An AI-based pre-training system for digital monitoring and analysis of facilities

Abstract

The present invention discloses a pre-training system for digital monitoring and analysis of facilities based on AI, specifically related to the field of AI, including a monitoring time division module, a data acquisition module, a data analysis and processing module, a sub-time region anomaly calculation module, a target time anomaly calculation module, a target time anomaly discrimination module, a target time warning module, and a target time optimization module. Through refined time region division and real-time collection of various sensor data, the present invention can comprehensively and accurately monitor pre-training changes, ensuring the comprehensiveness and accuracy of monitoring. Through the warning module, potential problems occurring in pre-training can be handled, avoiding the occurrence of system accidents, improving the performance of artificial intelligence, and reducing production costs and losses.

Description

Technical Field

[0001] The present invention relates to the technical field of facility digitization of AI. More specifically, the present invention relates to a pre-training system for facility digital monitoring and analysis based on AI. Background Art

[0002] With the development of the times, artificial intelligence (AI) is an important driving force for the new round of scientific and technological revolution and industrial transformation. It attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a way similar to human intelligence. AI technology has a wide range of applications in language recognition, image recognition, natural language processing, etc. And with the maturity of theory and the progress of technology, its application fields are constantly expanding.

[0003] The existing pre-training system for facility digital monitoring and analysis based on AI can make full use of the advantages of AI technology to conduct all-round monitoring and analysis of facilities. Through data pre-acquisition module, data pre-processing module, pre-training module, and early warning module, etc., the system can automatically identify and predict the operating status of facilities, timely discover potential problems and hidden dangers, and provide strong guarantee for the safe operation of facilities. The system can automatically analyze, identify, and understand the content through computer vision and deep learning algorithms, can detect abnormal situations in pre-training in real time, and immediately issue an alarm.

[0004] However, in actual use, the existing monitoring and analysis pre-training system still has deficiencies in data acquisition, analysis, and processing, and there are still some disadvantages. For example, for a large amount of real-time data, traditional data processing methods cannot process it in time, resulting in data redundancy. When detecting data, errors will occur, which leads to low efficiency and quality of data processing. The construction standards of the existing monitoring and analysis pre-training system are not unified, and the phenomenon of data islands is serious, thus restricting the performance of artificial intelligence. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a pre-training system for facility digital monitoring and analysis based on AI, through the facility digitization of AI, to solve the problems raised in the above background art.

[0006] To achieve the above object, the present invention provides the following technical solution: A pre-training system for facility digital monitoring and analysis based on AI, comprising:

[0007] Monitoring time division module: used to determine the monitoring time of the target pre-training as the target time region, divide the target time region into each sub-time region by equal time division, and sequentially mark them as 1, 2... n;

[0008] Data acquisition module: used to collect pre-trained learning efficiency fluctuation parameters, batch fluctuation parameters, training fluctuation parameters, and optimization fluctuation parameters for each sub-time region;

[0009] Data analysis and processing module: used to perform real-time analysis and processing on the data collected in the data analysis and processing module to obtain learning efficiency fluctuation coefficients, batch fluctuation coefficients, training fluctuation coefficients, and optimization fluctuation coefficients;

[0010] Sub-time region anomaly calculation module: used to import the learning efficiency fluctuation coefficient, batch fluctuation coefficient, training fluctuation coefficient, and optimization fluctuation coefficient into the mathematical model of the sub-time region anomaly index to obtain the sub-time region anomaly value;

[0011] Target time anomaly calculation module: used to import the sub-time region anomaly index into the mathematical model of the target time region anomaly index to obtain the target time anomaly value;

[0012] Target time anomaly discrimination module: used to precisely compare the target time anomaly value with the preset target time anomaly value, calculate the difference value between the target time anomaly value and the preset target time anomaly value, and when the difference value is greater than the preset difference value, transmit the difference value to the target time warning module for feedback;

[0013] Target time warning module: used to transmit the received data to the system, and then the system transmits the data obtained through intelligent processing to the target time optimization module. At the same time, corresponding warning reminders are made according to the received data warning instructions;

[0014] Target time optimization module: used to import the abnormal data in the target time warning module into the optimizer, and obtain the optimized data through fine-tuning processing and integration.

[0015] Preferably, the AI-based facility digital monitoring and analysis pre-training system is characterized in that: the data acquisition module is used to collect pre-trained learning efficiency fluctuation parameters, batch fluctuation parameters, training fluctuation parameters, and optimization fluctuation parameters for each sub-time region.

[0016] Preferably, the learning efficiency fluctuation parameters include the step size of weight update, weight step size, maximum weight step size, change rate of weight, and training speed, which are respectively denoted as S 1 、S 2 、S max 、S 3 , and S v ; the batch fluctuation parameters include the number of iterations, maximum memory storage, stored loss amount, and change rate of iteration, which are respectively denoted as N 1 、N max 、N 2 , and N3 ; The training fluctuation parameters include the number of times of traversing pre-training, the maximum number of epochs of pre-training, the traversing pre-training speed, and the epoch gradient change rate, which are denoted as J 1 、J max 、J v ,and J 2 ; The optimization fluctuation parameters include the current data loss, the maximum allowable data loss, the data momentum, and the optimization gradient change rate, which are denoted as H 1 、H max 、M, and H 0 where the data momentum is used to update the current data, which can help the model accelerate in the relevant direction and suppress oscillations.

[0017] Preferably, the data analysis and processing module is used to perform real-time analysis and processing on the data collected in the data analysis and processing module to obtain the learning efficiency fluctuation coefficient, batch fluctuation coefficient, training fluctuation coefficient, and optimization fluctuation coefficient.

[0018] Preferably, the data analysis and processing module includes a learning efficiency fluctuation coefficient calculation unit, a batch fluctuation coefficient calculation unit, a training fluctuation coefficient calculation unit, and an optimization fluctuation coefficient calculation unit.

[0019] Preferably, the learning efficiency fluctuation coefficient calculation unit imports the learning efficiency fluctuation parameters into the mathematical model of the learning efficiency fluctuation coefficient to obtain the learning efficiency fluctuation coefficient; the batch fluctuation coefficient calculation unit imports the batch fluctuation parameters into the mathematical model of the learning batch fluctuation coefficient to obtain the batch fluctuation coefficient; the training fluctuation coefficient calculation unit imports the training fluctuation parameters into the mathematical model of the training fluctuation coefficient to obtain the training fluctuation coefficient; the optimization fluctuation coefficient calculation unit imports the optimization fluctuation parameters into the mathematical model of the optimization fluctuation coefficient to obtain the optimization fluctuation coefficient.

[0020] Preferably, the mathematical model of the learning efficiency fluctuation coefficient for each sub-time is:

[0021]

[0022] La i represents the learning efficiency fluctuation coefficient of the i-th sub-time region, represents the training speed per hour of the i-th sub-time region, represents the step size of weight update in the i-th sub-time region, represents the weight step size of the i-th sub-time region, S max represents the maximum weight step size of the target time region, represents the change rate of the weight in the i-th sub-time region; the mathematical model of the batch fluctuation coefficient for each sub-time is:

[0023]

[0024] Ma i represents the batch fluctuation coefficient of the i-th sub-time region, represents the number of iterations of the i-th sub-time region, N max represents the maximum storage capacity in the target time region, represents the loss amount stored in the i-th sub-time region, the change rate of iteration in the i-th sub-time region; the mathematical model of the training fluctuation coefficient of each sub-time is:

[0025]

[0026] Xa i represents the training fluctuation coefficient of the i-th sub-time region, represents the number of times of traversing pre-training in the i-th sub-time region, J max represents the maximum number of epochs of pre-training in the target time region, represents the traversing pre-training speed of the i-th sub-time region, represents the rate of change of the number of epochs gradient in the i-th sub-time region; the mathematical model of the optimization fluctuation coefficient of each sub-time is:

[0027]

[0028] Ya i represents the optimization fluctuation coefficient of the i-th sub-time region, represents the current data loss amount in the i-th sub-time region, H max represents the maximum allowable data loss amount in the target time region, M i represents the data momentum of the i-th sub-time region, represents the rate of change of the optimization gradient in the i-th sub-time region, where i = 1, 2,..., n, and i represents the number of the i-th detection sub-region.

[0029] Preferably, the parsing method for the abnormal value of the sub-time region is: ξ i =(La i +Ma i +Xa i +Ya i )*υ, where ξ i represents the sub-time region abnormal index, and υ represents the weight of the abnormality in the target region.

[0030] Preferably, the specific acquisition method for the abnormal value of the target time is: where η is the abnormal value of the target time, Where μ is the mean value of the abnormal numerical value of the target time, λ is other influencing factors of the target time abnormality, and i = 1, 2,..., n, where i represents the number of the i-th detection sub-region.

[0031] Technical effects and advantages of the present invention:

[0032] 1. By dividing the target time region into the learning efficiency fluctuation coefficient, batch fluctuation coefficient, training fluctuation coefficient, and optimization fluctuation coefficient of each sub-time region, and judging the abnormal numerical value of the target time, the present invention ensures the comprehensiveness of data use and further ensures the rationality of target time abnormality discrimination;

[0033] 2. By digitally monitoring and analyzing the pre-training system through AI technical facilities, the present invention divides the system into multiple modules, accurately divides the target time region into equal-time sub-regions, and marks them in sequence, providing a basis for subsequent data collection and improving the accuracy and integrity of data analysis and processing;

[0034] 3. Through the warning module, the present invention can handle potential problems occurring in pre-training, avoid the occurrence of system accidents, improve the performance of artificial intelligence, and reduce industrial costs and losses;

[0035] 4. Through the analysis and trend of data, the present invention can accurately detect abnormal situations in pre-training, and there is a scientific basis for detecting abnormal situations in pre-training through the discrimination module of the system. Description of the Drawings

[0036] Figure 1 It is the overall flowchart of the present invention. Detailed Embodiments

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] The following explains the terms related to the embodiments of the present application.

[0039] Large pre-training model: Large pre-training uses specific deep learning algorithms to learn the internal connections contained in the data through self-supervised learning of large-scale data. The large pre-training model can perform transfer learning through fine-tuning and other methods to adapt to new specific requirements and specialized fields. There are corresponding large pre-training models in various branches of deep learning, such as computer vision, speech processing, natural language processing, and video processing.

[0040] Fine-tuning: The process of further training on a small amount of data based on a pre-trained model to adapt to new specific requirements and specialized fields. Fine-tuning is a transfer learning technique that can significantly reduce the time and resource consumption of pre-training and lower the corresponding industrial costs.

[0041] As shown in the appendix Figure 1 There is a pre-trained system for digital monitoring and analysis of AI-based facilities, including a monitoring time division module, a data collection module, a data analysis and processing module, a sub-time region anomaly calculation module, a target time anomaly calculation module, a target time anomaly discrimination module, a target time warning module, and a target time optimization module.

[0042] The output end of the monitoring time division module is connected to the input end of the data collection module, the output end of the data collection module is connected to the input end of the data analysis and processing module, the output end of the data analysis and processing module is connected to the input end of the sub-time region anomaly calculation module, the output end of the sub-time region anomaly calculation module is connected to the input end of the target time anomaly calculation module, the output end of the target time anomaly calculation module is connected to the input end of the target time anomaly discrimination module, the output end of the target time anomaly discrimination module is connected to the input end of the target time warning module, and the output end of the target time warning module is connected to the input end of the target time optimization module.

[0043] The monitoring time division module is used to determine the monitoring time of the target pre-training as the target time region, and divide the target time region into each sub-time region by equal time division, and sequentially mark them as 1, 2... n.

[0044] The data collection module is used to collect the pre-training learning efficiency fluctuation parameters, batch fluctuation parameters, training fluctuation parameters, and optimization fluctuation parameters of each sub-time region.

[0045] In this embodiment, it should be specifically noted that a GPU server and an optimizer are used to detect the pre-training learning efficiency fluctuation parameters, batch fluctuation parameters, training fluctuation parameters, and optimization fluctuation parameters. The model and manufacturer of the GPU server are: NVIDIA DGX Station and NVIDIA respectively.

[0046] In this embodiment, it should be specifically noted that the learning efficiency fluctuation parameters include the step size of weight update, weight step size, maximum weight step size, change rate of weight, and training speed, which are respectively denoted as S 1 、S 2 、S max 、S 3 , and S v ; the batch fluctuation parameters include the number of iterations, the maximum memory storage, the stored loss amount, and the change rate of iteration, which are respectively denoted as N1 , N max , N 2 , and N 3 ; The training fluctuation parameters include the number of times of traversing pre-training, the maximum number of epochs of pre-training, the speed of traversing pre-training, and the rate of change of the number of epochs gradient, which are respectively denoted as J 1 , J max , J v , and J 2 ; The optimization fluctuation parameters include the current data loss, the maximum allowable data loss, the data momentum, and the rate of change of the optimization gradient, which are respectively denoted as H 1 , H max , M, and H 0 , where the data momentum is used to update the current data, which can help the model accelerate in the relevant direction and suppress oscillations.

[0047] The data analysis and processing module is used to perform real-time analysis and processing on the data collected in the data analysis and processing module to obtain the learning efficiency fluctuation coefficient, batch fluctuation coefficient, training fluctuation coefficient, and optimization fluctuation coefficient.

[0048] In this embodiment, it should be specifically noted that the data analysis and processing module includes a learning efficiency fluctuation coefficient calculation unit, a batch fluctuation coefficient calculation unit, a training fluctuation coefficient calculation unit, and an optimization fluctuation coefficient calculation unit; the learning efficiency fluctuation coefficient calculation unit imports the learning efficiency fluctuation parameters into the mathematical model of the learning efficiency fluctuation coefficient to obtain the learning efficiency fluctuation coefficient; the batch fluctuation coefficient calculation unit imports the batch fluctuation parameters into the mathematical model of the learning batch fluctuation coefficient to obtain the batch fluctuation coefficient; the training fluctuation coefficient calculation unit imports the training fluctuation parameters into the mathematical model of the training fluctuation coefficient to obtain the training fluctuation coefficient; the optimization fluctuation coefficient calculation unit imports the optimization fluctuation parameters into the mathematical model of the optimization fluctuation coefficient to obtain the optimization fluctuation coefficient.

[0049] In this embodiment, it should be specifically noted that the mathematical model of the learning efficiency fluctuation coefficient for each sub-time is as follows: La i represents the learning efficiency fluctuation coefficient of the i-th sub-time region, represents the training speed per hour of the i-th sub-time region, represents the step size of weight update in the i-th sub-time region, represents the weight step size of the i-th sub-time region, S max represents the maximum weight step size of the target time region, represents the rate of change of the weight in the i-th sub-time region; the mathematical model of the batch fluctuation coefficient for each sub-time is as follows:

[0050]

[0051] Ma i represents the batch fluctuation coefficient of the i-th sub-time region, represents the number of iterations of the i-th sub-time region, N max represents the maximum storage capacity in the target time region, represents the loss amount stored in the i-th sub-time region, the change rate of iteration in the i-th sub-time region; the mathematical model of the training fluctuation coefficient of each sub-time is:

[0052]

[0053] Xa i represents the training fluctuation coefficient of the i-th sub-time region, represents the number of times of traversing pre-training in the i-th sub-time region, J max represents the maximum number of epochs of pre-training in the target time region, represents the traversing pre-training speed of the i-th sub-time region, represents the epoch gradient change rate of the i-th sub-time region; the mathematical model of the optimization fluctuation coefficient of each sub-time is:

[0054]

[0055] Ya i represents the optimization fluctuation coefficient of the i-th sub-time region, represents the current data loss amount of the i-th sub-time region, H max represents the maximum allowable data loss amount in the target time region, M i represents the data momentum of the i-th sub-time region, represents the optimization gradient change rate of the i-th sub-time region, where i = 1, 2,..., n, and i represents the number of the i-th detection sub-region.

[0056] The sub-time region anomaly calculation module is used to import the learning efficiency fluctuation coefficient, batch fluctuation coefficient, training fluctuation coefficient, and optimization fluctuation coefficient into the mathematical model of the sub-time region anomaly index to obtain the sub-time region anomaly value.

[0057] In this embodiment, it should be specifically noted that the parsing method of the sub-time region anomaly value is: ξ i =(La i +Ma i +Xa i +Ya i )*υ, where ξ i represents the sub-time region anomaly index, and υ represents the weight of the target region anomaly.

[0058] The target time anomaly calculation module is used to import the sub-time region anomaly index into the mathematical model of the target time region anomaly index to obtain the target time anomaly value.

[0059] In this embodiment, it should be specifically noted that the specific way to obtain the target time anomaly value is as follows: where η is the target time anomaly value, where μ is the mean value of the target time anomaly value, λ is other influencing factors of the target time anomaly, and i = 1, 2,..., n, where i represents the number of the i-th detection sub-region.

[0060] The target time anomaly discrimination module is used to precisely compare the target time anomaly value with the preset target time anomaly value, calculate the difference value between the target time anomaly value and the preset target time anomaly value, and when the difference value is greater than the preset difference value, transmit the difference value to the target time warning module for feedback.

[0061] In this embodiment, it should be specifically noted that the preset target time anomaly value is the mean value of the target time anomaly values in previous years, and the mean value excludes the maximum value and the minimum value, and both the maximum value and the minimum value are extreme values of external factors. The preset difference value is the difference between the theoretical target time anomaly value and the actual target time anomaly value.

[0062] The target time warning module is used to transmit the received data to the system, and then the data obtained by the intelligent processing of the system is transmitted to the target time optimization module. At the same time, corresponding warning reminders are made according to the received data warning instructions.

[0063] In this embodiment, it should be specifically noted that the warning reminder is: when the difference value is greater than the preset difference value, it indicates that the pre-training has an anomaly, and the system will perform corresponding optimization.

[0064] The target time optimization module is used to import the data with anomalies in the target time warning module into the optimizer, and obtain the optimized data through fine-tuning processing and integration.

[0065] The present invention analyzes the pre-training system through the digital monitoring of AI facilities, divides the system into multiple modules, precisely divides the time such as the target time region into each sub-region, and marks them in sequence, providing a basis for subsequent data collection, and improving the accuracy and integrity of data analysis and processing; through the analysis and trend of data, it can accurately detect the anomalies in pre-training, and through the discrimination module of the system, there is a good scientific basis for detecting the anomalies in pre-training; through the warning module, potential problems in pre-training can be processed, avoiding the occurrence of system accidents, improving the performance of artificial intelligence, and reducing production costs and losses.

[0066] Secondly, in the accompanying drawings of the disclosed embodiments of the present invention, only the structures related to the disclosed embodiments are involved. For other structures, reference can be made to the general design. Without conflict, the same embodiment and different embodiments of the present invention can be combined with each other;

[0067] The foregoing are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An AI-based facility digital monitoring and analysis pre-training system, comprising: Monitoring time division module: used to determine the monitoring time of target pre-training as the target time area, divide the target time area into various sub-time areas by equal time division, and mark them as 1, 2...n in sequence; Data collection module: used to collect the pre-trained learning efficiency fluctuation parameters, batch fluctuation parameters, training fluctuation parameters, and optimization fluctuation parameters of each sub-time zone; Data analysis and processing module: used to analyze and process the data collected in the data analysis and processing module in real time to obtain the learning efficiency fluctuation coefficient, batch fluctuation coefficient, training fluctuation coefficient, and optimization fluctuation coefficient; The data analysis and processing module includes a learning efficiency fluctuation coefficient calculation unit, a batch fluctuation coefficient calculation unit, a training fluctuation coefficient calculation unit, and an optimization fluctuation coefficient calculation unit; The learning efficiency fluctuation coefficient calculation unit imports the learning efficiency fluctuation parameter into the mathematical model of the learning efficiency fluctuation coefficient to obtain the learning efficiency fluctuation coefficient; the batch fluctuation coefficient calculation unit imports the batch fluctuation parameter into the mathematical model of the learning batch fluctuation coefficient to obtain the batch fluctuation coefficient; the training fluctuation coefficient calculation unit imports the training fluctuation parameter into the mathematical model of the training fluctuation coefficient to obtain the training fluctuation coefficient; the optimization fluctuation coefficient calculation unit imports the optimization fluctuation parameter into the mathematical model of the optimization fluctuation coefficient to obtain the optimization fluctuation coefficient; The mathematical model of the learning efficiency fluctuation coefficient of each sub-time is: La i represents the learning efficiency fluctuation coefficient of the ith sub-time region, represents the training speed of the ith sub-time zone, represents the step size of the weight update of the ith sub-time region, represents the weight step of the ith sub-time region, S max Indicates the maximum step length of the target time region weight, represents the rate of change of the weight of the ith sub-time region; The mathematical model of the fluctuation coefficient of each sub-time batch is: Ma i represents the batch fluctuation coefficient of the ith sub-time zone, Indicates the number of iterations of the ith sub-time region, N max Indicates the maximum memory storage capacity of the target time zone. represents the amount of loss stored in the i-th sub-time zone, The rate of change of the iteration of the i-th sub-time region; The mathematical model of the fluctuation coefficient of each sub-time training is: XA i represents the training fluctuation coefficient of the ith sub-time zone, represents the number of times the i-th sub-time region traverses the pre-training, J max Indicates the maximum number of rounds of pre-training in the target time region, represents the pre-training speed of the ith sub-time region, represents the gradient change rate of the round number in the i-th sub-time region; The mathematical model for optimizing the fluctuation coefficient of each sub-time is: Ya i represents the optimized volatility coefficient of the ith sub-time zone, represents the current data loss in the ith sub-time zone, H max Indicates the maximum amount of data loss allowed in the target time zone, M i represents the data momentum of the ith sub-time region, represents the optimization gradient change rate of the i-th sub-time region, where i = 1, 2, ..., n, i represents the number of the i-th detection sub-region; Sub-time region anomaly calculation module: used to import the learning efficiency fluctuation coefficient, batch fluctuation coefficient, training fluctuation coefficient, and optimization fluctuation coefficient into the mathematical model of the sub-time region anomaly index to obtain the sub-time region anomaly value; Target time anomaly calculation module: used to import the sub-time area anomaly index into the mathematical model of the target time area anomaly index to obtain the target time anomaly value; Target time anomaly discrimination module: used to accurately compare the target time anomaly value with the preset target time anomaly value, calculate the difference between the target time anomaly value and the preset target time anomaly value, and when the difference is greater than the preset difference, transmit the difference to the target time warning module for feedback; Target time warning module: used to transmit the received data to the system, and then the system transmits the data obtained after intelligent processing to the target time optimization module, and at the same time, the corresponding warning reminder is made according to the received data warning instruction; Target time optimization module: used to import abnormal data in the target time warning module into the optimizer, and obtain optimized data through fine-tuning and integration.

2. According to claim 1, the AI-based facility digital monitoring and analysis pre-training system is characterized by: The analysis method of the abnormal value of the sub-time region is: i =(La i +Ma i +Xa i +Ya i )*υ, where ξ i represents the sub-time region anomaly index, and υ represents the weight of the target region anomaly.

3. The AI-based facility digital monitoring and analysis pre-training system according to claim 2, characterized in that: The specific method for obtaining the target time abnormal value is: Where η is the target time anomaly value, Where μ is the mean value of the target time anomaly, λ is other influencing factors of the target time anomaly, and i = 1, 2, ..., n, i represents the number of the i-th detection sub-region.

4. The AI-based facility digital monitoring and analysis pre-training system according to claim 1, characterized in that: The warning reminder is: when the difference value is greater than the preset difference value, it means that there is an abnormality in the pre-training, and the system will perform corresponding optimization.

Images