System and method for improving water body sand content monitoring accuracy based on machine learning

Abstract

The invention relates to the technical field of water sediment measurement and monitoring, in particular to a system and a method for improving water sediment content monitoring accuracy based on machine learning. According to the method, a PLC program is designed through machine self-learning, the machine learning program is trained through examples (manual sampling data), PLC parameters are automatically calculated through machine learning, and parameter adjustment is conducted in time. Along with deep machine learning, the error between the measured value of the water body sand content online monitoring system and the measured value of manual sampling is gradually reduced, and the measurement accuracy of the water body sand content online monitoring system is improved. According to the method, the sensor is effectively close to a truth value under the training of a truth value sample, and the target of the method is that the numerical value of a certain measurement time is effectively close to the truth value, but the probability that the difference between the numerical value and the truth value is minimum in future measurement time is expected to be large; according to the machine learning algorithm based on artificial intelligence, the application range of the intelligent sensor field is effectively expanded, the application scene is wide, and the practicability is high.

Description

Technical Field

[0001] The present invention relates to the technical field of water body sediment measurement and monitoring, and in particular to a system and method for improving the accuracy of water body sediment content monitoring based on machine learning. Background Art

[0002] Hydroelectric power generation is a process in which water with a certain potential energy located at a relatively high place such as rivers, lakes and reservoirs is converted into mechanical energy of the turbine through a hydraulic turbine at a lower place, and then the turbine drives the generator to generate electricity, and finally the mechanical energy is converted into electrical energy. When high-sand water flows through a hydro-generator unit, it will cause serious cavitation and abrasion of the turbine runner, resulting in system instability, reduced efficiency and threats to the safety of the unit. In order to reduce the damage of sediment to the hydro-turbine unit of the power generation system, it is necessary to accurately monitor the sediment content passing through the machine, timely grasp the changes in the sediment content passing through the machine, and shut down the machine in time to avoid sand.

[0003] There are many methods for monitoring sediment content in water bodies, such as X-ray method, optical method, differential pressure method, acoustic method, turbidity method and spectral analysis method. Different methods correspond to different sensors, and they all show different advantages in practical applications, but they also have certain limitations.

[0004] When monitoring the sediment content, the sensor directly enters and contacts the water body, and the collected sediment content value will inevitably be affected by factors such as water temperature, flow rate, pressure pulsation, and the position in the water body; it will also inevitably cause certain interference to the measurement accuracy of the measuring sensor. Traditionally, manual sampling measurements are often used to compare the sensor's measurement values ​​to grasp and calibrate the measurement accuracy of the online water body sediment content, but there is a serious lag in the process, which poses a hidden danger to the safety of the generator set. Therefore, it is very meaningful and practical to study the use of machine learning program algorithms through artificial intelligence theory to improve the measurement accuracy of sediment content sensors and reduce the difference between sensor measurements and manual sampling measurements. Summary of the invention

[0005] The purpose of the present invention is to provide a system and method for improving the accuracy of water body sediment content monitoring based on machine learning, so as to solve the large error between the measured sediment content value of the water body sediment content online monitoring system and the sediment content value measured by manual sampling, and ensure the accuracy of online water body sediment content measurement.

[0006] To achieve the above-mentioned purpose, the present invention adopts the following technical scheme: a system for improving the monitoring accuracy of water body sediment content based on machine learning, including a linear sensor and a PLC, wherein the PLC is provided with a mathematical model of a machine learning program, and based on the machine self-learning method in artificial intelligence theory, the mathematical model of the machine learning program in the PLC is trained with the PLC measured sediment content data and the manually sampled sample data to obtain a new mathematical model, and the monitoring data of the linear sensor is calculated by the new mathematical model to obtain the data of the water body sediment content; with the deepening of machine learning, the error between the measured value of the water body sediment content online monitoring system and the measured value of the manually sampled value is gradually reduced, thereby improving the measurement accuracy of the water body sediment content online monitoring system.

[0007] Furthermore, the equivalent simplified mathematical model of the measurement program in the PLC is Y=kX+b; wherein Y is the sand content value; X is the current value of the linear sensor; k is the proportional coefficient; and b is the adjustment parameter for calculating the sand content.

[0008] Furthermore, the algorithm for training the mathematical model of the machine learning program in the PLC is based on the least squares method.

[0009] The method for improving the accuracy of monitoring the sediment content in water bodies comprises the following steps:

[0010] S1, build a PLC database, the database includes input array, output array and initial parameters;

[0011] S2, iterating the input array in S1 through the calculation function to obtain the output array, and continuously correcting the mathematical model of the machine learning program in the PLC through the output array to obtain a new mathematical model for sand content measurement;

[0012] S3. The monitoring data of the linear sensor is calculated through the new mathematical model obtained in S2 to obtain the data of the sediment content in the water body.

[0013] Furthermore, in the S1-S2, the input array includes a measured array and a sample array; the measured array includes a sensor sampling value array and a measured sediment content value array calculated linearly by PLC; the sample array is the value of samples manually sampled at the same time and location as the sensor; the measured sediment content value array and the sample array are used as input values ​​of the calculation function; the sensor sampling value array is used to calibrate the measured sediment content value array; the initial parameters include the function independent variable w1 and the learning depth a, the initial value of w1 is 1, and the value range of a is 0.01~0.0001.

[0014] Furthermore, in the S2, the calculation function is:

[0015] f(w1)=(y1w1-y1′)2 +(y2w1-y2′) 2 +……(y 10 w1-y 10 ′) 2 ;

[0016] f(w2)=(y1w2-y1′) 2 +(y2w2-y2′) 2 +……(y 10 w2-y 10 ′) 2 ;

[0017] Among them, w1 and w2 are function independent variables;

[0018] y n is the value in the numerical array of measured sediment content, (n=1,2,3…10);

[0019] y n ' is the value in the sample array, (n = 1, 2, 3...10);

[0020] Each machine learning calculation will infer the value of w2, which corresponds to the number of times the mathematical model of sediment content measurement is iterated w 2n .

[0021] Furthermore, the measured sediment content numerical array and the sample array are iterated through a calculation function, wherein w2=w1+a;

[0022] During the iteration, if f(w1)>f(w2), w2=w2+a;

[0023] Substitute f(w2) for f(w1), substitute w2 into f(w2) and recalculate to get a new f(w2), compare f(w2) with f(w1), repeat the above steps to get w2 corresponding to the minimum value of f(w2);

[0024] If f(w1)<f(w2), w2=w1-a;

[0025] Substitute f(w2) for f(w1), substitute w2 into f(w2) and recalculate to get a new f(w2), compare f(w2) with f(w1), repeat the above steps, and get w2 corresponding to the minimum value of f(w2).

[0026] Furthermore, in S2, the equivalent simplified theoretical mathematical model of the sediment content measurement after n iterations of machine learning is:

[0027] Y=(W 21 W 22 W 23 ...W2n )kX+b;

[0028] Among them, W 21 Parameters inferred for the first machine learning;

[0029] W 22 Parameters inferred for the second machine learning;

[0030] W 23 Parameters inferred for the third machine learning;

[0031] W 2n Parameters inferred for the nth machine learning;

[0032] k and b are theoretical coefficients in the mathematical model of the PLC measurement program.

[0033] Beneficial effects of the present invention:

[0034] The present invention enables the sensor to effectively approach the true value under the training of true value samples. Its goal is not only to make the value of a certain measurement effectively approach the true value, but also to expect that the probability of minimizing the gap with the true value in future measurements is relatively high. This machine learning algorithm based on artificial intelligence effectively expands the application scope of the field of smart sensors, and its application scenarios are broad and its practicality is strong. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a graph showing the reduction in error ratio after multiple iterations of machine learning in the method of the present invention;

[0036] Figure 2 It is a logic diagram of the PLC learning program of the present invention;

[0037] Figure 3 It is the theoretical curve diagram of the program application when f(w1)>f(w2) and f(w1)<f(w2);

[0038] Figure 4 It is a diagram of the human-computer interaction interface of the present invention;

[0039] Figure 5 It is a diagram of the group situation before the static test program of the method of the present invention is executed;

[0040] Figure 6 This is a diagram of the execution of a machine learning program in a static test of the method of the present invention. DETAILED DESCRIPTION

[0041] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0042] The principle of the present invention is:

[0043] The present invention applies artificial intelligence theory and adopts machine self-learning to design PLC programs. Theoretically, a linear operation model is designed based on the least square method; then the model is solved to find a straight line in linear regression so that the sum of the Euclidean distances of all samples to the straight line is minimized.

[0044] First, compile the PLC program according to the design theory; second, use examples (artificially sampled data) to train the machine program, and calculate the PLC parameters by machine learning; then adjust the parameters in a timely manner so that the error probability between the sensor's measured value and the artificial sample in future measurements is minimized.

[0045] With the deepening of machine learning, that is, multiple iterative operations, the difference between the measured values ​​of the online monitoring system for sediment content passing through the machine and the measured values ​​of manual sampling can be effectively reduced, thereby improving the measurement accuracy and precision of the online monitoring system for sediment content passing through the machine; reducing manual configuration and adjustment of parameters to avoid the parameters causing greater errors in subsequent measurements; effectively reducing measurement errors caused by changes in ambient temperature and turbulent pressure pulsations in water bodies; and effectively avoiding small error fluctuations such as zero drift of sensor electronic components during the sensor calibration period, causing the system measured values ​​to exceed the allowable error.

[0046] Example 1

[0047] 1. System design

[0048] Machine self-learning used to implement artificial intelligence requires hardware support with a certain level of computing power.

[0049] In this embodiment, an independently controllable domestic PLC is used to build the system hardware core; the on-site measurement and control cabinet is configured according to the hardware configuration and the on-site environment.

[0050] PLC hardware should meet the following conditions:

[0051] User program of no less than 200K steps; no less than 1MByte of custom variables, including no less than 128kByte supporting power-off retention; instruction speed no less than 12k steps / ms; support for EtherNet / IP, ModbusTCP and other communication protocols to facilitate program download and firmware upgrade; support for expansion; terminal input power rated voltage 24V DC±10%; support short circuit protection, support reverse connection protection; no less than 12 channels of AI input, the input form should support 4~20mA current input, and the resolution should be no less than 20K.

[0052] The machine self-learning method in artificial intelligence theory is adopted to "learn" the sample data of manual sampling through PLC, and generate a new calculation model from the data. The sensor of the online monitoring system of sand content in the machine is a differential pressure online density meter sensor, and the output current value of the sensor is 4-20mA, which is a linear sensor.

[0053] 2. Mathematical model of measurement program in PLC

[0054]

[0055] I = INT_TO_REAL (var0 = IN);

[0056] OUTR=TO_REAL((OSH-OSL)*(I-ISL) / (ISH-ISL)+OSL);

[0057]

[0058] OT = (IN - INQSMD) * 1590;

[0059] Among them, IN is the data input current value of the model;

[0060] OUTR is the data output of the model;

[0061] ISH is the high limit of the input quantity domain;

[0062] ISL is the lower limit of the input range;

[0063] OSH is the high limit of the output range;

[0064] OSL is the lower limit of the output range;

[0065] I is the data type conversion of the input quantity;

[0066] INQSMD is the standard quantity for the density of clean water;

[0067] OT is the sand content of the model’s data output;

[0068] 1590 is the sensor range conversion parameter.

[0069] The equivalent simplified theoretical mathematical model is:

[0070] Y = kX + b;

[0071] Among them, Y is the value of sand content;

[0072] X is the current value of the linear sensor;

[0073] k is the proportionality coefficient;

[0074] b is the adjustment parameter for calculating the sediment content.

[0075] 3. Build the database

[0076] Design a database in the local PLC according to the mathematical model and algorithm, including input array, output array, and initial parameters;

[0077] Design the retention time, data type and application range of arrays and parameters;

[0078] Design the initial parameter values, set the initial value of the function independent variable w1 to 1, and set the learning depth a value in the range of 0.01 to 0.0001.

[0079] 3.1 Constructing an array of measured data

[0080] The design uses 10 sensor sampling values ​​as the array of measured data, X is the sensor sampling value, Y is the engineering value (the sand content value calculated by the mathematical model in the PLC measurement program based on the sensor sampling value (current value)). Array Y is used to enter the function calculation, and X is used to calibrate Y.

[0081] X={x1,x2,x3……x 10}

[0082] Y={y1,y2,y3……y 10}

[0083] 3.2 Constructing sample array

[0084] The design uses representative values ​​from 10 manually sampled sample values ​​(manually measured) to construct a sample array, and the sample array is used as the input value of the function operation.

[0085] Y′={y1′,y2′,y3′……y 10 ′}

[0086] 3.3 Design and calculation functions f(w1) and f(w2)

[0087] f(w1)=(y1w1-y1′) 2 +(y2w1-y2′) 2 +……(y 10 w1-y 10 ′) 2 ;

[0088] f(w2)=(y1w2-y1′) 2 +(y2w2-y2′) 2 +……(y 10 w2-y 10 ′) 2 ;

[0089] Here, w1 and w2 are function variables, which have different meanings from those in the mathematical model of sediment content measurement after subsequent iterations. Each machine learning calculation will infer the value of w2, which corresponds to the number of iterations of the mathematical model of sediment content measurement.

[0090] 3.4 Mathematical Theory and Algorithm Model of Machine Learning

[0091] The calculation model based on the least square method is used in accordance with the theoretical algorithm of machine learning and computer programming language to realize the self-learning algorithm.

[0092]

[0093] 3.5 Mathematical model for sediment content measurement after n iterations of machine learning

[0094] An equivalent simplified theoretical mathematical model is used here to express it.

[0095] Y=(W 21 W 22 W 23 ...W 2n )kX+b;

[0096] Among them, W 21 Parameters inferred for the first machine learning;

[0097] W 22 Parameters inferred for the second machine learning;

[0098] W 23 Parameters inferred for the third machine learning;

[0099] W 2n Parameters inferred for the nth machine learning;

[0100] k and b are theoretical coefficients in the mathematical model of the PLC measurement program.

[0101] It should be noted that this only means that the parameters are being adjusted, not multiplied.

[0102] 3.6 Machine Learning Operation Model Design

[0103] 3.6.1f(w1) model design

[0104]

[0105]

[0106] Calculation results:

[0107] f(w1)=((y1w1-y1′)2 +(y2w1-y2′) 2 +(y3w1-y3′) 2 +(y4w1-y4′) 2 +(y5w1-y5′) 2 +(y6

[0108] w1-y6′) 2 +(y7w1-y7′) 2 +(y8w1-y8′) 2 +(y9w1-y9′) 2 +(y 10 w1-y 10 ′) 2 );

[0109] Completed = 1.

[0110] 3.6.2f(w2) model design

[0111]

[0112]

[0113] w2=w1+a;

[0114] Calculation results:

[0115] f(w2)=((y1w2-y1′) 2 +(y2w2-y2′) 2 +(y3w2-y3′) 2 +(y4w2-y4′) 2 +(y5w2-y5′) 2 +(y6

[0116] w2-y6′) 2 +(y7w2-y7′) 2 +(y8w2-y8′) 2 +(y9w2-y9′) 2 +(y 10 w2-y 10 ′) 2 );

[0117] Completed = 1.

[0118] 3.6.3 Design of calculation model when f(w1)>f(w2)

[0119]

[0120] w2=w2+a;

[0121] Calculation results:

[0122] f(w2)=((y1w2-y1′) 2 +(y2w2-y2′) 2 +(y3w2-y3′) 2 +(y4w2-y4′) 2 +(y5w2-y5′) 2 +(y6

[0123] w2-y6′) 2 +(y7w2-y7′) 2 +(y8w2-y8′) 2 +(y9w2-y9′) 2 +(y 10 w2-y 10 ′) 2 );

[0124] Completed = 1.

[0125] 3.6.4 Calculation model design when f(w1)<f(w2)

[0126]

[0127] w2=w1-a;

[0128] Calculation results:

[0129] f(w2)=((y1w2-y1′) 2 +(y2w2-y2′) 2 +(y3w2-y3′) 2 +(y4w2-y4′) 2 +(y5w2-y5′) 2 +(y6

[0130] w2-y6′) 2 +(y7w2-y7′) 2 +(y8w2-y8′) 2 +(y9w2-y9′) 2 +(y 10 w2-y 10 ′) 2 );

[0131] Completed = 1.

[0132] 3.7 Programming

[0133] Design PLC learning program based on least squares method; Figure 2 The program logic diagram is shown in the figure.

[0134] The single neural network structure is simulated to solve the weight parameters. The target result is the minimum error value under multiple measurement results. The expectation is that the probability of obtaining the minimum error in future measurements will be higher.

[0135] The program starts when the machine learning variable = 1; the program calculation ends when the limit value in the function is found, and the value of w2 is automatically calculated.

[0136] 3.7.1 Program execution action when f(w1)>f(w2)

[0137] When f(w1)>f(w2), the action executed by the process is equivalent to Figure 3 Find the extreme values ​​from left to right in the curve shown, corresponding to w2.

[0138] 3.7.2 Program execution actions when f(w1)<f(w2)

[0139] When f(w1)<f(w2), the action executed by the process is equivalent to Figure 3 Find the extreme value from right to left in the curve shown, corresponding to w2.

[0140] 3.8 Human-computer interface design

[0141] like Figure 4 As shown in the figure, a human-machine interface is designed on a modern PLC touch screen for parameter setting, data collection, sample input and display of learning results. A sampling button is designed on the left side for collecting data from the sampling sensor, and the sample input permission is turned on at the same time, waiting for the entry of the sample value; a sample value input box is designed on the right side, where the manually collected sample value can be entered into the system. A confirmation learning button is designed at the bottom. After ten samplings and sample inputs are completed, confirm to enter the machine self-learning program here, that is, the machine learning variable = 1.

[0142] 4. System debugging, operation and evaluation

[0143] 4.1 System static debugging and effectiveness test

[0144] The system static debugging uses simulated test data to verify the effectiveness of the program. According to the theoretical analysis of program design, we select two different sets of data to verify the program effectiveness. The first set selects a set of data greater than the measured value; the second set selects data less than the measured value.

[0145] 4.1.1 The first set of data experiments

[0146] Input data:

[0147] X={x1,x2,x3……x 10}={12000,12000,12000……12000}

[0148] Y={y1,y2,y3……y 10}={318,318,318,……318}

[0149] Y′={y1′,y2′,y3′……y 10 '}={338,338,338,……338}

[0150] Program data input is as follows Figure 5 .

[0151] Through machine learning, it is estimated that w2 = 1.06

[0152] How machine learning programs perform Figure 6 .

[0153] Through the calculation and analysis of the machine learning results of the first set of data, see Table 1 for details; it was found that the error ratio was reduced by 95%; the experimental effectiveness was evaluated as effective.

[0154] Table 1 The first set of data machine learning static test data

[0155]

[0156] 4.1.2 The second set of data experiments

[0157] Input data:

[0158] X={x1,x2,x3……x 10}={12000,12000,12000……12000}

[0159] Y={y1,y2,y3……y 10}={318,318,318……318}

[0160] Y′={y1′,y2′,y3′……y 10 ′}={298,298,298……298}

[0161] Through machine learning, we can estimate w2 = 0.93

[0162] Through the calculation and analysis of the machine learning results of the second set of data, see Table 2 for details; it was found that the error ratio was reduced by 88.7%; the experimental effectiveness was evaluated as effective.

[0163] Table 2 The second set of data machine learning static test data

[0164]

[0165] 4.2 System truth test and effectiveness evaluation

[0166] The system is tested using historical data from 2024.

[0167] A small amount of observation data from the tailwater of Unit 1 of Xiaolangdi Power Plant was used for system testing to evaluate the error reduction rate after machine self-learning. From the results of machine learning, the maximum error reduction rate is 98.77% for the comparison data No. 5928; the minimum error reduction rate is 35.88% for the comparison data No. 6638. The system runs effectively, and detailed data are shown in Table 3.

[0168] Table 3 System truth data machine learning experiments and effectiveness evaluation

[0169]

[0170] The present invention is not limited to the above-mentioned optimal implementation mode. Anyone can derive other various forms of products under the inspiration of the present invention. However, no matter what changes are made in the shape or structure, all technical solutions that are the same or similar to those of the present application fall within the protection scope of the present invention.

Claims

1. A system for improving the accuracy of monitoring sediment content in water bodies based on machine learning, including a linear sensor and a PLC, characterized in that: The PLC is provided with a mathematical model of a machine learning program. Based on the machine self-learning method in artificial intelligence theory, the mathematical model of the machine learning program in the PLC is trained with the PLC measured sediment content data and the manually sampled sample data to obtain a new mathematical model. The monitoring data of the linear sensor is calculated through the new mathematical model to obtain the data of the sediment content in the water body. With the deepening of machine learning, the error between the measured value of the online monitoring system for the sediment content in the water body and the measured value of the manual sampling is gradually reduced, thereby improving the measurement accuracy of the online monitoring system for the sediment content in the water body.

2. The system for improving the accuracy of monitoring water sediment content based on machine learning according to claim 1 is characterized in that: The equivalent simplified mathematical model of the measurement program in the PLC is Y=kX+b, wherein Y is the sand content value; X is the current value of the linear sensor; k is the proportional coefficient; and b is the adjustment parameter for calculating the sand content.

3. The system for improving the accuracy of monitoring sediment content in water bodies based on machine learning according to claim 1 is characterized in that: The algorithm for training the mathematical model of the machine learning program in the PLC is based on the least squares method.

4. The method for improving the accuracy of monitoring sediment content in water according to any one of claims 1 to 3, characterized in that: The following steps are involved: S1, build a PLC database, the database includes input array, output array and initial parameters; S2, iterating the input array in S1 through the calculation function to obtain the output array, and continuously correcting the mathematical model of the machine learning program in the PLC through the output array to obtain a new mathematical model for sand content measurement; S3. The monitoring data of the linear sensor is calculated through the new mathematical model obtained in S2 to obtain the data of the sediment content in the water body.

5. The method for improving the accuracy of monitoring sediment content in water bodies according to claim 4 is characterized in that: In the S1-S2 described above, the input array includes a measured array and a sample array; the measured array includes a sensor sampling value array and a measured sediment content value array calculated linearly by PLC; the sample array is the value of samples manually sampled at the same time and location as the sensor; the measured sediment content value array and the sample array are used as input values ​​of the calculation function; the sensor sampling value array is used to calibrate the measured sediment content value array; the initial parameters include the function independent variable w1 and the learning depth a, the initial value of w1 is 1, and the value range of a is 0.01 to 0.0001.

6. The method for improving the accuracy of monitoring sediment content in water bodies according to claim 5, characterized in that: In the above S2, the calculation function is: f(w1)=(y1w1-y1′) 2 +(y2w1-y2′) 2 +……(y 10 w1-y 10 ′) 2 ; f(w2)=(y1w2-y1′) 2 +(y2w2-y2′) 2 +……(y 10 w2-y 10 ′) 2 ; Among them, w1 and w2 are function independent variables; y n is the value in the numerical array of measured sediment content, (n=1,2,3…10); y n ' is the value in the sample array, (n = 1, 2, 3...10); Each machine learning calculation will infer the value of w2, which corresponds to the number of times the mathematical model of sediment content measurement is iterated w 2n .

7. The method for improving the accuracy of monitoring sediment content in water bodies according to claim 6, characterized in that: The measured sediment content value array and the sample array are iterated through a calculation function, where w2 = w1 + a; During the iteration, if f(w1)>f(w2), w2=w2+a; Substitute f(w2) for f(w1), substitute w2 into f(w2) and recalculate to get a new f(w2), compare f(w2) with f(w1), repeat the above steps to get w2 corresponding to the minimum value of f(w2); If f(w1)<f(w2), w2=w1-a; Substitute f(w2) for f(w1), substitute w2 into f(w2) and recalculate to get a new f(w2), compare f(w2) with f(w1), repeat the above steps, and get w2 corresponding to the minimum value of f(w2).

8. The method for improving the accuracy of monitoring sediment content in water bodies according to any one of claims 4 to 7, characterized in that: In S2, the equivalent simplified theoretical mathematical model of the sediment content measurement after n iterations of machine learning is: Y=(W 21 W 22 W 23 ……W 2n )kX+b; Among them, W 21 Parameters inferred for the first machine learning; W 22 Parameters inferred for the second machine learning; W 23 Parameters inferred for the third machine learning; W 2n Parameters inferred for the nth machine learning; k and b are theoretical coefficients in the mathematical model of the PLC measurement program.

Images