CN-121980763-A - Virtual sensor configuration optimization method and system for industrial soft measurement

CN121980763ACN 121980763 ACN121980763 ACN 121980763ACN-121980763-A

Abstract

A virtual sensor configuration optimization method and system for industrial soft measurement comprise the steps of generating a plurality of different initial virtual sensor configuration schemes based on a candidate physical sensor set of equipment to be monitored, calculating a synthesis evaluation index for each virtual sensor configuration scheme, determining an optimal synthesis index and an optimal configuration scheme, selecting, exchanging and mutating the virtual sensor configuration schemes under any iteration times, updating the virtual sensor configuration schemes, determining the corresponding optimal synthesis index and the optimal configuration scheme, and outputting the corresponding optimal configuration scheme until the difference of the optimal synthesis indexes under adjacent iteration times is smaller than a preset threshold value. According to the invention, the configuration scheme is subjected to iterative optimization by constructing the multidimensional synthesis evaluation index integrating the detection performance, the information efficiency, the cost complexity and the system robustness, so that the optimization target meeting the requirements of industrial soft measurement precision and stability is realized.

Inventors

WANG YOUYUAN
GUO YONGBO
GUO JIALE
ZOU JIAHAO
HONG XINYUAN
HUANG DONGNING
Leng Yuxiang
NIE YAQI

Assignees

重庆大学

Dates

Publication Date: 20260505
Application Date: 20251225

Claims (12)

1. The virtual sensor configuration optimization method for industrial soft measurement is characterized by comprising the following steps of: s1, generating a plurality of different initial virtual sensor configuration schemes based on a candidate physical sensor set of equipment to be monitored; S2, for each virtual sensor configuration scheme, collecting corresponding historical samples, sequentially calculating detection confidence coefficient, characteristic contribution entropy, cost complexity coefficient and vulnerability index, taking the maximum value of the characteristic contribution entropy subtracted from 0 and 1 as information efficiency, taking the product of the detection confidence coefficient and the information efficiency as a molecule, multiplying the vulnerability index by the cost complexity coefficient as a denominator, and calculating a synthesis evaluation index; S3, selecting, exchanging and mutating the virtual sensor configuration scheme under any iteration times, updating the virtual sensor configuration scheme, recalculating the synthesized evaluation index of the updated configuration scheme, determining the corresponding optimal synthesis index and the optimal configuration scheme, and outputting the corresponding optimal configuration scheme until the difference of the optimal synthesis indexes under the adjacent iteration times is smaller than a preset threshold value.
2. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 1, wherein the method comprises the following steps: In S1, initializing the number of each sensor in a candidate physical sensor set, and generating N different initial virtual sensor configuration schemes; For each virtual sensor configuration scheme, analyzing the contribution degree of each input physical signal to the output of the virtual sensor to obtain the original contribution degree of K input signals; when all original contribution degrees are 0 or all prediction results of the historical samples are wrong, the virtual sensor configuration is indicated to be incapable of extracting effective information from input signals, and is regarded as invalid configuration, and the corresponding virtual sensor configuration scheme is reinitialized.
3. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 1, wherein the method comprises the following steps: the process of calculating the detection confidence in S2 is as follows: For each virtual sensor configuration scheme, collecting sensor output data of a corresponding virtual sensor within M preset time lengths to serve as M historical samples of the corresponding sensor, operating a virtual sensor model to obtain M predicted output values, accumulating absolute values of differences between the predicted output of each historical sample and corresponding real labels, dividing the absolute values by the total number of the historical samples to obtain average calibration errors, and subtracting the maximum value between the value obtained by the average calibration errors and 0 from 1 to serve as detection confidence of the corresponding scheme.
4. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 1, wherein the method comprises the following steps: the process of calculating the characteristic contribution entropy in the S2 is as follows: The method comprises the steps of obtaining original contribution degrees of all input signals in each virtual sensor configuration scheme, taking absolute values of the original contribution degrees of the corresponding input signals as molecules for each input signal, accumulating the contribution degrees of all input signals as denominators, calculating relative contribution ratios of the corresponding input signals, calculating products of logarithms of the relative contribution ratios and the relative contribution ratios of each input signal respectively, accumulating the corresponding products of all input signals, and taking negative numbers to obtain characteristic contribution entropy.
5. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 1, wherein the method comprises the following steps: The process of calculating the cost complexity coefficient in S2 is: The method comprises the steps of obtaining single reasoning time of a virtual sensor model under each virtual sensor configuration scheme, taking the ratio of the single reasoning time to a preset maximum reasoning time threshold value as a calculation cost component of the corresponding configuration scheme, dividing the total cost of physical sensors in the configuration scheme by the total cost when all candidate sensors reach the maximum allowable configuration quantity to obtain a hardware cost component of the corresponding configuration scheme, and weighting the calculation cost component and the hardware cost component under the virtual sensor configuration scheme to obtain a cost complexity coefficient.
6. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 1, wherein the method comprises the following steps: The process of calculating the vulnerability index in S2 is: For each virtual sensor configuration scheme, running a virtual sensor model under a normal working condition, outputting a corresponding prediction result based on historical samples under M corresponding configuration schemes, dividing the number of correctly predicted samples by the total number of samples to obtain the prediction accuracy under the normal working condition, designing N typical pressure test scenes, running the virtual sensor model under each pressure scene, calculating the prediction accuracy under the scene, dividing the difference between the prediction accuracy under the normal working condition and the prediction accuracy under different pressure test scenes by the prediction accuracy under the normal working condition to obtain the relative performance reduction rate under the corresponding pressure test scenes, and averaging the square of the relative performance reduction rates under all the pressure test scenes to obtain the vulnerability index of the corresponding configuration scheme.
7. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 1, wherein the method comprises the following steps: the process of selecting the virtual sensor configuration scheme in the S3 is as follows: For each virtual sensor configuration scheme under each iteration number, taking the synthesis evaluation index under the corresponding configuration scheme as a numerator, taking the synthesis evaluation index accumulation of all the virtual sensor configuration schemes under the current iteration number as a denominator, calculating the selection probability of the configuration scheme selected under the next iteration number of the corresponding configuration scheme, and selecting a preset number of virtual sensor configuration schemes as the optimal configuration scheme based on the selection probability of all the schemes to obtain an optimal configuration scheme set.
8. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 1, wherein the method comprises the following steps: The process of exchanging the virtual sensor configuration scheme in S3 is as follows: for any two preferred configuration schemes in the preferred configuration scheme set, averaging absolute values of original contribution degrees of each sensor position in the two preferred configuration schemes to obtain comprehensive contribution degrees of the corresponding sensors; The method comprises the steps of selecting a preset exchange number position with the highest comprehensive contribution degree from all exchange point positions in two corresponding preferable configuration schemes as an intersection point, and exchanging sensor configuration values of the two corresponding preferable configuration schemes at the intersection point to generate two new preferable configuration sub-schemes.
9. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 1, wherein the method comprises the following steps: the process of performing mutation operation on the virtual sensor configuration scheme in the S3 is as follows: for each sensor position in the preferred configuration sub-scheme, calculating a trend factor for adjusting the number of corresponding sensor configurations; if the trend factor is larger than a preset first adjustment threshold value, multiplying the trend factor by the maximum configuration quantity of the corresponding sensors, rounding up the trend factor, and adding the original configuration quantity of the corresponding sensor positions to obtain the quantity to be updated; If the number of the sensor positions is smaller than a preset second adjustment threshold value, multiplying the trend factor by the maximum configuration number of the corresponding sensor and rounding downwards to obtain the number of the sensor positions to be updated as the reduction number of the original configuration number of the corresponding sensor positions; and updating the configuration quantity of all the sensor positions of all the preferred configuration sub-schemes to obtain all the updated preferred configuration sub-schemes under the current iteration times.
10. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 9, wherein the trend factor is calculated as follows: ; In the formula, A trend factor of the sensor position k in the ith preferred configuration sub-scheme; Original contribution of the corresponding sensor physical input to sensor position k; Configuring the maximum value of the original contribution degree of all the sensors in the sub-scheme for preference; For the remaining cost budget of the current solution, Hardware cost for the ith preferred configuration sub-scheme; is a preset cost tolerance threshold; 、 is a weight coefficient and is used for balancing performance optimization and cost utilization.
11. The method for optimizing virtual sensor configuration for industrial soft measurement according to claim 1 or 10, wherein: the process of updating the virtual sensor configuration scheme in S3 is: Directly copying the schemes with the preset reserved quantity and highest synthesis evaluation index value in the virtual sensor configuration scheme set before the selection operation under the current iteration times into the scheme set of the next round; and merging all the updated preferable configuration sub-schemes with the reserved configuration schemes, if the total number exceeds N, sequencing from high to low according to the combined evaluation index value, and reserving the first N schemes to obtain an updated virtual sensor configuration scheme set, namely a scheme set of the next iteration number.
12. A virtual sensor configuration optimization system for industrial soft measurements using the method of any one of claims 1-11, comprising: The method comprises the steps of initializing a configuration scheme module, and generating a plurality of different initial virtual sensor configuration schemes based on a candidate physical sensor set of equipment to be monitored; The configuration scheme evaluation index calculation module is used for collecting corresponding historical samples for each virtual sensor configuration scheme, sequentially calculating detection confidence coefficient, characteristic contribution entropy, cost complexity coefficient and vulnerability index, taking the maximum value of the characteristic contribution entropy subtracted by 0 and 1 as information efficiency, taking the product of the detection confidence coefficient and the information efficiency as a molecule, multiplying the vulnerability index by the cost complexity coefficient as a denominator, and calculating a synthesis evaluation index; The configuration scheme optimizing module is used for selecting, exchanging and mutating the virtual sensor configuration scheme under any iteration times and updating the virtual sensor configuration scheme, calculating the synthesis evaluation index of the updated configuration scheme again and determining the corresponding optimal synthesis index and the optimal configuration scheme until the difference of the optimal synthesis index under the adjacent iteration times is smaller than a preset threshold value, and outputting the corresponding optimal configuration scheme.

Description

Virtual sensor configuration optimization method and system for industrial soft measurement Technical Field The invention belongs to the technical field of sensor optimal configuration, and particularly relates to a virtual sensor configuration optimization method and system for industrial soft measurement. Background The virtual sensor technology is a key means in the field of industrial equipment state monitoring and fault diagnosis, and is characterized in that a data driving or mechanism model is constructed through a group of easily-deployed and low-cost actual physical sensor signals, so that on-line inference and inversion of key state parameters which are difficult to directly measure or have extremely high measurement cost are realized. In the engineering design and implementation process of the virtual sensor, the optimal configuration scheme is screened from the candidate physical sensor set, and the method is a core link for determining the final performance, economy and reliability of the whole system. At present, the existing sensor configuration optimization method comprises the steps that a wireless sensor configuration optimization method based on a constraint multi-objective optimization algorithm is provided by a patent with a publication number of CN114139459A, a double-objective optimization model with information effectiveness and energy consumption is established, network connectivity and reliability are used as constraint conditions, an improved double-population algorithm is adopted to solve the problem, synchronous optimization of wireless sensors and sink positions in spacecraft structure health monitoring is achieved, the sensor optimization configuration method of an electromechanical product is disclosed by a Chinese patent CN115616919A, the number of sensors, cost and fault probability are used as targets based on an incidence matrix of a fault mode and the sensors, a multi-objective optimization model is established by taking the number of the sensors, the cost and the fault probability as targets, the fault detection rate, the isolation rate and the like as constraints, a group of pareto optimal configuration scheme is provided for physical sensors of electromechanical products by means of discrete multi-objective particle swarm algorithm, an heterogeneous detection sensor optimization configuration method based on an improved subtraction average optimizer is provided by the patent with the publication number of CN120633213A, a gain ratio of a sensor can be optimized by a quantitative remote detection, recognition and a gain ratio of a sensor is improved, and a gain ratio of a sensor can be optimized by a machine is established by adopting a gain optimization mode. While the prior art provides a basic idea for configuration selection of virtual sensors, they all suffer from systematic drawbacks that are inherent and cannot be overcome by simple modification. The method based on the single performance index pursues prediction precision on one side, ignores key engineering dimensions such as hardware cost, calculation complexity, information redundancy among signals, robustness of the system under abnormal working conditions and the like, and makes the selected scheme difficult to stably operate in an actual industrial environment due to overhigh cost, low efficiency or poor anti-interference capability. Another comprehensive evaluation method using weighted summation, although considering a plurality of targets, has a linear compensation mechanism that allows serious defects in one dimension to be masked by the advantages in other dimensions, so that a scheme with fatal short plates can be selected, and meanwhile, the setting of each index weight is highly dependent on subjective experience of a designer, so that the evaluation result lacks objectivity and repeatability. Disclosure of Invention The invention provides a virtual sensor configuration optimization method and a system for industrial soft measurement, which are used for solving the technical problems that in the prior art, when the virtual sensor configuration for the industrial soft measurement is carried out, the multi-dimensional performances such as inference precision, information efficiency, implementation cost, abnormal robustness and the like are difficult to objectively, uniformly quantify and automatically comprehensively optimize. The method comprises the steps of generating a plurality of different initial virtual sensor configuration schemes based on a candidate physical sensor set of equipment to be monitored, calculating a synthesis evaluation index for each virtual sensor configuration scheme, determining an optimal synthesis index and an optimal configuration scheme, selecting, exchanging and mutating the virtual sensor configuration schemes under any iteration times, updating the virtual sensor configuration schemes, determining the corresponding optimal synthesis index and the optimal configurati