CN117609703B

CN117609703B - Complex equipment health management method and system integrating multi-source heterogeneous information

Info

Publication number: CN117609703B
Application number: CN202410073360.8A
Authority: CN
Inventors: 周志杰; 陈满林; 冯志超; 胡昌华; 姚鑫智; 韩晓霞; 胡明
Original assignee: Rocket Force University of Engineering of PLA
Current assignee: Rocket Force University of Engineering of PLA
Priority date: 2024-01-18
Filing date: 2024-01-18
Publication date: 2024-04-26
Anticipated expiration: 2044-01-18
Also published as: CN117609703A

Abstract

The invention discloses a complex equipment health management method and system integrating multisource heterogeneous information, and relates to the technical field of complex equipment health management. The data interpretation module performs abnormal recognition and preprocessing on the operation parameters of the complex equipment, and determines a plurality of preprocessed operation parameters and the change trend of each operation parameter; the health evaluation module determines the real-time health state of the complex equipment by utilizing an ER algorithm according to various preprocessed operation parameters; the health trend analysis module determines the predicted health state of the complex equipment by using a prediction algorithm according to various preprocessed operation parameters; and the decision support module determines the standby quantity of each device in the complex equipment according to the operation parameters after the various preprocessing. The invention can carry out health management on the complex equipment based on the multi-source heterogeneous information of the complex equipment, and improves the rationality of the health management of the complex equipment.

Description

Complex equipment health management method and system integrating multi-source heterogeneous information

Technical Field

The invention relates to the technical field of complex equipment health management, in particular to a complex equipment health management method and system integrating multi-source heterogeneous information.

Background

The complex equipment is an industrial product with complex structure and multiple parts, and combines optical, electric and mechanical power technologies. With the rapid advancement of the industrial process of china, more and more complex devices are appearing in various large-scale devices to become device cores. Therefore, to ensure the normal operation of large-scale equipment, developing health management on complex equipment has become an important component of the daily maintenance of large-scale equipment. Health management is a method which covers multiple aspects such as monitoring, diagnosis, prediction and maintenance to ensure the normal operation of a managed object and prolong the service life of the managed object, and is often applied to the production and use processes of various large-scale equipment. The method is characterized in that health management research is carried out on complex equipment, and aims to carry out comprehensive, complete and interpretable health management operation on the complex equipment by utilizing historical data and related expert information accumulated in the working process of the complex equipment and combining various methods such as data analysis, prediction and evaluation, so that the working state of the complex equipment is obtained, and safe and effective operation of the complex equipment is ensured. The method has the advantages that the comprehensive performance and the health situation of the complex equipment can be mastered in time, the potential risk of the equipment is avoided in time, the equipment maintenance cost is reduced, and the preventive maintenance of the equipment is realized.

However, the internal structure of the complex equipment is complex and consists of various types of components, and the operation conditions of the components must be comprehensively considered when the complex equipment is subjected to health management so as to obtain complete, accurate and interpretable health state information. This requires the information of each component to be incorporated into a system, and the health status of the complex equipment is analyzed by a comprehensive utilization method. However, since the components constituting the complex device originate from different disciplines, the index structure representing the state thereof is not uniform, and therefore, how to fuse multi-source heterogeneous information when performing complex device health management becomes a primary problem that must be solved. In addition, because the complex equipment is in the key position of various large-scale equipment, when the complex equipment is in health management, a certain interpretation is required from the process to the result so as to ensure the reliability of health management. Currently, most of the common health management methods are based on model-based or data-driven methods. The model-based method needs to establish an accurate mathematical model on the basis of grasping the working mechanism of the study object, and then utilizes the model and data to analyze the health condition. However, due to the complexity of the complex equipment, the working mechanism of the complex equipment needs more expertise and accumulated working experience for many years, and the complex equipment is difficult to achieve. Moreover, even if the working mechanism is mastered, since the health management of the complex equipment needs to consider various indexes, it is difficult to construct the mathematical model to achieve the required accuracy, and even the mathematical model cannot be constructed, so that the interpretability of the health management cannot be ensured. The data-driven based approach requires training the model with a large amount of data covering all modalities to improve the accuracy of the model. However, when the complex device is applied to some high-precision tip devices, data meeting the training requirements of the model are difficult to collect, so that the model obtained by training often has a lack of fitting phenomenon. Meanwhile, the training process is difficult to understand, so that the interpretability of health management is not guaranteed. Therefore, finding a method of complex device health management with interpretability is another problem that is currently urgently needed to be solved.

Health management of complex equipment involves multiple links such as data interpretation, health assessment, health trend analysis, decision support and the like. The current researches are scattered and independent, which results in complicated basic data patterns, messy result expression modes and lack of support among each other. Therefore, constructing a unified, complete platform-level health management system for complex devices is a current urgent need.

Disclosure of Invention

The invention aims to provide a complex equipment health management method and system integrating multi-source heterogeneous information, which can be used for carrying out health management on complex equipment based on the multi-source heterogeneous information of the complex equipment and improving the rationality of the health management of the complex equipment.

In order to achieve the above object, the present invention provides the following solutions:

a complex device health management system that fuses multisource heterogeneous information, comprising: the system comprises a data interpretation module, a health assessment module, a health trend analysis module and a decision support module;

the health evaluation module, the health trend analysis module and the decision support module are all connected with the data interpretation module;

The data interpretation module is used for carrying out abnormal recognition and preprocessing on the operation parameters of the complex equipment and determining various preprocessed operation parameters and the change trend of each operation parameter;

The health evaluation module is used for determining the real-time health state of the complex equipment by utilizing an ER algorithm according to various preprocessed operation parameters;

The health trend analysis module is used for determining the predicted health state of the complex equipment by using a prediction algorithm according to various preprocessed operation parameters; the prediction algorithm comprises a gray prediction method, an ARIMA prediction method and an LSTM prediction method;

the decision support module is used for determining the standby quantity of each device in the complex equipment according to various preprocessed operation parameters.

A complex device health management method integrating multisource heterogeneous information comprises the following steps:

acquiring various operation parameters of complex equipment;

performing anomaly identification and preprocessing on the operation parameters of the complex equipment, and determining various preprocessed operation parameters and the change trend of each operation parameter;

Determining the real-time health state of the complex equipment by utilizing an ER algorithm according to various preprocessed operation parameters;

According to the operation parameters after various preprocessing, the predicted health state of the complex equipment is determined by using a prediction algorithm; the prediction algorithm comprises a gray prediction method, an ARIMA prediction method and an LSTM prediction method;

determining the standby quantity of each device in the complex equipment according to the operation parameters after various pretreatment;

the trend of each operation parameter, the real-time health status of the complex equipment, the predicted health status and the standby quantity of each device in the complex equipment are stored and displayed.

According to the specific embodiment provided by the invention, the invention discloses the following technical effects:

According to the complex equipment health management method and system integrating the multi-source heterogeneous information, provided by the invention, the evidence reasoning algorithm is taken as a core, and the health management of the complex equipment can be completed by combining a plurality of prediction algorithms and data processing algorithms, so that limited data can be comprehensively utilized, and health management works such as health assessment, health trend analysis, life prediction and the like with interpretability are carried out on the complex equipment, so that the current and future health states of the complex equipment are further reflected.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings that are needed in the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

Fig. 1 is a schematic structural diagram of a complex device health management system for merging multi-source heterogeneous information in embodiment 1 of the present invention;

FIG. 2 is a flowchart of a complex device health management method for merging multi-source heterogeneous information in embodiment 2 of the present invention;

FIG. 3 is a flow chart of the laser inertial measurement unit data interpretation method based on evidence reasoning in embodiment 2 of the invention;

FIG. 4 is a diagram showing the status of K1bz index data in embodiment 2 of the present invention;

FIG. 5 is a graph showing reliability versus quality factor in example 2 of the present invention;

FIG. 6 is a schematic diagram of experimental data of an accelerometer and gyroscope of example 2 of the present invention;

FIG. 7 is a diagram showing the result of health evaluation in example 2 of the present invention;

FIG. 8 is a schematic diagram of the evaluation result of BP neural network in embodiment 2 of the present invention;

FIG. 9 is a schematic diagram of a fuzzy inference evaluation result in embodiment 2 of the present invention;

FIG. 10 is a diagram of a system for analyzing health trend in example 2 of the present invention;

FIG. 11 is a flow chart of a method for analyzing health trend of a laser inertial measurement unit based on evidence reasoning in embodiment 2 of the present invention;

FIG. 12 is a schematic diagram of the variation of the drift coefficient and the prediction result of the 1 st order item of the gyro in embodiment 2 of the present invention;

FIG. 13 is a schematic diagram of the variation of the primary drift coefficient and the prediction result of the gyro 1 in embodiment 2 of the present invention;

FIG. 14 is a diagram showing the variation of the zero-order term drift coefficient and the prediction result of the accelerometer 1 in example 2 of the present invention;

FIG. 15 is a graph showing the positive scale factor change and predicted result of the accelerometer 1 according to example 2 of the present invention;

FIG. 16 is a schematic illustration of scale factor variation and prediction results for accelerometer 1;

FIG. 17 is a graph showing the predicted health of the laser inertial measurement unit in example 2 of the present invention;

FIG. 18 is a graph showing the health of an inertial sensor over time in example 2 of the present invention;

FIG. 19 is a flow chart of a preferred decision making method of the inertial navigation system in embodiment 2 of the present invention;

FIG. 20 is a diagram of the test data of the monitor index according to the embodiment 2 of the present invention;

FIG. 21 is a schematic diagram of the result of the reduced evaluation of the preferred decision model of the inertial navigation system according to embodiment 2 of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.

Example 1

As shown in fig. 1, this embodiment provides a complex device health management system that merges heterogeneous multi-source information, including: the system comprises a data interpretation module, a health assessment module, a health trend analysis module and a decision support module. The health evaluation module, the health trend analysis module and the decision support module are all connected with the data interpretation module. The data interpretation module is used for carrying out abnormal recognition and preprocessing on the operation parameters of the complex equipment and determining various preprocessed operation parameters and the change trend of each operation parameter. The health evaluation module is used for determining the real-time health state of the complex equipment by utilizing an ER algorithm according to various preprocessed operation parameters. The health trend analysis module is used for determining the predicted health state of the complex equipment by using a prediction algorithm according to various preprocessed operation parameters. The prediction algorithm includes gray prediction, ARIMA prediction, and LSTM prediction. The decision support module is used for determining the standby quantity of each device in the complex equipment according to the operation parameters after the pretreatment.

The complex device health management system for fusing multi-source heterogeneous information provided in this embodiment further includes: the system comprises an operation data acquisition module, a display module and a storage module; the operation data acquisition module is connected with the data interpretation module. The operation data acquisition module is used for acquiring various operation parameters of the complex equipment. The display module is respectively connected with the health evaluation module, the health trend analysis module, the decision support module and the data interpretation module. The display module is used for displaying the change trend of each operation parameter, the real-time health state of the complex equipment, the predicted health state of the complex equipment and the standby quantity of each device in the complex equipment. The storage module is respectively connected with the health evaluation module, the health trend analysis module, the decision support module and the data interpretation module. The storage module is used for storing the change trend of each operation parameter, the real-time health state of the complex equipment, the predicted health state of the complex equipment and the standby quantity of each device in the complex equipment.

The data interpretation is a basis and an entry of health management, and is used for completing data reading and preprocessing operations and displaying the change trend and abnormal (exceeding a threshold value) condition of the data, in the process of data interpretation, according to health indexes of specific complex equipment, index data of the complex equipment are read from a data source, and upper and lower limits (threshold values) of each index are obtained, wherein the thresholds can be used as references to judge whether the indexes are abnormal or not. The data source is usually a file under a specified path, and after index data is acquired from the file, a data file with a specified format is generated so as to facilitate the use of subsequent preprocessing operation; and applying a preset pretreatment method to carry out pretreatment operation on the obtained index data. After the pretreatment operation is finished, a pretreated data file is generated under a specified path and is used for data abnormality judgment, subsequent health trend analysis, life prediction and other operations; and reading the preprocessed data file, performing abnormality judgment on the data in the preprocessed data file, judging whether each index data exceeds a corresponding threshold value, and returning an abnormality judgment result. In this step, the preprocessed data within the specified range is compared with the threshold value of each index. If the value of a certain index exceeds the set threshold range, judging the current index as data abnormality, and generating a corresponding abnormality judgment result; in the abnormal judging process, recording the number (out-of-tolerance times) of data exceeding a threshold value in index data, and storing a recording result into an out-of-tolerance times record file under a specified path so as to facilitate the reading of the out-of-tolerance times of the index by subsequent operation; and displaying the preprocessed data in the form of a line graph so as to observe the change trend of the data. By drawing the line graph, the change trend of the index along with time can be intuitively displayed. This helps to observe the overall trend of the health data, identify any significant rising or falling trends, and which data exceeds the threshold.

The health state evaluation aims at evaluating the performance state of the complex equipment according to the monitored equivalent data of each index of the complex equipment, and provides a basis for the next analysis of the health state; and according to specific technical standards of the complex equipment and combining expert knowledge, compiling a performance state table corresponding to the running precision error of the complex equipment. Then, converting the error into a corresponding performance grade according to the reference value corresponding to the formulated table; calculating the weight of each input by a coefficient of variation method; calculating the reliability of each input; combining the reliability and the weight into a new quality factor; converting the uncertain input into a form of belief distribution and carrying out evidence fusion by considering the evidence quality factors; and calculating the output confidence, and outputting the performance grade corresponding to the current complex equipment.

The health trend analysis is to fuse analysis results by utilizing a evidence reasoning algorithm on the basis of predicting and analyzing the existing data, and finally, the health trend is displayed in a quantitative mode to acquire all index data preprocessed by a data interpretation part; calculating the index data by using a prediction model to obtain a prediction result of each index; converting the prediction result into probability distribution according to the reference grade of each index; and establishing an information fusion model based on the evidence reasoning algorithm. The probability distribution of each index prediction result is used for representing the estimation of each index on the health state of the complex equipment, and the probability distribution is fused by utilizing an evidence reasoning algorithm to obtain the prediction result of the health state of the complex equipment; and according to the health state evaluation grade, the utility of the prediction result obtained in the last step is converted to obtain the prediction value of the health trend of the complex equipment, and the quantitative display of the health trend is completed.

The decision support comprises spare part management and optimal maintenance time, wherein the optimal decision is to establish an optimal decision method of complex equipment based on a confidence rule base expert system, obtain optimal performance selection, provide decision support for subsequent use, and comprise calculation of complex road condition influence factors; and (5) converting key characteristic index multielement information. And the expert combines the actual working conditions of the complex equipment to select a proper number of quality state characteristic indexes. Because the acquired multi-element monitoring information has different formats, the multi-element monitoring information cannot be directly used and needs to be converted into a unified frame; and (5) calculating the matching degree under the condition of the influence factors. After the matching degree of each index in each rule is obtained, the matching degree of all key characteristic indexes in the corresponding rule can be calculated; activation of decision method rules is preferred. In the constructed preferred decision model, different monitoring information has different effects on different rules, so that the weights of the rules need to be activated; and evaluating the performance results of the complex equipment. The activated rule produces a feature vector of the system quality state that represents the results produced by the rule diagnosis. And the decision support executes different tasks to select objects with different performances according to the performance evaluation result.

Spare part management is a act of ensuring equipment maintenance requirements. The purpose is to use minimum spare part fund, reasonable stock storage, ensure the maintenance requirement of equipment and continuously improve the reliability, maintainability and economy of the equipment. In order to achieve the above purpose, a method for determining the number of spare parts of complex equipment based on evidence reasoning is provided, a performance change rule of the complex equipment is determined, performance monitoring information of complex equipment in the same batch is collected and arranged, health degrees of all parts in the same batch are evaluated, and a time change rule chart of the health degrees is determined according to an evidence fusion idea; the performance of each component in the complex devices currently to be used is evaluated. And determining the interval range in the time-varying law of the health value of each component. And determining the number of spare parts according to the determined change rule of the health degree of each part along with time based on the health degree of each part in the complex equipment to be used.

Example 2

As shown in fig. 2, this embodiment discloses a complex device health management method for fusing multi-source heterogeneous information, which includes:

step 101: various operating parameters of the complex device are obtained.

Step 102: and carrying out anomaly identification and pretreatment on the operation parameters of the complex equipment, and determining various pretreated operation parameters and the change trend of each operation parameter.

Step 103: and determining the real-time health state of the complex equipment by utilizing an ER algorithm according to various preprocessed operation parameters.

Step 104: and determining the predicted health state of the complex equipment by using a prediction algorithm according to the various preprocessed operation parameters. The prediction algorithm includes gray prediction, ARIMA prediction, and LSTM prediction.

Step 105: and determining the standby quantity of each device in the complex equipment according to the operation parameters after various pretreatment.

Step 106: the trend of each operation parameter, the real-time health status of the complex equipment, the predicted health status and the standby quantity of each device in the complex equipment are stored and displayed.

Step 102, including:

Step 102-1: and preprocessing the operation parameters of the complex equipment to obtain a plurality of preprocessed operation parameters.

Step 102-2: and determining any one of the operation parameters as the current operation parameter.

Step 102-3: and acquiring a normal value interval of the current operation parameter.

Step 102-4: counting the number of out-of-normal intervals in the current operation parameters after pretreatment in a preset time period as out-of-tolerance times.

Step 102-5: and determining the out-of-tolerance times as an abnormal identification result of the current operation parameters.

Step 102-6: and obtaining a history fitting result of the current operation parameters.

Step 102-7: and fitting the preprocessed current operation parameters in a preset time period by taking time as an abscissa and the current operation parameters as an ordinate to obtain a real-time fitting result of the current operation parameters.

Step 102-8: and updating the normal value interval of the current operation parameters according to the real-time fitting result of the current operation parameters.

Step 102-9: and adding the real-time fitting result of the current operation parameters into the history fitting result to obtain the change trend of the current operation parameters.

Step 102-10: updating the current operation parameters, and returning to the step 102-3 until all the operation parameters are traversed, so as to obtain a plurality of preprocessed operation parameters and the change trend of each operation parameter.

Step 103, including:

step 103-1: and determining the weight and reliability of each operation parameter by using a coefficient of variation method.

Step 103-2: a quality factor for each of the operating parameters is determined based on the plurality of weights and the plurality of reliabilities.

Step 103-3: and determining the confidence coefficient by using an ER algorithm according to the operation parameters after the pretreatment and the quality factors corresponding to each operation parameter.

Step 103-4: real-time health status of the complex device is determined based on the confidence. Confidence intervals corresponding to different real-time health states are different.

Step 104, including:

step 104-1: and inputting various preprocessed operation parameters into the gray prediction model to obtain a first predicted health state of the complex equipment.

Step 104-2: and inputting the various preprocessed operation parameters into an ARIMA prediction model to obtain a second predicted health state of the complex equipment.

Step 104-3: and inputting the various preprocessed operation parameters into the LSTM prediction model to obtain a third predicted health state of the complex equipment.

Step 104-4: and carrying out weighted summation on the first predicted health state, the second predicted health state and the third predicted health state to obtain a fourth predicted health state.

Step 104-5: determining the first predicted health state, the second predicted health state, the third predicted health state, and the fourth predicted health state as predicted health states of the complex device.

The gray prediction model is obtained by taking various historical operation parameters of complex equipment as input, taking the health state of the complex equipment at the next historical moment as output, establishing a gray differential equation by using a gray prediction method, whitening the gray differential equation and solving the gray differential equation.

The ARIMA prediction model is obtained by fitting by using an ARIMA prediction method by taking various historical operation parameters of the complex equipment as independent variables and the health state of the complex equipment at the next historical moment as dependent variables.

The LSTM prediction model is obtained by training the LSTM network by taking various historical operation parameters of the complex equipment as input and taking the health state of the complex equipment at the next historical moment as output.

Step 105, comprising:

Step 105-1: and acquiring the service time of each device in the complex equipment.

Step 105-2: and determining any model as the current model.

Step 105-3: let the current model spare number i=0.

Step 105-4: and determining any device of the current model in the complex equipment as the current device.

Step 105-5: and when the difference value between the service time of the current device and the service time threshold value of the current device is obtained, obtaining the current device position as the position of the to-be-replaced device.

Step 105-6: and determining the current model as a standby model.

Step 105-7: the number of spare numbers of the current model is increased by 1.

Step 105-8: updating the current device, and returning to step 105-5 until all devices of the current model are traversed, and determining the standby number of the current model and the positions of a plurality of to-be-replaced devices.

Step 105-9: updating the model devices, returning to step 105-3 until the models of all devices in the complex equipment are traversed, and determining the standby number of different models and a plurality of to-be-replaced parts.

The invention will be further described with reference to the accompanying drawings and examples, which are representative of complex devices, namely, laser inertial measurement units. The laser inertial measurement unit is a high-precision device widely applied to the aerospace field, is often used as a key component of large-scale equipment such as an airplane, a carrier rocket and the like, completes the tasks of positioning and attitude determination, and plays a vital role in the normal operation of the equipment.

The embodiment takes a laser inertial measurement unit as a research object, provides a complex equipment health management method integrating multi-source heterogeneous information, and carries out health management on complex equipment through 4 parts of data interpretation, health assessment, health trend analysis and decision support. And finally, fusing the obtained predicted values of all indexes of the complex equipment by adopting an evidence reasoning algorithm to obtain the future health state information.

In this example, each part of the complex device health management method for fusing multi-source heterogeneous information specifically includes the following steps.

A first part: and (5) data interpretation.

As shown in fig. 3-4, the data interpretation is based on the original data file, and the reading and preprocessing of the laser inertial measurement unit index data, and the analysis and feedback of the preprocessed data change trend and abnormal state are completed. The method specifically comprises the following steps:

And 1, reading the original file. According to the requirement of the health management of the laser inertial measurement unit, reading the index data (such as scale factors, drift coefficients and the like) and the threshold values of the indexes in the index system of the laser inertial measurement unit from a data source. Where the data source is typically the original data file under the specified path. After each index data is obtained from the data source, a specified format data file which can be processed by the system is generated.

And 2, preprocessing data. Index data are obtained from the generated data file, and a preprocessing method (such as mean square error solving, extremely poor solving and the like) is selected to perform preprocessing operation on each index data. The purpose of the preprocessing operation is to improve the data quality. After the pretreatment is finished, a pretreated data file is generated under a specified path, and the data file is used for operations such as index abnormality judgment and subsequent health assessment of data.

Step 3: and judging the index abnormality. Traversing the data in the data file, and judging whether the data exceeds a threshold value. In this step, the preprocessed data within the specified range is typically compared to a threshold. If the value of one of the indexes exceeds the set threshold range, the index is judged as abnormal data, and an abnormal judgment result is generated, wherein the result is used as a reference index for health management.

Step 4: and judging the out-of-tolerance times. Traversing index data, and recording the number of data exceeding the corresponding threshold value as the out-of-tolerance times. For example, the index x has n data, where m data exceeds the upper and lower thresholds, and the number of times of out-of-tolerance is recorded as "x: m". Meanwhile, in order to avoid inconsistent out-of-tolerance times caused by multiple interpretation, time is added in the records so as to distinguish out-of-tolerance times records at different times. The generated record result is stored in the out-of-tolerance times record file under the specified path, so that the out-of-tolerance times of the index can be conveniently read by subsequent operation.

Step 5: and (5) data state feedback. In this step, the preprocessed data is visually displayed, and the trend and change of the data are presented in various statistical charts (such as line diagrams, bar charts, pie charts and the like), so that the state and distribution of the data can be more intuitively known. Through the charts, the change trend and distribution condition of the data and whether an abnormal state exists can be clearly observed, and more accurate analysis and judgment can be assisted.

A second part: health assessment.

As shown in fig. 5-9, the health state evaluation aims at evaluating the performance state of the current laser inertial measurement unit by accumulating pulse equivalent data of the accelerometer and the gyroscope after data interpretation processing, and mainly adopts ER algorithm for evidence fusion, and the specific implementation steps are as follows.

Step 1: a reference level is set. And (3) according to a navigation equation, pre-calculating a static navigation error corresponding to the accumulated pulse equivalent data of the accelerometer and the gyroscope which are processed by the data interpretation. According to the precision evaluation method of the GJB 729-89, an inertial navigation system performance state table corresponding to the static navigation error is compiled by combining expert knowledge. And then, converting the static navigation error into the performance grade of laser inertial navigation according to the reference value corresponding to the formulated table.

Step 2: and calculating the weight of each input value by a coefficient of variation method. The weight is used for explaining the importance degree of the index and is a key parameter when the ER algorithm performs evidence fusion. Therefore, a calculation method of determining weights is required before fusion calculation. The coefficient of variation is an important parameter for measuring the relative fluctuations between different data sets. It is explained that when the coefficient of variation of a certain attribute is large, its relative fluctuation is large, and the attribute should be given a large weight. The coefficient of variation is calculated as follows:

(1)。

In the middle of The coefficient of variation of the ith attribute after run time t is represented. /(I)Representing the mean of the ith attribute. /(I)Representing the mean of the ith attribute over time t.

(2)。

The above equation represents the Mean Square Error (MSE) of the ith attribute after the runtime t, reflecting the data discrete trend of the data set.

(3)。

The above equation represents the average value of the ith attribute after the running time t, reflecting the data dispersion trend of the data set. The formula for calculating the attribute weight according to the variation coefficient is as follows:

(4)。

wherein, Representing the weight of a given attribute i.

Step 3: and calculating the reliability degree of each index. Evidence reliability is different from evidence weight, which refers to the externality of evidence and the relative importance of different pieces of evidence to the outcome, while evidence reliability refers to the intrinsic nature of evidence, describing the degree of confidence of evidence itself. According to Fan and left studies, the laser inertial navigation input reliability is defined as follows:

(5)。

wherein, Representing the reliability of the first item of evidence. /(I)Is a weight coefficient. /(I)The sufficiency index representing the ith evidence may be measured by a fuzzy relationship function.

Step 4: reliability and weight are combined into a new quality factor. Considering the combined effect of reliability and weight on evidence, the following new formula for evidence quality factor calculation for expressing the combined effect is defined in the formula:

(6)。

(7)。

wherein, To combine evidence weight and evidence confidence quality factor,/>For the evidence weight calculated in step 2,/>Is the confidence of evidence i. The reliability and quality factor show a positive correlation as shown in fig. 5. The practical significance of larger quality factor value is satisfied as the reliability is larger.

Step 5: the uncertain input is converted into a form of belief distribution taking into account evidence quality factors. The uncertain input is converted into a form that considers a belief distribution of evidence quality factors (QBD). A general input form corresponding to all attributes is shown as:

(8)。

wherein, Representing the confidence in the input value xi of the ith attribute representing the uncertainty of the input information. For example, (zero order term drift coefficient out of tolerance, 0.5) means 50% certainty, i.e. the input can obtain the value of the property (zero order term drift coefficient). To meet the requirements of ER entry forms, input/>Conversion into belief distribution form is as follows:

(9)。

wherein, Is x _i points to the result/>Is included in the reference value of (2). This QBD format can be described as follows:

(10)。

wherein, The nth possible result is shown. /(I)Indicating that the ith evidence produces a result/>Is to be determined. N is the number of attributes; /(I)Representing recognition frameworks,/>To identify the frame/>A power set of (a).

Step 6: evidence fusion. After obtaining the quality factor of the evidence according to the weight and the reliability of the evidence, the evidence can be combined through an ER algorithm, and the confidence relative to the evaluation result D _j is calculated. The course of the study for health assessment can be divided into two parts: the confidence of the output portion is translated into a base probability mass and then evidence is combined. In the reasoning process of the assessment model, an iterative ER algorithm is used to represent the aggregation process of evidence. It can be divided into two main parts: a first part that converts the confidence level of the output part into a basic probability mass:

(11)。

The basic probability mass of L pieces of evidence can be aggregated by the following iterative process:

(12)。

(13)。

(14)。

(15)。

(16)。

In the method, in the process of the invention, An aggregate base probability mass for the first belief rule representing the kth state. /(I)Is a normalization factor intended to be ensured.

(17)。

And a second part, calculating and determining the joint confidence of the model result.

(18)。

Step 7: and calculating output confidence coefficient, and outputting the performance grade corresponding to the current laser inertial navigation. The final output utility is given by the following function:

(19)。/>

wherein, Representing the integrated performance state of the laser inertial navigation as assessed by the collected input information S (x). D= { D ₁、D₂、···、D_n } represents the label utility value of the output result.

Step 8: and optimizing parameters of the health evaluation model. In the operation process of laser inertial navigation, the operation condition changes along with the change of the system performance. In order to guarantee the accuracy of the model, the parameters of the model need to be optimized. The invention updates the parameters of the evaluation model by adopting an optimization algorithm. Mean Square Error (MSE) is a measure reflecting the degree of difference between an estimated value and an actual value. The MSE between the actual output utility and the estimated output utility in the present invention may reflect the accuracy of the assessment model. Thus, taking the minimum value of the MSE as the objective function:

(20)。

(21)。

In terms of model parameter updating, ER is an expert system that has stringent requirements on the physical meaning of model parameters. Thus, in performing model parameter optimization, the following constraints should be followed:

(22)。

Third section: and (5) analyzing health trend.

As shown in fig. 10-17, the health trend analysis is based on the laser inertial measurement unit operation data after the data interpretation processing, and the prediction results of each index are obtained by using any one of gray prediction, arima prediction and LSTM prediction algorithms, and then the prediction results are fused, so as to finally obtain the quantized health trend. The specific process is shown in fig. 11, and comprises the following steps:

Step 1: and establishing a laser inertial measurement unit health trend analysis index system. According to the specific model of the laser inertial measurement unit, the mutual relation of all the component parts in the laser inertial measurement unit needs to be established, and the index range is indicated. The collection and arrangement of information was completed using the index system table and the index range table as shown in tables 1 and 2. The health management is divided into four parts, wherein the data interpretation is the basis of each part and is used for completing the data processing in advance, and the following three parts belong to parallel relations and complete the respective health management tasks.

TABLE 1 index System Table

Table 2 index range table

The above tables 1 and 2 are distributed to different specialists, and the specialists combine their own experiences to complete the construction of the index system and the setting of the index value range, weight and reliability.

Step 2: and reading the data and performing data preprocessing operation. When the trend analysis is carried out on the laser inertial measurement unit, various indexes of the gyroscope and the accelerometer which are integrated in the laser inertial measurement unit, such as a zero-order drift coefficient and a first-order drift coefficient of the gyroscope, a zero-order drift coefficient of the accelerometer, a positive scale factor and a negative scale factor, are needed. The data of these indices cannot be directly applied to prediction, and are typically grouped first in a sliding window and then the mean square error is calculated for the grouped data, as shown in equation (23).

（23）。

Where s represents the mean square error, x _i is the index data,Is the average value of index data, and n is the number of index data.

Step 3: the prediction is done using a prediction model. The method provides three prediction algorithms, namely: gray prediction, ARIMA prediction, and LSTM prediction.

The Gray Model (GM) is a method of predicting time series containing uncertainty. GM is used for identifying the degree of dissimilarity of the development trend among the time sequences, namely carrying out association analysis, carrying out generation processing on the original data to find the law of system variation, generating a data sequence with stronger regularity, and then establishing a corresponding differential equation model so as to predict the condition of future development trend of things. The method is suitable for short-term prediction, and is a system with short time sequence and incomplete information. The procedure for GM was as follows:

The time series data processing according to certain requirements is called gray generation, and the generation of the original data is to try to find the internal rule from the random phenomenon. The commonly used generation method is accumulation generation.

The gray differential equation defining GM (1, 1) is:

（24）。

wherein, Is the gray derivative,/>For the development coefficient,/>For whitening background value,/>The ash is used as an action amount.

Establishing a whitened GM (1, 1) model:

The data obtained at this time is a discrete state scatter diagram, and the gray model needs to be whitened to obtain an exact relational expression and calculate a result, so that the whitening process is a continuous process by discrete regression. The white differential variance is:

（25）。

And solving the differential equation to obtain a prediction result.

The ARIMA model is a method for modeling a time sequence and is widely applied to the fields of finance, medical treatment, machinery and the like. The ARIMA model is utilized to analyze and research the time sequence and find the change rule, so that the future trend and development process of the time sequence can be predicted, and the method is suitable for a system with long-term prediction and small time sequence fluctuation. The ARIMA model generally has the following form:

（26）。

wherein, Is the time sequence of each index,/>Is a white noise sequence expected to be 1,/>Is a white noise sequence expected to be 0,/>Is an autoregressive part of ARIMA model,/>As part of the sliding average of the ARIMA model,Is/>A time series sequence after the step difference. /(I)For the difference times/>Is of lag order,/>Is the moving average order, and therefore, the model is generally denoted/>. The ARIMA model was used as follows:

stationarity test of time series:

The time series must meet the stationary non-randomness requirements of the ARIMA model, so a stationary check of the time series is required before modeling. The stability is usually checked by the Dickey-Fuller test, Statistical/>The value is the key to measuring stationarity. If the requirements are not met, differential processing is performed.

Determination of model order:

Calculating a time series Autocorrelation Coefficient (ACF) and a Partial Autocorrelation Coefficient (PACF), determining model orders by ACF and PACF And/>Selecting the model with the minimum AIC according to the erythro pool information criterion (Akaike information criterion, AIC), and determining the optimal order/>。

And (3) significance test of the model:

After determining the model order, a saliency test is required for the model. The model is significantly valid if the residual sequence of the fitted model is a purely random sequence. Otherwise, the representation model is not valid, and a more appropriate order needs to be selected for re-fitting.

A Long Short-Term Memory (LSTM) network has strong processing and predicting capabilities on time sequences, can solve the problem of Short-time Memory in a cyclic neural network (Recurrent Neural Network, RNN), and is suitable for systems with medium and Long Term and large time sequence fluctuation.

Key to LSTM are cell status and "gate" structure. The cell state is used as a transmission chain of information, only partial linear interaction is participated, and the integrity of the information is maintained. The door structure includes a forget door, an input door, and an output door.

The retention and deletion of information is determined by the forget gate. Inputting the information of the hidden state of the upper layer and the information of the current moment into a forgetting gate, establishing a nonlinear mapping relation through a sigmoid function, and finally obtaining an output vector。/>Between 0 and 1, 0 indicating total deletion and 1 indicating total retention. The forgetting gate formula is as follows:

（27）。

wherein, Representing a sigmoid function,/>Representing a weight matrix,/>Information indicating the hidden state of the previous layer,Information indicating the current time,/>Representing the bias.

The update of the cell state is determined by the input gate. Obtaining an output vector by a sigmoid functionInputting the information of the hidden state of the previous layer and the information of the current moment into a tanh function, and creating a new candidate value vector/>. Old State of cells/>And/>Multiplying and deleting unnecessary information. /(I)And/>Multiplying, retaining the needed information, and finally adding to obtain updated cell state/>. The input gate formula is as follows:

（28）。

（29）。

（30）。

the hidden state is determined by the output gate and contains information in the previous time sequence. Obtaining an output vector by a sigmoid function And determining information to be output. Treatment of cell states/>, using tanh functionAnd with output vector/>Multiplying to obtain new hidden state information/>Output to the next layer. The output gate formula is as follows:

（31）。

（32）。

The LSTM prediction is used as follows:

Pretreatment of a time sequence:

The training data needs to be normalized before prediction, and alternate time steps are selected. The training data is processed by using the alternate time step, and the observed data of the previous moment is used as the input for predicting the observed data of the next moment.

Definition and training of the network: the input layer is determined by the dimension of the input, the full link layer is determined by the dimension of the output, and the implicit cell number of the LSTM layer is determined by the data. In order to improve the learning effect, it is necessary to select an appropriate optimizer after the structure determination. And the learning rate is updated in the training process by using an adaptive moment estimation (Adaptive Moment Estimation, adam) optimizer, so that the learning efficiency and the convergence rate are improved.

Prediction of the index:

The last step of training data is used to predict the first predicted value. To complete the prediction of the index, it is necessary to determine the number of prediction steps and input the first predicted value into the network.

Step 4: the prediction result is converted into a probability distribution. Since the evidence reasoning algorithm is needed for health trend analysis, the prediction result needs to be converted into probability distribution, and the formula is as follows:

（33）。

Where m represents that the prediction result x _i has m reference levels, Representing the probability distribution of the ith prediction result relative to the p-th reference level,/>The interval value representing the q-th reference level.

Step 5: and fusing probability distribution of the prediction results of the indexes at the same time point by using a evidence reasoning algorithm. The evidence reasoning algorithm (Evidence Reasoning) is based on the D-S evidence theory, and can perform multi-attribute fusion operation on the decision problem. In this example, the prediction result of each index at the same time point can be regarded as a piece of evidence, and the probability distribution of the prediction result is fused to obtain the probability distribution of the laser inertial measurement unit health trend result. The formula of the ER algorithm is as follows:

（34）。/>

（35）。

wherein, Weights representing the ith index involved in fusion,/>Representing the confidence of the h possible result after fusion; h represents the total number of results.

Step 6: the confidence distribution of the results is converted into a quantized score. In order to facilitate the demonstration of the health trend of the laser inertial measurement unit, the confidence distribution of the result can be converted into a health score through utility calculation, and the formula is as follows:

（36）。

wherein, And (3) representing the preset score of the h result, wherein y is the utility, and the utility is the health score.

Fourth part: and (5) decision support.

As in fig. 18-21, the preferred decision is: decision support is an integral part of health management.

Step 1: and calculating the complex road condition influence factor y. The sources of the inertial navigation vibration impact are mainly railway transportation and road transportation, generally, along with the increase of transportation mileage, the influence of the vibration impact on the inertial navigation is larger, and the vibration impact degrees of the railway transportation and the road transportation are different. The maximum transportation mileage of the inertial navigation device is required, so that a standardized formula of the transportation mileage can be established:

（37）。

Wherein: And/> The coefficients corresponding to the road transportation mileage x ₁ and the railway transportation mileage x ₂. Through looking up the related data, a ₂ is generally taken as the maximum value or theoretical limit value of the inertial navigation regulation transportation mileage conversion value, and a ₁ is taken as the minimum value of the inertial navigation regulation transportation mileage conversion value. y is the influence degree of vibration impact on inertial navigation in railway transportation and road transportation.

Step 2: and (5) calculating the matching degree of the key characteristic index multivariate information conversion and influence factor of the inertial navigation system. Therefore, an expert is first required to select a proper number of quality state characteristic indexes by combining the actual working conditions of the object inertial navigation system. On the other hand, because the acquired multi-element monitoring information has different formats, the multi-element monitoring information cannot be directly used, and the multi-element monitoring information needs to be converted into a unified frame by the following formula:

（35）。

Wherein, R _ik and R _i(k+1) are reference levels of the ith key feature index in the kth rule and the kth+1th rule, and the reference levels need to be determined by combining information distribution and types of the features. L' is the number of rules after the model is adaptively adjusted. The matching degree in the j-th rule after index conversion is adopted; /(I)The actual value of index i at time t is indicated.

After the matching degree of each index in each rule is obtained, the matching degree of all key characteristic indexes in the rule can be obtained through the following formula:

（36）。

（37）。

wherein, Indicating the relative weight of the index. /(I)The weight of the i-th index is represented.

Step 3: activation of the rules of the preferred decision method and evaluation of the performance results of the inertial navigation system. In the constructed preferred decision model, different monitoring information can have different effects on different rules, and the rule activation weight is represented by the following formula:

（38）。

wherein, The rule weight in the dynamic adjustment process of the model is represented, namely, when the rule importance degree of the rule is not satisfied, the rule weight is reduced. /(I)The weight of the first rule is represented.

The activated rule produces a feature vector of the system quality state that represents the results produced by the rule diagnosis. The quality state feature vectors output by all rules can be fused through a evidence reasoning (EVIDENTIAL REASONING, ER) algorithm to obtain a final output quality state feature vector. The ER algorithm resolution format is as follows:

（39）。（40）。

wherein, An output quality state feature vector is generated for the model. /(I)Confidence of nth output result level D _n obtained after fusion of input index monitoring data,/>And/>。

Assuming a single evaluation resultUtility of (1)/>The output inertial navigation system performance results are as follows:

（41）。

wherein, And the final output result of the inertial navigation system quality state evaluation model constructed based on the BRB, namely the quality state evaluation grade of the inertial navigation system obtained by monitoring data.

The reasons for model optimization are added: because the BRB initial model is given by an expert and is influenced by the limitation of the cognitive ability of the expert, the parameters of the initial BRB model have certain deviation, so that the actual modeling effect cannot meet the requirements, and therefore, an optimization model needs to be constructed to optimize the parameters of the BRB model, and meanwhile, the fusion of data and knowledge is achieved. In terms of model parameter updating, because the BRB belongs to an expert system, strict requirements are imposed on the physical meaning of the model parameters. Therefore, the following constraints need to be obeyed in the model parameter optimization process:

（42）。

Spare part management:

Step 1: in order to accurately evaluate the performance of the inertial sensor, the detection samples of the inertial sensor in the same batch in factory sampling inspection are fully utilized, indexes capable of reflecting the performance of the inertial sensor are selected, and the change of the indexes along with time is recorded. The decision support of health management is divided into two components, namely, optimal decision and spare part management, wherein the optimal decision and the spare part management are respectively used for providing advice for selection of equipment, the spare part management is used for equipment maintenance, and the spare part management is guaranteed to have economy and reliability while maintenance requirements are guaranteed.

1. The batch detects the health of the sample. According to the index system and the index value range of each component part in the laser inertial measurement unit established in the step 1 of the third part, calculating probability distribution (excellent: Good:/> Difference/(I)). And calculates the health h of the device according to equation (43). The value range of the health degree is [0,1], the greater the health degree is, the more healthy the device is, the less the device needs to be replaced, and the smaller the health degree is, the more the device needs to be replaced.

(43)。

2. And fusing a plurality of detection samples to obtain a change curve which can most represent the health degree of the batch of inertial sensors along with time. The n detection samples are at the momentN different health degrees exist, each detection sample is an evidence, so that the change rule of the health degree of the inertia devices in the batch along with time can be best represented according to the numerical relation among the evidences at each moment.

2.1 And (3) judging the accuracy of the health degree of each detection sample at any moment:

Positioning all health values on a real axis, and setting Health value/>, of time ith detection sampleHealth value/>, to the j-th test sampleDistance of/>。/>Representing health value/>Distance to all remaining health values/>Average value of (2). /(I)

(44)。

2.2 Conversion from the health value of each detection sample to evidence at moment:

the present invention adopts the formula (45) and is based on Calculating the weight of each health value at the current moment when participating in fusion. A large amount of measured data calculation proves that when the power value of the transformation function is taken as 3, the influence of the wild value can be well reduced. When the power value of the transformation function takes a larger value, the effect of reducing the outlier increases less significantly, so the power value of the transformation function in the formula takes 3.

(45)。

2.3 Generating a final fusion value:

For time of day Each health value is subjected to weighted fusion to generate a final fusion result, namely the best characterization moment/>Health value of the batch of inertial devices.

(46)。

According to the method in step 1, a time-dependent graph of the health of the batch of inertial sensors can be obtained, as shown in fig. 18.

Step 2: the performance of each inertial sensor in the currently-to-be-used inertial measurement unit is evaluated. And determining the interval range in the time-varying law of the health value of each inertial sensor.

And assuming that the equipment is in a working state, evaluating according to test data of a certain input index to obtain a performance evaluation result of the equipment. The health value of the inertial sensor can be obtained by adopting the method in the step 1Corresponding to time/>, in the inertial sensor health change curveHealth of/>. Because of the uncertainty between the performance of the same batch of sensors, the batch sensor performance change law is already determined. Therefore, the health value is/>May be in the time interval/>, of the batch of inertial sensor health change curves. HandleThe range serves as an interval window of uncertainty in the health of the inertial sensor.

Step 3: time of dayAnd determining the number of spare parts of the inertial sensor in the inertial measurement unit.

Based on the determined interval window of uncertainty of the health degree of the inertial sensor at the initial moment, the method can obtainThe time interval corresponding to the health degree of the time sensor is/>The health interval is/>. Assuming that m laser inertial sensors are in use at this time, a corresponding number of spare parts need to be prepared in advance in order to ensure that the equipment has enough spare parts to replace when a problem occurs. According to the health degree interval, the number of required spare parts is as follows:

(47)。

Wherein, Time/>And the health degree interval corresponding to the kth laser inertial sensor. /(I)Representing no more than/>Is the largest integer of (a).

To ensure that enough spare parts are replaced when equipment is in trouble, the equipment should be replaced。

Example application:

the first portion of data is interpreted.

1. Preprocessing and generating a data file.

The original file contains all index data of the internal constitution of the laser inertial measurement unit. According to the index system, the index data to be evaluated are taken out, a new data file is generated for facilitating the subsequent operation, and the read index data (the index of the laser gyroscope and the accelerometer on the internal x axis, the y axis and the z axis of the three-axis laser inertial measurement unit of a certain model is taken as an example, and the index is shown in table 3) are stored. And reading the newly generated data file, selecting a preprocessing method to perform preprocessing operation on the index data, and storing the preprocessing result into the new data file.

2. And judging the abnormal and out-of-tolerance times.

And reading data in a specified time range from the preprocessed file to perform abnormality judgment, taking an index D _0x as an example. Assuming that data within 1 day is taken for judgment, if abnormal data occurs in the data, the abnormal state is marked as yes, and if 7 data exceeds a threshold value, the out-of-tolerance count is marked as 7. At this time, the abnormal state of the entire assembly (top 1) is also noted as "yes". Based on the judgment result, the index abnormality judgment result of the following table 3 is established.

Table 3 shows the index abnormality determination results.

3. And (5) data state feedback.

According to the preprocessed file data, a statistical chart is constructed, a threshold value is marked in the statistical chart, K1b _z is taken as an example (data is imaginary), a line graph of the index is drawn, wherein two indexes exceed the threshold value, so that the out-of-tolerance frequency is 2, and the index state is abnormal.

And a second portion health assessment.

The case study mainly comprises the following steps:

A. Experimental setup.

Because the axial rigidity and the radial rigidity of the gyro support are unequal, the elastic deformation along the axial direction and the radial direction are unequal, so that interference moment is generated along the transverse axis, and coaxial drift is generated. However, elastic deformation of the vibration device is limited by the storage time and the history maintenance condition. Long-term fixed storage and maintenance damage can reduce deformation of the vibration device, thereby affecting navigation accuracy. According to the experiment, static navigation errors corresponding to accumulated pulse equivalent data of 18000 groups of accelerometers and gyroscopes are calculated in advance according to a navigation equation. According to the accuracy evaluation method and expert knowledge of GJB 729-89, an INS performance state table corresponding to the static navigation error shown in Table 4 is compiled. Then, according to the reference values corresponding to table 4, the static navigation error is converted into four performance levels of INS as follows:

。

Table 4 reference values for static navigation errors and corresponding performance states are shown schematically.

A total of 800 sets of data were selected from 18000 sets of data, with 200 sets of data having 4 levels of performance status. As shown in fig. 6, 400 sets of data are selected as training data sets and the remaining 400 sets of data sets are test data sets by uniform sampling. The experimental data selected in fig. 6 are pre-processed accelerometer and gyroscope pulse accumulation data for INS. The preprocessing method is to read 5 ms pulse increment data and meta information first, and then convert the data into 5 ms pulse increment data in a measurement coordinate system.

B. model parameter sets of the health assessment model.

The initial reference values of the input information are shown in table 5. The first layer index weight is calculated by CVBW and the second layer index weight is given by an expert.

In table 5, assuming that the input value is x, if x < A4, it is evaluated as a G4 level. If A4< x < A3, then it is evaluated as grade G3. If A3< x < A2, then it is evaluated as class G2. If A2< x < A1, then it is evaluated as class G1. For example, for accelerometer x: when its cumulative pulse value is 3300, the accelerometer will be considered as level G3, participating in the assessment, weighting 0.41, at 3210-3455, i.e., A4-A3. The distribution of the reference values after function optimization is shown in table 6.

Table 5 initial reference levels and reference values are schematically shown.

Table 6 reference levels and reference values after optimization are schematically shown.

C. a performance state is determined.

Based on the final decision result, the equation (26) is combined to obtain the final utility corresponding to the evaluation result. The black (x) broken line points in fig. 7 represent the level of final output utility, which enables accurate assessment of the INS performance state, and the straight lines represent the expected results. Thus, the closer the distribution is to the straight line, the higher the accuracy of the model.

Meanwhile, more accurate judgment can be made on the part of the monitoring point which cannot be accurately judged by the expert, uncertainty and partial lack of expert knowledge are effectively overcome, effective update of the expert knowledge is realized through monitoring data, and an expert system is further improved. The final utility points obtained by the evaluation are consistent with the expected result and distributed on the upper side and the lower side of the expected result.

D. and (5) comparing experiments.

The result of the BP neural network (BPNN) prediction is utilized, as shown in FIG. 8. The output of BPNN is very different from the expected results, indicating the rationality of the health assessment herein.

In contrast, the method used in the present invention has a lower MSE than other methods, which means that the evaluation method proposed in the present invention has higher accuracy in determining the inertial navigation system performance state than other methods.

And (3) analyzing health trend in the third part.

1. And (5) establishing an index system.

Taking a certain type of triaxial laser inertial measurement unit as an example according to the internal constitution and operation characteristics of the laser inertial measurement unit, according to the index system shown in fig. 10, the index system shown in the following table 7 and the index value range shown in the table 8 are established (the index system and the value range of the gyroscope are only shown here, and the space is limited).

Table 7 index system table.

Table 8 shows the range of index values.

2. And (5) preprocessing data.

In this example, index data for each second over 1 day for gyroscopes and accelerometers was collected. And setting the sliding window size as 200, performing grouping treatment on the data, and calculating the mean square error of each group of data.

3. And selecting a prediction model for prediction.

The data of 10 time points using the history samples were set, and prediction of data of 3 time points in the future was performed by using gray prediction, ARIMA prediction, and LSTM prediction in this order, and the prediction results are shown in table 9 (for a short period, only the first two sets of prediction results are shown here).

Table 9 a schematic representation of the prediction results of each prediction algorithm.

As can be seen from Table 9, the predicted results vary more closely under the three prediction algorithms. At this time, any prediction algorithm may be selected. As shown in fig. 12 to 16, the results of predicting two indexes of the gyro 1 and three indexes of the accelerometer 1 using the gray prediction algorithm are shown. The first 10 points in the graph are the history data for prediction after preprocessing, and the last 3 points are prediction results.

4. The prediction results of the indexes are converted into probability distribution.

The prediction results of the respective indices in table 9 are converted into probability distributions with reference level intervals of the respective indices shown in table 8. Taking the predicted result 4.4529 ×10 ^-3 of the gray prediction model as an example in the first set of predicted results of D0 _x, the probability distribution is given by the following equation (48):

（48）

According to the description of the values of the reference levels in table 9, when the reference level is 0, the index is optimal. Therefore, the result shown in the formula (48) shows that when the D0 _x prediction result is 4.4529 ×10 ^-3, the probability of being excellent is 0.1613, the probability of being excellent is 0.8387, and the probability of being poor is 0. This result is consistent with the case of the data itself.

5. And (5) carrying out fusion of the probability distribution of the prediction result of each index at the same time point by using an ER algorithm.

The probability distribution of the predicted results of each index at the same time point is used as a evidence to represent the possibility of various predicted results, and the predicted results of each index at the same time point are fused by using the formula (34) and the formula (35), so that the confidence degree distribution of the health state of the laser inertial measurement unit at each time point can be obtained, as shown in table 10.

Table 10 fuses the confidence profiles of the results.

6. The health probability distribution is converted into component values.

The probability distribution of the health state at each time point is converted by using the formula (36) and the setting of the laser inertial measurement unit health degree score in table 8, taking the time point 1 as an example:

（49）

The final prediction results at three time points are shown in fig. 17, and the detailed results are shown in table 11 below.

Table 11 health values for laser inertial measurement unit predictions are shown.

As can be seen from fig. 17 and table 11, the health status of the laser inertial measurement unit is in a superior state, and the overall trend is slightly declining, so that the laser inertial measurement unit meets the practical situation.

And the fourth part is decision support.

1. Problem description and influence factor calculation.

In the case of rockets, the position and velocity of the rocket are variable due to the long flight time, and these data are initial parameters of the launched rocket, directly affecting the accuracy of the rocket's flight, thus requiring the provision of high accuracy position, velocity and vertical alignment signals. The inertial navigation is carried out independently and independently by means of the carrier equipment, does not depend on external information, has the advantages of good concealment, and high accuracy, and works are not affected by meteorological conditions and artificial interference. The inertial technology is gradually popularized to the fields of aerospace, aviation, navigation, petroleum development, geodetic survey, marine survey, geological drilling control, robot technology, railway and the like, and is applied to automobile industry and medical electronic equipment along with the appearance of novel inertial sensitive devices. Therefore, the inertia technology not only plays a very important role in national defense modernization, but also increasingly shows great roles in various fields of national economy. In the experiment, the interference suffered in the real working environment is simulated by the simulation interference device, the selected index is the output result of three accelerometers, and the monitoring data of each index collected in the experiment is shown in fig. 20.

2. And establishing an optimal decision model of the inertial navigation system under the condition of complex road condition transportation.

In the monitoring data of the three accelerometers, the reference levels of the monitoring information of the three accelerometers are respectively determined to be 4 by combining the data quantity, the complexity of the model, the diagnosis precision, the diagnosis real-time performance and the like, as shown in table 12. In combination with the rule construction mode, the constructed preferred decision model has 64 rules in total. Since accelerometers are susceptible to environmental influences during actual use, three accelerometer monitoring information needs to be considered simultaneously in determining the rule output confidence, and an initial diagnostic model is shown in table 13. In the initial model, it is assumed that the rule is equally important, i.e. the rule weight is set to 1.

Table 12 accelerometer reference levels and reference values are schematically shown.

Table 13 schematic representation of preferred decision initial model of inertial navigation system.

3. And under the condition of complex road condition transportation, the inertial navigation system prefers the training and testing of the decision model.

Under the condition of complex road condition transportation, the inertial navigation system has the following four stages: no transportation stage, highway transportation stage, railway transportation stage, highway and railway transportation stage. In the experimental process, the data 1300 sets are collected together, wherein the accelerometer 1 is subjected to shaking and collecting 300 sets, the accelerometer 2 is subjected to shaking and collecting 300 sets, and the accelerometer 3 is subjected to shaking and collecting 300 sets. The set 650 is randomly screened from the dataset as training data to train the constructed preferred decision model.

As can be seen from fig. 21, after the initial model is optimized, the blue line can accurately evaluate the quality state of the inertial navigation system, and can accurately evaluate the quality state of the inertial navigation system at partial monitoring points which cannot be accurately judged by an expert, thereby effectively overcoming the uncertainty and partial unknowing of expert knowledge, achieving the effective back feeding of the monitoring data to the expert knowledge, perfecting the expert system, realizing the effective fusion of the monitoring data and the expert knowledge, and accurately performing the optimal decision sequencing. After training, the index weights of the influence factors are respectively 0.9. The MSE of the trained model is 0.0237, which is far smaller than the mean value of safety evaluation, and the evaluation accuracy is higher.

Spare part management:

1. assume that the detection sample has 5 laser inertial sensors, and the 5 sensors are at the same time The probability of each level distribution at the time is shown in table 14. Sample health was calculated according to equation (43), and each sample health is shown in table 14.

Table 14.

5 Samples atAnd (5) distributing probabilities and health degrees of each grade at the moment.

Calculating the weight of each health degree to obtainWeighted summation to obtain/>The fusion health of the time sample is 0.9368290. I.e. the batch inertial sensor has a use time/>The health degree was 0.9368290. And fusing the health degrees of all samples at all times to obtain the change rule of the health degrees of the batch of sensors along with time. It is assumed here that the resulting health rule corresponds to the function/>。

2. Assuming that the currently used inertial sensor equipment comprises 4 inertial sensors, the health degree of each sensor is respectively as follows. The 4 sensors may be in the use time interval as/>, according to the time-dependent health degree change rule obtained in the step 1，[427.84065，522.91635]，[193.10922，236.02235]，[579.46518，708.23522]。

3. According to the determined range of the use time of the inertial sensor at the initial time, the use time of the sensor at 200 units of time can be obtained as,[627.84065, 722.91635],[393.10922, 436.02235],[779.46518, 908.23522]. At this time, the health intervals of the inertial sensors are [0.6584104,0.6269460], [0.3721594,0.2770837], [0.6068908,0.5639777], and [0.2205348,0.0917648]. /(I)

And according to the formula (46) and the formula (47), the number of required spare parts is not less than 2.

In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.

The principles and embodiments of the present invention have been described herein with reference to specific examples, the description of which is intended only to facilitate an understanding of the method of the present invention and its core ideas. Also, it is within the scope of the present invention to be modified by those of ordinary skill in the art in light of the present teachings. In view of the foregoing, this description should not be construed as limiting the invention.

Claims

1. A complex device health management system incorporating multi-source heterogeneous information, comprising: the system comprises a data interpretation module, a health assessment module, a health trend analysis module and a decision support module;

The data interpretation module is used for carrying out abnormal recognition and preprocessing on the operation parameters of the complex equipment and determining various preprocessed operation parameters and the change trend of each operation parameter; the data interpretation module is used for preprocessing the operation parameters of the complex equipment to obtain a plurality of preprocessed operation parameters; determining any one of the operation parameters as a current operation parameter; acquiring a normal value interval of a current operation parameter; counting the number exceeding the normal value interval in the current operation parameters after pretreatment in a preset time period as out-of-tolerance times; determining the number of times of out-of-tolerance as an abnormal recognition result of the current operation parameters; acquiring a history fitting result of the current operation parameters; fitting the preprocessed current operation parameters in a preset time period by taking time as an abscissa and the current operation parameters as an ordinate to obtain real-time fitting results of the current operation parameters; updating a normal value interval of the current operation parameters according to the real-time fitting result of the current operation parameters; adding the real-time fitting result of the current operation parameters into the history fitting result to obtain the variation trend of the current operation parameters; updating the current operation parameters, and returning to the step of acquiring the normal value interval of the current operation parameters until all the operation parameters are traversed, so as to obtain a plurality of preprocessed operation parameters and the change trend of each operation parameter;

the health evaluation module is used for determining the real-time health state of the complex equipment by utilizing an ER algorithm according to various preprocessed operation parameters; the health evaluation module is used for determining the weight and the reliability of each operation parameter by using a variation coefficient method;

Determining a quality factor for each operating parameter based on a plurality of the weights and a plurality of the reliabilities; determining the confidence coefficient by using an ER algorithm according to the operation parameters after the pretreatment and the quality factors corresponding to each operation parameter; determining a real-time health status of the complex device based on the confidence level; confidence intervals corresponding to different real-time health states are different;

2. The complex device health management system of claim 1, wherein the system further comprises: a data acquisition module is operated;

The operation data acquisition module is connected with the data interpretation module;

the operation data acquisition module is used for acquiring various operation parameters of the complex equipment.

3. The complex device health management system of claim 1, wherein the system further comprises: a display module;

the display module is respectively connected with the health evaluation module, the health trend analysis module, the decision support module and the data interpretation module;

The display module is used for displaying the change trend of each operation parameter, the real-time health state of the complex equipment, the predicted health state of the complex equipment and the standby quantity of each device in the complex equipment.

4. The complex device health management system of claim 1, wherein the system further comprises: a storage module;

the storage module is respectively connected with the health evaluation module, the health trend analysis module, the decision support module and the data interpretation module;

The storage module is used for storing the change trend of each operation parameter, the real-time health state of the complex equipment, the predicted health state of the complex equipment and the standby quantity of each device in the complex equipment.

5. A complex equipment health management method integrating multisource heterogeneous information is characterized by comprising the following steps:

acquiring various operation parameters of complex equipment;

storing and displaying the variation trend of each operation parameter, the real-time health state of the complex equipment, the predicted health state and the standby quantity of each device in the complex equipment;

Performing anomaly identification and preprocessing on the operation parameters of the complex equipment, and determining various preprocessed operation parameters and the change trend of each operation parameter, wherein the method comprises the following steps:

preprocessing the operation parameters of the complex equipment to obtain a plurality of preprocessed operation parameters;

Determining any one of the operation parameters as a current operation parameter;

Acquiring a normal value interval of a current operation parameter;

Counting the number exceeding the normal value interval in the current operation parameters after pretreatment in a preset time period as out-of-tolerance times;

determining the number of times of out-of-tolerance as an abnormal recognition result of the current operation parameters;

acquiring a history fitting result of the current operation parameters;

Fitting the preprocessed current operation parameters in a preset time period by taking time as an abscissa and the current operation parameters as an ordinate to obtain real-time fitting results of the current operation parameters;

Updating a normal value interval of the current operation parameters according to the real-time fitting result of the current operation parameters;

adding the real-time fitting result of the current operation parameters into the history fitting result to obtain the variation trend of the current operation parameters;

Updating the current operation parameters, and returning to the step of acquiring the normal value interval of the current operation parameters until all the operation parameters are traversed, so as to obtain a plurality of preprocessed operation parameters and the change trend of each operation parameter;

Determining the real-time health status of the complex device using an ER algorithm based on the plurality of preprocessed operating parameters, comprising:

Determining the weight and reliability of each operation parameter by using a coefficient of variation method;

determining a quality factor for each operating parameter based on a plurality of the weights and a plurality of the reliabilities;

Determining the confidence coefficient by using an ER algorithm according to the operation parameters after the pretreatment and the quality factors corresponding to each operation parameter;

Determining a real-time health status of the complex device based on the confidence level; confidence intervals corresponding to different real-time health states are different.

6. The method for health management of a complex device incorporating multi-source heterogeneous information of claim 5, wherein determining a predicted health status of the complex device using a predictive algorithm based on a plurality of pre-processed operating parameters comprises:

Inputting a plurality of preprocessed operation parameters into a gray prediction model to obtain a first predicted health state of complex equipment;

inputting a plurality of preprocessed operation parameters into an ARIMA prediction model to obtain a second predicted health state of the complex equipment;

inputting a plurality of preprocessed operation parameters into the LSTM prediction model to obtain a third predicted health state of the complex equipment;

the first predicted health state, the second predicted health state and the third predicted health state are weighted and summed to obtain a fourth predicted health state;

Determining that the first predicted health state, the second predicted health state, the third predicted health state, and the fourth predicted health state are predicted health states of a complex device.

7. The complex equipment health management method based on multi-source heterogeneous information fusion according to claim 6, wherein the gray prediction model is obtained by taking various historical operation parameters of complex equipment as input, taking the health state of the complex equipment at the next historical moment as output, establishing a gray differential equation by using a gray prediction method, whitening the gray differential equation and solving;

The ARIMA prediction model is obtained by taking various historical operation parameters of the complex equipment as independent variables, taking the health state of the complex equipment at the next historical moment as dependent variables, and performing fitting by using an ARIMA prediction method;

The LSTM prediction model is obtained by training an LSTM network by taking various historical operation parameters of the complex equipment as input and taking the health state of the complex equipment at the next historical moment as output.

8. The method for health management of a complex device incorporating multi-source heterogeneous information according to claim 5, wherein determining the number of spares for each device in the complex device based on the plurality of preprocessed operating parameters comprises:

acquiring the service time of each device in the complex equipment;

determining any model as the current model;

Let the spare number i=0 for the current model;

determining any device of the current model in the complex equipment as the current device;

when the difference value between the service time of the current device and the service time threshold value of the current device reaches a difference value threshold value, acquiring the current device position as the position of the to-be-replaced device;

Determining the current model as a standby model;

Increasing the number of standby numbers of the current model by 1;

Updating the current device, and returning to the step of acquiring the current device position as the to-be-replaced device position until all devices of the current model are traversed when the difference value between the service time of the current device and the service time threshold of the current device reaches the difference value threshold, and determining the standby number of the current model and a plurality of to-be-replaced device positions;

updating the model devices, returning to the step of enabling the standby number i=0 of the current model until the model of all devices in the complex equipment is traversed, and determining the standby number of different models and a plurality of positions of the to-be-replaced pieces.