确定综合模型参数是否满足当前时间片对应的结束条件，若满足，则执行步骤d 1，若不满足，则执行步骤d 2。
其中，任一时间片对应的结束条件可以为设定时段未接收到模型参数、模型训练次数大于或等于设定次数、模型训练时间大于或等于设定时间中的任意一项或任意多项。
步骤d 1，模型管理设备确定当前时间片的模型训练已结束，综合模型参数即为当前时间片对应的识别模型的模型参数，综合模型参数对应的模型即为当前时间片对应的识别模型，因此模型管理设备可以向当前时间片对应的各个移动设备发送模型训练结束指令。
步骤d 2，模型管理设备确定当前时间片的模型训练还未结束，综合模型参数不是当前时间片对应的识别模型参数，综合模型参数对应的模型不是当前时间片对应的识别模型，因此模型管理设备可以不作特殊处理。由于当前时间片对应的各个移动设备重复执行采集训练数据、训练中间模型和上报模型参数的操作，因此模型管理设备可以重复执行步骤a至步骤d 1或步骤a至步骤d 2，直至型管理设备确定当前时间片的模型训练已结束。
本发明实施例中，在确定当前时间片对应的识别模型后，模型管理设备还可以获取各个移动设备使用当前时间片对应的识别模型的识别效果，并确定识别效果是否满足模型训练的结束条件。若是，则模型管理设备可以将当前时间片对应的识别模型作为目标识别模型，并向当前时间片对应的各个移动模型发送模型训练结束指令。若否，则模型管理设备可以启动下一时间片的模型训练。
在一个示例中，在需要增加新的巡检区域或巡检机房时，模型管理设备可以先跟新的巡检区域或巡检机房中的移动设备建立通信连接，再将模型管理设备中存储的最新模型参数下发给新接入的移动设备，以便于新接入的移动设备使用最新模型参数进行识别。相应地，在下一时间片启动时，模型管理设备可以重新从全部设备(包括新接入的移动设备)中选取下一时间片对应的移动设备，并联合执行下一时间片的模型训练，从而快速实现对新巡检区域的巡检操作，提高协同训练过程的灵活性。
本发明实施例中，模型管理设备可以为服务器或移动设备，当模型管理设备不同时，启动下一时间片的模型训练的方式也不同，具体为：
当模型管理设备为服务器时，模型管理设备可以与机房中的每个移动设备通信连接，而机房中的各个移动设备之间可以不通信连接。具体实施中，模型管理设备若确定启动下一时间片的模型训练，则可以先从全部移动设备中选取部分移动设备作为下一时间片对应的移动设备，然后再联合下一时间片对应的移动设备训练得到下一时间片对应的识别模型。
当模型管理设备为移动设备时，机房中的任意两个移动设备可以通信连接，如此，机房中的各个移动设备可以构成去中心化分布式集群。具体实施中，若确定启动下一时间片的模型训练，则全部移动设备可以先通过通信交互选择某一移动设备作为下一时间片对应的模型管理设备，再经由下一时间片对应的模型管理设备从其它移动设备中选取部分移动设备作为下一时间片对应的其他移动设备，然后下一时间片对应的模型管理设备再联合下一时间片对应的其它移动设备训练得到下一时间片对应的识别模型。其中，选取下一时间片对应的模型管理设备的方式可以有多种，比如可以采用投票方式将得票数最多的移动设备作为下一时间片对应的模型管理设备，或者也可以将计算能力最强的移动设备作为下一时间片对应的模型管理设备，或者还可以轮流或随机选取未做过模型管理设备的移动设备作为下一时间片对应的模型管理设备，不作限定。
为了便于理解，下面分别描述采用服务器作为模型管理设备和采用移动设备作为模型管理设备时巡检方法的具体实现过程。
图4为本发明实施例提供的一种巡检方法对应的交互流程示意图，该方法可以适用于模型管理设备和移动设备，如图2所示意出的模型管理设备110和移动设备121～移动设备123。在该示例中，模型管理设备可以是指服务器。如图4所示，该方法包括：
步骤401，模型管理设备使用初始训练数据训练得到初始模型。
步骤402，模型管理设备将初始模型的模型参数下发给机房的全部移动设备。
步骤403，模型管理设备从机房的全部移动设备中选取当前时间片对应的移动设备。
步骤404，模型管理设备向当前时间片对应的移动设备发送模型训练指示。
步骤405，针对于当前时间片对应的任一移动设备(即接收到模型训练指示的任一移动设备)，该移动设备根据模型训练指示，循环执行如下操作：按照该移动设备对应的巡检路线在机房中行进时采集训练数据，基于本地模型和训练数据进行模型训练，得到移动设备对应的中间模型。
其中，本地模型可以为初始模型、上一时间片对应的识别模型和综合模型参数对应的模型中的任意一项或任意多项。若当前时间片为第一时间片，且该次训练为第一时间片中的第一次训练，则本地模型为初始模型。若当前时间片为第一时间片之后的任一时间片，且该次训练为任一时间片中的第一次训练，则本地模型为上一时间片对应的识别模型。若当前时间片不为任一时间片中的第一次训练，则本地模型为综合模型参数对应的模型。
相应地，针对于不是当前时间片的任一移动设备(即未接收到模型训练指示的任一移动设备)，该移动设备不参与当前时间片的模型训练，但是可以执行识别操作。比如在机房中按照对应的巡检路线行进时，还可以采集待识别数据，然后使用本地模型对待识别数据进行识别操作，或者也可以执行模型更新操作，比如在机房中按照对应的巡检路线行进时，还可以同采集训练数据，使用训练数据更新移动设备的本地模型，并使用本地模型对待识别数据进行识别操作。
步骤406，当前时间片对应的任一移动设备按照模型训练指示，循环上报中间模型的模型参数给模型管理设备。
步骤407，模型管理设备每接收到设定数量的模型参数，即根据设定数量的模型参数计算得到综合模型参数。
步骤408，模型管理设备将综合模型参数下发给机房中的全部移动设备。
步骤409，机房中的任一移动设备接收到模型管理设备发送的综合模型参数后，使用 综合模型参数更新本地模型，以使用更新后的本地模型对巡检路线上采集到的待识别数据进行识别操作。其中，机房中的任一移动设备可以为当前时间片对应的任一移动设备，也以为任一其它移动设备，如未接收到模型训练指示的任一移动设备。
步骤410，模型管理设备判断综合模型参数对应的模型是否满足当前时间片的结束条件，若是，则执行步骤411，若否，则执行步骤407。
其中，当前时间片的结束条件可以为训练次数大于或等于预设训练次数、训练时长大于或等于预设训练时长、综合模型参数对应的模型效果满足预设模型效果、设定时长内未接收到移动设备发送的模型参数中的任意一项或任意多项，不作限定。
步骤411，模型管理设备向当前时间片对应的移动设备发送模型训练结束指令。
步骤412，模型管理设备确定综合模型参数对应的模型为当前时间片对应的识别模型，判断当前时间片对应的识别模型是否满足模型训练的结束条件，若是，则执行步骤413，若否，则执行步骤414。
其中，模型训练的结束条件可以为模型效果满足预设效果、模型训练时长大于或等于预设时长、时间片数量大于或等于预设时间片数量中的任意一项或任意多项，不作限定。
举例来说，若模型训练的结束条件为模型效果满足预设效果，则模型管理设备可以获取各个移动设备(当前时间片对应的移动设备，或者还包括其它移动设备)使用当前时间片对应的识别模型执行识别操作的识别效果，然后根据各个移动设备的识别效果得到综合识别效果。若综合识别效果未达到预设效果，则确定当前时间片对应的识别模型不满足模型训练的结束条件。若综合识别效果已达到预设效果，则确定当前时间片对应的识别模型满足模型训练的结束条件。
步骤413，模型管理设备将当前时间片对应的识别模型作为目标识别模型，结束模型训练。
步骤414，模型管理设备确定当前时间片的模型训练结束，将下一时间片作为当前时间片，并执行步骤403，以启动下一时间片的模型训练。
本发明实施例中，通过将服务器作为模型管理设备，能通过服务器对各个移动设备的模型训练过程进行统一管理，这不仅有助于降低移动设备的压力，还能提高移动设备训练中间模型的效率。且，通过在每个时间片启动时重新选取执行训练的移动设备，能够有效分配移动设备，充分考虑到不同移动设备的差异，提高识别模型的准确性。
图5为本发明实施例提供的又一种巡检方法对应的交互流程示意图，该方法适用于模型管理设备和移动设备，如图2所示意的模型管理设备110和移动设备121～移动设备123。其中，该示例中的模型管理设备可以是指移动设备。如图5所示，该方法包括：
步骤501，任一移动设备使用初始训练数据训练得到初始模型。
步骤502，任一移动设备将初始模型的模型参数同步给机房中的其它移动设备，如除任一移动设备以外的其它移动设备。
步骤503，机房中的各个移动设备进行通信交互，从各个移动设备中选取出某一移动设备作为当前时间片对应的模型管理设备。
此处，选取的方式可以有多种：比如可以设置任意两个移动设备相互发送各自的资源占用情况，通过各个移动设备对资源占用率最低的移动设备进行共识，以选取出处理能力最强的移动设备作为模型管理设备；或者可以为各个移动设备设置编号，每个移动设备均保存有其它移动设备的编号和国际互联协议(Internet Protocol，IP)地址的对应关系，当 某一移动设备执行完对应时间片的模型训练后，可以查询该对应关系确定下一编号的移动设备的IP地址，然后向下一编号的移动设备的IP地址发送指令，该指令用于指示下一编号的移动设备启动下一时间片的模型训练，等等。
步骤504，当前时间片对应的模型管理设备从其它移动设备(即除当前时间片对应的模型管理设备以外的移动设备)中选取出当前时间片对应的其它移动设备。
步骤505，当前时间片对应的模型管理设备向当前时间片对应的其它移动设备发送模型训练指示。
步骤506，针对于当前时间片对应的任一其它移动设备，该移动设备接收到模型训练指示后，根据模型训练指示，循环执行：按照对应的巡检路线在机房中行进时采集训练数据，并基于本地模型和训练数据进行模型训练，得到移动设备对应的中间模型。
相应地，当前时间片对应的模型管理设备也循环执行：按照对应的巡检路线在机房中行进时采集训练数据，并基于本地模型和训练数据进行模型训练，得到当前时间片对应的模型管理设备对应的中间模型。
步骤507，当前时间片对应的任一其它移动设备循环执行：将中间模型的模型参数同步给当前时间片对应的模型管理设备。
步骤508，当前时间片对应的模型管理设备每接收到设定数量的模型参数，即根据当前时间片对应的模型管理设备训练得到的中间模型的模型参数和设定数量的模型参数计算得到综合模型参数。
步骤509，当前时间片对应的模型管理设备将综合模型参数下发给机房中的各个移动设备。
步骤510，机房中的任一移动设备使用综合模型参数更新本地模型，并使用更新后的本地模型对对应的巡检路线上采集到的待识别数据进行识别操作，确定机房的健康状态。
步骤511，当前时间片对应的模型管理设备判断综合模型参数对应的模型是否满足当前时间片的结束条件，若是，则执行步骤512，若否，则执行步骤508。
步骤512，当前时间片对应的模型管理设备向当前时间片对应的其它移动设备发送模型训练结束指令。
步骤513，当前时间片对应的模型管理设备确定综合模型参数对应的模型为当前时间片对应的识别模型，判断当前时间片对应的识别模型是否满足模型训练的结束条件，若是，则执行步骤514，若否，则执行步骤515。
步骤514，当前时间片对应的模型管理设备将当前时间片对应的识别模型作为目标识别模型。
步骤515，当前时间片对应的模型管理设备确定当前时间片的模型训练结束，将下一时间片作为当前时间片，执行步骤503。
其中，执行步骤503的方式可以有多种。比如当前时间片对应的模型管理设备可以向其它移动设备(除当前时间片对应的模型管理设备之外的移动设备)发送交互指令，该交互指令用于任一移动设备向其它移动设备发送各自的资源占用情况，通过各个移动设备对资源占用率最低的移动设备进行共识，以选取出处理能力最强的移动设备作为模型管理设备。或者，当前时间片对应的模型管理设备可以向下一编号的移动设备的IP地址发送启动指令，该启动指令用于指示下一编号的移动设备启动下一时间片的模型训练，等等，不作限定。
本发明实施例中，通过将移动设备作为模型管理设备，可以无需再额外设置服务器，从而可以降低模型训练的成本。且，通过在每个时间片启动时重新选取执行训练的移动设备，能够有效分配移动设备，充分考虑到不同移动设备的差异，提高识别模型的准确性。
需要说明的是，图4和图5中的步骤编号仅为执行流程的一种示例，并不构成对各个步骤的执行先后顺序的限定，比如步骤409或步骤510可以发生在流程中的任一时刻，不作限定。
本发明的上述实施例中，模型管理设备向各个移动设备发送模型训练指示，以使各个移动设备根据模型训练指示按照各自的巡检路线在机房中行进时采集训练数据，并基于本地模型和所述训练数据训练得到中间模型并上报给模型管理设备，如此，模型管理设备在接收各个移动设备上报的中间模型的模型参数后，可以基于各个移动设备上报的中间模型的模型参数训练得到识别模型，该识别模型用于确定机房的健康状态。本发明实施例中，通过将中间模型的训练过程放置在移动设备侧执行，使得移动设备可以仅上报中间模型的模型参数给模型管理设备，而无需上报全量的训练数据，由于模型参数相对于训练数据来说具有较小的数据量，因此该种方式可以节省通信开销，提高巡检效率。且，通过在移动设备的巡检过程中使用识别模型确定机房的健康状态，实现了模型训练和模型识别的联合操作，有助于进一步提高巡检效率。
针对上述方法流程，本发明实施例还提供一种巡检装置，该装置的具体内容可以参照上述方法实施。
图6为本发明实施例提供的一种巡检装置的结构示意图，如图6所示，该装置包括：
收发模块601，用于向各个移动设备发送模型训练指示，以及接收所述各个移动设备上报的中间模型的模型参数；其中，所述模型训练指示用于：所述各个移动设备按照各自的巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到中间模型并上报；
训练模块602，用于基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型，所述识别模型用于确定所述机房的健康状态。
可选地，所述识别模型为当前时间片对应的识别模型。在这种情况下，在所述收发模块601向各个移动设备发送模型训练指示之前，所述训练模块602还用于：从机房的全部移动设备中选取出当前时间片对应的移动设备，作为所述各个移动设备。
可选地，所述模型训练指示用于：所述各个移动设备循环按照各自的巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到中间模型并上报。在这种情况下，所述训练模块602具体用于：每接收到设定数量的模型参数，则至少基于所述设定数量的模型参数构建得到综合模型参数。对应的，所述收发模块601具体用于：将所述综合模型参数下发给所述各个移动设备，当某一综合模型参数对应的模型满足当前时间片的结束条件时，向所述各个移动设备发送模型训练结束指令，所述综合模型参数对应的模型即为所述识别模型。其中，所述综合模型参数用于所述各个移动设备更新本地模型。
可选地，所述装置为服务器。在这种情况下，所述训练模块602基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还用于：判断所述识别模型是否满足所述模型训练的结束条件；若不满足，则从所述机房的全部移动设备中选取下一时间片对应的移动设备，向所述下一时间片对应的移动设备发送模型训练指示；若满足，则将所述识别模型作为目标识别模型。
可选地，所述装置为移动设备。在这种情况下，所述训练模块602基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还用于：判断所述识别模型是否满足所述模型训练的结束条件；若不满足，则通过与其他移动设备的通信选取某一移动设备作为下一时间片对应的模型管理设备，向所述下一时间片对应的模型管理设备发送模型启动指示，以指示所述下一时间片对应的模型管理设备从其它移动设备中选取下一时间片对应的其它移动设备并向所述下一时间片对应的其它移动设备发送模型训练指示；若满足，则将所述识别模型作为目标识别模型。
图7为本发明实施例提供的一种巡检装置的结构示意图，如图7所示，该装置包括：
收发模块701，用于接收模型管理设备发送的模型训练指示；
训练模块702，用于根据所述模型训练指示，按照对应的巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到移动设备对应的中间模型；
所述收发模块701，还用于将所述中间模型的模型参数上报给所述模型管理设备。
其中，所述各个移动设备上报的中间模型的模型参数用于所述模型管理设备训练得到识别模型，所述识别模型用于确定所述机房的健康状态。
可选地，所述训练模块702具体用于：根据所述模型训练指示，循环执行：按照巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到所述移动设备对应的中间模型。所述收发模块701具体用于：根据所述模型训练指示，循环执行：将所述中间模型的模型参数上报给所述模型管理设备。其中，在每次循环中，在所述收发模块701将所述中间模型的模型参数上报给所述模型管理设备之后：所述收发模块701还用于接收所述模型管理设备发送的综合模型参数，所述训练模块702还用于使用所述综合模型参数更新所述本地模型。
可选地，所述装置还包括识别模块703。在所述训练模块702使用所述综合模型参数更新所述本地模型之后，所述识别模块703用于：按照所述巡检路线在所述机房中行进时采集待识别数据，使用更新后的本地模型对所述待识别数据进行识别，以确定所述机房的健康状态。
从上述内容可以看出：本发明的上述实施例中，模型管理设备向各个移动设备发送模型训练指示，以指示各个移动设备根据模型训练指示按照各自的巡检路线在机房中行进时采集训练数据，并基于本地模型和所述训练数据训练得到中间模型并上报给模型管理设备，如此，模型管理设备在接收各个移动设备上报的中间模型的模型参数后，可以基于各个移动设备上报的中间模型的模型参数训练得到识别模型。其中，识别模型用于确定机房的健康状态。本发明实施例中，通过将中间模型的训练过程放置在移动设备侧执行，使得移动设备可以仅上报中间模型的模型参数给模型管理设备，而无需上报全量的训练数据，由于模型参数相对于训练数据来说具有较小的数据量，因此该种方式可以节省通信开销，提高巡检效率。且，通过在移动设备的巡检过程中使用识别模型确定机房的健康状态，实现了模型训练和模型识别的联合操作，有助于进一步提高巡检效率。
基于相同的技术构思，本发明实施例提供了一种计算设备，如图8所示，包括至少一个处理器801，以及与至少一个处理器连接的存储器802，本发明实施例中不限定处理器801与存储器802之间的具体连接介质，图8中处理器801和存储器802之间通过总线连接为例。总线可以分为地址总线、数据总线、控制总线等。
在本发明实施例中，存储器802存储有可被至少一个处理器801执行的指令，至少一 个处理器801通过执行存储器802存储的指令，可以执行上述任意步骤所述的巡检方法。
其中，处理器801是计算设备的控制中心，可以利用各种接口和线路连接计算设备的各个部分，通过运行或执行存储在存储器802内的指令以及调用存储在存储器802内的数据，从而实现数据处理。可选的，处理器801可包括一个或多个处理单元，处理器801可集成应用处理器和调制解调处理器，其中，应用处理器主要处理操作系统、用户界面和应用程序等，调制解调处理器主要处理下发指令。可以理解的是，上述调制解调处理器也可以不集成到处理器801中。在一些实施例中，处理器801和存储器802可以在同一芯片上实现，在一些实施例中，它们也可以在独立的芯片上分别实现。
处理器801可以是通用处理器，例如中央处理器(CPU)、数字信号处理器、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件，可以实现或者执行本发明实施例中公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者任何常规的处理器等。结合巡检实施例所公开的方法的步骤可以直接体现为硬件处理器执行完成，或者用处理器中的硬件及软件模块组合执行完成。
存储器802作为一种非易失性计算机可读存储介质，可用于存储非易失性软件程序、非易失性计算机可执行程序以及模块。存储器802可以包括至少一种类型的存储介质，例如可以包括闪存、硬盘、多媒体卡、卡型存储器、随机访问存储器(Random Access Memory，RAM)、静态随机访问存储器(Static Random Access Memory，SRAM)、可编程只读存储器(Programmable Read Only Memory，PROM)、只读存储器(Read Only Memory，ROM)、带电可擦除可编程只读存储器(Electrically Erasable Programmable Read-Only Memory，EEPROM)、磁性存储器、磁盘、光盘等等。存储器802是能够用于携带或存储具有指令或数据结构形式的期望的程序代码并能够由计算机存取的任何其他介质，但不限于此。本发明实施例中的存储器802还可以是电路或者其它任意能够实现存储功能的装置，用于存储程序指令和/或数据。
基于同一发明构思，本发明实施例还提供了一种计算机可读存储介质，其存储有可由计算设备执行的计算机程序，当所述程序在所述计算设备上运行时，使得所述计算设备执行图3至图5任意所述的巡检方法。
本领域内的技术人员应明白，本发明的实施例可提供为方法、或计算机程序产品。因此，本发明可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本发明可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本发明是参照根据本发明实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方 式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
尽管已描述了本发明的优选实施例，但本领域内的技术人员一旦得知了基本创造性概念，则可对这些实施例作出另外的变更和修改。所以，所附权利要求意欲解释为包括优选实施例以及落入本发明范围的所有变更和修改。
显然，本领域的技术人员可以对本发明进行各种改动和变型而不脱离本发明的精神和范围。这样，倘若本发明的这些修改和变型属于本发明权利要求及其等同技术的范围之内，则本发明也意图包含这些改动和变型在内。

Claims (18)

  1. 一种巡检方法，其特征在于，应用于模型管理设备，所述方法包括：
    向各个移动设备发送模型训练指示；所述模型训练指示用于所述各个移动设备按照各自的巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到中间模型；
    接收所述各个移动设备上报的中间模型的模型参数；
    基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型，所述识别模型用于确定所述机房的健康状态。
  2. 根据权利要求1所述的方法，其特征在于，所述识别模型为当前时间片对应的识别模型；
    所述向各个移动设备发送模型训练指示之前，还包括：
    从机房的全部移动设备中选取出当前时间片对应的移动设备，作为所述各个移动设备。
  3. 根据权利要求2所述的方法，其特征在于，所述模型训练指示用于所述各个移动设备循环按照各自的巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到中间模型并上报；
    所述接收所述各个移动设备上报的中间模型的模型参数，基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型，包括：
    每接收到设定数量的模型参数，则至少基于所述设定数量的模型参数构建得到综合模型参数，将所述综合模型参数下发给所述各个移动设备；所述综合模型参数用于所述各个移动设备更新本地模型；
    当某一综合模型参数对应的模型满足当前时间片的结束条件时，向所述各个移动设备发送模型训练结束指令，所述综合模型参数对应的模型即为所述识别模型。
  4. 根据权利要求2或3所述的方法，其特征在于，所述模型管理设备为服务器；
    所述基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还包括：
    判断所述识别模型是否满足所述模型训练的结束条件，若不满足，则从所述机房的全部移动设备中选取下一时间片对应的移动设备，向所述下一时间片对应的移动设备发送模型训练指示；若满足，则将所述识别模型作为目标识别模型。
  5. 根据权利要求2或3所述的方法，其特征在于，所述模型管理设备为移动设备；
    所述基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还包括：
    判断所述识别模型是否满足所述模型训练的结束条件，若不满足，则通过与其他移动设备的通信选取某一移动设备作为下一时间片对应的模型管理设备，向所述下一时间片对应的模型管理设备发送模型启动指示；所述模型启动指示用于所述下一时间片对应的模型管理设备从其它移动设备中选取下一时间片对应的其它移动设备，向所述下一时间片对应的其它移动设备发送模型训练指示；若满足，则将所述识别模型作为目标识别模型。
  6. 一种巡检方法，其特征在于，应用于移动设备，所述方法包括：
    接收模型管理设备发送的模型训练指示；
    根据所述模型训练指示，按照对应的巡检路线在机房中行进时采集训练数据，基于本 地模型和所述训练数据训练得到所述移动设备对应的中间模型；
    将所述中间模型的模型参数上报给所述模型管理设备；所述各个移动设备上报的中间模型的模型参数用于所述模型管理设备训练得到识别模型，所述识别模型用于确定所述机房的健康状态。
  7. 根据权利要求6所述的方法，其特征在于，所述根据所述模型训练指示，按照巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到所述移动设备对应的中间模型，将所述中间模型的模型参数上报给所述模型管理设备，包括：
    根据所述模型训练指示，循环执行：按照巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到所述移动设备对应的中间模型，将所述中间模型的模型参数上报给所述模型管理设备；
    其中，在每次循环中，所述将所述中间模型的模型参数上报给所述模型管理设备之后，还包括：
    接收所述模型管理设备发送的综合模型参数；
    使用所述综合模型参数更新所述本地模型。
  8. 根据权利要求7所述的方法，其特征在于，所述使用所述综合模型参数更新所述本地模型之后，还包括：
    按照所述巡检路线在所述机房中行进时采集待识别数据，使用更新后的本地模型对所述待识别数据进行识别，以确定所述机房的健康状态。
  9. 一种巡检装置，其特征在于，所述装置包括：
    收发模块，用于向各个移动设备发送模型训练指示；所述模型训练指示用于所述各个移动设备按照各自的巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到中间模型；以及，接收所述各个移动设备上报的中间模型的模型参数；
    训练模块，用于基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型，所述识别模型用于确定所述机房的健康状态。
  10. 根据权利要求9所述的装置，其特征在于，所述识别模型为当前时间片对应的识别模型；
    所述收发模块向各个移动设备发送模型训练指示之前，所述训练模块还用于：
    从机房的全部移动设备中选取出当前时间片对应的移动设备，作为所述各个移动设备。
  11. 根据权利要求10所述的装置，其特征在于，所述模型训练指示用于所述各个移动设备循环按照各自的巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到中间模型并上报；
    所述训练模块具体用于：每接收到设定数量的模型参数，则至少基于所述设定数量的模型参数构建得到综合模型参数；
    所述收发模块具体用于：将所述综合模型参数下发给所述各个移动设备；所述综合模型参数用于所述各个移动设备更新本地模型；以及，当某一综合模型参数对应的模型满足当前时间片的结束条件时，向所述各个移动设备发送模型训练结束指令，所述综合模型参数对应的模型即为所述识别模型。
  12. 根据权利要求10或11所述的装置，其特征在于，所述装置为服务器；
    所述训练模块基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还用于：
    判断所述识别模型是否满足所述模型训练的结束条件，若不满足，则从所述机房的全部移动设备中选取下一时间片对应的移动设备，向所述下一时间片对应的移动设备发送模型训练指示；若满足，则将所述识别模型作为目标识别模型。
  13. 根据权利要求10或11所述的装置，其特征在于，所述装置为移动设备；
    所述训练模块基于所述各个移动设备上报的中间模型的模型参数训练得到识别模型之后，还用于：
    判断所述识别模型是否满足所述模型训练的结束条件，若不满足，则通过与其他移动设备的通信选取某一移动设备作为下一时间片对应的模型管理设备，向所述下一时间片对应的模型管理设备发送模型启动指示；所述模型启动指示用于所述下一时间片对应的模型管理设备从其它移动设备中选取下一时间片对应的其它移动设备，向所述下一时间片对应的其它移动设备发送模型训练指示；若满足，则将所述识别模型作为目标识别模型。
  14. 一种巡检装置，其特征在于，所述装置包括：
    收发模块，用于接收模型管理设备发送的模型训练指示；
    训练模块，用于根据所述模型训练指示，按照对应的巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到所述移动设备对应的中间模型；
    所述收发模块，还用于将所述中间模型的模型参数上报给所述模型管理设备；所述各个移动设备上报的中间模型的模型参数用于所述模型管理设备训练得到识别模型，所述识别模型用于确定所述机房的健康状态。
  15. 根据权利要求14所述的装置，其特征在于，
    所述训练模块具体用于：根据所述模型训练指示，循环执行：按照巡检路线在机房中行进时采集训练数据，基于本地模型和所述训练数据训练得到所述移动设备对应的中间模型；
    所述收发模块具体用于：根据所述模型训练指示，循环执行：将所述中间模型的模型参数上报给所述模型管理设备；
    其中，在每次循环中，所述收发模块将所述中间模型的模型参数上报给所述模型管理设备之后，还用于：接收所述模型管理设备发送的综合模型参数；
    所述训练模块还用于：使用所述综合模型参数更新所述本地模型。
  16. 根据权利要求15所述的装置，其特征在于，所述装置还包括识别模块，所述训练模块使用所述综合模型参数更新所述本地模型之后，所述识别模块用于：
    按照所述巡检路线在所述机房中行进时采集待识别数据，使用更新后的本地模型对所述待识别数据进行识别，以确定所述机房的健康状态。
  17. 一种计算设备，其特征在于，包括至少一个处理器以及至少一个存储器，其中，所述存储器存储有计算机程序，当所述程序被所述处理器执行时，使得所述处理器执行权利要求1～8任一权利要求所述的方法。
  18. 一种计算机可读存储介质，其特征在于，其存储有可由计算设备执行的计算机程序，当所述程序在所述计算设备上运行时，使得所述计算设备执行权利要求1～8任一权利要求所述的方法。
PCT/CN2020/137411 2020-02-20 2020-12-17 一种巡检方法及装置 WO2021164404A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010103868.X 2020-02-20
CN202010103868.XA CN111352799A (zh) 2020-02-20 2020-02-20 一种巡检方法及装置

Publications (1)

Publication Number Publication Date
WO2021164404A1 true WO2021164404A1 (zh) 2021-08-26

Family

ID=71192396

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/137411 WO2021164404A1 (zh) 2020-02-20 2020-12-17 一种巡检方法及装置

Country Status (3)

Country Link
CN (1) CN111352799A (zh)
TW (1) TWI770749B (zh)
WO (1) WO2021164404A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114434456A (zh) * 2022-01-19 2022-05-06 山东新一代信息产业技术研究院有限公司 一种机房巡检机器人及其巡检方法
CN116841301A (zh) * 2023-09-01 2023-10-03 杭州义益钛迪信息技术有限公司 巡检机器人巡检模型训练方法、装置、设备及介质
CN118506026A (zh) * 2024-07-16 2024-08-16 杭州海康威视数字技术股份有限公司 一种巡检设备以及巡检方法

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111352799A (zh) * 2020-02-20 2020-06-30 中国银联股份有限公司 一种巡检方法及装置
CN116887290A (zh) * 2022-03-28 2023-10-13 华为技术有限公司 一种用于机器学习模型训练的通信方法及装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109711556A (zh) * 2018-12-24 2019-05-03 中国南方电网有限责任公司 机巡数据处理方法、装置、网级服务器和省级服务器
CN110298185A (zh) * 2019-06-28 2019-10-01 北京金山安全软件有限公司 一种模型训练方法、装置、电子设备及存储介质
CN110334633A (zh) * 2019-06-27 2019-10-15 北京御航智能科技有限公司 识别巡检数据并更新识别模型的方法、装置及存储介质
CN110442457A (zh) * 2019-08-12 2019-11-12 北京大学深圳研究生院 基于联邦学习的模型训练方法、装置及服务器
CN110766169A (zh) * 2019-10-31 2020-02-07 深圳前海微众银行股份有限公司 强化学习的迁移训练优化方法、装置、终端及存储介质
CN111352799A (zh) * 2020-02-20 2020-06-30 中国银联股份有限公司 一种巡检方法及装置

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180089587A1 (en) * 2016-09-26 2018-03-29 Google Inc. Systems and Methods for Communication Efficient Distributed Mean Estimation
TWI587152B (zh) * 2016-10-26 2017-06-11 先智雲端數據股份有限公司 用於延長雲端服務系統中磁碟預期壽命值的方法及使用該方法的系統
CN110309847B (zh) * 2019-04-26 2021-09-17 深圳前海微众银行股份有限公司 一种模型压缩方法及装置
CN110490738A (zh) * 2019-08-06 2019-11-22 深圳前海微众银行股份有限公司 一种混合联邦学习方法及架构

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109711556A (zh) * 2018-12-24 2019-05-03 中国南方电网有限责任公司 机巡数据处理方法、装置、网级服务器和省级服务器
CN110334633A (zh) * 2019-06-27 2019-10-15 北京御航智能科技有限公司 识别巡检数据并更新识别模型的方法、装置及存储介质
CN110298185A (zh) * 2019-06-28 2019-10-01 北京金山安全软件有限公司 一种模型训练方法、装置、电子设备及存储介质
CN110442457A (zh) * 2019-08-12 2019-11-12 北京大学深圳研究生院 基于联邦学习的模型训练方法、装置及服务器
CN110766169A (zh) * 2019-10-31 2020-02-07 深圳前海微众银行股份有限公司 强化学习的迁移训练优化方法、装置、终端及存储介质
CN111352799A (zh) * 2020-02-20 2020-06-30 中国银联股份有限公司 一种巡检方法及装置

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
NADIGER CHETAN; KUMAR ANIL; ABDELHAK SHERINE: "Federated Reinforcement Learning for Fast Personalization", 2019 IEEE SECOND INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE AND KNOWLEDGE ENGINEERING (AIKE), IEEE, 3 June 2019 (2019-06-03), pages 123 - 127, XP033592870, DOI: 10.1109/AIKE.2019.00031 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114434456A (zh) * 2022-01-19 2022-05-06 山东新一代信息产业技术研究院有限公司 一种机房巡检机器人及其巡检方法
CN116841301A (zh) * 2023-09-01 2023-10-03 杭州义益钛迪信息技术有限公司 巡检机器人巡检模型训练方法、装置、设备及介质
CN116841301B (zh) * 2023-09-01 2024-01-09 杭州义益钛迪信息技术有限公司 巡检机器人巡检模型训练方法、装置、设备及介质
CN118506026A (zh) * 2024-07-16 2024-08-16 杭州海康威视数字技术股份有限公司 一种巡检设备以及巡检方法

Also Published As

Publication number Publication date
TWI770749B (zh) 2022-07-11
CN111352799A (zh) 2020-06-30
TW202132989A (zh) 2021-09-01

Similar Documents

Publication Publication Date Title
WO2021164404A1 (zh) 一种巡检方法及装置
TWI547817B (zh) 叢集運算架構的資源規劃方法、系統及裝置
CN105446860B (zh) 基于异步并发机制的压力测试系统和测试方法
CN112929187B (zh) 网络切片管理方法、装置和系统
CN107925612A (zh) 网络监视系统、网络监视方法和程序
WO2017084016A1 (zh) 模型参数融合方法及装置
EP3818446A1 (en) Methods and systems for dynamic service performance prediction using transfer learning
CN110784929B (zh) 一种接入资源分配方法、装置、设备及系统
WO2023098374A1 (zh) 网络资源部署方法、装置、电子设备及存储介质
CN110532021A (zh) 分布式控制系统的组态文件的处理方法及装置
CN111753875A (zh) 一种电力信息系统运行趋势分析方法、装置及存储介质
CN111966289A (zh) 基于Kafka集群的分区优化方法和系统
CN113692021A (zh) 一种基于亲密度的5g网络切片智能资源分配方法
CN107943697A (zh) 问题分配方法、装置、系统、服务器和计算机存储介质
CN107004003B (zh) 模型参数融合方法及装置
WO2024066346A1 (zh) 一种告警处理方法、装置、存储介质及电子装置
CN111865681A (zh) 核心网切片端到端的时延优化方法、系统与存储介质
WO2024066720A1 (zh) 指标阈值的确定方法、装置、存储介质及电子装置
CN113114480A (zh) 一种数据的上报方法及相关设备
CN112672298A (zh) 一种基于分布式边缘网络结构的农业数据采集系统及其方法
JP5768722B2 (ja) 検証試験のためのプログラム、情報処理方法及び情報処理装置
CN107479977B (zh) 一种确定设备性能的方法和设备
CN108989116A (zh) 大数据收集分析系统及方法
CN115037625A (zh) 网络切片处理方法、装置、电子设备及可读存储介质
CN115460700A (zh) 基于联邦学习的网络资源配置方法、装置、电子设备及介质

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: 20919824

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: 20919824

Country of ref document: EP

Kind code of ref document: A1