Distributed device cooperative predictive maintenance method based on federated learning

Abstract

The application discloses a distributed device cooperative prediction maintenance method based on federal learning, and relates to the technical field of prediction maintenance.The application comprehensively acquires chip operation state data, provides multi-dimensional feature support for accurate diagnosis, establishes a chip-level multi-dimensional evaluation system, realizes more refined chip health state evaluation, improves the comprehensiveness of fault prediction, innovates an aging trend prediction algorithm, combines historical data and a theoretical model to calculate a power supply deviation coefficient, realizes accurate evaluation and early abnormal early warning capability of a chip power supply system, constructs an intelligent temperature control evaluation model, accurately quantifies chip overheating risk, improves abnormal detection sensitivity of a heat dissipation system, identifies mutation points by comparing actual and simulated time sequence images, effectively captures signal integrity abnormalities, improves digital circuit fault diagnosis capability, and comprehensively uses quantity deviation and feature index deviation dual parameters to realize intelligent grading early warning of chip fault risk.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance technology, and more specifically to a distributed device collaborative predictive maintenance method based on federated learning. Background Technology

[0002] As the core infrastructure of modern data centers, the reliability of network switching equipment directly affects the stability of global Internet services. The internal chips of the switch face multiple challenges such as power supply stability, thermal management, and signal integrity. The lack of sharing of private data between companies can easily lead to data silos, making it impossible to accurately predict failures. Federated learning technology can establish data connections with other companies without leaving the local area. Therefore, it is necessary to use federated learning technology for predictive maintenance of switches from various companies.

[0003] Existing technology, such as the invention patent application with publication number CN120124861A, discloses a lifecycle management system for electromechanical equipment based on digital twin technology. This system includes: a physical layer monitoring self-powered sensors; a data acquisition layer using improved evaluation formulas; a model layer using adaptive calibration formulas; an analysis and decision-making layer using improved prediction and cost formulas; and an interaction layer using comprehensive feedback formulas to evaluate user experience. This invention enables accurate fault prediction and scientific maintenance; facilitates resource recovery during decommissioning; provides a superior user experience; and promotes good supply chain collaboration, comprehensively improving the level and efficiency of lifecycle management for electromechanical equipment.

[0004] Existing technologies, such as the AI-based predictive maintenance system and method for distributed energy storage devices disclosed in patent application CN119919125B, include: a system that achieves precise monitoring of the operating status of distributed energy storage devices by integrating comprehensive collection of multi-dimensional operational data with in-depth analysis of AI predictive models. This generates targeted predictive maintenance instructions, significantly improving the timeliness and accuracy of equipment maintenance, extending equipment lifespan, and optimizing the efficiency of maintenance resource allocation.

[0005] As can be seen from the above solutions, the current distributed device collaborative predictive maintenance methods (1) mostly adopt a centralized data collection and analysis architecture, which requires uploading the original operating data of the device to the cloud for processing, which poses a risk of data privacy leakage and cannot meet the real-time requirements.

[0006] (2) Current mainstream technologies mainly focus on single-dimensional monitoring indicators, such as analyzing power supply parameters or temperature data, which makes it difficult to fully reflect the actual operating status of the chip.

[0007] (3) Existing solutions typically use a generic evaluation model, which fails to fully consider the differences between equipment from different manufacturers.

[0008] (4) Existing technologies do not adequately consider the impact of environmental factors, especially cosmic rays, and lack effective compensation mechanisms.

[0009] (5) Traditional methods can usually only locate faults at the board level and cannot accurately identify problems with specific chips. Summary of the Invention

[0010] The purpose of this invention is to provide a distributed device collaborative prediction and maintenance method based on federated learning, which solves the problems existing in the background technology.

[0011] To solve the above technical problems, the present invention adopts the following technical solution: The present invention provides a distributed device collaborative prediction and maintenance method based on federated learning, including: Step 1. At the target monitoring time point, start the switch data acquisition function of each company to obtain the feature values ​​of each privacy data of each switch belonging to each company.

[0012] Step 2. Based on the feature values ​​of each privacy data of each switch belonging to each company, analyze the privacy feature index of each chip type of each company's switch.

[0013] Step 3. Send the privacy feature indices of each company's switches to the cloud database, and build a joint data model of the switches to obtain the parameter tuning deviation ratio of each privacy feature index of each chip type of each company's switches.

[0014] Step 4. Based on the parameter tuning deviation ratio of each privacy feature index of each chip type in each company's switches, screen each estimated faulty chip in each switch belonging to each company.

[0015] Step 5. Send the estimated faulty chips of each switch belonging to each company to the switch management person in charge.

[0016] The beneficial effects of this invention are as follows: By comprehensively acquiring chip operating status data, this invention provides multi-dimensional feature support for accurate diagnosis, establishes a chip-level multi-dimensional evaluation system, achieves more refined chip health status assessment, improves the comprehensiveness of fault prediction, innovates aging trend prediction algorithms, calculates power supply deviation coefficients by combining historical data and theoretical models, realizes accurate assessment and early anomaly warning capabilities for chip power supply systems, constructs an intelligent temperature control evaluation model, accurately quantifies chip overheating risks, improves the sensitivity of heat dissipation system anomaly detection, identifies mutation points by comparing actual and simulated time-series images, effectively captures signal integrity anomalies, improves digital circuit fault diagnosis capabilities, enhances the rationality of cross-regional equipment status assessment through a geographic feature-based association matching mechanism, establishes a dynamic fault assessment system, integrates both quantitative deviation and characteristic index deviation parameters to achieve intelligent graded early warning of chip fault risks, and provides quantitative basis for equipment configuration rationality assessment by comparing actual deployment volume with industry benchmark values. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 As shown, the present invention provides a distributed device collaborative prediction and maintenance method based on federated learning, including: Step 1. At the target monitoring time point, start the switch data acquisition function of each company to acquire the feature values ​​of each privacy data of each switch belonging to each company.

[0021] In one specific embodiment, the method for obtaining the power supply voltage and current values ​​of each chip in each switch belonging to each company at each monitoring time point is as follows: the power supply voltage and current values ​​of each chip in each switch belonging to each company at each monitoring time point are obtained through current sensors, voltage sensors and PMIC registers installed on the switch motherboard.

[0022] Thermodynamic images of each chip in each switch belonging to each company were acquired at each monitoring time point using infrared cameras.

[0023] In one specific embodiment, the method for obtaining the thermodynamic images of each chip of each switch belonging to each company at each monitoring time point is as follows: an infrared sensor is located on the slide rail of the switch cabinet, and the outer surface of the location of each chip of each switch is scanned at the target monitoring time point to obtain the thermodynamic images of each chip of each switch belonging to each company at each monitoring time point.

[0024] The timing images of each target pin of each chip of each switch belonging to each company are obtained by the probe during the monitoring period.

[0025] In one specific embodiment, the method for acquiring the timing images of each target pin of each chip of each switch belonging to each company during the monitoring period is as follows: the probe is installed on the slide rail of the switch cabinet, and reserved points are reserved on the outer shell of the switch for the oscilloscope probe to measure each target pin of each chip. At the target monitoring time point, the probe uses the slide rail and AI to automatically locate and draw the timing diagram of the reserved points of the switch one by one, thereby acquiring the timing images of each target pin of each chip of each switch belonging to each company during the monitoring period.

[0026] It should be noted that, for NAND Flash chips, the target pins are CLE, ALE, RE, WE, etc.

[0027] The timing images of each target pin of each chip of each switch belonging to each company during the monitoring period, as well as the power supply voltage, power supply current, and thermodynamic images of each chip of each switch belonging to each company at each monitoring time point, are used as the feature values ​​of each privacy data.

[0028] Step 2. Based on the feature values ​​of each privacy data of each switch belonging to each company, analyze the privacy feature index of each chip type of each company's switch.

[0029] In a specific embodiment of the present invention, the analysis method for the privacy feature index of each chip type of each company's switch is as follows: based on the power supply voltage and power supply current values ​​of each chip of each company's switch at each monitoring time point, a power supply evaluation model is constructed to obtain the power supply feature index of each chip type of each company's switch.

[0030] It should be noted that the various chip types mentioned include NAND Flash, EEPROM, and DRAM.

[0031] Based on the thermodynamic images of each chip in each switch of each company at each monitoring time point, a chip temperature control evaluation model is constructed to obtain the temperature control characteristic index of each chip type in the switches of each company.

[0032] Based on the timing images of each target pin of each chip in each switch of each company during the monitoring period, a chip timing evaluation model is constructed to obtain the timing characteristic index of each chip type in each company's switch.

[0033] The chip power supply characteristic index, chip temperature control characteristic index, and chip timing characteristic index of each company's switch chip type are used as privacy characteristic indices.

[0034] In a specific embodiment of the present invention, the method for constructing a power supply evaluation model to obtain the power supply characteristic index of each chip type of each company's switch is as follows: obtain the usage time of each chip of each company's switch and the historical power supply voltage and historical power supply current values ​​at each historical monitoring time point from the local database, perform aging trend prediction, and calculate the power supply deviation coefficient of each chip of each company's switch accordingly.

[0035] It should be noted that the local database is used to store the usage time of each chip in each switch belonging to each company, the historical power supply voltage and current values ​​at each historical monitoring time point, the theoretical voltage fluctuation value and theoretical current fluctuation value of each chip in each usage time interval, the standard power supply voltage value and standard power supply current value of each chip, the chip type of each chip in each switch belonging to each company, the suitable operating temperature of each chip at each chip location in each usage time interval, the binary values ​​of the transmit and receive data of each target pin of each chip in each switch belonging to each company, the range of cosmic radiation intensity values, the suitable increase ratio value for fault assessment corresponding to each range of quantity deviation parameters, and the fault assessment coefficient threshold.

[0036] In one specific embodiment, the method for predicting aging trends and calculating the power supply deviation coefficient of each chip in each switch belonging to each company is as follows: The theoretical voltage fluctuation value and theoretical current fluctuation value of each chip in each usage time interval are obtained from a local database; based on the usage time of each chip in each switch belonging to each company, the theoretical voltage fluctuation value 'a' of each chip in each switch belonging to each company is mapped to the actual usage time. xni and theoretical current fluctuation value b xni Here, x represents the company number, x = 1, 2, ..., y, where y is a positive integer greater than 2; n represents the switch number, n = 1, 2, ..., m, where m is a positive integer greater than 2; and i represents the chip number, i = 1, 2, ..., j, where j is a positive integer greater than 2. The standard power supply voltage value c of each chip is obtained from the local database. i and standard supply current value d i Based on the historical power supply voltage values ​​f of each chip in each switch belonging to each company at each historical monitoring time point. xnip and historical power supply current value g xnip Where p represents the number of each historical monitoring time point, p = 1, 2, ..., q, and q is a positive integer greater than 2. Calculate the allowable voltage fluctuation value of each chip in each switch belonging to each company. and allowable fluctuation current value Where p represents the number of historical monitoring time points.

[0037] Based on the power supply voltage h of each chip in each switch belonging to each company at each monitoring time point. xniand supply current value k xni Calculate the power supply deviation coefficient of each chip in each switch of each company.

[0038] The chip types of each chip in each switch belonging to each company are obtained from the local database. Based on the power supply deviation coefficient of each chip in each switch belonging to each company, the power supply characteristic index of each chip type in each company's switch is evaluated.

[0039] In one specific embodiment, the method for evaluating the power supply characteristic index of each chip type of each company's switches is as follows: based on the chip type of each chip of each company's switches, and based on the power supply deviation coefficient of each chip of each company's switches, the power supply deviation coefficient of each chip type of each company's switches is mapped to obtain the power supply deviation coefficient of each chip of each company's switches. Through a statistical averaging algorithm, the average value of the power supply deviation coefficient of each chip type of each company's switches is obtained, and this average value is used as the power supply characteristic index of each chip type of each company's switches.

[0040] In a specific embodiment of the present invention, the method for constructing a chip temperature control evaluation model to obtain the temperature control characteristic index of each chip type of each company's switch is as follows: based on the thermodynamic images of each chip of each company's switch at each monitoring time point, and based on the usage time of each chip of each company's switch, the overheating threat coefficient of each chip of each company's switch is calculated.

[0041] In one specific embodiment, the method for calculating the overheating threat coefficient of each chip in each switch belonging to each company is as follows: obtain the suitable operating temperature of each chip at each chip location in each usage time interval from the local database, and map the suitable operating temperature of each chip location of each chip in each switch belonging to each company to the usage time of each chip in each switch belonging to each company.

[0042] Based on the thermodynamic images of each chip in each switch belonging to each company at each monitoring time point, AI is used to identify the chip locations and obtain the actual operating temperature of each chip location at each monitoring time point for each chip in each switch belonging to each company. xnirt Where r represents the number of each monitoring time point, r = 1, 2, ..., s, where s is a positive integer greater than 2, and t represents the number of each chip location, t = 1, 2, ..., w, where w is a positive integer greater than 2, based on the suitable operating temperature u of each chip location of each chip in each switch belonging to each company. xnit Calculate the overheating threat coefficient of each chip in each switch belonging to each company. Where s is the number of monitoring time points and w is the number of chip locations.

[0043] Based on the chip type of each chip in each switch of each company, and based on the overheat threat coefficient of each chip in each switch of each company, the temperature control characteristic index of each chip type in each company's switches is evaluated.

[0044] In one specific embodiment, the method for evaluating the temperature control characteristic index of each chip type of each company's switch is as follows: based on the method for evaluating the power supply characteristic index of each chip type of each company's switch, the temperature control characteristic index of each chip type of each company's switch is evaluated in the same way.

[0045] In a specific embodiment of the present invention, the method for constructing a chip timing evaluation model to obtain the timing characteristic index of each chip type of each company's switch is as follows: obtain the binary values ​​of the transmit and receive data of each target pin of each chip of each company's switch from the local database, and extract the timing mutation time points of each target pin of each chip of each company's switch within the monitoring time period based on the timing image of each target pin of each chip of each company's switch, and calculate the timing mutation hazard coefficient of each chip of each company's switch.

[0046] In one specific embodiment, the method for extracting the timing mutation time points of each target pin of each chip of each switch belonging to each company and calculating the timing mutation hazard coefficient of each chip of each switch belonging to each company is as follows: Based on the binary values ​​of the transmit and receive data of each target pin of each chip of each switch belonging to each company, a timing diagram simulation is performed to obtain the simulated timing image of each target pin of each chip of each switch belonging to each company during the monitoring period. This simulated timing image is then compared with the timing images of each target pin of each chip of each switch belonging to each company during the monitoring period. Inconsistent time points in the timing images are extracted and used as the timing mutation time points of each target pin of each chip of each switch belonging to each company. The total number A of timing mutation time points of each target pin of each chip of each switch belonging to each company is then counted. xniv Where v represents the number of each target pin, v = 1, 2, ..., v′, where v′ is a positive integer greater than 2, and the total number B of timing points of each target pin of each chip of each switch belonging to each company is counted. xniv Calculate the timing mutation hazard factor of each chip in each switch of each company. Where e represents the natural constant.

[0047] Based on the chip type of each chip in each switch of each company, and based on the timing mutation hazard coefficient of each chip in each switch of each company, the timing characteristic index of each chip type in each company's switches is evaluated.

[0048] In one specific embodiment, the method for evaluating the timing characteristic index of each chip type of each company's switch is as follows: based on the method for evaluating the power supply characteristic index of each chip type of each company's switch, the timing characteristic index of each chip type of each company's switch is evaluated in the same way.

[0049] Step 3. Send the privacy feature indices of each company's switches to the cloud database, and build a joint data model of the switches to obtain the parameter tuning deviation ratio of each privacy feature index of each chip type of each company's switches.

[0050] In a specific embodiment of the present invention, the method for activating the data acquisition function of each company's switch to acquire the feature values ​​of each privacy data of each switch belonging to each company is as follows: the power supply voltage and power supply current values ​​of each chip of each switch belonging to each company are acquired by sensors at each monitoring time point.

[0051] In a specific embodiment of the present invention, the method for constructing a data federation model of the switch to obtain the parameter deviation ratio of each privacy feature index of each chip type of the switch of each company is as follows: obtain the publicly available geographical altitude and publicly available latitude and longitude of each company from the cloud database, and perform correlation matching of each company accordingly to obtain each company in each relevant correlation division.

[0052] Based on the privacy feature indices of each chip type of the switches from each company, and based on each company in each relevant matching division, the privacy feature indices of each chip type of the switches from each relevant matching division are mapped to each company, and the parameter tuning deviation ratio of each privacy feature index of each chip type of the switches from each company is analyzed accordingly.

[0053] In one specific embodiment, the method for analyzing the parameter tuning deviation ratio of each privacy feature index of each chip type of switches from various companies is as follows: based on the values ​​D of each privacy feature index of each chip type of switches from various companies, as determined by each relevant matching division. NntTP Let N represent the number of each relevant matching partition, N = 1, 2, ..., M, where M is a positive integer greater than 2; T represent the number of each privacy feature index, T = 1, 2, ..., W, where W is a positive integer greater than 2; and P represent the number of each privacy feature index, P = 1, 2, ..., Q, where Q is a positive integer greater than 2. Calculate the variance of each privacy feature index for each chip type of each company's switches in each relevant matching partition. Where Q represents the number of values, and based on this, the parameter tuning deviation ratio of each privacy feature index for each chip type of each company's switch is calculated.

[0054] In a specific embodiment of the present invention, the method for performing correlation matching of companies to obtain companies with relevant correlations is as follows: obtain the cosmic radiation intensity values ​​at various altitude ranges and latitude and longitude ranges at the target monitoring time point from the cosmic ray monitoring platform, and map the cosmic radiation intensity values ​​of each company at the target monitoring time point based on the publicly disclosed geographical altitude and publicly disclosed regional latitude and longitude of each company.

[0055] The range of cosmic radiation intensity values ​​is obtained from the local database. Based on the cosmic radiation intensity values ​​of each company at the target monitoring time point, the companies corresponding to each cosmic radiation intensity value range are obtained and used as the companies in each relevant association division.

[0056] Step 4. Based on the parameter tuning deviation ratio of each privacy feature index of each chip type in each company's switches, screen each estimated faulty chip in each switch belonging to each company.

[0057] In a specific embodiment of the present invention, the method for screening the estimated faulty chips of each company's switches is as follows: based on the chip type of each chip in each company's switches, calculate the quantity deviation parameter of each chip type in each company's switches.

[0058] Based on the quantity deviation parameters of each chip type in the switches of each company and the parameter adjustment deviation ratio of each privacy feature index, the fault assessment coefficient of each chip in each switch of each company is calculated.

[0059] In one specific embodiment, the calculation method for the fault assessment coefficient of each chip in each switch belonging to each company is as follows: based on the chip type of each chip in each switch belonging to each company, the quantity deviation parameters of each chip type in the switch, and the parameter adjustment deviation ratio of each privacy feature index, the parameter adjustment deviation ratio J of the quantity deviation parameters and power supply deviation coefficient of each chip in each switch belonging to each company is mapped to obtain the parameter adjustment deviation ratio J of the power supply deviation coefficient of each chip in each switch belonging to each company. xni The overheating threat factor adjustment deviation ratio L xni The parameter tuning deviation ratio K of the time-series sudden change hazard coefficient xni .

[0060] The appropriate increase ratio for fault assessment corresponding to each quantity deviation parameter range is obtained from the local database. Based on the quantity deviation parameters of each chip in each switch of each company, the appropriate increase ratio F for fault assessment of each chip in each switch of each company is mapped to the appropriate increase ratio F. xni Based on the power supply deviation coefficient ε of each chip in each switch belonging to each company xni Overheating threat factor and the risk factor λ of temporal mutation xni Calculate the fault assessment coefficients for each chip in each switch belonging to each company.

[0061] It should be noted that the larger the quantity deviation parameter, the larger the appropriate increase ratio for fault assessment. When the quantity deviation parameter is large, it indicates that the amount of data is small. In this case, the fault assessment coefficient should be appropriately increased to reduce the number of cases where faults cannot be detected. For example, when the quantity deviation parameter is 0.2, the appropriate increase ratio for fault assessment is 1.02, and when the quantity deviation parameter is 0.5, the appropriate increase ratio for fault assessment is 1.05.

[0062] The fault assessment coefficient threshold is obtained from the local database. Based on the fault assessment coefficient of each chip in each switch of each company, the estimated fault chips of each switch of each company are screened.

[0063] In one specific embodiment, the screening method for each estimated fault chip of each switch belonging to each company is as follows: if the fault assessment coefficient of a certain chip of a certain switch belonging to a certain company is greater than the fault assessment coefficient threshold, then the chip is marked as an estimated fault chip, thereby screening each estimated fault chip of each switch belonging to each company.

[0064] In a specific embodiment of the present invention, the method for calculating the quantity deviation parameter of each chip type of each company's switches is as follows: based on the chip type of each chip in each switch of each company, the quantity of each chip type of each chip in each switch of each company is counted.

[0065] The chip comparison count for each chip type is obtained from the cloud database. Based on the chip count for each chip type in each switch of each company, the quantity deviation parameter for each chip type in each company's switches is calculated.

[0066] It should be noted that the number of chips to be compared for each chip type is set by the researchers and can be set according to the shipment volume of each chip type. For example, if the shipment volume of a certain chip type is 10,000, then its chip comparison number is 100.

[0067] In one specific embodiment, the method for calculating the quantity deviation parameter of each chip type of each company's switches based on the chip quantity of each chip type of each company's switches is as follows: based on the chip comparison quantity G of each chip type... t Based on the number of chips of each chip type in each switch of each company, calculate the total number H of each chip type in the switches of each company. xt Calculate the quantity deviation parameters of each chip type in the switches of each company.

[0068] Step 5. Send the estimated faulty chips of each switch belonging to each company to the switch management person in charge.

[0069] This invention provides multi-dimensional feature support for accurate diagnosis by comprehensively acquiring chip operating status data, establishing a chip-level multi-dimensional evaluation system to achieve more refined chip health status assessment, improve the comprehensiveness of fault prediction, innovate aging trend prediction algorithm, calculate power supply deviation coefficient by combining historical data and theoretical models to achieve accurate assessment and early anomaly warning capability of chip power supply system, construct intelligent temperature control evaluation model to accurately quantify chip overheating risk, improve the sensitivity of heat dissipation system anomaly detection, identify mutation points by comparing actual and simulated time series images to effectively capture signal integrity anomalies and improve digital circuit fault diagnosis capability, improve the rationality of cross-regional equipment status assessment based on geographical feature-based association matching mechanism, establish a dynamic fault assessment system, comprehensively consider both quantitative deviation and characteristic index deviation parameters to achieve intelligent graded early warning of chip fault risk, and provide quantitative basis for equipment configuration rationality assessment by comparing actual deployment volume with industry benchmark values.

[0070] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A distributed device collaborative predictive maintenance method based on federated learning, characterized in that, include: Step 1. At the target monitoring time point, activate the data acquisition function of each company's switch to obtain the feature values ​​of each privacy data of each switch belonging to each company; Step 2. Based on the feature values ​​of each privacy data of each switch belonging to each company, analyze the privacy feature index of each chip type of each company's switch; The specific analysis method for the privacy feature indices of various chip types in the switches of various companies is as follows: Based on the power supply voltage and current values ​​of each chip in each switch of each company at each monitoring time point, a power supply evaluation model is constructed to obtain the power supply characteristic index of each chip type in the switches of each company. Based on the thermodynamic images of each chip in each switch of each company at each monitoring time point, a chip temperature control evaluation model is constructed to obtain the temperature control characteristic index of each chip type in the switches of each company. Based on the timing images of each target pin of each chip in each switch of each company during the monitoring period, a chip timing evaluation model is constructed to obtain the timing characteristic index of each chip type in each company's switch. The chip power supply characteristic index, chip temperature control characteristic index, and chip timing characteristic index of each chip type in the switches of each company are used as privacy characteristic indices. The method for constructing the chip timing evaluation model to obtain the timing characteristic index of each chip type in the switches of various companies is as follows: Obtain the binary values ​​of the transmit and receive data of each target pin of each chip of each switch belonging to each company from the local database, and extract the timing change time points of each target pin of each chip of each switch belonging to each company from the timing images of each target pin of each chip of each switch belonging to each company during the monitoring period, and calculate the timing change hazard coefficient of each chip of each switch belonging to each company. Based on the chip type of each chip in each switch of each company, and based on the timing mutation hazard coefficient of each chip in each switch of each company, the timing characteristic index of each chip type in each company's switches is evaluated. Step 3. Send the privacy feature indices of each company's switches to the cloud database, and build a data joint model of the switches to obtain the parameter tuning deviation ratio of each privacy feature index of each chip type of each company's switches. The method for constructing a joint data model for the switches and obtaining the parameter tuning deviation ratios of various privacy feature indices for various chip types of switches from different companies is as follows: The publicly available geographical altitude and latitude and longitude of each company are obtained from the cloud database, and the correlation between the companies is matched accordingly to obtain the companies in each relevant correlation classification. Based on the privacy feature indices of each chip type of each company's switches, and based on each company in each relevant matching division, the privacy feature indices of each chip type of each company's switches are mapped to each relevant matching division, and the parameter tuning deviation ratio of each privacy feature index of each chip type of each company's switches is analyzed accordingly. The specific method for performing correlation matching among companies to obtain the companies in each relevant correlation classification is as follows: The cosmic radiation intensity values ​​at various altitude and latitude intervals at the target monitoring time point are obtained from the cosmic ray monitoring platform. Based on the publicly disclosed geographical altitude and latitude and longitude of each company, the cosmic radiation intensity values ​​of each company at the target monitoring time point are mapped. The range of cosmic radiation intensity values ​​is obtained from the local database. Based on the cosmic radiation intensity values ​​of each company at the target monitoring time point, the companies corresponding to each range of cosmic radiation intensity values ​​are obtained and used as the companies in each relevant association division. Step 4. Based on the parameter tuning deviation ratio of each privacy feature index of each chip type in each company's switches, screen each estimated faulty chip of each company's switches. Step 5. Send the estimated faulty chips of each switch belonging to each company to the switch management person in charge.

2. The distributed device collaborative prediction and maintenance method based on federated learning according to claim 1, characterized in that, The method for activating the data acquisition function of each company's switches to obtain the feature values ​​of each privacy data of each company's switches is as follows: The power supply voltage and current values ​​of each chip in each switch belonging to each company are obtained through sensors at each monitoring time point. Thermodynamic images of each chip in each switch belonging to each company at each monitoring time point are obtained using infrared cameras. The timing images of each target pin of each chip of each switch belonging to each company are obtained by the probe during the monitoring period. The timing images of each target pin of each chip of each switch belonging to each company during the monitoring period, as well as the power supply voltage, power supply current, and thermodynamic images of each chip of each switch belonging to each company at each monitoring time point, are used as the feature values ​​of each privacy data.

3. The distributed device collaborative prediction and maintenance method based on federated learning according to claim 1, characterized in that, The method for constructing the power supply evaluation model to obtain the power supply characteristic index of each chip type in the switches of various companies is as follows: The system retrieves the usage time of each chip in each switch belonging to each company and the historical power supply voltage and current values ​​at each historical monitoring time point from the local database, performs aging trend prediction, and calculates the power supply deviation coefficient of each chip in each switch belonging to each company based on this. The chip types of each chip in each switch belonging to each company are obtained from the local database. Based on the power supply deviation coefficient of each chip in each switch belonging to each company, the power supply characteristic index of each chip type in each company's switch is evaluated.

4. The distributed device collaborative prediction and maintenance method based on federated learning according to claim 3, characterized in that, The method for constructing a chip temperature control evaluation model to obtain the temperature control characteristic index of each chip type in switches from various companies is as follows: Based on the thermodynamic images of each chip in each switch belonging to each company at each monitoring time point, and based on the usage time of each chip in each switch belonging to each company, the overheating threat coefficient of each chip in each switch belonging to each company is calculated. Based on the chip type of each chip in each switch of each company, and based on the overheat threat coefficient of each chip in each switch of each company, the temperature control characteristic index of each chip type in each company's switches is evaluated.

5. The distributed device collaborative prediction and maintenance method based on federated learning according to claim 3, characterized in that, The specific screening method for each estimated faulty chip in each switch belonging to each company is as follows: Based on the chip type of each chip in each switch of each company, calculate the quantity deviation parameter of each chip type in the switches of each company. Based on the quantity deviation parameters of each chip type in the switches of each company and the parameter adjustment deviation ratio of each privacy feature index, the fault assessment coefficient of each chip in each switch of each company is calculated. The fault assessment coefficient threshold is obtained from the local database. Based on the fault assessment coefficient of each chip in each switch of each company, the estimated fault chips of each switch of each company are screened.

6. The distributed device collaborative prediction and maintenance method based on federated learning according to claim 5, characterized in that, The specific calculation method for the quantity deviation parameter of each chip type in the switches of each company is as follows: Based on the chip type of each chip in each switch of each company, count the number of chips of each chip type in each switch of each company; obtain the chip comparison number of each chip type from the cloud database, and calculate the quantity deviation parameter of each chip type in each switch of each company based on the chip number of each chip type in each switch of each company.

Images