A general instrument health assessment method based on neural networks

Abstract

This invention relates to a general instrument health assessment method based on neural networks, belonging to the field of instrumentation. The invention acquires calibration data from M newly purchased instruments and M in-use instruments that have undergone at least one periodic calibration. The data is then shuffled and merged, and a suitable data structure for training is designed. A backpropagation (BP) algorithm is designed to distinguish between new and old attributes. If the calculation results indicate that a health baseline does not exist, the instrument's health status in the current operating environment is the same as when it was purchased, and its performance remains good. If the conclusion indicates that a health baseline exists, a state-of-the-art (SOFM) algorithm is designed to classify the instrument's health status through the topological structure of the calibration data. This invention, starting from data analysis, brings a new technical means to metrological assurance, effectively promoting the informatization and intelligentization of metrological testing.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of instruments and meters, and particularly relates to a general instrument and meter health condition evaluation method based on a neural network. BACKGROUND

[0002] Metrology guarantee, which contains management of measurement equipment and measurement process, is an indispensable part in engineering construction.

[0003] Traditional methods can only provide compliance certification for single instruments and meters according to verification, calibration and test requirements in metrology technical specifications, that is, rely on mandatory verification to provide use permission within a certain period. The method cannot investigate and analyze the health condition in use, and the risk prevention and control based thereon cannot be carried out, and unit engineering construction is plagued by possible instrument and meter damage at any time.

[0004] Investigation shows that in the field of metrology, there is no method to analyze instrument and meter original data by using a neural network to obtain the health condition in use. The closest existing technology is a method and system for evaluating metrological process capability of an automatic verification system. The method obtains original data of basic errors, analyzes to obtain metrological process capability of the verification system, and realizes quantitative evaluation. The method is only for checking of low-voltage current transformers, has strong limitations, still relies on traditional metrology technical specifications in principle, and needs additional checking operation, cannot analyze using existing data, and cannot evaluate the health level of instruments and meters in use. SUMMARY

[0005] (I) Technical problem to be solved

[0006] The technical problem to be solved by the application is how to provide a general instrument and meter health condition evaluation method based on a neural network to solve the problem that traditional methods cannot analyze using existing data and cannot evaluate the health level of instruments and meters in use.

[0007] (II) Technical scheme

[0008] In order to solve the above technical problem, the application provides a general instrument and meter health condition evaluation method based on a neural network, which comprises the following steps:

[0009] S1, obtaining verification data of M pieces of newly purchased instruments and meters and verification data of M pieces of instruments and meters in use which have experienced at least one periodical verification, dispersing and merging the two data, and designing a data structure suitable for training;

[0010] S2, designing a BP algorithm, using the algorithm to distinguish new and old attributes, training the data in a parallel algorithm in the BP algorithm, verifying the data in a serial algorithm in the BP algorithm, and proving whether there is a baseline according to the conclusion.

[0011] S3, as the calculation result shows that the health condition baseline does not exist, the study is ended and it is concluded that the health condition of the instrument is the same as that when it is purchased, and the performance is good;

[0012] S4, as the conclusion shows that the health condition baseline exists, the SOFM algorithm is designed to realize the topological structure classification of the test data, the cluster center point is found on the appropriate scale, and then the health condition level division of the instrument is studied and it is concluded that the health condition of the instrument in the current use environment has a clear threshold, and the health level can be divided into several categories according to the threshold.

[0013] (Three) beneficial effects

[0014] The application provides a general instrument health condition evaluation method based on a neural network, and the key point of the application is that:

[0015] Data system design. The application takes a pressure gauge as an example to perform scientific, rigorous and concise data reconstruction, the example can be directly put into use, and the example can be used as a template to bring a universal method design to the data system of various instruments;

[0016] Custom BP algorithm combination. The application designs a parallel-serial training combination relying on a BP neural network, and the optimal learning rate, momentum factor and iteration number and other parameters are assigned with clear expressions, which is a necessary means for judging the health condition baseline of the general instrument;

[0017] Threshold division of the custom SOFM algorithm. The application takes a pressure gauge as an example to design a set of scientific, rigorous and intuitive clustering drawing and threshold division method, which is a necessary means for dividing the health condition baseline of the general instrument, and the generated result fully refers to and respects the working experience and industry habit, is highly practical, and the algorithm itself also has excellent adaptability and flexibility;

[0018] Overall evaluation method. The application forms a complete and sufficient instrument health condition evaluation algorithm after a large number of tests and test verifications, and the algorithm has high completion degree and excellent generalization ability.

[0019] The application provides a complete and scientific general instrument health condition evaluation method by introducing a neural network algorithm, taking the health condition quantitative research process of a pressure gauge as an example, and the research result has excellent generalization ability.

[0020] The pressure gauge example introduced in the present application can be directly used as a method basis for evaluating the health status of the pressure gauge in use by various enterprises, factories and other production units in the society, and the modular execution structure provides a scientific and rigorous guideline for health status evaluation of other general instruments and meters, effectively solves the risk prevention and control problem in the current use of instruments and meters from the data analysis point of view, and better fills in a blank of safety control of in-use equipment;

[0021] The evaluation method introduced in the present application is simple and effective, the analysis starting point is conventional and general, has good generalization ability. In the application process, only the instrument and meter data itself needs to be concerned, without additional operation equipment and field detection, and all execution processes only rely on common PC software such as MATLAB and EXCEL, the generated result form can also be determined by the judgment of the detection personnel according to their industry habit, the overall evaluation method has almost 0 learning cost, almost 0 economic cost, 0 special person and special equipment demand, and the benefit is very excellent;

[0022] The general research method, modular analysis process, open data system design and result-oriented of the present application not only provide reference basis for the health status of general instruments and meters, but also provide auxiliary decision for periodic verification and evaluation of instruments and meters based on the present application, and the research idea for in-situ verification is also inspired.

[0023] In summary, the present application is one of the few technical ability breakthroughs in the field of measurement from the informationization point of view, the present application takes data analysis as the starting point, brings a new technical means for measurement guarantee, and effectively promotes the informationization and intelligence of measurement and detection.

[0024] The present application solves the health status evaluation method of general instruments and meters in use, realizes the progress from nothing to something, and there is no alternative scheme. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The flowchart of the present application;

[0026] Figure 2 The original record format of the pressure gauge;

[0027] Figure 3 The basic structure diagram of the BP neural network in the prior art;

[0028] Figure 4 The BP network flowchart;

[0029] Figure 5 The training data form diagram;

[0030] Figure 6 The initial structure diagram of the SOFM in the prior art;

[0031] Figure 7The flow chart of SOFM network algorithm is shown in the figure;

[0032] Figure 8 The schematic diagram of data initial drawing and clustering is shown in the figure;

[0033] Figure 9 The schematic diagram of 3*5 cellular topology network clustering and center point is shown in the figure;

[0034] Figure 10 The schematic diagram of SOFM clustering center point is shown in the figure. DETAILED DESCRIPTION

[0035] In order to make the purpose, content and advantages of the present application more clear, the specific embodiment of the present application is described in further detail below in combination with the drawings and examples.

[0036] The present application aims at the health condition of in-use instruments and meters in metrological guarantee, introduces two kinds of neural networks of BP and SOFM, designs sample data system and network algorithm through different characteristics of the network, and generates a scientific and complete evaluation method through algorithm practice and data analysis, so as to provide an effective analysis scheme for the health condition research of in-use instruments and meters.

[0037] As shown in the figure, Figure 1 The analysis process of the present application is as follows:

[0038] S1, obtaining the calibration data of M pieces of newly purchased instruments and meters and the calibration data of M pieces of in-use instruments and meters which have experienced at least one periodical calibration, dispersing and merging the two data, and designing a data structure suitable for training;

[0039] S2, designing BP algorithm (Back-Propagation), using the algorithm to distinguish new and old attributes. The data is trained in the parallel algorithm in the BP algorithm and verified in the serial algorithm in the BP algorithm, and whether the health condition has a baseline is proved according to the conclusion;

[0040] S3, if the calculation result shows that the health condition baseline does not exist, the research is ended and the conclusion is drawn that the health condition of the instrument and meter is not different from that when it is purchased, and the performance is good;

[0041] S4, if the conclusion shows that the health condition baseline exists, SOFM algorithm (Self-organizing Feature Map) is designed to realize the topological structure classification of the calibration data, the clustering center point is found on a suitable scale, and then the health condition level division of the instrument and meter is researched and the conclusion is drawn that the health condition of the instrument and meter in the current use environment has a clear threshold, and the health level can be divided into several categories according to the threshold.

[0042] The specific implementation steps of the present application are as follows:

[0043] 1. Data system design (taking a pressure gauge as an example, and demonstrating an excellent data system design for pressure gauges; this is hoped to be used as a point of innovation).

[0044] Taking a general pressure gauge (hereinafter referred to as a pressure gauge) as an example, this paper introduces the process of data reconstruction and provides a universal method design for the data system of various instruments and meters.

[0045] The calibration items for pressure gauges mainly include the indication error, hysteresis error, and tap displacement during both the pressure increase and decrease strokes. The pressure increase and decrease indication data directly characterize the pressure gauge's performance during operation. The recording format is as follows: Figure 2 As shown:

[0046] Considering the generalization ability of the research method, the collected pressure gauge sample data needs to be reconstructed. Pressure gauges in use have different ranges and accuracy classes, and their error levels cannot be directly compared. Therefore, the relative error is designed as follows:

[0047]

[0048] Where r represents the relative error based on the maximum permissible error limit, e represents the error value of a single data point, and MPE represents the maximum permissible error. The algorithm for the maximum permissible error is as follows:

[0049] MPE=±(FS×LOA%) (2)

[0050] FS represents the full-scale value, which is the range for a typical pressure gauge; LOA represents the accuracy class.

[0051] The relative error *r* can adequately represent the current error level of the pressure gauge, but if a clearer picture is desired in network training and result visualization, the training data samples need to be organized in a two-dimensional or three-dimensional format. Several data points from each pressure gauge are summarized into two typical feature values, calculated as follows:

[0052] r 11 =Max(r) (2) ,r (3) ,…,r (m) (3)

[0053]

[0054] Where m represents the number of inspection points, r 11 r 12 The subscript form represents the first and second typical characteristic values ​​of the first pressure gauge. The zero point, r, is removed from the value. (1) The reason for this is that the pressure gauge often has a stop pin at this point, which makes the zero-point error unable to accurately express the error level. 11 r 12The array forms a basic description of the health status of the pressure gauge, where r 11 It can effectively reflect the linear error of a single pressure gauge or nonlinear data jumps caused by structural damage, etc. 12 It can systematically describe the indication error level of the pressure gauge; r 11 r 12 The data sets were constructed without deep processing to provide bias-free input to the neural network as much as possible. All data sets also underwent preliminary normalization during the reconstruction process described above, resulting in the following simplified form:

[0055] R = {r} 11 ,r 12 ),(r 21 ,r 22 ),…,(r n1 ,r n2 (5)

[0056] (Here, n represents the total number of pressure gauge data points, and R represents the data structure suitable for training.)

[0057] 2. Design and Implementation of a Custom BP Network Algorithm

[0058] Backpropagation (BP) neural networks are a type of feedforward network. They are widely used in classification, approximation, regression, and compression. In practice, approximately 80% of neural network models utilize BP networks or their variations.

[0059] like Figure 3 As shown, the basic structure of a BP network in the prior art is as follows:

[0060] For sample set

[0061] S = {(X1,Y1),(X2,Y2),…,(X...} n ,Y n )}

[0062] Network according to (X) n ,Y n The actual output was calculated to be 0. n And error measure E n , for W ⑴ W ⑵ ,…,W (M) Perform the nth adjustment. This process is equivalent to iteratively processing each sample in the sample set. This loop needs to be repeated until the sum of the error measures for the entire sample set meets the system requirements.

[0063] BP Neural Network Design and Implementation

[0064] First, a batch training method is used, where all the reconstructed training samples R are input simultaneously, and the overall error is calculated. The network flow is as follows: Figure 4 As shown.

[0065] S21. Data Input and Partitioning. Store the data structure R in an XLS format table, listing the number, attribute, and r... n1 and r n2 ,like Figure 5 As shown in the diagram, MATLAB is used to read the above table and assign values ​​to the set parameter `data`, where 1 represents the new attribute table and 0 represents the old attribute table. This attribute is stored in `label`. The `randperm` function is used to generate random positive integers to shuffle the data, dividing it into training data (`train`) and test data (`test`).

[0066] S22. Initialize the BP network and calculate the sample error. The neural network contains one hidden layer, and the training method uses the steepest descent method with momentum, performed in batches. Since the output value of the output layer is either 0 or 1, the transfer function of both the hidden and output layers uses the Log-Sigmoid function:

[0067]

[0068] The threshold of the BP neural network is incorporated into the weights of the BP neural network (each input of the BP neural network corresponds to a corresponding weight, and the boundary for judging whether the input wins is called its threshold, which is its basic property and is the structural description in the basic method). A function is written to construct the network, and the weights are initialized to a small random number. In order to speed up the training, the number of hidden layer neurons is set to 3.

[0069] To offset the relative error rate of more than 100% that may exist in the data structure R, the samples must be normalized again: the data is moved to the center of the coordinate axis by subtracting the sample mean, and then the variance is standardized by dividing by the sample standard deviation. After completion, the samples are input into the BP network to calculate the error.

[0070] S23. Determine if the error has converged and adjust the weights accordingly. Define an error tolerance, here set to 0.01. The algorithm converges when the sum of squares of the sample errors is less than this tolerance. Also, give a maximum number of iterations, set to 2000. The iteration stops when this number is reached.

[0071] The weights are adjusted according to the following formula:

[0072]

[0073] Where Δω is the weight modification amount, Let η be the local gradient and η be the learning rate. This is the input signal for the entire network. Further optimization of the formula is achieved using the steepest descent method with a momentum factor. Therefore, except for the first iteration, subsequent iterations must consider the weight modifications from the previous iteration.

[0074]

[0075] Where α is the momentum factor. This represents the local gradient.

[0076] S24. Testing. Use a BP network for attribute recognition, and obtain sample distribution, misclassifications, number of iterations, and accuracy. During testing, the learning rate, momentum factor, and number of iterations need to be continuously adjusted.

[0077] S25. Algorithm Optimization and Validation. The above method uses batch training (i.e., parallel algorithm), which is prone to local optima, leading to a lack of further training correction. Designing a serial algorithm, which randomly inputs samples one by one, can avoid this problem to some extent. This method can also improve the ability to partition training samples while validating the baseline determination of the parallel algorithm. During testing of the serial training method, the learning rate, momentum factor, and number of iterations also need to be continuously adjusted.

[0078] S26. Algorithm Practice. The above-mentioned parallel and serial algorithms have been sufficiently trained to obtain the health status of the instruments at the time of manufacture. When the measured data of the instruments in use are substituted into the BP parallel algorithm and the serial algorithm, if both can obtain definite and stable calculation results, it can be concluded that under the current usage conditions, this batch of instruments has a clear health status baseline and its health status can be assessed (Note: If the calculation result is close to 0 or 100%, it indicates that the health status of this batch of instruments is very consistent, and there is no need to further divide its health status threshold).

[0079] 3. Design and Implementation of a Custom SOFM Neural Network Algorithm

[0080] The custom SOFM algorithm still follows the process of input-network construction-iterative update-training classification, introduces a learning radius, and uses the similarity and topological structure of the data itself for clustering.

[0081] Introduction to SOFM Neural Networks

[0082] The SOFM self-organizing competitive neural network employs a completely different approach from feedforward neural networks. It utilizes the idea of ​​competitive learning, where the network's output neurons compete with each other, and only one output neuron wins at any given time; this is called the winning neuron. The biological basis of this neural network is the lateral inhibition phenomenon between neurons. Its initial structure is as follows: Figure 6 As shown.

[0083] The basic algorithm can be summarized as follows: In each computation, only one output neuron wins, and the winning neuron is marked as 1, while the remaining neurons are marked as 0. Initially, the weights from the input layer to the core layer are randomly assigned, so each core layer neuron has an equal probability of winning, but eventually, there will be a neuron with the strongest excitation. This strongest neuron is further strengthened in the weight modulation, and through this competitive method, the distribution information of the training samples, i.e., the sample category, can be obtained. When a new sample is input, pattern classification can be performed based on the excited neuron.

[0084] The method is designed as follows:

[0085] S41. Data Loading and Initial Plotting. Use the dataset function to load the original data structure R, create a self-organizing map network (selformap) to perform initial clustering of the data, and introduce the hist function to perform initial plotting of the clustering results (when plotting initially, the hist category can be set to a larger value to facilitate the judgment of clustering features);

[0086] S42. Data Topology Clustering and Centroid Plotting. The `plotsomhits` function is introduced to perform more accurate cellular topology division of the data, and the `plotsompos` function is introduced to plot the centroid data points for each category. The threshold is determined by connecting the centroids to define the hierarchical levels.

[0087] S43. Threshold Division. Create a self-organizing competitive neural network (SOFM), and train and test the data based on the threshold definition levels given in the above process to divide the data into scientifically reasonable threshold ranges.

[0088] like Figure 7 As shown, the custom SOFM network calculation process is as follows:

[0089] S431. Construct the SOFM network. The data structure R to be analyzed is defined according to the above thresholds, and the learning rate is set.

[0090] S432. Normalize the data structure R that needs to be input into the network for analysis;

[0091] S433. The sample input network is trained directly using a serial training method since no verification process is required.

[0092] S434. Adjust the algorithm weights of the SOFM input layer to make the algorithm result converge;

[0093] S435. Convergence Assessment and Testing. If the training process converges, it indicates that the neural network is arriving at a clear quantitative conclusion, the final result points to a definite classification, and the classification result remains constant after several tests.

[0094] At this point, the classification of health thresholds for pressure gauges has been completed.

[0095] 4. Algorithm Examples

[0096] A batch of 168 pressure gauges was selected as the actual measurement data. After system design, the data was successively substituted into the parallel and serial algorithms of a custom BP network for training. After several repetitions, the results fluctuated around 37%, indicating that under the current usage conditions, the selected batch of pressure gauges has a clear health baseline and its health threshold can be further defined.

[0097] Therefore, the evaluation algorithm is continued, and an SOFM network is introduced to plot the above 168 sets of data, using an 8*8 topology network for clustering:

[0098] Figure 8 Among the 64 classification frames, some categories had only 1 or even 0 categories, indicating that the data dispersion was not at that level. Considering that the relative error rate scale has 5 regions in increments of 0.2 from 0 to 1, and that r can be derived from formulas (3) and (4) n2 ≤r n1 Considering the aforementioned degree of dispersion, the network size is reduced to 3*5, allowing for the creation of a more intuitive cellular topology diagram (based on...). Figure 6 The interpretation of the clustering results on the right aims to detail the analytical process of correct clustering (from dividing into multiple classes to determining if there are too many classes, reducing the number of classes to obtain a suitable classification scale), and to plot the data points accordingly. Figure 9 As shown:

[0099] The gradient formed by connecting the center points does not reflect the health trend of a single pressure gauge, but several radial networks can provide a threshold division radius for the health status of the pressure gauges in use.

[0100] The SOFM algorithm was redesigned. Based on the three data types defined during baseline finding—"new," "old," and linearly inseparable—and combined with the commonly used evaluation methods of "good," "medium," and "poor," the threshold categories were divided into three corresponding categories (the preceding steps started with data reconstruction, first examining whether the data could distinguish between the new and old states, and then, combining the common good, medium, and poor classification patterns, the intermediate areas outside these two states were separately classified into a new category, thus clarifying the category division into three types: good, medium, and poor, corresponding to new, inseparable, and old, respectively). Finally, with 2000 iterations and the learning rate and learning radius adaptively decreasing with the number of iterations, a relatively rigorous clustering result was obtained, as shown below. Figure 10 As shown.

[0101] In conclusion, under the current sampling conditions, the health status of the pressure gauge can be quantified and can be divided into three thresholds to represent the health level: "good", "moderate" and "poor".

[0102] The key point of this invention is:

[0103] Data system design. This invention uses a pressure gauge as an example to conduct a scientific, rigorous, and concise data reconstruction. This example can be directly put into use and serves as a template for the design of a universal method for the data systems of various instruments and meters.

[0104] Custom BP algorithm combination. This invention designs a parallel-serial training combination based on BP neural network, and clearly describes the optimal learning rate, momentum factor and number of iterations, which is a necessary means to judge the baseline problem of the health status of general instruments and meters;

[0105] This invention presents a custom thresholding method for the SOFM algorithm. Using pressure gauges as an example, it designs a scientific, rigorous, and intuitive clustering and thresholding method, which is a necessary means to define the baseline of health status for general instruments and meters. The results fully reference and respect professional experience and industry practices, are highly practical, and the algorithm itself has excellent adaptability and flexibility.

[0106] The overall assessment method. After extensive testing and experimental verification, this invention has developed a complete and comprehensive algorithm for assessing the health status of instruments and meters. The algorithm has a high degree of completeness and excellent generalization ability.

[0107] This invention introduces a neural network algorithm and, taking the quantitative research process of the health status of in-use pressure gauges as an example, provides a complete, scientific, and universal method for assessing the health status of instruments and meters. The research results have excellent generalization ability.

[0108] The pressure gauge examples described in this invention can be directly used as a method for assessing the health status of pressure gauges in use by various enterprises, factories and other production units. Moreover, the modular execution structure provides a scientific and rigorous guide for assessing the health status of other general-purpose instruments and meters. From the perspective of data analysis, it effectively solves the risk prevention and control problem in the current use of instruments and meters, and fills a gap in the safety management of equipment in use.

[0109] The evaluation method introduced in this invention is simple and effective, with a conventional and universal starting point for analysis and good generalization ability. During application, only the instrument data itself needs to be considered; no additional equipment operation or on-site testing is required. Furthermore, all execution processes rely solely on common PC software such as MATLAB and Excel, and the format of the results can be determined by the testing personnel based on their industry practices. The overall evaluation method has virtually zero learning cost, virtually zero economic cost, and requires zero specialized personnel or equipment, resulting in excellent benefits.

[0110] The invention’s universal research method, modular analysis process, open data system design and results orientation not only provide a reference for the health status of general instruments and meters, but also provide auxiliary decision-making for the periodic verification and evaluation of instruments and meters based on these results, and also inspire research ideas for in-situ verification.

[0111] In summary, this invention represents one of the few technological breakthroughs in the field of metrology that takes an information technology perspective. Starting with data analysis, this invention brings new technical means to metrological assurance and effectively promotes metrological testing towards informatization and intelligence.

[0112] This invention solves the problem of a method for assessing the health status of general-purpose instruments during use, representing a significant advancement from scratch, with no alternative solution available at present.

[0113] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A neural network based general purpose instrument health assessment method, characterized by, The method comprises the following steps: S1, obtaining the calibration data of M newly purchased instruments and meters and the calibration data of the in-use instruments and meters that have experienced at least one periodic calibration, merging the two data, and designing a data structure suitable for training; S2, designing a BP algorithm to distinguish new and old attributes using the algorithm; The data is trained in the parallel algorithm in the BP algorithm and verified in the serial algorithm in the BP algorithm, and whether the baseline of the health status exists is verified according to the conclusion; S3, if the result of S2 indicates that the baseline of the health status does not exist, the research is ended and the conclusion is drawn that the health status of the instrument and meter under the current use environment is the same as that when the instrument and meter is purchased, and the performance is good; S4, if the conclusion indicates that the baseline of the health status exists, a SOFM algorithm is designed to realize the topological structure classification of the calibration data; The cluster center point is found on a suitable scale, and the health status level of the instrument and meter is further studied, and the conclusion is drawn that the health status of the instrument and meter under the current use environment has a clear threshold, and the health level can be divided into several categories according to the threshold; Wherein, The step S2 specifically comprises: S21, data reading and division: save the data structure R in an XLS format table, respectively list the number, attribute, r n1 and r n2 ; S22, initializing the BP network and calculating the sample error: the neural network comprises a hidden layer, the training method adopts the steepest descent method containing momentum, and the batch mode is adopted; since the output value of the output layer is 0 or 1, the transmission functions of the hidden layer and the output layer both use the Log-Sigmoid function: （6） The threshold value of the BP neural network is merged into the weight value of the BP neural network, and the weight value is initialized as a small random number; S23, judging whether the error converges and adjusting the weight value according to the judgment: defining an error tolerance, when the sum of squares of the sample error is less than the tolerance, the algorithm converges; in addition, a maximum iteration number is given, and the iteration is stopped when the number is reached; The weight value is adjusted according to the following formula: （7） where, is the weight modification amount, δ i j is the local gradient, η is the learning rate, is the input signal of the whole network; the steepest descent method with momentum factor is used for optimization, thus, in addition to the first iteration, the weight modification amount of the previous iteration needs to be considered in the subsequent iterations: （8） wherein is a momentum factor, is a local gradient; S24, testing: using the BP network to identify the attributes, obtaining the sample distribution, error classification, iteration number and accuracy; the learning rate, momentum factor and iteration number need to be adjusted constantly in the test; S25, algorithm optimization and verification: the above step uses the batch training mode, that is, the parallel algorithm; at the same time, the serial algorithm is designed, the samples are input randomly one by one, the baseline determination of the parallel algorithm is verified, and the ability to divide the training samples is improved; the learning rate, momentum factor and iteration number also need to be adjusted constantly in the test of the serial training method; S26, algorithm practice; the above parallel algorithm and serial algorithm are two over-training methods that have fully obtained the health status of the instrument and meter when it is shipped, if the measured data of the in-use instrument and meter can obtain a certain and stable calculation result when it is brought into the parallel algorithm and the serial algorithm, the conclusion is drawn that the instrument and meter has a clear baseline of the health status under the current use condition, and the health status can be evaluated.

2. The neural network based universal instrument health assessment method as claimed in claim 1, wherein, The instrument and meter is a pressure gauge.

3. The neural network based universal instrument health assessment method as claimed in claim 1, wherein, In the step S1, the pressure value sample data collected is reconstructed, and the relative error is: （1） Wherein, r represents the relative error based on the maximum allowable error limit, e represents the error value of a single data, and MPE represents the maximum allowable error; the algorithm of the maximum allowable error is: （2） Wherein, FS represents the full scale value, for the pressure gauge, it means the range, and LOA represents the accuracy level; The data of each pressure gauge is summarized into two typical characteristic values, and the calculation method is as follows: （3） （4） where m represents the number of calibration points, r 11 , r 12 The subscripted form of r 11 , r 12 performs a preliminary normalization, reducing the form to: （5） Here n represents the total of n pressure gauge data, and R represents the data structure suitable for training.

4. The neural network based universal instrument health assessment method as claimed in claim 1, wherein, In step S21, the table is read using MATLAB and is assigned to the set parameter data, wherein 1 represents the attribute new table, 0 represents the attribute old table, and the attribute is saved in label; the data is scattered by using the randperm function to generate a random positive integer, and is divided into two parts of training data train and test data test.

5. The neural network based universal instrument health assessment method as claimed in claim 1, wherein, In step S22, the data is moved to the center of the coordinate axis by subtracting the sample average value in the data structure R, and is divided by the sample standard deviation to standardize the variance, and after completion, the sample is input into the BP network.

6. The neural network based universal instrument health assessment method as claimed in claim 1, wherein, In step S22, the neural network includes one hidden layer, and the number of hidden layer neurons is 3.

7. The neural network based universal instrument health assessment method as claimed in claim 1, wherein, In step S23, the error tolerance is set to 0.01, and the maximum iteration number is set to 2000 times.

8. The neural network based universal instrument health assessment method according to any one of claims 4-7, characterized in that, The step S4 specifically includes: S41, data loading and initial drawing: the dataset function is used to load the original data structure R, the self-organizing mapping network selformap is created to preliminarily cluster the data, and the hist function is introduced to initially draw the clustering results; S42, data topology clustering and center point drawing: the plotsomhits function is introduced to divide the data in a honeycomb topology, the plotsompos function is introduced to draw the center data points of the classification, and the threshold value is determined by the connection between the center points; S43, threshold value division: the self-organizing competitive neural network SOFM is created, the threshold value hierarchy given by S42 is used to train and test the data, and a scientific and reasonable threshold interval is divided.

9. The neural network based universal instrument health assessment method as claimed in claim 8, wherein, The step S43 specifically includes: S431, constructing the SOFM network, setting the learning rate according to the threshold value hierarchy of the data structure R; S432, normalizing the data structure R; S433, inputting the sample into the network, and directly using the serial training method for training; S434, adjusting the algorithm weight of the SOFM input layer to make the algorithm result point to convergence; S435, judging convergence and testing: if the training process converges, it indicates that the neural network is deriving a clear quantitative conclusion, and the final result points to a certain classification, and after several tests, the classification result still has the ability to remain constant.

Images