A method for calculating the working hours of large instruments based on DP-SVM

The current data is preprocessed and feature extraction based on DP-SVM, which solves the accuracy and efficiency of state calculation of large instruments, and realizes automated state duration calculation, which is suitable for many types of large instruments.

CN114580503BActive Publication Date: 2024-12-03HANGZHOU THINGCOM INFORMATION TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210138038.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-15
Publication Date
2024-12-03
Estimated Expiration
2042-02-15

AI Technical Summary

Technical Problem

The prior art is difficult to accurately and automatically record and calculate the working state of large instruments, especially the shutdown and standby states, which have errors and cumbersome problems, and manual labeling of current state data is high cost and has large errors.

Method used

Using a DP-SVM-based method, the current data is pre-transformed and feature extraction is performed, and the SVM classifier is optimized using dynamic programming algorithms to generate a state model, and the working time of the meter is automatically calculated, which is compatible with various types of instrument states.

Benefits of technology

It realizes accurate and meticulous time calculation of the instrument status, reduces manual labeling costs, improves calculation accuracy and efficiency, and is suitable for many types of large instruments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114580503B_ABST
    Figure CN114580503B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for calculating the working hours of large instruments based on the DP-SVM algorithm. The present invention first collects the current data of each large instrument, respectively intercepts the prominent current of each state of each large instrument as the training set of SVM, generates the state model of the specified instrument, and obtains the data features of each state corresponding to each large instrument. In the actual classification and recognition process, the DP algorithm is first used to perform window segmentation on the current interval, and the best data feature area is segmented for each current point. The best data feature of each point is classified and predicted to achieve the purpose of classifying the current data of any state.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data analysis, and in particular to a method for calculating the working hours of a large instrument based on DP-SVM. Background Art

[0002] Instrument time (power-off time, standby time, working time) can reflect the working status of the instrument and is an important prerequisite for improving the instrument management level. As the "Internet +" thinking penetrates into all walks of life in society, the demand for obtaining instrument time of large instruments and equipment through automation needs to be urgently addressed.

[0003] At present, except for a very small number of professional instruments that automatically record the working status of the instrument during operation, most instruments still require employees to manually record the working hours on paper. This method is inevitably prone to errors, omissions, and record loss, and the shutdown time and standby time can only be inferred from the working time, which is very cumbersome. The machine time is actually a manifestation of the duration of each state of the instrument. If the current signal of the instrument can be classified and identified, the cumulative duration of each state can be obtained.

[0004] Current analysis belongs to the category of signal analysis, so some methods of signal analysis can also be used in the current scenario. At present, common current analysis algorithm models include those based on the SVM support vector machine algorithm, which uses the device current and device status as sample data, trains the state function, and uses the binary tree structure to form multiple classifiers to identify the current state. There is also an image recognition algorithm based on Faster R-CNN to identify the state of power equipment, classify and predict it. However, the above method requires first obtaining the current characteristics of the instrument in different states, and then manually labeling the samples before training. Due to the large amount of current state data, manual labeling is more complicated and the cost is too high. Calculating the threshold is also a method. A fixed current threshold is set, and the state of the instrument is judged by comparing the current value and the threshold. However, this method has errors for instruments with overlapping currents in different states, and the threshold is difficult to determine, which has large errors. It can only be applied to situations with a small number of instruments. Summary of the invention

[0005] The purpose of the present invention is to provide a method, according to the current data of a large instrument collected on site, by pre-transforming the current data, classifying the data using an SVM classifier optimized by a DP (dynamic programming) algorithm, and processing the classification results, so as to obtain the working time of the large instrument.

[0006] To achieve the above purpose, the present invention adopts the following technical solution: a method for calculating the working hours of a large instrument based on current, which is implemented with the aid of an instrument management system. The instrument management system includes an instrument operation status query and a terminal for collecting current data. The terminal collects the current data of the instrument at a certain frequency and reports it to a management platform. The management platform is used to summarize and display the instrument operation status, state current selection, etc. The method comprises the following steps:

[0007] A. Dataset creation steps:

[0008] A-1. Obtain the current data of a large instrument from the instrument management system, wherein the current data includes shutdown current, standby current, and working current.

[0009] A-2. Label each current data in A-1 as the initial training set.

[0010] B. Data analysis steps:

[0011] B-1. Perform feature transformation on the initial training set obtained in A-2 to convert the initial training set into a sample set.

[0012] B-2. Perform multi-state classification based on SVM on the sample set obtained in B-1 to obtain the state model of a large instrument.

[0013] B-3. ​​Encapsulate the state model obtained in B-2 and save it to the cloud database for the next calculation.

[0014] C. Working hours calculation steps:

[0015] C-1. Obtain the current data saved in compressed format from the cloud, decompress the current data, and analyze the current data in minutes.

[0016] C-2. The current data obtained in C-1 is divided into windows, and the best data feature extraction window is selected for each current point through the DP algorithm.

[0017] C-3. According to the optimal data feature extraction window of each current point in C-2, data feature extraction is performed to generate the optimal sample set and perform state classification prediction.

[0018] C-4. Bind the prediction result obtained in C-3 with the initial current timestamp, perform padding and compression, and obtain the final machine time state result.

[0019] In the present invention, the current data corresponding to the large instrument is first obtained, and the current data includes the complete operating current cycle of the large instrument. After the current data is preprocessed and transformed, the state label is performed, and a sample set is generated, and the sample set is feature extracted and classified by the SVM classifier. When the current data is analyzed again later, the data set needs to be converted into a sample set according to the DP algorithm to select the best data feature extraction window for each current point, and the state of each current point is obtained. The state of the large instrument corresponding to each sample can be quickly obtained. Compared with the threshold method for dividing the state of the large instrument, it can be compatible with more types of large instruments, and preprocessing is performed before data classification, so that the sample set of data feature extraction is more accurate and more representative.

[0020] By adopting the present invention, the operating status classification of the large instrument can be determined manually according to the actual situation, and the accurate and detailed duration of each state of the large instrument can be obtained. The model characteristics corresponding to the state of each large instrument are recorded in the database, and the subsequent statistical machine time does not require algorithm intervention, and the DP algorithm is directly optimized and the state prediction is performed, which solves the problem of unclear and inaccurate instrument working hours and has certain engineering value. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of the method of the present invention.

[0022] Figure 2 It is a flowchart for training set selection and classification model generation.

[0023] Figure 3 It is the optimal data feature extraction window extraction flow chart.

[0024] Figure 4 This is the flow chart of the work time analysis of Dayi. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, further detailed description is given below with reference to the accompanying drawings and examples. The unit for describing the current magnitude hereinafter is ampere (A).

[0026] The present invention is based on a large amount of current data collected by the instrument management system, taking into account the fact that the current value of the instrument has the characteristic of random fluctuation under different conditions, and the current data of different instruments are different. Figure 1 As shown, the specific process is as follows:

[0027] A. Dataset creation steps:

[0028] A-1. Current data acquisition steps.

[0029] A-2. Initial training set and classification model generation steps.

[0030] In step A-1, the current data is provided by the instrument management system and is manually selected to ensure that the data is sufficient and reasonable, including the complete operating current cycle of the instrument, and the current is clear, with obvious differences between different states.

[0031] Step A-2 process is as follows Figure 2 First, according to the current data obtained in A-1, all current values ​​appearing in the current data are stored in a candidate set C, and the candidate set C is sorted according to the current value.

[0032] Embodiment: Assume that the current data S = {0.01, 0.02, 0.01, 0.02, 0.01, 0.01, 0.02, 1.34, 1.37, 1.43, 1.34, 1.34, 1.38, 1.36, 1.35}, after the processing of A-2 above, the reference current data set C = {{0.01, 0.02, 0.01, 0.02, 0.01, 0.01, 0.02}, {1.34, 1.37, 1.43, 1.34, 1.34, 1.38, 1.36, 1.35}}, and the label set K = {0, 1}.

[0033] B. Data analysis mainly includes the following three steps.

[0034] B-1. Initial training set feature conversion.

[0035] B-2. Train the SVM classifier.

[0036] B-3. ​​Encapsulate the state model and upload it to the cloud database.

[0037] Step B-1 mainly converts the initial training set obtained in step A-2 into a sample set to facilitate the extraction of data features and training of the SVM classifier. Each data in the data set is a continuous index sequence. First, the sequence is converted into the corresponding current value to form a current sequence; secondly, the mean and standard deviation of the current sequence are calculated, and these two values ​​are used as the features of the sample; finally, the features of all samples are normalized to obtain the final sample set T, which is in the following form:

[0038] T={t0,t1,t2,…t n}

[0039] Where t is a specific sample, and the specific form of t is as follows:

[0040] t=[mean,var]

[0041] Where mean represents the mean and var represents the variance.

[0042] Feature standardization uses the maximum and minimum normalization method to avoid the large gap between different features, which may cause some features to account for too much in the calculation process. The formula for maximum and minimum normalization is as follows:

[0043]

[0044] Among them, x min represents the minimum value of this feature among all samples, x max Represents the maximum value of this feature in all samples.

[0045] Step B-2 is to train the SVM classifier using the sample set converted in step B-1. Here, the mathematical theory involved in the SVM classifier is explained, and the classification surface expression is as follows:

[0046] W T X+b=0

[0047] W T =(w1+w2)

[0048] Find each classification surface in the sample set and perform optimization calculation. T After the value of and b, the distance from any point in the sample set to the classification surface is calculated as follows:

[0049]

[0050] In this way, for any point, we only need to determine the sign of w1x1+w2x2+b and y i Whether the classification result is correct can be judged by whether it is consistent or not. Extended to a higher dimension, the formula is as follows:

[0051]

[0052] Then, we only need to ask for W T The sign of X+b tells us whether the result of the classifier is correct. The positive sample corresponds to W T X+b≥0, negative samples correspond to W T X+b≤0, the distance between the two categories is γ, and the formula is as follows:

[0053]

[0054] Then, the optimal distance between the hyperplane and the two classes can be deduced as: γ / 2. In fact, since b is a constant, the maximum distance between the hyperplane and the two classes is:

[0055]

[0056] Convert it to solve the minimum value problem, and the formula changes to:

[0057]

[0058] This is a complex quadratic programming problem, which can be solved by introducing Lagrange multipliers. The formula is as follows.

[0059]

[0060] C. The working hours calculation steps mainly include the following four steps.

[0061] C-1. Obtain current data.

[0062] C-2. Obtain the optimal data feature extraction window.

[0063] C-3. Obtain the data characteristics of each current point.

[0064] C-4. State mapping, complete the calculation.

[0065] In step C-2, the DP algorithm is used to obtain the best data feature extraction window, and the current data is randomly segmented, feature extracted, and predicted. The one with the largest score is found among multiple segmentation methods. The state transition equation is as follows:

[0066]

[0067] Among them, l is the length of the last sub-signal, k is the number of sub-signals that the first (nl) part of the signal is divided into, and score l Represents the recognition score of the last sub-signal. In the solution process of the DP algorithm, g(n) is defined as the size of the sub-signal ending with the nth data, and the g(n) transfer equation is as follows:

[0068]

[0069] Moreover, in the actual process, the window size needs to be constrained to ensure that the result is more realistic. Therefore, after adding the constraints, the formula is as follows:

[0070]

[0071] s,t,l min ≤l≤min(l max ,n)

[0072] Among them, l min is the minimum length of the sub-signal, l max is the maximum length of the sub-signal. At this time, the initial value of the state transfer equation can be set to f(l min ), the first search position is l min +1.

[0073] In step C-3, data feature extraction is performed at each current point according to the length of the optimal data feature extraction window obtained in step C-2, and step C-4 is performed to perform SVM classification state prediction to achieve state mapping.

[0074] So far, a method for calculating the working hours of the instrument based on current has been realized. This method first converts the current data into a sample set after state labeling and feature transformation, and then uses the SVM algorithm to extract and classify the current data, and finally saves the model to the cloud database. The current data collected subsequently is first selected through the DP algorithm for the best data feature extraction window, and the state mapping is performed on each current point. The current data of all working states are counted, and finally the working time of the instrument is obtained.

Claims

1. A method for calculating the working hours of a large instrument based on the DP-SVM algorithm, characterized in that The method comprises the following steps: A. Dataset creation steps: A-1. Obtain the current data of a large instrument from the instrument management system, wherein the current data includes shutdown current, standby current, and working current; A-2, label each current data in A-1 as the initial training set; B. Data analysis steps: B-1, perform feature transformation on the initial training set obtained in A-2, and convert the initial training set into a sample set; B-2, perform multi-state classification based on SVM on the sample set obtained in B-1 to obtain the state model of a certain large instrument; B-3. ​​Encapsulate the state model obtained in B-2 and save it to the cloud database for the next calculation; C. Working hours calculation steps: C-1. Obtain the current data saved in compressed format from the cloud, decompress the current data, and analyze the current data in minutes; C-2, divide the current data obtained in C-1 into windows, and select the best data feature extraction window for each current point through the DP algorithm; C-3, extract data features according to the best data feature extraction window of each current point in C-2, generate the best sample set, and perform state classification prediction; C-4. Bind the prediction result obtained in C-3 with the initial current timestamp, perform padding and compression, and obtain the final machine time state result.

2. The method for calculating the working hours of a large instrument based on the DP-SVM algorithm according to claim 1 is characterized in that: The steps for converting the current data into the initial training set in step A-2 are as follows: First, according to the current data obtained in A-1, all current values ​​appearing in the current data are stored in the candidate set C, and the candidate set C is manually labeled. The selected candidate set has the following requirements: moderate length, including the complete operating current cycle of the specified equipment as a whole, and the difference between the currents of each state data is relatively obvious.

3. The method for calculating the working hours of a large instrument based on the DP-SVM algorithm according to claim 1 is characterized in that: Step B-1 The specific steps of converting the initial training set into the sample set are as follows: The continuous current data of each state is intercepted from the initial sample set obtained in step A-2, and the current sequence is extracted, normalized and rearranged; secondly, the mean and standard deviation of the current sequence are calculated, and these two values ​​are used as the characteristics of the sample; finally, the characteristics of all data are normalized to obtain the final sample set T.

4. The method for calculating the working hours of a large instrument based on the DP-SVM algorithm according to claim 1 is characterized in that: The specific steps of extracting the optimal current data feature window in step C-2 are as follows; The initial current value in step C-1 is used to extract the current sequence; the region is traversed within the preset region, and data features are pre-extracted and pre-predicted for each region. Through the DP algorithm, the current region with the highest prediction score and the most accurate result centered on a certain current point is selected as the optimal data extraction window for the current point; Step C-3 uses this window as the data feature extraction of the current point to accurately classify the current point.