CN114266944B - Rapid model training result checking system - Google Patents
Rapid model training result checking system Download PDFInfo
- Publication number
- CN114266944B CN114266944B CN202111589142.2A CN202111589142A CN114266944B CN 114266944 B CN114266944 B CN 114266944B CN 202111589142 A CN202111589142 A CN 202111589142A CN 114266944 B CN114266944 B CN 114266944B
- Authority
- CN
- China
- Prior art keywords
- model
- module
- model training
- data
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Debugging And Monitoring (AREA)
Abstract
The invention discloses a rapid model training result inspection system, which relates to the technical field of data processing and comprises an upper computer, a cloud platform, an error inspection module, a data distribution module and a training analysis module; when the error detection module detects a plurality of model training results in batch, the data distribution module is used for classifying the model training results to be detected and sequentially distributing corresponding detection units for processing, so that the data processing is more hierarchical and orderly; the error checking module compares a corresponding model training result with an actual result after receiving the model training data to obtain a training error, and uploads the training error to the upper computer through the bus communication unit for display and storage; the training analysis module is used for obtaining the inspection data with the same model identification generated by the error inspection module to perform model correction analysis, judging whether the corresponding model needs to be corrected or not, and reminding management personnel to correct the relevant parameters of the corresponding model in time so as to improve the detection precision of the corresponding model.
Description
Technical Field
The invention relates to the technical field of data processing, in particular to a rapid model training result inspection system.
Background
With the development of computer technology and network technology, deep learning technology has been widely applied in many fields; for example: detecting the text in the image by adopting a deep learning technology to determine the position of the text in the image; in some target detection scenarios, a smaller target needs to be detected from a larger picture, and the like; the detection method based on deep learning generally depends on a trained deep learning model, a large number of model training results can be generated in the model training process, the method is limited by the current technical capability, recognition errors are easy to generate when a large number of models are trained, the accuracy is reduced, and the effect stability cannot be ensured;
the existing model training result inspection system has the problems that the model training results cannot be reasonably distributed to corresponding selected terminals for data inspection, so that the data inspection efficiency and speed are low, and whether the trained deep learning model meets the standard or not cannot be reasonably judged according to the inspection results, so that the detection precision is improved.
Disclosure of Invention
The present invention is directed to solving at least one of the problems of the prior art. Therefore, the invention provides a rapid model training result checking system.
In order to achieve the above object, an embodiment according to the first aspect of the present invention provides a rapid model training result checking system, which includes an upper computer, a cloud platform, an error checking module, a data distribution module, a training analysis module, and an alarm module;
the data acquisition module is used for acquiring model training results of a plurality of models and transmitting the acquired model training results to the error inspection module;
when the error checking module checks a plurality of model training results in batch, the data distribution module is used for classifying the model training results to be checked and distributing corresponding checking units to process according to the loss value SH; wherein each model training result is provided with a model identifier;
the error detection module comprises a control unit, a bus communication unit and a plurality of detection units, wherein the detection units are used for comparing corresponding model training results with actual results after receiving model training data to obtain training errors, and uploading the training errors to an upper computer through the bus communication unit for displaying and storing;
the training analysis module is connected with the error inspection module and used for obtaining inspection data with the same model identification generated by the error inspection module to perform model correction analysis and judging whether the corresponding model needs to be corrected or not according to a correction coefficient XZ; wherein the test data is represented as corresponding training errors.
Further, the specific allocation process of the data allocation module is as follows:
marking the model training results of the same model identification as model training data; marking a deep learning model corresponding to the model training data as a target model, and collecting application information of the target model to evaluate an application coefficient Y1; the application information comprises application times and corresponding application time;
the model training data are arranged in a descending order according to the size of an application coefficient Y1, and corresponding test units are sequentially distributed according to the order of the model training data for processing, specifically:
aiming at the first-ranked model training data, selecting a test unit with the smallest loss value SH as a selection unit of the model training data; and aiming at the model training data ranked the second, selecting a test unit with the loss value SH times as a selected unit of the model training data, and so on.
Further, the specific evaluation process of the application coefficient Y1 is as follows:
marking the application times of the target model as C1, and calculating the time difference of adjacent application times to obtain an application interval GTi; counting the times that GTi is less than or equal to the interval threshold value as C2; when the GTi is not more than the interval threshold, summing the difference value of the corresponding GTi and the interval threshold to obtain a total difference value CZ; calculating a difference coefficient CX by using a formula CX-C2 × d1+ CZ × d2, wherein d1 and d2 are coefficient factors;
counting the time difference between the latest application time and the current time of the system as a delay time length H1, and calculating an application coefficient Y1 of the target model by using a formula Y1 of (C1 × d3+ CX × d4)/(H1 × d5+ u), wherein d3, d4 and d5 are coefficient factors, and u is a compensation factor.
Further, the method for calculating the loss value SH includes:
collecting historical processing records of the inspection unit; marking the processing time length of each processing of the inspection unit as Ti, and marking the standby time length of each processed inspection unit as DTi;
setting a plurality of standby time length thresholds, wherein each standby time length threshold corresponds to a preset processing time length range, and marking the standby time length threshold corresponding to the processing time length Ti as Yr;
comparing the standby duration DTi with a corresponding standby duration threshold Yr; when the standby time DTi is less than Yr, the inspection unit is judged to continuously work at the moment, and extra loss is generated;
counting the number of times that DTi is smaller than YR as loss frequency P1; when DTi is smaller than Yr, summing the difference between YR and DTi to obtain a difference standby total value TZ; marking the time difference between the latest processing end time and the current time of the system as a slow time length H2; the loss value SH of the test cell is calculated using the formula SH (P1 × a1+ TZ × a2)/(H2+ u), where a1 and a2 are both coefficient factors and u is a compensation factor.
Further, the specific analysis process of the training analysis module is as follows:
acquiring inspection data which is generated by an error inspection module and has the same model identification, wherein the inspection data carries a qualified identification and an unqualified identification; when unqualified marks are detected, automatically counting down, wherein the counting down is D1, and D1 is a preset value; every time one inspection data is collected, the count-down is reduced by one;
the occurrence frequency of unqualified marks in the counting down stage is P2, the frequency of automatic returning of counting down to the original value is P3, and the length of the counting down stage is L1;
calculating a correction coefficient XZ of the corresponding model by using a formula XZ (P2 × a3+ P3 × a4)/(L1 × a5+ u), wherein a3, a4 and a5 are coefficient factors, and u is a compensation factor; if the XZ is larger than or equal to the correction threshold, judging that the corresponding model needs to be corrected, and generating a correction signal;
the training analysis module is used for transmitting the correction signal to an upper computer, and the upper computer controls the alarm module to give an alarm after receiving the correction signal so as to remind a manager to correct the relevant parameters of the corresponding model.
Further, if the training error is within the allowable range, marking a qualified mark on the corresponding training error; otherwise, printing an unqualified mark.
Further, monitoring unqualified marks continuously in a countdown stage, and if a new unqualified mark is monitored, automatically returning countdown to an original value and counting again; otherwise, the count-down returns to zero and the counting is stopped.
Furthermore, the control unit and the inspection unit are respectively in communication connection with the upper computer through the bus communication unit; the control unit is used for transmitting the model training data to the corresponding checking unit through the bus communication unit.
Compared with the prior art, the invention has the beneficial effects that:
1. when the error checking module checks a plurality of model training results in batch, the data distribution module is used for classifying the model training results to be checked, marking the model training results identified by the same model as model training data, collecting application information of the corresponding model, calculating by combining application times and application intervals to obtain an application coefficient Y1 of the corresponding model, and sequentially distributing corresponding checking units according to the size of the application coefficient Y1 for processing, so that the data processing is more hierarchical and orderly; when the inspection units are distributed, historical processing records of the inspection units are collected to evaluate the loss values SH of the inspection units, and the inspection unit with the minimum loss value SH is selected as a selection unit, so that the data inspection efficiency and speed are improved;
2. the error detection module is used for comparing a corresponding model training result with an actual result to obtain a training error; if the training error is within the allowable range, marking a qualified mark on the corresponding training error; otherwise, marking an unqualified mark; the training analysis module is used for obtaining the inspection data with the same model identification generated by the error inspection module to perform model correction analysis, when an unqualified identification is monitored, automatically counting down, calculating to obtain a correction coefficient of the corresponding model by combining the occurrence frequency of the unqualified identification in the counting down stage, the frequency of the counting down automatically returning to an original value and the length of the counting down stage, and reminding a manager to correct the relevant parameters of the corresponding model if the correction coefficient is larger than or equal to a correction threshold value so as to improve the detection precision of the corresponding model.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a block diagram of the system of the present invention.
Detailed Description
The technical solutions of the present invention will be described clearly and completely with reference to the following embodiments, and it should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
As shown in fig. 1, the rapid model training result inspection system includes an upper computer, a cloud platform, a data acquisition module, an error inspection module, a data distribution module, a training analysis module and an alarm module;
the upper computer in the embodiment preferably selects an industrial PC, the industrial PC is responsible for displaying, storing and uploading inspection data to the cloud platform, the industrial PC provides a uniform display interface for data obtained after the error inspection module inspects model training results in batches, so that result data concerned by a user can be displayed simply and clearly, application software of various devices can be unified, all links of the whole detection system use the same software, users can adapt to operation software of different links quickly, and each model training result is provided with a model identifier;
the data acquisition module is used for acquiring model training results of a plurality of models and transmitting the acquired model training results to the error checking module, wherein the models are deep learning models, and the model training results are output results obtained by substituting input data into the deep learning models;
the error detection module comprises a control unit, a bus communication unit and a plurality of detection units, wherein the control unit and the detection units are respectively in communication connection with the upper computer through the bus communication unit; the testing unit is a data processing terminal and is used for testing the error of the model training result;
the data distribution module is connected with the error checking module, and when the error checking module checks a plurality of model training results in batches, the data distribution module is used for classifying the model training results to be checked and distributing corresponding checking units for processing, and the specific process is as follows:
classifying the model training results according to the model identifications, and marking the model training results of the same model identification as model training data;
marking a deep learning model corresponding to the model training data as a target model, and collecting application information of the target model, wherein the application information comprises application times and corresponding application time; the application of the target model is expressed as applying the target model to a certain scene, for example, detecting a text in an image by using a deep learning model to determine the position of the text in the image, and the like;
marking the application times of the target model as C1, sequencing the application time according to the time sequence, and calculating the time difference of adjacent application time to obtain an application interval GTi;
comparing the GTi with an interval threshold value, and counting the number of times that the GTi is less than or equal to the interval threshold value as C2; when GTi is not more than the interval threshold, summing the difference between the corresponding GTi and the interval threshold to obtain a total difference value CZ, and calculating by using a formula CX-C2 × d1+ CZ × d2 to obtain a difference coefficient CX, wherein d1 and d2 are coefficient factors;
counting the time difference between the latest application time and the current time of the system as a delay time H1, normalizing the application times, the difference coefficient and the delay time and taking the numerical value, and calculating the application coefficient Y1 of the target model by using a formula Y1 as (C1 × d3+ CX × d4)/(H1 × d5+ u), wherein d3, d4 and d5 are coefficient factors, and u is a compensation factor;
the model training data are arranged in a descending order according to the size of an application coefficient Y1, and corresponding test units are sequentially distributed according to the sequence of the model training data for processing, and the method specifically comprises the following steps:
collecting historical processing records of the inspection unit; marking the processing time length of each processing of the inspection unit as Ti, and calculating the time difference between the corresponding processing ending time and the next processing starting time to obtain standby time length DTi, wherein Ti and DTi are in one-to-one correspondence;
setting a plurality of standby time length thresholds, wherein each standby time length threshold corresponds to a preset processing time length range, and marking the standby time length threshold corresponding to the processing time length Ti as Yr; comparing the standby duration DTi with a corresponding standby duration threshold Yr; when the standby time DTi is less than Yr, the inspection unit is considered not to have sufficient rest, and at the moment, the inspection unit continuously works, and extra loss is generated;
counting the number of times that DTi is smaller than YR as loss frequency P1; when DTi is smaller than Yr, carrying out difference calculation on YR and DTi, and summing all the differences to obtain a difference standby total value TZ; marking the time difference between the latest processing end time and the current time of the system as a slow time length H2;
calculating the loss value SH of the inspection unit by using a formula SH (P1 × a1+ TZ × a2)/(H2+ u), wherein a1 and a2 are coefficient factors, and u is a compensation factor;
distributing the first-ranked model training data, and selecting a test unit with the smallest loss value SH as a selection unit of the model training data; marking the model training data with the first sequence as distributed data, distributing the model training data with the second sequence, selecting a test unit with the loss value of SH (rank) as a selected unit of the model training data, and so on;
the control unit is used for transmitting the model training data to the corresponding inspection unit through the bus communication unit, the inspection unit compares the corresponding model training result with the actual result after receiving the model training data to obtain a training error, and the training error is uploaded to the upper computer through the bus communication unit to be displayed and stored; if the training error is within the allowable range, marking the corresponding training error with a qualified mark; otherwise, marking an unqualified mark;
the training analysis module is connected with the error detection module and used for obtaining detection data which are generated by the error detection module and have the same model identification, performing model correction analysis and judging whether the corresponding model needs to be corrected or not, wherein the detection data are expressed as corresponding training errors; the specific analysis process is as follows:
acquiring inspection data with the same model identification, which is generated by an error inspection module, wherein the inspection data carries a qualified identification and an unqualified identification;
when unqualified marks are detected, automatically counting down, wherein the counting down is D1, and D1 is a preset value; for example, D1 takes the value 10; every time one inspection data is collected, the count-down is reduced by one;
in the countdown stage, monitoring the unqualified identification continuously, if a new unqualified identification is monitored, automatically returning the countdown to the original value, and performing countdown again according to D1, otherwise, returning the countdown to zero and stopping counting;
the occurrence frequency of unqualified marks in the counting down stage is P2, the frequency of automatic returning of counting down to the original value is P3, and the length of the counting down stage is L1;
calculating a correction coefficient XZ of the corresponding model by using a formula XZ (P2 × a3+ P3 × a4)/(L1 × a5+ u), wherein a3, a4 and a5 are coefficient factors, and u is a compensation factor;
comparing the correction coefficient XZ with a correction threshold, and if the XZ is larger than or equal to the correction threshold, judging that the error of the model training result of the corresponding model is larger, the model parameters need to be corrected, and generating a correction signal;
the training analysis module is used for transmitting the correction signal to the upper computer, and the upper computer controls the alarm module to give an alarm after receiving the correction signal so as to remind a manager to correct the relevant parameters of the corresponding model, so that the detection precision of the corresponding model is improved.
The above formulas are all calculated by removing dimensions and taking numerical values thereof, the formula is a formula which is obtained by acquiring a large amount of data and performing software simulation to obtain the closest real situation, and the preset parameters and the preset threshold value in the formula are set by the technical personnel in the field according to the actual situation or obtained by simulating a large amount of data.
The working principle of the invention is as follows:
when the rapid model training result inspection system works, when the error inspection module inspects a plurality of model training results in batch, the data distribution module is used for classifying the model training results to be inspected and distributing corresponding inspection units for processing, marking the model training results of the same model identification as model training data, collecting application information of corresponding models, calculating by combining application times and application intervals to obtain application coefficients Y1 of the corresponding models, performing descending order arrangement on the model training data according to the size of the application coefficients Y1, sequentially distributing the corresponding inspection units for processing, and selecting the inspection unit with the smallest loss value SH as a selected unit aiming at the first ordered model training data; aiming at the model training data of the second order, selecting a checking unit with the loss value SH to be the next as a selecting unit, and so on, thereby improving the data checking efficiency and speed;
the error testing module is used for comparing the corresponding model training result with the actual result to obtain a training error, namely testing data; the training analysis module is used for obtaining the inspection data with the same model identification generated by the error inspection module to perform model correction analysis, monitoring the unqualified identification in the countdown stage, calculating a correction coefficient of the corresponding model by combining the occurrence frequency of the unqualified identification in the countdown stage, the frequency of automatic return of countdown to an original value and the length of the countdown stage, and reminding a manager to correct the related parameters of the corresponding model if the correction coefficient is larger than or equal to a correction threshold value so as to improve the detection precision of the corresponding model.
In the description herein, references to the description of "one embodiment," "an example," "a specific example" or the like are intended to mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
The preferred embodiments of the invention disclosed above are intended to be illustrative only. The preferred embodiments are not intended to be exhaustive or to limit the invention to the precise forms disclosed. Obviously, many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the invention and the practical application, to thereby enable others skilled in the art to best utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims (2)
1. The rapid model training result inspection system is characterized by comprising an upper computer, an error inspection module, a data acquisition module, a data distribution module, a training analysis module and an alarm module;
the data acquisition module is used for acquiring model training results of a plurality of models and transmitting the acquired model training results to the error inspection module;
when the error checking module checks a plurality of model training results in batch, the data distribution module is used for classifying the model training results to be checked and distributing corresponding checking units for processing according to the loss value SH; each model training result is provided with a model identifier; the specific distribution process is as follows:
marking the model training results of the same model identification as model training data; marking a deep learning model corresponding to the model training data as a target model;
acquiring application information of a target model to evaluate an application coefficient Y1, wherein the application information comprises application times and corresponding application time; the specific evaluation procedure is as follows:
marking the application times of the target model as C1, and calculating the time difference of adjacent application times to obtain an application interval GTi; counting the times that GTi is less than or equal to the interval threshold value as C2; when the GTi is not more than the interval threshold, summing the difference value of the corresponding GTi and the interval threshold to obtain a total difference value CZ; calculating a difference coefficient CX by using a formula CX = C2 × d1+ CZ × d2, wherein d1 and d2 are coefficient factors;
counting the time difference between the latest application time and the current time of the system as a delay time H1, and calculating an application coefficient Y1 of the target model by using a formula Y1= (C1 × d3+ CX × d4)/(H1 × d5+ u), wherein d3, d4 and d5 are coefficient factors, and u is a compensation factor;
the model training data are arranged in a descending order according to the size of an application coefficient Y1, and corresponding test units are sequentially distributed according to the order of the model training data for processing, specifically:
aiming at the first-ranked model training data, selecting a test unit with the smallest loss value SH as a selection unit of the model training data; aiming at the second model training data, selecting a test unit with the loss value SH times as a selected unit of the model training data, and so on;
the method for calculating the loss value SH comprises the following steps:
collecting historical processing records of the inspection unit; marking the processing time length of each processing of the inspection unit as Ti, and marking the standby time length of each processed inspection unit as DTi;
setting a plurality of standby time length thresholds, wherein each standby time length threshold corresponds to a preset processing time length range, and marking the standby time length threshold corresponding to the processing time length Ti as Yr;
comparing the standby duration DTi with a corresponding standby duration threshold Yr; when the standby time DTi is less than Yr, the inspection unit is judged to continuously work at the moment, and extra loss is generated;
counting the number of times that DTi is smaller than YR as loss frequency P1; when DTi is smaller than Yr, summing the difference between YR and DTi to obtain a difference standby total value TZ; marking the time difference between the latest processing end time and the current time of the system as a slow time length H2; calculating the loss value SH of the inspection unit by using a formula SH = (P1 × a1+ TZ × a2)/(H2+ u), wherein a1 and a2 are coefficient factors, and u is a compensation factor;
the error detection module comprises a control unit, a bus communication unit and a plurality of detection units, wherein the detection units are used for comparing corresponding model training results with actual results after receiving model training data to obtain training errors, and uploading the training errors to an upper computer through the bus communication unit for displaying and storing;
if the training error is within the allowable range, marking a qualified mark on the corresponding training error; otherwise, marking an unqualified mark;
the training analysis module is connected with the error inspection module and used for obtaining inspection data with the same model identification generated by the error inspection module to perform model correction analysis and judging whether the corresponding model needs to be corrected or not according to a correction coefficient XZ; the specific analysis process is as follows:
acquiring test data with the same model identification generated by an error test module, wherein the test data is expressed as corresponding training errors; when unqualified marks are detected, automatically counting down, wherein the counting down is D1, and D1 is a preset value; every time one inspection data is collected, the count-down is reduced by one;
continuously monitoring the unqualified identification in the countdown stage, and if a new unqualified identification is monitored, automatically returning the countdown to the original value and counting again; otherwise, the count-down returns to zero, and the counting is stopped;
the occurrence frequency of unqualified marks in the counting down stage is P2, the frequency of automatic returning of counting down to the original value is P3, and the length of the counting down stage is L1;
calculating a correction coefficient XZ of the corresponding model by using a formula XZ = (P2 × a3+ P3 × a4)/(L1 × a5+ u), wherein a3, a4 and a5 are coefficient factors, and u is a compensation factor; if the XZ is larger than or equal to the correction threshold, judging that the corresponding model needs to be corrected, and generating a correction signal;
the training analysis module is used for transmitting the correction signal to an upper computer, and the upper computer controls the alarm module to give an alarm after receiving the correction signal so as to remind a manager to correct the relevant parameters of the corresponding model.
2. The system for checking the training result of the rapid model according to claim 1, wherein the control unit and the checking unit are respectively connected with the upper computer through a bus communication unit; the control unit is used for transmitting the model training data to the corresponding checking unit through the bus communication unit.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111589142.2A CN114266944B (en) | 2021-12-23 | 2021-12-23 | Rapid model training result checking system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111589142.2A CN114266944B (en) | 2021-12-23 | 2021-12-23 | Rapid model training result checking system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114266944A CN114266944A (en) | 2022-04-01 |
| CN114266944B true CN114266944B (en) | 2022-08-12 |
Family
ID=80829147
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111589142.2A Active CN114266944B (en) | 2021-12-23 | 2021-12-23 | Rapid model training result checking system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114266944B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114844482A (en) * | 2022-04-25 | 2022-08-02 | 安徽铜峰电子股份有限公司 | Wafer for high-frequency crystal resonator and production control system thereof |
| CN115201444A (en) * | 2022-07-14 | 2022-10-18 | 安徽强邦新材料股份有限公司 | Aluminum substrate defect detection screening system and method |
| CN115540987B (en) * | 2022-09-19 | 2023-07-14 | 蚌埠高灵传感系统工程有限公司 | Dynamic vehicle overload detection method and control device based on load sensor |
| CN115601230A (en) * | 2022-11-09 | 2023-01-13 | 科讯嘉联信息技术有限公司（Cn） | Digital human synthesis method based on multi-task learning |
| CN116663777A (en) * | 2023-06-05 | 2023-08-29 | 重庆翰海睿智大数据科技股份有限公司 | Experimental training system and method based on knowledge graph |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104951846A (en) * | 2015-06-02 | 2015-09-30 | 国电南京自动化股份有限公司 | Microgrid short-term power and load prediction system and dynamic error classification and correction method |
| CN110163368A (en) * | 2019-04-18 | 2019-08-23 | 腾讯科技（深圳）有限公司 | Deep learning model training method, apparatus and system based on mixed-precision |
| US10628755B1 (en) * | 2018-12-14 | 2020-04-21 | Sas Institute Inc. | Distributable clustering model training system |
| CN111650829A (en) * | 2020-05-18 | 2020-09-11 | 瑞纳智能设备股份有限公司 | Embedded PID module parameter adjusting method, system and device based on artificial intelligence |
| CN112199885A (en) * | 2020-09-09 | 2021-01-08 | 北京达佳互联信息技术有限公司 | Distributed model training system and application method |
| US10963792B1 (en) * | 2020-03-26 | 2021-03-30 | StradVision, Inc. | Method for training deep learning network based on artificial intelligence and learning device using the same |
| CN112860520A (en) * | 2021-02-23 | 2021-05-28 | 合肥大多数信息科技有限公司 | Information data formatting assembly based on artificial intelligence |
| CN113110104A (en) * | 2021-04-25 | 2021-07-13 | 同济大学 | Intelligent cascade test system for braking of single marshalling train |
| CN113282011A (en) * | 2021-04-25 | 2021-08-20 | 同济大学 | Prediction model construction and implementation method applied to single-marshalling train braking intelligent cascade test |
| CN113792373A (en) * | 2021-11-17 | 2021-12-14 | 中化学建设投资集团北京科贸有限公司 | Personnel behavior monitoring emergency disposal method based on machine vision |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20210048895A (en) * | 2019-10-24 | 2021-05-04 | 엘지전자 주식회사 | Training aritificial neural network model based on generative adversarial network |
-
2021
- 2021-12-23 CN CN202111589142.2A patent/CN114266944B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104951846A (en) * | 2015-06-02 | 2015-09-30 | 国电南京自动化股份有限公司 | Microgrid short-term power and load prediction system and dynamic error classification and correction method |
| US10628755B1 (en) * | 2018-12-14 | 2020-04-21 | Sas Institute Inc. | Distributable clustering model training system |
| CN110163368A (en) * | 2019-04-18 | 2019-08-23 | 腾讯科技（深圳）有限公司 | Deep learning model training method, apparatus and system based on mixed-precision |
| US10963792B1 (en) * | 2020-03-26 | 2021-03-30 | StradVision, Inc. | Method for training deep learning network based on artificial intelligence and learning device using the same |
| CN111650829A (en) * | 2020-05-18 | 2020-09-11 | 瑞纳智能设备股份有限公司 | Embedded PID module parameter adjusting method, system and device based on artificial intelligence |
| CN112199885A (en) * | 2020-09-09 | 2021-01-08 | 北京达佳互联信息技术有限公司 | Distributed model training system and application method |
| CN112860520A (en) * | 2021-02-23 | 2021-05-28 | 合肥大多数信息科技有限公司 | Information data formatting assembly based on artificial intelligence |
| CN113110104A (en) * | 2021-04-25 | 2021-07-13 | 同济大学 | Intelligent cascade test system for braking of single marshalling train |
| CN113282011A (en) * | 2021-04-25 | 2021-08-20 | 同济大学 | Prediction model construction and implementation method applied to single-marshalling train braking intelligent cascade test |
| CN113792373A (en) * | 2021-11-17 | 2021-12-14 | 中化学建设投资集团北京科贸有限公司 | Personnel behavior monitoring emergency disposal method based on machine vision |
Non-Patent Citations (2)
| Title |
|---|
| 10米重型数控立车误差建模、测量及实时补偿;沈岳熙;《中国优秀博硕士学位论文全文数据库(硕士)工程科技Ⅰ辑》;20120715(第07期);第B022-494页 * |
| Boosted MMI for model and feature-space discriminative training;Daniel Povey等;《2018 IEEE International Conference on Accoustics，Speech and Signal Processing》;20081231;第4057-4060页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114266944A (en) | 2022-04-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114266944B (en) | Rapid model training result checking system | |
| US20240087104A1 (en) | Method for monitoring manufacture of assembly units | |
| CN112598343B (en) | Risk inspection method, device, equipment and storage medium for data center | |
| KR101953558B1 (en) | Apparatus and Method for Fault Management of Smart Devices | |
| CN115269342A (en) | Monitoring cloud platform based on edge computing and monitoring method thereof | |
| CN114325405A (en) | Battery pack consistency analysis method, modeling method, device, equipment and medium | |
| CN113065538A (en) | Pressure sensor detection method, device and equipment based on image recognition | |
| CN112729884B (en) | Equipment fault diagnosis method and device based on big data | |
| CN116563841B (en) | Detection method and detection device for power distribution network equipment identification plate and electronic equipment | |
| CN117434440A (en) | Automatic detection method and system for measuring switch | |
| CN117406027A (en) | Distribution network fault distance measurement method and system | |
| CN117007175A (en) | Sensor automatic test method and test device thereof | |
| CN111212434A (en) | WIFI module quality prediction method, device, equipment and storage medium | |
| CN114666774B (en) | Identification matching system and method of Bluetooth equipment | |
| CN116184060A (en) | Abnormal monitoring method and system suitable for porcelain insulator live working | |
| CN115617784A (en) | Data processing system and processing method for informationized power distribution | |
| CN115618286A (en) | Transformer partial discharge type identification method, system, equipment, terminal and application | |
| CN115035481A (en) | Image object distance fusion method, device, equipment and storage medium | |
| CN114897006A (en) | Intelligent ammeter error classification method, system, equipment and storage medium | |
| CN114218430A (en) | Remote cooperative equipment operation and maintenance system, method and device | |
| CN111176931A (en) | Operation monitoring method, operation monitoring device, server and storage medium | |
| CN110688273A (en) | Classification model monitoring method and device, terminal and computer storage medium | |
| CN116908533B (en) | Power consumer electricity consumption information acquisition equipment with metering function | |
| CN110389947A (en) | A kind of blacklist generation method, device, equipment and medium | |
| CN116229700B (en) | Method and system for analyzing error of monitoring data of lithium battery energy storage box |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |