Data processing and quality evaluation method for efficiency test data of electric drive system

Abstract

The invention discloses a data processing and quality evaluation method for efficiency test data of an electric drive system. The method comprises the following steps: acquiring an original electric drive efficiency test data set; performing noise reduction on the original electric driving efficiency test data set through a probability distribution-based data noise reduction method to obtain a noise-reduced electric driving efficiency test data set; carrying out the first abnormity detection and processing of the noise reduction electric drive efficiency test data set through a DBSCAN model, and obtaining a preliminary screening electric drive efficiency test data set; carrying out the second anomaly detection and processing of the preliminary screening electric drive efficiency test data set through an LOF algorithm model and an iForest algorithm model, and obtaining a preprocessed electric drive efficiency test data set; and performing quality evaluation on the preprocessed electric drive efficiency test data set through the constructed six-tuple data quality model. According to the invention, effective noise reduction and abnormity detection can be carried out on the efficiency test data of the electric drive system, so that the data quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of quality assessment of electric drive system efficiency test data, and in particular to a data processing and quality assessment method for electric drive system efficiency test data. Background Art

[0002] As the power source of electric vehicles, the efficiency of the electric drive system directly determines the power and economy of the vehicle. Therefore, the efficiency test of the electric drive system has become an indispensable and important part of its development. However, when conducting electric drive system tests, a large amount of high-dimensional data will be generated, and there will be a lot of noise and outliers, which are very difficult to process and have low effectiveness. Traditional data preprocessing methods are difficult to process and evaluate accurately, resulting in a large amount of data accumulation and low utilization rate.

[0003] Traditional data processing methods are mostly based on mathematical statistical theories and are mainly used for statistical analysis of small-scale sample sets with obvious data correlation. It is very difficult and ineffective to use traditional processing methods to process large and complex sample sets.

[0004] In order to solve the problem of data preprocessing, methods such as dimensionality reduction and normalization, data attribute clustering, and correlation analysis in data reduction are often used to achieve data compression and extraction of main influencing factors. In the prior art, there are schemes that extract data features with high correlation in the original data through principal component analysis and Pearson correlation coefficient analysis, which compresses the data set and improves efficiency, but does not fully consider the processing of outliers. There are schemes that use the kernel principal component analysis method to extract data features, but it is easy to fall into local solutions and the calculation is complex when extracting data features. There are schemes that use the Pearson correlation coefficient to analyze the correlation between the characteristic parameters of the capacitor, but the analysis is not sensitive enough to nonlinear analysis and is more sensitive to the processing of outliers and is prone to distortion. There are schemes that use the median rank method to weight the speed of each gear to evaluate the stable transmission efficiency in the straight-through mode, but this method will lose some information in the original data.

[0005] In summary, how to design a method that can effectively reduce noise and detect anomalies in the electric drive system efficiency test data to improve data quality is a technical problem that needs to be solved urgently. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the technical problem to be solved by the present invention is: how to provide a data processing and quality assessment method for the efficiency test data of an electric drive system, firstly, the original data set is denoised by a data denoising method based on probability distribution, secondly, the denoised data set is subjected to a first abnormality check by a DBSCAN model, and finally, the data set is subjected to a second abnormality check by a LOF algorithm model and an iForest algorithm model, thereby improving the data quality of the efficiency test data of the electric drive system. At the same time, a six-tuple data quality model is constructed to perform quality assessment on the processed data set, thereby improving the accuracy and comprehensiveness of the data quality assessment.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] Data processing and quality assessment methods for electric drive system efficiency test data, including:

[0009] S1: Get the original electric drive efficiency test data set;

[0010] S2: De-noising the original electric drive efficiency test data set by a data de-noising method based on probability distribution to obtain a de-noised electric drive efficiency test data set;

[0011] S3: Perform the first abnormality inspection and processing on the noise reduction electric drive efficiency test data set through the DBSCAN model to obtain the preliminary screening electric drive efficiency test data set;

[0012] S4: Performing a second abnormality inspection and processing on the preliminary screening electric drive efficiency test data set through the LOF algorithm model and the iForest algorithm model to obtain a preprocessed electric drive efficiency test data set;

[0013] S5: Quality assessment of the preprocessed electric drive efficiency test dataset is performed using the constructed six-tuple data quality model.

[0014] Preferably, the electric drive efficiency test data includes the electric operation efficiency and the power generation operation efficiency;

[0015] The calculation formula of electric working efficiency is expressed as:

[0016]

[0017] The calculation formula of power generation efficiency is as follows:

[0018]

[0019] Wherein: η1 represents the electric driving efficiency of the electric drive system; η2 represents the power generation efficiency of the electric drive system; r1 represents the speed of the left dynamometer; T1 represents the torque of the left dynamometer; r2 represents the speed of the right dynamometer; T2 represents the torque of the right dynamometer; P1 and P2 are the power of the DC bus of the electric drive system controller.

[0020] Preferably, in step S2, the processing steps of the data denoising method based on probability distribution are as follows:

[0021] S201: performing probability distribution analysis on the original electric drive efficiency test data set; dividing the data points in the original electric drive efficiency test data set into a plurality of intervals according to the law of probability distribution;

[0022] S202: Select the median of the data points in each interval as the data representative value of the interval;

[0023] S203: Calculate the mean and standard deviation of the data representative values ​​of all intervals;

[0024] S204: Calculate the noise lower limit judgment value according to the mean and standard deviation of the data representative value combined with the Laida criterion;

[0025] S205: treating data points in the original electric drive efficiency test data set that are lower than the noise lower limit judgment value as noise points;

[0026] S206: Process all noise points in the original electric drive efficiency test data set to obtain a noise-reduced electric drive efficiency test data set.

[0027] Preferably, in step S3, the step of performing the first abnormality inspection and processing by using the DBSCAN model includes:

[0028] S301: Determine the neighborhood radius r of the DBSCAN model and the minimum number of samples Minpts required to form a dense area;

[0029] S302: Initialize the unvisited data set D, the core object set, and the number of clusters of the DBSCAN model; add all data points in the noise reduction electric drive efficiency test data set to the unvisited data set D;

[0030] S303: Randomly select a data point p from the unvisited data set D as the current observation point;

[0031] S304: Calculate all sample points within the neighborhood radius r of the current observation point p, that is, |N r (p)|;

[0032] S305: If | N r(p)|≥Minpts, then the data point p is regarded as a core point; otherwise, the data point p is regarded as an outlier and added to the outlier list;

[0033] S306: repeating steps S303 to S305 until a core point q is found, and adding the core point q to the core object set; if no core point is found, adding the uncollected data points to the outlier list;

[0034] S307: Establish cluster C at the core point q, and add all data points within the neighborhood radius r of the core point q to cluster C, that is, N r (p) = {q∈D|dist(p,q)≤r}; at the same time, the data points added to cluster C are removed from the unvisited data set D; whenever a new cluster is formed, the cluster number counter is increased by 1;

[0035] S308: Repeat steps S303 to S307 until the unvisited data set D is empty, and output the clustering results and the list of outliers;

[0036] S309: Processing the data points in the noise reduction electric drive efficiency test data set that belong to the abnormal value list to obtain a preliminary screening electric drive efficiency test data set.

[0037] Preferably, in step S301, the optimal neighborhood radius and minimum number of samples of the DBSCAN model are solved by a particle swarm optimization algorithm as input parameters of the DBSCAN model;

[0038] The processing steps of the particle swarm optimization algorithm include:

[0039] S3011: Set the relevant parameters of the particle swarm optimization algorithm; set the search range of the neighborhood radius r and the minimum number of samples Minpts as the constraints on the position and speed of the particles;

[0040] S3012: Randomly initialize the speed, position, individual optimal solution and global optimal solution of all particles within the set search range;

[0041] S3013: Calculate the fitness function value of each particle in each evolution process through the fitness function;

[0042] The calculation formula of the fitness function is as follows:

[0043]

[0044] Where: S represents the silhouette coefficient; A represents the average distance of the sample point of the same category to other samples; B represents the average distance between the sample and other clusters;

[0045] S3014: Update the speed and position of particles;

[0046] The formula is:

[0047]

[0048] Where: represents the updated particle velocity; ω represents the inertia weight; represents the current particle speed; c1 and c2 represent learning factors; r1 and r2 represent random data of [0,1], which are used for the initial random position; represents the individual optimal value at time t; Indicates the current particle position; represents the global optimal value at time t; Indicates the particle update position;

[0049] S3015: Compare the fitness values ​​of each particle and update the individual optimal solution and the global optimal solution of each particle;

[0050] S3016: Determine whether the set algorithm termination condition is met: if so, the global optimal solution, that is, the optimal neighborhood radius and the minimum number of samples, is used as the output result; otherwise, return to step S3013.

[0051] Preferably, in step S3011, the inertia weight of the particle swarm optimization algorithm is optimized by linear reduction, that is, the inertia weight of the particle swarm optimization algorithm is designed to gradually decrease linearly from ω1 to ω2 with the number of iterations;

[0052] The formula is:

[0053]

[0054] Where: ω represents the inertia weight; ω1 and ω2 represent the upper and lower limits of the inertia weight; i represents the current number of iterations; iter max Indicates the maximum number of iterations.

[0055] Preferably, in step S4, the step of performing a second abnormality inspection and processing by using the LOF algorithm model and the iForest algorithm model includes:

[0056] S401: constructing LOF algorithm model and iForest algorithm model;

[0057] S402: performing anomaly detection on the preliminary screening electric drive efficiency test data set by using the LOF algorithm model to obtain a first anomaly score for each data point in the preliminary screening electric drive efficiency test data set;

[0058] S403: performing anomaly detection on the preliminary screening electric drive efficiency test data set by using the iForest algorithm model to obtain a second anomaly score for each data point in the preliminary screening electric drive efficiency test data set;

[0059] S404: performing weighted calculation on the first anomaly scores and the second anomaly scores of all data points to obtain a final anomaly score of the data point;

[0060] S405: taking the data points whose final anomaly scores are lower than the preset score threshold as outliers for the second outlier test;

[0061] S406: Processing the outliers of the second outlier test in the initial screening electric drive efficiency test data set to obtain a preprocessed electric drive efficiency test data set.

[0062] Preferably, in step S402, the processing steps of the LOF algorithm model for anomaly detection include:

[0063] S4021: Determine a neighborhood parameter N of the LOF algorithm model;

[0064] S4022: For each data point in the preliminary screening electric drive efficiency test data set, calculate the distance to the Nth nearest neighbor data point, i.e., the k distance;

[0065] S4023: Calculate the reachable distance of each data point according to its k distance;

[0066] The formula is:

[0067] reach_dist k (p,o)=max{d k (o),d k (p,o)};

[0068] Where: o represents the kth observation point closest to data point p; d k (o) represents the k distance of observation point o; d k (p,o) represents the distance between data point p and observation point o;

[0069] S4024: Calculate the local reachability density ρ of each data point based on its reachable distance k (p);

[0070] The formula is:

[0071]

[0072] Where: N k (p) represents the number of samples of data point p in the k-distance neighborhood;

[0073] S4025: Calculate the local reachability density ratio LOF of each data point based on its local reachability density k (o);

[0074] The formula is:

[0075]

[0076] Where: k (o) represents the local reachability density of observation point o; N k (o) represents the number of samples of observation point o within the k-distance neighborhood;

[0077] S4026: Calculate the LOF score, i.e., the first anomaly score, of each data point according to its local reachable density ratio.

[0078] Preferably, in step S403, the processing steps of the iForest algorithm model for anomaly detection include:

[0079] S4031: Determine the number of isolated trees t of the iForest algorithm model and the number of samples T for training each isolated tree;

[0080] S4032: For each isolated tree, perform the following steps:

[0081] 1) Randomly select T data points from the initial screening electric drive efficiency test data set as a subsample set and put them into the root node of the isolation tree;

[0082] 2) Randomly select a feature dimension and randomly generate a cutting point p in the current node data on this dimension;

[0083] 3) Generate a hyperplane with the cutting point p to divide the current node data space into two subspaces: put the data points less than p in the specified dimension in the left child node of the current node, and put the data points greater than or equal to p in the right child node of the current node;

[0084] 4) Recursively execute steps 2) and 3) in the child node, and continuously construct new child nodes until there is only one data in the child node or the child node has reached the specified height;

[0085] S4033: Repeat step S4032 until t isolated trees are generated to form an isolated forest;

[0086] S4034: for each data point x in the preliminary screening electric drive efficiency test data set, traverse each isolated tree in the isolation forest;

[0087] S4035: In each isolated tree, calculate the path length h(x) of the data point x from the root node to the leaf node;

[0088] The formula is:

[0089] h(x)=e+C(T);

[0090] Where: e represents the number of nodes that sample x has traveled from the root node to the leaf node; T represents the number of samples at the same leaf node as sample x; C represents a correction value;

[0091] S4036: Calculate the average path length E(h(x)) of the data point x in t isolated trees;

[0092] The formula is:

[0093]

[0094] In the formula: n represents the total number of isolated numbers;

[0095] S4037: Calculate the second anomaly score of the data point x by using the average path length E(h(x));

[0096] The formula is:

[0097] s(x,n)=2-c(n)E(h(x));

[0098]

[0099] Where: s(x,n) represents the second anomaly score of data point x; c(n) represents the average path length when the binary sort tree search is unsuccessful; n represents the number of samples used to generate each isolated tree; H(n-1) represents the harmonic number.

[0100] Preferably, in step S5, the six-tuple data quality model calculates the quality comprehensive score of the pre-processed electric drive efficiency test data set by the following formula to achieve quality assessment;

[0101] S=0.2554a+0.0850b+0.1185c+0.1076d+0.1649e+0.2683f;

[0102] in:

[0103]

[0104] Where: S represents the comprehensive quality score of the data set; a, b, c, d, e, and f represent the data accuracy, data integrity, data uniqueness, data trend consistency, data stability, and data efficiency of the data set, respectively; X1 represents the number of outliers; X2 represents the number of missing values ​​and null values; X3 represents the number of duplicate values; X4 represents the total amount of data with consistent trends; X5 represents the total amount of noise value data; h η represents the proportion of high-efficiency area; p η Indicates the peak efficiency ratio; X6 indicates the total amount of data whose computing efficiency exceeds the preset value; s n represents the score of each peak data; N represents the total amount of data in the data set.

[0105] Compared with the prior art, the data processing and quality assessment method for the electric drive system efficiency test data in the present invention has the following beneficial effects:

[0106] The present invention reduces the noise of the original electric drive efficiency test data set by a data denoising method based on probability distribution. First, the data denoising method based on probability distribution can effectively remove random noise and outliers in the original data set (these noise and outliers may be caused by factors such as measurement errors, equipment failures or external environmental interference, which will interfere with the accuracy of subsequent data analysis), and the denoised data is closer to the true value, thereby improving the accuracy and effectiveness of the overall data set, that is, improving the data quality. Secondly, the denoising process makes the data points inside the data set smoother and more continuous, reduces the jumps and mutations between the data points, thereby enhancing the consistency and predictability of the data, which is crucial for subsequent data modeling and analysis, and can reduce the complexity of model fitting and improve the generalization ability of the model. Finally, the denoising process reduces the redundant information and noise points in the data set, so that the subsequent data processing and analysis process can be executed faster, especially for large-scale data sets such as electric drive system efficiency test data, the denoising process can significantly reduce the amount of calculation and optimize the use efficiency of computing resources.

[0107] After the present invention performs noise reduction on the original data set, the first anomaly test is further performed on the noise-reduced data set through the DBSCAN model. DBSCAN can effectively identify the abnormal points in the data set (i.e., those points with significant density differences from most data points). By performing the first anomaly test on the noise-reduced data set through the DBSCAN model, those abnormal data points that may be caused by equipment failure, operating errors, or extreme working conditions can be accurately identified; by removing these abnormal data points, a purer and more reliable data set can be obtained, which not only improves the data quality of the data set, but also provides a more solid foundation for subsequent data analysis and modeling. At the same time, the presence of abnormal data may have a negative impact on subsequent data analysis and modeling, causing the results to deviate from the actual situation or produce misleading conclusions. Through the abnormal inspection and processing of the DBSCAN model, this risk can be significantly reduced, and the robustness and reliability of subsequent data analysis can be enhanced.

[0108] After the present invention performs the first anomaly check through the DBSCAN model, it further performs a second anomaly check on the data set through the LOF (local anomaly factor) algorithm model and the iForest (isolation forest) algorithm model. First, the LOF algorithm model and the iForest algorithm model each have unique detection mechanisms and advantages. By using these two algorithms at the same time to perform a second anomaly check on the primary screening data set, complementary verification between multiple algorithms can be achieved, thereby further improving the accuracy and comprehensiveness of anomaly detection, and then improving data quality. Secondly, the LOF algorithm detects abnormal points by calculating the local anomaly factor of each data point, while the iForest algorithm detects abnormal values ​​in the data by constructing an isolated tree. The combined use of the two algorithms can more deeply explore the potential abnormal patterns in the data and reduce missed detections and false detections. Finally, after the second anomaly check and processing, the obtained pre-processed electric drive efficiency test data set will be purer and more reliable, which not only improves the quality of the data set, but also provides more accurate and effective input for subsequent data analysis and modeling.

[0109] The present invention performs quality assessment on the preprocessed data set through a six-tuple data quality model. First, the six-tuple data quality model includes the accuracy, completeness, uniqueness, trend consistency, stability and efficiency of the data, and can perform a comprehensive quality assessment on the preprocessed data set, not only focusing on the surface characteristics of the data (such as accuracy and completeness), but also deeply exploring the intrinsic properties and relevance of the data (such as trend consistency and stability), thereby improving the accuracy and comprehensiveness of the quality assessment of the electric drive system efficiency test data. Secondly, the data quality assessment results can provide important guidance for subsequent data optimization and improvement. If it is found that the data is insufficient in some aspects (such as low accuracy or poor stability), corresponding measures can be taken to improve and optimize (such as re-collecting data or improving data processing algorithms), thereby improving the quality and availability of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0111] Figure 1 A logic block diagram of a data processing and quality assessment method for an electric drive system efficiency test data;

[0112] Figure 2 It is an electric drive efficiency test system;

[0113] Figure 3 is the efficiency test operating point;

[0114] Figure 4 Efficiency MAP diagram: (a) is the electric mode; (b) is the power generation mode;

[0115] Figure 5 Calculate and analyze the efficiency values ​​at different speeds and combine them with box plots;

[0116] Figure 6 The efficiency test data distribution diagram: (a) is the efficiency frequency histogram; (b) is the efficiency probability distribution diagram;

[0117] Figure 7 This is the flowchart of the secondary outlier test;

[0118] Figure 8 is the K-dist graph;

[0119] Fig. 9 DBSCAN outlier test: (a) is the data outlier score; (b) is the test result;

[0120] Fig.10 It is the convergence process of IPSO-DBSCAN;

[0121] Fig.11 is the result of the secondary outlier test;

[0122] Fig.12 Part of the preprocessed data: (a) is 5000 rpm; (b) is 8000 rpm;

[0123] Fig.13 Evaluation flow chart of the six-tuple data quality model. DETAILED DESCRIPTION

[0124] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0125] The following is a further detailed description through specific implementation methods:

[0126] Example:

[0127] This embodiment discloses a data processing and quality assessment method for electric drive system efficiency test data.

[0128] like Figure 1As shown, the data processing and quality assessment method for the electric drive system efficiency test data includes:

[0129] S1: Get the original electric drive efficiency test data set;

[0130] S2: De-noising the original electric drive efficiency test data set by a data de-noising method based on probability distribution to obtain a de-noised electric drive efficiency test data set;

[0131] S3: Perform the first abnormality inspection and processing on the noise reduction electric drive efficiency test data set through the DBSCAN model to obtain the preliminary screening electric drive efficiency test data set;

[0132] S4: Performing a second abnormality inspection and processing on the preliminary screening electric drive efficiency test data set through the LOF algorithm model and the iForest algorithm model to obtain a preprocessed electric drive efficiency test data set;

[0133] S5: Quality assessment of the preprocessed electric drive efficiency test dataset is performed using the constructed six-tuple data quality model.

[0134] The present invention reduces the noise of the original electric drive efficiency test data set by a data denoising method based on probability distribution. First, the data denoising method based on probability distribution can effectively remove random noise and outliers in the original data set (these noise and outliers may be caused by factors such as measurement errors, equipment failures or external environmental interference, which will interfere with the accuracy of subsequent data analysis), and the denoised data is closer to the true value, thereby improving the accuracy and effectiveness of the overall data set, that is, improving the data quality. Secondly, the denoising process makes the data points inside the data set smoother and more continuous, reduces the jumps and mutations between the data points, thereby enhancing the consistency and predictability of the data, which is crucial for subsequent data modeling and analysis, and can reduce the complexity of model fitting and improve the generalization ability of the model. Finally, the denoising process reduces the redundant information and noise points in the data set, so that the subsequent data processing and analysis process can be executed faster, especially for large-scale data sets such as electric drive system efficiency test data, the denoising process can significantly reduce the amount of calculation and optimize the use efficiency of computing resources.

[0135] After the present invention performs noise reduction on the original data set, the first anomaly test is further performed on the noise-reduced data set through the DBSCAN model. DBSCAN can effectively identify the abnormal points in the data set (i.e., those points with significant density differences from most data points). By performing the first anomaly test on the noise-reduced data set through the DBSCAN model, those abnormal data points that may be caused by equipment failure, operating errors, or extreme working conditions can be accurately identified; by removing these abnormal data points, a purer and more reliable data set can be obtained, which not only improves the data quality of the data set, but also provides a more solid foundation for subsequent data analysis and modeling. At the same time, the presence of abnormal data may have a negative impact on subsequent data analysis and modeling, causing the results to deviate from the actual situation or produce misleading conclusions. Through the abnormal inspection and processing of the DBSCAN model, this risk can be significantly reduced, and the robustness and reliability of subsequent data analysis can be enhanced.

[0136] After the present invention performs the first anomaly check through the DBSCAN model, it further performs a second anomaly check on the data set through the LOF (local anomaly factor) algorithm model and the iForest (isolation forest) algorithm model. First, the LOF algorithm model and the iForest algorithm model each have unique detection mechanisms and advantages. By using these two algorithms at the same time to perform a second anomaly check on the primary screening data set, complementary verification between multiple algorithms can be achieved, thereby further improving the accuracy and comprehensiveness of anomaly detection, and then improving data quality. Secondly, the LOF algorithm detects abnormal points by calculating the local anomaly factor of each data point, while the iForest algorithm detects abnormal values ​​in the data by constructing an isolated tree. The combined use of the two algorithms can more deeply explore the potential abnormal patterns in the data and reduce missed detections and false detections. Finally, after the second anomaly check and processing, the obtained pre-processed electric drive efficiency test data set will be purer and more reliable, which not only improves the quality of the data set, but also provides more accurate and effective input for subsequent data analysis and modeling.

[0137] The present invention performs quality assessment on the preprocessed data set through a six-tuple data quality model. First, the six-tuple data quality model includes the accuracy, completeness, uniqueness, trend consistency, stability and efficiency of the data, and can perform a comprehensive quality assessment on the preprocessed data set, not only focusing on the surface characteristics of the data (such as accuracy and completeness), but also deeply exploring the intrinsic properties and relevance of the data (such as trend consistency and stability), thereby improving the accuracy and comprehensiveness of the quality assessment of the electric drive system efficiency test data. Secondly, the data quality assessment results can provide important guidance for subsequent data optimization and improvement. If it is found that the data is insufficient in some aspects (such as low accuracy or poor stability), corresponding measures can be taken to improve and optimize (such as re-collecting data or improving data processing algorithms), thereby improving the quality and availability of the data.

[0138] In order to better introduce the technical solution of the present invention, this embodiment is described through the following parts.

[0139] 1. Collecting electric drive efficiency test data

[0140] The drive system efficiency test mainly simulates the real load through the efficiency bench and uses various sensors to measure the prototype components and overall performance. The core idea of ​​the efficiency test is to analyze the ratio of output energy to input energy. The electric drive system is mainly divided into power generation mode and electric mode during the test. The efficiency test data is mainly collected in real time through various sensors and power analyzers, so the electric drive efficiency test data includes electric working efficiency and power generation working efficiency.

[0141] The calculation formula of electric working efficiency is expressed as:

[0142]

[0143] The calculation formula of power generation efficiency is as follows:

[0144]

[0145] Wherein: η1 represents the electric working efficiency of the electric drive system; η2 represents the power generation efficiency of the electric drive system; r1 represents the left dynamometer speed (r / min); T1 represents the left dynamometer torque (N·m); r2 represents the right dynamometer speed (r / min); T2 represents the right dynamometer torque (N·m); P1 and P2 are the power of the DC bus of the electric drive system controller (kW).

[0146] The drive system efficiency test in this embodiment refers to GB / T 18488.1-2015 and GB / T 18488.2-2015, and a bench efficiency test is conducted on a certain electric drive assembly as the research object. This test is conducted in a dedicated electric drive system test room. Figure 2 shown.

[0147] The electric drive assembly efficiency bench test system is a complex system, which is mainly composed of hardware and software parts. This embodiment mainly selects and designs the hardware part of the test system, and the software part is assisted by professionals. The hardware part selects the magnetoelectric speed sensor, T40B torque sensor, and DS18B20 digital temperature sensor in turn; the power analyzer uses the Yokogawa WT5000 power analyzer and the Watson EVWS series battery simulator to ensure the smooth progress of the test process and the reliability of the collected test data. After the selection of key measurement components is completed, the test software system needs to be configured at the host computer, and professional assemblers will perform installation and calibration and debugging of the efficiency bench.

[0148] In order to ensure the safety and reliability of the test, a static test is first carried out, and the measurement results are compared with the data provided by the motor manufacturer. Then a dynamic test is carried out to detect the different efficiency conditions of the electric drive system under all working conditions.

[0149] It is stipulated that the test of a certain type of electric drive system starts at 500rpm and the step length is 500rpm until the maximum speed; the torque is divided between 0 and the maximum torque, with a step length of 5Nm below 30Nm and 10Nm above 30Nm. The specific efficiency condition of the electric drive system is obtained as follows Figure 3 shown. Figure 4 It is an efficiency MAP diagram of a certain type of electric drive system drawn based on the efficiency data obtained from the efficiency bench test.

[0150] 2. Data denoising method based on probability distribution

[0151] Since the electric drive efficiency test data is collected according to the test operating conditions and is also affected by the bench test environment, it is necessary to perform feature analysis on the collected test data.

[0152] Through the analysis of data characteristics, it can be seen that the data is seriously dispersed and unstable, the gap between the two extreme values ​​of the data is too large, the quartiles show that the data are mainly concentrated between 84 and 92, the dispersion coefficient is too large, the data is unstable (0.6 is stable), and the skewness / kurtosis: the data is positively skewed and noisy as shown in Table 1.

[0153] Table 1 Data feature analysis table

[0154]

[0155] By calculating and analyzing the efficiency values ​​at different speeds of the test data and combining it with the box plot, such as Figure 5As shown, it can be found that the data is severely discrete and unstable, the data is right-skewed and not flat, and the data contains large noise values. Since the efficiency test data is obtained through the torque and speed sensors at the output end and the power analyzer is used to obtain the real-time input electric power, output mechanical power and calculate the efficiency of the electric drive assembly. It also contains real-time bus current, voltage, temperature and other sensor signals. Therefore, the collected data is a real-time high-dimensional flow data set containing multiple influencing factors.

[0156] In summary, combined with the analysis of the source and characteristics of the test data, it can be seen that the subsequent data processing algorithm should meet the requirements of low time complexity, unsupervised or semi-supervised, universality of abnormal measurement, and strong robustness. According to the above analysis of the characteristics of the test data, it can be seen that the noise value in the efficiency data is large and the data is mainly concentrated in the interval with larger values, but there are also many efficiency values ​​with smaller values ​​that are also normal values. Therefore, the present invention proposes a new denoising method suitable for efficiency test data based on the above analysis of the characteristics of efficiency test data: a data denoising method based on probability distribution.

[0157] Specifically, the processing steps of the data denoising method based on probability distribution are as follows:

[0158] S201: performing probability distribution analysis on the original electric drive efficiency test data set; dividing the data points in the original electric drive efficiency test data set into a plurality of intervals according to the law of probability distribution;

[0159] In this embodiment, according to the characteristics of the electric drive efficiency test data (such as skewness, kurtosis, etc.), a suitable probability distribution model (such as normal distribution, lognormal distribution, Beta distribution, etc.) is selected for fitting, and statistical methods (such as maximum likelihood estimation, moment estimation, etc.) are used to estimate the selected distribution parameters (such as mean, standard deviation, shape parameters, etc.). The data is divided into several intervals according to the distribution parameters to ensure that each interval contains a sufficient number of data points for statistical analysis.

[0160] S202: Select the median of the data points in each interval as the data representative value of the interval;

[0161] S203: Calculate the mean and standard deviation of the data representative values ​​of all intervals;

[0162] In this embodiment, the medians of all intervals are added and then divided by the number of intervals to obtain the mean of the median representative value. The standard deviation is calculated using a statistical method (such as a sample standard deviation formula) to measure the degree of dispersion of the data.

[0163] S204: Calculate the noise lower limit judgment value (i.e., the mean minus three times the standard deviation) according to the mean and standard deviation of the data representative values ​​combined with the Laida criterion;

[0164] In this embodiment, the Laida criterion (3σ criterion) is an outlier detection method based on the normal distribution assumption, which defines that the probability of a data point falling outside the mean plus or minus three times the standard deviation is extremely small and can be regarded as an anomaly.

[0165] S205: treating data points in the original electric drive efficiency test data set that are lower than the noise lower limit judgment value as noise points;

[0166] S205: Process all noise points in the original electric drive efficiency test data set to obtain a noise-reduced electric drive efficiency test data set.

[0167] In this embodiment, the noise points may be deleted or replaced (eg, replaced with the median of an interval).

[0168] The data denoising method based on probability distribution is to perform probability distribution analysis on the test data, re-divide the data interval according to the distribution law, and select the median of each interval as the data representative value of the interval; then combine the median representative value of each interval and the 3σ criterion to calculate the mean and standard deviation of the interval representative value, and finally obtain the noise lower limit judgment value of the efficiency test data.

[0169] By analyzing the frequency histogram and probability density distribution of the efficiency test data, such as Figure 6 And the data are mainly concentrated above 60%, among which the data above 90% account for the largest proportion, followed by 80 to 90%, as shown in Table 2 of the probability distribution of efficiency test data.

[0170] Table 2 Probability distribution table of efficiency test data

[0171]

[0172]

[0173] Through the above analysis, the test data are divided into three intervals as [-∞,80], [80,90], and [90,+∞], and the medians of each interval are calculated to be 70.85, 86.99, and 92.48, respectively. The lower limit of the test data noise judgment (μ-3σ) is calculated to be 35.74 according to the 3σ principle, and the final normal data interval is [35.74,100], which realizes the effective noise reduction of the efficiency test data.

[0174] In order to verify the universality of the method of the present invention, this embodiment processes multiple groups of test data under different working conditions in the same way to achieve noise reduction for the efficiency test data and obtain a data set without noise values. The results are shown in Table 3.

[0175] Table 3 Test data denoising results

[0176]

[0177] 3. Outlier Test of Secondary Detection Method

[0178] According to the analysis of the test data characteristics, in addition to the noise values ​​with large numerical differences in the efficiency test data, there are also outliers within the normal numerical range. These outliers are shown as values ​​that are close to the overall normal data, but the distribution trend of the test point is obviously inconsistent with other normal points in the current working condition. They are conditional outliers. Since the original data is unlabeled data, it is very difficult to directly process the conditional outliers. However, the distribution intervals of efficiency test data under different speed conditions will be different, and some outliers in the efficiency test data are more obvious while the distribution of some outliers is very complex. Therefore, the present invention proposes a secondary outlier test method based on this data characteristic: first, the data is clustered as a whole and some obvious efficiency outlier test data with smaller out-of-cluster data in the efficiency test data are tested; then, a separate outlier test is performed on the efficiency data within the cluster. The overall test process is as follows: Figure 7 shown.

[0179] 1. DBSCAN model

[0180] DBSCAN (Density-Based Spatial Clustering of Applications with Noise), also known as density-based clustering method with noise, is an algorithm for spatial density clustering of data. By combining the test data feature analysis and outlier detection requirements, this embodiment uses the DBSCAN model to perform overall clustering on the efficiency test data, divides the data into cluster intervals according to different speeds, and deletes some obvious abnormal points outside the cluster.

[0181] The main idea of ​​DBSCAN is density clustering. Different clusters are obtained by defining the density of the test data and obtaining the data sample set with the maximum density connection according to the density reachable relationship. The main parameters in the DBSCAN algorithm are: the minimum domain radius (eps) and the minimum number of samples (Min-points). The core points of the data are determined by these two parameters and the density reachable relationship. The core point refers to the point where the number of sample data within the minimum domain radius is greater than or equal to the minimum number of samples; the density reachable relationships are: density direct, density reachable and density connected. The DBSCAN algorithm uses the core point to determine the density relationship of the data points in the data set and perform data clustering. Density direct means that the measurement point is within the minimum domain radius of the core point; density reachable means that there is a density direct relationship between multiple core points, and the first core point has a density reachable relationship with the last core point; density connected means that one core point is density reachable to the other two points, and the other two points are density connected; in the test data set, if the data does not meet the data connection relationship, it can be determined that the data is a different cluster or an outlier.

[0182] In this embodiment, the steps of performing the first abnormality inspection and processing by using the DBSCAN model include:

[0183] S301: Determine the neighborhood radius r of the DBSCAN model and the minimum number of samples Minpts required to form a dense area;

[0184] S302: Initialize the unvisited data set D, the core object set, and the number of clusters of the DBSCAN model; add all data points in the noise reduction electric drive efficiency test data set to the unvisited data set D;

[0185] In this embodiment, the core object set refers to creating an empty set for storing data points identified as core points. The number of clusters refers to setting a counter with an initial value of 0 to record the number of clusters found. The unvisited data set refers to adding all data points in the noise reduction electric drive efficiency test data set to an unvisited set, which will gradually decrease during the execution of the algorithm.

[0186] S303: Randomly select a data point p from the unvisited data set D as the current observation point;

[0187] S304: Calculate all sample points within the neighborhood radius r of the current observation point p, that is, |N r (p)|;

[0188] S305: If | N r (p)|≥Minpts, then the data point p is regarded as a core point; otherwise, the data point p is regarded as an outlier and added to the outlier list;

[0189] S306: repeating steps S303 to S305 until a core point q is found, and adding the core point q to the core object set; if no core point is found, adding the uncollected data points to the outlier list;

[0190] S307: Establish cluster C at the core point q, and add all data points within the neighborhood radius r of the core point q to cluster C, that is, N r (p) = {q∈D|dist(p,q)≤r}; at the same time, the data points added to cluster C are removed from the unvisited data set D; whenever a new cluster is formed, the cluster number counter is increased by 1;

[0191] S308: Repeat steps S303 to S307 until the unvisited data set D is empty, and output the clustering results and the list of outliers;

[0192] S309: Processing the data points in the noise reduction electric drive efficiency test data set that belong to the abnormal value list to obtain a preliminary screening electric drive efficiency test data set.

[0193] This embodiment uses scikit-learn in Python to build a DBSCAN outlier detection model according to the above process. The minimum area radius (eps) and the minimum number of samples (Min-points) of the model itself are determined by using the traditional K-dist graph. Figure 8 shown.

[0194] The K-dist graph is a graph that calculates the distance between all data and its Kth nearest neighbor by the DBSCAN model, sorts the results by distance, and selects parameter values ​​using the elbow method. The DBSCAN model outlier test determines the number of anomalies based on the model anomaly test score. A score close to -1 indicates that the data is normal, otherwise it is considered abnormal data. Fig. 9 shown.

[0195] Through the analysis of the test results, it can be found that the DBSCAN model can realize the detection of abnormal data using the clustering principle, but it can also be found that its outlier detection effect is not good, and some outliers outside the cluster have not been detected; through analysis, it is found that although the traditional K-dist is used when selecting the main parameters of the model, when using the inflection point method, due to certain subjective and unstable factors in the inflection point determination method, it may not select the optimal parameter combination of the model in the selection of the main parameters, and it is necessary to improve the selection method of the main parameters.

[0196] 3. Particle Swarm Optimization (IPSO)

[0197] Particle Swarm Optimization (PSO) is an intelligent optimization algorithm that simulates the foraging of bird flocks. Its main principle is to regard the problem to be optimized as the insects that the bird flock needs to find. The particles in the particle swarm represent different birds. Different particles "forage" within the set optimization space at their own speed, and different particles will exchange information and share the location information of food in real time. Then each particle adjusts its speed to move closer to the local optimal solution by judging the distance between itself and the optimal solution. Finally, through continuous iterative optimization, the global optimal solution is obtained to solve the target problem.

[0198] In this embodiment, the particle swarm optimization algorithm is used to solve the optimal neighborhood radius and minimum number of samples of the DBSCAN model as input parameters of the DBSCAN model; the processing steps of the particle swarm optimization algorithm include:

[0199] S3011: Set the relevant parameters of the particle swarm optimization algorithm, including the particle swarm size, the maximum number of iterations, the acceleration constant and the inertia weight; set the search range of the neighborhood radius r and the minimum number of samples Minpts as the constraints on the position and speed of the particles;

[0200] S3012: Randomly initialize the speed, position, individual optimal solution and global optimal solution of all particles within the set search range;

[0201] S3013: Calculate the fitness function value of each particle in each evolution process through the fitness function;

[0202] The calculation formula of the fitness function is as follows:

[0203]

[0204] Where: S represents the silhouette coefficient; A represents the average distance of the sample point of the same category to other samples; B represents the average distance between the sample and other clusters;

[0205] In this embodiment, the fitness function receives the neighborhood radius and the minimum number of samples of the DBSCAN model as input, performs DBSCAN clustering, and returns the evaluation index of the clustering effect as output. The specific implementation of the fitness function should be based on the selected clustering effect evaluation index.

[0206] S3014: Update the speed and position of particles;

[0207] The formula is:

[0208]

[0209] Where: Indicates the updated particle speed; ω indicates the inertia weight, which is in the range of 0.4 to 0.9; represents the current particle speed; c1 and c2 represent learning factors; r1 and r2 represent random data of [0,1], which are used for the initial random position; represents the individual optimal value at time t; Indicates the current particle position; represents the global optimal value at time t; Indicates the particle update position;

[0210] S3015: Compare the fitness values ​​of each particle and update the individual optimal solution and the global optimal solution of each particle;

[0211] S3016: Determine whether the set algorithm termination condition is met: if so, the global optimal solution, that is, the optimal neighborhood radius and the minimum number of samples, is used as the output result; otherwise, return to step S3013.

[0212] The present invention solves the optimal parameters of the DBSCAN model, namely the minimum domain radius and the minimum number of samples, by constructing a particle swarm optimization algorithm. However, the particle swarm optimization algorithm is prone to falling into a local optimal solution and convergence in advance, and the inertia weight factor dominates the optimization ability of all particles, but in the standard particle swarm optimization algorithm, the inertia weight is fixed, and the optimization ability of all particles cannot be balanced, which reduces the optimization effect of the model. Therefore, an improved particle swarm algorithm (Improved Particle Swarm Optimization, IPSO) is obtained by improving the inertia weight value method.

[0213] In order to improve the efficiency of obtaining the optimal solution set, the inertia weight is applied with a larger value in the early stage of model optimization, so that the particles can search the entire space evenly; in the later stage of model optimization, the model needs to conduct a detailed search of the local space where the optimal solution set is located to improve the accuracy of the final target solution value. Therefore, this embodiment optimizes the inertia weight of the particle swarm optimization algorithm by linearly reducing it, that is, the inertia weight of the particle swarm optimization algorithm is designed to gradually decrease linearly from ω1 (0.9) to ω2 (0.4) with the number of iterations;

[0214] The formula is:

[0215]

[0216] Where: ω represents the inertia weight; ω1 and ω2 represent the upper and lower limits of the inertia weight; i represents the current number of iterations; iter max Indicates the maximum number of iterations.

[0217] In the constructed improved particle swarm optimization algorithm, the learning factor adopts the standard formula of particle swarm optimization algorithm, i.e. c1=c2=2; according to the principle of selecting the amount of efficiency test data and the number of particles, the initial particles are selected as N=30 and the number of iterations is 50; the fitness function selects the silhouette coefficient of the DBSCAN model as the fitness value calculation criterion; and the final optimal minimum area radius (eps) and minimum number of samples (Min-points) are used as the input parameters of the DBSCAN model for outlier test.

[0218] In order to verify that the IPSO-DBSCAN model used in the present invention is superior to the traditional DBSCAN model, this embodiment introduces unsupervised model evaluation indicators: silhouette coefficient, Calinski-Harabaz index and Davidson-Borg index as comparative indicators for model performance verification. The silhouette coefficient reflects the distance between samples, the closer to 1 the better; the Calinski-Harabaz index measures the compactness by calculating the sum of the squares of the distances of each point. The larger the CH index, the tighter the clusters and the more dispersed the clusters; the smaller the Davidson-Borg index, the tighter the intra-cluster distance and the cluster distance. The specific formula is as follows:

[0219]

[0220] Where: S is the silhouette coefficient; A is the average distance between the sample point and other samples of the same category; B is the average distance between the sample and other clusters; CH is the Calinski-Harabaz index; M k is the inter-cluster covariance matrix; k is the intra-cluster covariance matrix; N is the sample size; K is the number of clusters; DB is the Davidson-Boting index; l i is the average distance from the sample point to the cluster centroid; d ij is the centroid distance between clusters i and j.

[0221] By comparing the values ​​of various evaluation indicators, it is found that the performance of the DBSCAN model improved by IPSO is better than that of the traditional DBSCAN model, as shown in Table 4. The model convergence process is as follows Fig.10 As shown in the figure, the DBSCAN model based on IPSO realizes the clustering of efficiency test data and completes the test of outliers outside the cluster.

[0222] Table 4. Performance comparison of DBSCAN models

[0223]

[0224] 3. LOF (Local Outlier Factor) Algorithm Model and iForest (Isolation Forest) Algorithm Model

[0225] In this embodiment, the steps of performing a second abnormality inspection and processing using the LOF algorithm model and the iForest algorithm model include:

[0226] S401: constructing LOF algorithm model and iForest algorithm model;

[0227] S402: performing anomaly detection on the preliminary screening electric drive efficiency test data set by using the LOF algorithm model to obtain a first anomaly score for each data point in the preliminary screening electric drive efficiency test data set;

[0228] S403: performing anomaly detection on the preliminary screening electric drive efficiency test data set by using the iForest algorithm model to obtain a second anomaly score for each data point in the preliminary screening electric drive efficiency test data set;

[0229] S404: performing weighted calculation on the first anomaly scores and the second anomaly scores of all data points to obtain a final anomaly score of the data point;

[0230] In this embodiment, the weights of the first anomaly score and the second anomaly score are both set to 0.5.

[0231] S405: taking the data points whose final anomaly scores are lower than the preset score threshold as outliers for the second outlier test;

[0232] S406: Processing the outliers of the second outlier test in the initial screening electric drive efficiency test data set to obtain a preprocessed electric drive efficiency test data set.

[0233] The local outlier factor (LOF) is a complex outlier detection method based on density. It calculates the density of each observation point and its domain points and calculates the outlier factor. Finally, it determines whether the observation point is outliers by judging whether the outlier factor is close to 1. The main definitions in the LOF algorithm are: the Kth distance refers to the Kth farthest distance from the observation point; the Kth distance neighborhood is the distance of the Kth farthest point from the observation point; the reachable distance is at least the Kth distance of a point or the true distance between two points; the local reachable density refers to the reciprocal of the average reachable distance from the observation point to the observation point in the Kth domain of the observation point; the local outlier factor is the average of the ratio of the local reachable density of the neighborhood points to the local reachable density of the observation point.

[0234] Specifically, the processing steps of the LOF algorithm model for anomaly detection include:

[0235] S4021: Determine a neighborhood parameter N of the LOF algorithm model;

[0236] S4022: For each data point in the initial screening electric drive efficiency test data set, calculate its distance to the Nth nearest neighbor data point, i.e., k distance, expressed as k_dist k ;

[0237] In the sample space, there is an object (observation point) o, and the distance between it and the object p is recorded as d(p,o). Under the conditions of the algorithm, it is considered that k_dist k =d(p,o).

[0238] S4023: Based on the k distance k_dist of each data point k Calculate its reachable distance;

[0239] The formula is:

[0240] reach_dist k (p,o)=max{d k (o),d k (p,o)};

[0241] Where: o represents the kth observation point closest to data point p; d k (o) represents the reachable distance of observation point o; d k (p,o) represents the distance between data point p and observation point o;

[0242] In this embodiment, for each point data p and one of its neighboring points o, the reachable distance of point o to point p is the larger value of the k distance of point o and the distance from point p to point o, which ensures that the reachable distance will not be less than the k distance of any point.

[0243] S4024: Calculate the local reachability density ρ of each data point based on its reachable distance k (p);

[0244] The formula is:

[0245]

[0246] Where: N k (p) represents the number of samples of data point p in the k-distance neighborhood; the local reachability density of object p is defined as the inverse of the average reachable distance of p's k nearest neighbors (essentially, it is the derivative of the average of all reachable distances in p's k-th distance neighborhood. The larger the better, representing a more compact data cluster).

[0247] In this embodiment, for each data point p, its local reachability density is the average of the inverses of the reachability distances of all its neighbors. The higher the local reachability density, the greater the density around point p, and the less likely it is an outlier.

[0248] S4025: Calculate the local reachability density ratio LOF of each data point based on its local reachability density k (o);

[0249] The formula is:

[0250]

[0251] Where: k (o) represents the local reachability density of observation point o; N k (o) represents the number of samples of data point o in the k-distance neighborhood;

[0252] If this ratio is closer to 1, it means that the density of p is similar to that of the points in its k-distance neighborhood, and p may belong to the same cluster as the points in its k-distance neighborhood; if this ratio is less than 1, it means that the density of p is higher than the density of other points in its neighborhood, and p is a dense point; if this ratio is greater than 1, it means that the density of p is lower than the density of other points in its neighborhood, and p is more likely to be an outlier. In this way, outliers can be accurately found even when the sample space data is unevenly distributed.

[0253] In this embodiment, for each neighbor o of each data point p, the ratio of the local reachable density of point p to the local reachable density of point o is calculated, which is the local reachable density ratio. This reflects whether the local density of point p is higher or lower relative to its neighbors.

[0254] S4026: Calculate the LOF score, i.e., the first anomaly score, of each data point according to its local reachable density ratio.

[0255] In this embodiment, for each data point p, its LOF score is the geometric mean of the average values ​​of the local reachable density ratios of all its neighbors. The higher the LOF score, the more isolated the point p is relative to its neighbors, and the more likely it is to be an outlier.

[0256] Isolation Forest (iForest) is an algorithm that does not need to describe normal sample points, but only needs to isolate outliers. Its main idea is to use the characteristics of outliers being sparsely distributed in the overall data set and far away from high-density groups to perform outlier detection. In the Isolation Forest algorithm, an isolation tree (binary search tree structure) is used to recursively and randomly split the data set until all data points are isolated. Among these isolated data points, the outlier data is sparsely distributed, so the number of splits required to completely isolate it will be significantly less than that of normal data points. Finally, the average path length of the observation point in each isolation tree is calculated to obtain the anomaly score, based on which the outliers in the data set are judged.

[0257] Specifically, the processing steps of the iForest algorithm model for anomaly detection include:

[0258] S4031: Determine the number of isolated trees t of the iForest algorithm model and the number of samples T for training each isolated tree;

[0259] S4032: For each isolated tree, perform the following steps:

[0260] 1) Randomly select T data points from the initial screening electric drive efficiency test data set as a subsample set and put them into the root node of the isolation tree;

[0261] 2) Randomly select a feature dimension and randomly generate a cutting point p in the current node data on this dimension (the cutting point is generated between the maximum and minimum values ​​of the specified dimension in the current node data);

[0262] 3) Generate a hyperplane with the cutting point p to divide the current node data space into two subspaces: put the data points less than p in the specified dimension in the left child node of the current node, and put the data points greater than or equal to p in the right child node of the current node;

[0263] 4) Recursively execute steps 2) and 3) in the child node, and continuously construct new child nodes until there is only one data in the child node (no further cutting is possible) or the child node has reached the specified height;

[0264] S4033: Repeat step S4032 until t isolated trees are generated to form an isolated forest;

[0265] S4034: for each data point x in the preliminary screening electric drive efficiency test data set, traverse each isolated tree in the isolation forest;

[0266] S4035: In each isolated tree, calculate the path length h(x) of the data point x from the root node to the leaf node;

[0267] The formula is:

[0268] h(x)=e+C(T);

[0269] Where: e represents the number of nodes that sample x has traveled from the root node to the leaf node; T represents the number of samples at the same leaf node as sample x; C represents a correction value;

[0270] In this embodiment, the path length from the root node to the counting node is calculated by recursively traversing the tree, which is equal to the sum of the depths of all branches passed. If a data point x falls on a node containing multiple training data, a formula can be used to correct the height calculation of x.

[0271] S4036: Calculate the average path length E(h(x)) of the data point x in t isolated trees;

[0272] The formula is:

[0273]

[0274] In the formula: n represents the total number of isolated numbers;

[0275] S4037: Calculate the second anomaly score of the data point x by using the average path length E(h(x));

[0276] The formula is:

[0277] s(x,n)=2-c(n)E(h(x));

[0278]

[0279] Where: s(x,n) represents the second anomaly score of data point x; c(n) represents the average path length when the binary sorted tree (BST) search is unsuccessful; n represents the number of samples used to generate each isolated tree; H(n-1) represents the harmonic number, which can be estimated by ln(n-1)+0.5772156649 (Euler constant).

[0280] According to the different principles of LOF and iForest algorithms, the present invention can respectively realize the detection of obvious outliers and complex outliers within a cluster. The local outlier factor algorithm model and the isolation forest algorithm model are respectively constructed by the scikit-learn library, and the outlier value scores finally obtained by them are weighted and added to obtain the outlier test results of the in-cluster test data, so as to realize the in-cluster outlier test of the efficiency test data. The test results are as follows: Fig.11 shown.

[0281] The noise values, abnormal values, repeated values, and empty values ​​in the efficiency test data are processed, and the Lagrangian method is selected for value filling according to the characteristics of the test data. After the preprocessing is completed, a "clean" data set is obtained. Some data such as Fig.12 As shown, it can be found that the preprocessed data is more stable and has an obvious trend and consistency.

[0282] IV. Quality Assessment

[0283] 1. Six-tuple (SVMRWD) data quality model

[0284] The present invention constructs a six-tuple (SVMRWD) data quality model based on the efficiency data characteristics and actual needs. The six-tuple evaluation model consists of the test data set (D), dimension layer (V), indicator layer (M), rule layer (R), weight layer (W), and final score layer (S). By gradually building each tuple, the quality evaluation of the efficiency test data can be achieved. The specific evaluation process of the six-tuple data quality model is as follows: Fig.13 shown.

[0285] Dimension layer: It is composed of internal quality, operational quality, and representation quality. Internal quality refers to the data characteristics of the efficiency test data itself. Operational quality refers to the data characteristics that are expressed by further feature analysis of the efficiency test data using visualization tools or scientific computing. Representation quality refers to the new feature expression obtained by scientifically processing the efficiency test data in combination with the actual R&D needs of the enterprise. Indicator layer: It is constructed according to the distribution of different dimensional layers in combination with data characteristics, and belongs to the upper and lower layer relationship. The rule layer describes the indicator layer and belongs to the same layer description relationship.

[0286] (1) The internal quality dimension includes three indicators: accuracy, completeness, and uniqueness. Accuracy is used to evaluate whether the efficiency test data can accurately reflect the actual performance of the sample. The outliers in the test data cannot accurately express the characteristics of the sample. Therefore, the calculation rule of accuracy is: use the outlier test model to test how many outliers there are in the data and calculate the proportion of normal data in the data. Integrity is used to reflect whether the test data is complete, that is, whether there are null values, missing values, etc. in the test data (in this embodiment, missing values, null values ​​and duplicate values ​​are detected using the info function). The calculation rule is: calculate the total amount of null value data in the test data. Uniqueness is used to reflect the degree of data repetition in the data. The calculation rule of uniqueness is to detect the proportion of repeatedly collected data in the data. The specific formula is as follows:

[0287]

[0288] In the formula: a is data accuracy; X1 is the number of outliers; N is the total amount of efficiency test data; b is data completeness; X2 is the number of missing values ​​and null values; c is data uniqueness; X3 is the number of repeated values.

[0289] (2) Operation quality includes two evaluation indicators: trend consistency and stability. Trend consistency is used to evaluate the overall trend consistency of efficiency test data under different speed conditions. Data with inconsistent trends may have quality problems. The evaluation rule for trend consistency is to visually check the proportion of data with consistent trends. Stability is used to evaluate whether the overall discreteness of the data is stable, which is mainly judged based on the distribution of data noise values. The evaluation rule is to calculate the proportion of noise values ​​in the efficiency data. The specific formula is as follows:

[0290]

[0291]

[0292] In the formula: d is the consistency of data trend; X4 is the total amount of data with consistent trend; e is data stability; X5 is the total amount of noise value data.

[0293] (3) The high efficiency evaluation index (f) in the characterization quality is proposed based on actual R&D needs and is used to evaluate the energy conversion capacity of the prototype. The high efficiency index is divided into two parts: the high efficiency area ratio and the peak efficiency ratio according to actual needs. The weight of the two parts is 50% each when calculating. Among them, the high efficiency area ratio refers to the overall ability of the prototype to efficiently convert energy. The rule is to calculate the proportion of data above 80% in the efficiency data; the peak efficiency ratio reflects the maximum energy conversion capacity of the prototype. The calculation rule is to calculate the score based on the peak efficiency interval and refer to the designed peak efficiency interval score table. The score table is shown in Table 5.

[0294] Table 5 Peak efficiency score table

[0295]

[0296] 2. Evaluation indicator weight

[0297] The weight layer can directly affect the final score of data quality assessment. Since the importance of each evaluation indicator is different, the weight layer assigns different weights to different evaluation indicators to achieve accurate assessment of data quality. In order to combine subjectivity and objectivity, this embodiment uses a comprehensive weight that combines the subjective analytic hierarchy process weight and the objective entropy weight method weight as the evaluation weight of each evaluation indicator.

[0298] (1) The analytic hierarchy process (AHP) is a subjective method in the evaluation method. It can combine qualitative analysis and quantitative analysis, and divide the complex problem system into levels. Finally, the final weight is obtained by calculating each factor in the level. Its main process is as follows:

[0299] 1) A number of electric drive system test experts participated in constructing a judgment matrix based on the evaluation indicators. Since the importance of each indicator is similar, the 1-5 scaling method is used to improve the weight accuracy, as shown in Table 6.

[0300] Table 6 Evaluation index judgment matrix

[0301]

[0302] 2) Matrix consistency test: Check the rationality of the judgment matrix. Calculate the matrix consistency index CI; obtain the matrix average random consistency index RI by looking up the table, as shown in Table 7; finally calculate the consistency ratio CR. If CR<0.1, the matrix consistency test passes, otherwise adjust the judgment matrix.

[0303]

[0304] Where: maxis the maximum eigenvalue of the judgment matrix; n is the number of indicators.

[0305] Table 7 Average random consistency index table

[0306]

[0307] 3) Matrix weight calculation: The matrix weight calculation methods include arithmetic mean method, geometric mean method, and maximum eigenvalue method. This embodiment adopts the maximum eigenvalue method and normalizes the result to obtain the analytic hierarchy process weight vector of the judgment matrix W1=(0.2958, 0.0701, 0.0463, 0.1115, 0.1801, 0.2958) T .

[0308] (2) The entropy weight method is an objective weighting method in the field of statistics. It has the characteristics of high accuracy and good adaptability. The entropy weight of each evaluation index is obtained by calculating the information entropy of each data, and the entropy weight is normalized to obtain the weight of each index. The main process is as follows:

[0309] 1) According to the constructed evaluation index system, calculate the index scores and construct the horizontal matrix;

[0310] 2) If there are negative indicators in the evaluation indicators, each indicator needs to be positive. The evaluation indicators constructed in this embodiment are all positive indicators, so this step is skipped;

[0311] 3) Data normalization: used to narrow the dimensional gap between data indicators and improve the accuracy and efficiency of weight calculation; this embodiment uses the maximum-minimum normalization method to normalize the indicator level matrix.

[0312] The announcement is as follows:

[0313]

[0314] Where: y is the normalized result; x min is the minimum value of the indicator; x max is the maximum value of the indicator.

[0315] 4) Entropy value solution: Calculate the information entropy of each evaluation indicator. The smaller the information entropy value, the greater the amount of information;

[0316] The announcement is as follows:

[0317]

[0318]

[0319] Where: H j is the entropy value; k is a constant; f ij is the proportion of indicator information; r ijis the evaluation index level matrix; m is the number of evaluation indicators.

[0320] 5) Calculate information entropy redundancy and weight calculation: Calculate redundancy based on the information entropy of the indicator, and normalize the obtained information utility value to obtain the weight of each evaluation indicator;

[0321] The announcement is as follows:

[0322] d j =1-H j ;

[0323]

[0324] Where: d j is the information entropy redundancy; W j is the weight of index j.

[0325] This embodiment solves the constructed judgment matrix and level matrix respectively through the hierarchical analysis method and the entropy weight method, and finally obtains the hierarchical analysis method weight W1 and the entropy weight method weight W2 = (0.2150, 0.0998, 0.1907, 0.1037, 0.1497, 0.2408) T ; Combining W1 and W2, W1 accounts for 60% and W2 accounts for 40%, we get the comprehensive weight of each evaluation index of the data quality assessment model W0 = (0.2635, 0.0820, 0.1041, 0.1084, 0.1679, 0.2738) T .

[0326] In summary, the six-tuple data quality model realizes quality assessment by calculating the comprehensive quality score of the preprocessed electric drive efficiency test data set through the following formula;

[0327] S=0.2554a+0.0850b+0.1185c+0.1076d+0.1649e+0.2683f;

[0328] in:

[0329]

[0330] Where: S represents the comprehensive quality score of the data set; a, b, c, d, e, and f represent the data accuracy, data integrity, data uniqueness, data trend consistency, data stability, and data efficiency of the data set, respectively; X1 represents the number of outliers; X2 represents the number of missing values ​​and null values; X3 represents the number of duplicate values; X4 represents the total amount of data with consistent trends; X5 represents the total amount of noise value data; h η represents the proportion of high-efficiency area; p η Indicates the peak efficiency ratio; X6 indicates the total amount of data whose computing efficiency exceeds the preset value (80%); sn represents the score of each peak data; N represents the total amount of data in the data set.

[0331] It should be noted that improving the proportion of the high-efficiency area of ​​the electric drive system is helpful for energy conversion efficiency, endurance, cost-effectiveness and driving experience. Among them, the proportion of the high-efficiency area of ​​the electric drive system directly affects the energy conversion efficiency of the electric vehicle. Within the high efficiency range, the electric motor can convert more electrical energy into mechanical energy, thereby improving the power and cruising range of the electric vehicle; the electric drive system with a high proportion of the high-efficiency area can provide a longer driving distance under the same battery energy to meet the needs of users; improving the proportion of the high-efficiency area of ​​the electric drive system can reduce energy consumption, thereby reducing the use cost and maintenance cost of the battery and extending the service life of the electric vehicle; the efficient system can provide a more sensitive and smooth control experience and enhance the driving experience. Therefore, the present invention sets the weight of the high-efficiency area ratio to 60% and the peak efficiency ratio to 40%, with the purpose of better highlighting the comprehensive performance of the electric drive system.

[0332] 3. Data quality assessment

[0333] The final scoring level of the data quality assessment model is formulated based on actual enterprise needs by a number of new energy vehicle performance research and development experts, based on the comprehensive performance of each evaluation indicator of the efficiency test data and the actual performance capabilities. The specific levels are shown in Table 8.

[0334] Table 8 Data quality assessment levels

[0335]

[0336] After the data quality assessment model is constructed, this embodiment uses two groups of original efficiency test data as test samples to perform quality assessment on the data before and after preprocessing. From the results, it can be found that the initial scores of various indicators of the original data are not high, but after preprocessing, the accuracy, completeness, uniqueness, etc. of the efficiency data are greatly improved, and the final scoring levels of the two groups of test data are improved to varying degrees, as shown in Table 9, which realizes the effective evaluation of the efficiency test data.

[0337] Table 9 Example of efficiency data quality assessment

[0338]

[0339] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. A method for data processing and quality assessment of electric drive system efficiency test data, characterized in that: include: S1: Get the original electric drive efficiency test data set; S2: De-noising the original electric drive efficiency test data set by a data de-noising method based on probability distribution to obtain a de-noised electric drive efficiency test data set; S3: Perform the first abnormality inspection and processing on the noise reduction electric drive efficiency test data set through the DBSCAN model to obtain the preliminary screening electric drive efficiency test data set; S4: Performing a second abnormality inspection and processing on the preliminary screening electric drive efficiency test data set through the LOF algorithm model and the iForest algorithm model to obtain a preprocessed electric drive efficiency test data set; S5: Quality assessment of the preprocessed electric drive efficiency test dataset is performed using the constructed six-tuple data quality model.

2. The method for data processing and quality assessment of electric drive system efficiency test data according to claim 1, characterized in that: The electric drive efficiency test data includes the efficiency of electric operation and the efficiency of power generation; The calculation formula of electric working efficiency is expressed as: The calculation formula of power generation efficiency is as follows: Wherein: η1 represents the electric driving efficiency of the electric drive system; η2 represents the power generation efficiency of the electric drive system; r1 represents the speed of the left dynamometer; T1 represents the torque of the left dynamometer; r2 represents the speed of the right dynamometer; T2 represents the torque of the right dynamometer; P1 and P2 are the power of the DC bus of the electric drive system controller.

3. The method for data processing and quality assessment of electric drive system efficiency test data according to claim 1, characterized in that: In step S2, the processing steps of the data denoising method based on probability distribution are as follows: S201: performing probability distribution analysis on the original electric drive efficiency test data set; dividing the data points in the original electric drive efficiency test data set into a plurality of intervals according to the law of probability distribution; S202: Select the median of the data points in each interval as the data representative value of the interval; S203: Calculate the mean and standard deviation of the data representative values ​​of all intervals; S204: Calculate the noise lower limit judgment value according to the mean and standard deviation of the data representative value combined with the Laida criterion; S205: treating data points in the original electric drive efficiency test data set that are lower than the noise lower limit judgment value as noise points; S206: Process all noise points in the original electric drive efficiency test data set to obtain a noise-reduced electric drive efficiency test data set.

4. The method for data processing and quality assessment of electric drive system efficiency test data according to claim 1, characterized in that: In step S3, the steps of performing the first anomaly inspection and processing by using the DBSCAN model include: S301: Determine the neighborhood radius r of the DBSCAN model and the minimum number of samples Minpts required to form a dense area; S302: Initialize the unvisited data set D, the core object set, and the number of clusters of the DBSCAN model; add all data points in the noise reduction electric drive efficiency test data set to the unvisited data set D; S303: Randomly select a data point p from the unvisited data set D as the current observation point; S304: Calculate all sample points within the neighborhood radius r of the current observation point p, that is, |N r (p)|; S305: If | N r (p)|≥Minpts, then the data point p is regarded as a core point; otherwise, the data point p is regarded as an outlier and added to the outlier list; S306: repeating steps S303 to S305 until a core point q is found, and adding the core point q to the core object set; if no core point is found, adding the uncollected data points to the outlier list; S307: Establish cluster C at the core point q, and add all data points within the neighborhood radius r of the core point q to cluster C, that is, N r (p) = {q∈D|dist(p,q)≤r}; at the same time, the data points added to cluster C are removed from the unvisited data set D; whenever a new cluster is formed, the cluster number counter is increased by 1; S308: Repeat steps S303 to S307 until the unvisited data set D is empty, and output the clustering results and the list of outliers; S309: Processing the data points in the noise reduction electric drive efficiency test data set that belong to the abnormal value list to obtain a preliminary screening electric drive efficiency test data set.

5. The method for data processing and quality assessment of electric drive system efficiency test data according to claim 1, characterized in that: In step S301, the optimal neighborhood radius and minimum number of samples of the DBSCAN model are solved by the particle swarm optimization algorithm as input parameters of the DBSCAN model; The processing steps of the particle swarm optimization algorithm include: S3011: Set the relevant parameters of the particle swarm optimization algorithm; set the search range of the neighborhood radius r and the minimum number of samples Minpts as the constraints on the position and speed of the particles; S3012: Randomly initialize the speed, position, individual optimal solution and global optimal solution of all particles within the set search range; S3013: Calculate the fitness function value of each particle in each evolution process through the fitness function; The calculation formula of the fitness function is as follows: Where: S represents the silhouette coefficient; A represents the average distance of the sample point of the same category to other samples; B represents the average distance between the sample and other clusters; S3014: Update the speed and position of particles; The formula is: Where: represents the updated particle velocity; ω represents the inertia weight; represents the current particle speed; c1 and c2 represent learning factors; r1 and r2 represent random data of [0,1], which are used for the initial random position; represents the individual optimal value at time t; Indicates the current particle position; represents the global optimal value at time t; Indicates the particle update position; S3015: Compare the fitness values ​​of each particle and update the individual optimal solution and the global optimal solution of each particle; S3016: Determine whether the set algorithm termination condition is met: if so, the global optimal solution, that is, the optimal neighborhood radius and the minimum number of samples, is used as the output result; otherwise, return to step S3013.

6. The method for data processing and quality assessment of electric drive system efficiency test data according to claim 5, characterized in that: In step S3011, the inertia weight of the particle swarm optimization algorithm is optimized by linear reduction, that is, the inertia weight of the particle swarm optimization algorithm is designed to gradually decrease linearly from ω1 to ω2 with the number of iterations; The formula is: Where: ω represents the inertia weight; ω1 and ω2 represent the upper and lower limits of the inertia weight; i represents the current number of iterations; iter max Indicates the maximum number of iterations.

7. The method for data processing and quality assessment of electric drive system efficiency test data according to claim 1, characterized in that: In step S4, the steps of performing a second abnormality inspection and processing by using the LOF algorithm model and the iForest algorithm model include: S401: constructing LOF algorithm model and iForest algorithm model; S402: performing anomaly detection on the preliminary screening electric drive efficiency test data set by using the LOF algorithm model to obtain a first anomaly score for each data point in the preliminary screening electric drive efficiency test data set; S403: performing anomaly detection on the preliminary screening electric drive efficiency test data set by using the iForest algorithm model to obtain a second anomaly score for each data point in the preliminary screening electric drive efficiency test data set; S404: performing weighted calculation on the first anomaly scores and the second anomaly scores of all data points to obtain a final anomaly score of the data point; S405: taking the data points whose final anomaly scores are lower than the preset score threshold as outliers for the second outlier test; S406: Processing the outliers of the second outlier test in the initial screening electric drive efficiency test data set to obtain a preprocessed electric drive efficiency test data set.

8. The method for data processing and quality assessment of electric drive system efficiency test data according to claim 7, characterized in that: Step S402, the processing steps of the LOF algorithm model for abnormality detection include: S4021: Determine a neighborhood parameter N of the LOF algorithm model; S4022: For each data point in the preliminary screening electric drive efficiency test data set, calculate the distance to the Nth nearest neighbor data point, i.e., the k distance; S4023: Calculate the reachable distance of each data point according to its k distance; The formula is: reach_dist k (p,o)=max{d k (o),d k (p,o)}; Where: o represents the kth observation point closest to data point p; d k (o) represents the k distance of observation point o; d k (p,o) represents the distance between data point p and observation point o; S4024: Calculate the local reachability density ρ of each data point based on its reachable distance k (p); The formula is: Where: N k (p) represents the number of samples of data point p in the k-distance neighborhood; S4025: Calculate the local reachability density ratio LOF of each data point based on its local reachability density k (o); The formula is: Where: k (o) represents the local reachability density of observation point o; N k (o) represents the number of samples of observation point o within the k-distance neighborhood; S4026: Calculate the LOF score, i.e., the first anomaly score, of each data point according to its local reachable density ratio.

9. The method for data processing and quality assessment of electric drive system efficiency test data according to claim 7, characterized in that: Step S403, the iForest algorithm model performs anomaly detection processing steps including: S4031: Determine the number of isolated trees t of the iForest algorithm model and the number of samples T for training each isolated tree; S4032: For each isolated tree, perform the following steps: 1) Randomly select T data points from the initial screening electric drive efficiency test data set as a subsample set and put them into the root node of the isolation tree; 2) Randomly select a feature dimension and randomly generate a cutting point p in the current node data on this dimension; 3) Generate a hyperplane with the cutting point p to divide the current node data space into two subspaces: put the data points less than p in the specified dimension in the left child node of the current node, and put the data points greater than or equal to p in the right child node of the current node; 4) Recursively execute steps 2) and 3) in the child node, and continuously construct new child nodes until there is only one data in the child node or the child node has reached the specified height; S4033: Repeat step S4032 until t isolated trees are generated to form an isolated forest; S4034: for each data point x in the preliminary screening electric drive efficiency test data set, traverse each isolated tree in the isolation forest; S4035: In each isolated tree, calculate the path length h(x) of the data point x from the root node to the leaf node; The formula is: h(x)=e+C(T); Where: e represents the number of nodes that sample x has traveled from the root node to the leaf node; T represents the number of samples at the same leaf node as sample x; C represents a correction value; S4036: Calculate the average path length E(h(x)) of the data point x in t isolated trees; The formula is: In the formula: n represents the total number of isolated numbers; S4037: Calculate the second anomaly score of the data point x by using the average path length E(h(x)); The formula is: s(x,n)=2-c(n)E(h(x)); Where: s(x,n) represents the second anomaly score of data point x; c(n) represents the average path length when the binary sort tree search is unsuccessful; n represents the number of samples used to generate each isolated tree; H(n-1) represents the harmonic number.

10. The method for data processing and quality assessment of electric drive system efficiency test data according to claim 1, characterized in that: In step S5, the six-tuple data quality model calculates the quality comprehensive score of the pre-processed electric drive efficiency test data set by the following formula to achieve quality assessment; S=0.2554a+0.0850b+0.1185c+0.1076d+0.1649e+0.2683f; in: Where: S represents the comprehensive quality score of the data set; a, b, c, d, e, and f represent the data accuracy, data integrity, data uniqueness, data trend consistency, data stability, and data efficiency of the data set, respectively; X1 represents the number of outliers; X2 represents the number of missing values ​​and null values; X3 represents the number of duplicate values; X4 represents the total amount of data with consistent trends; X5 represents the total amount of noise value data; h η represents the proportion of high-efficiency area; p η Indicates the peak efficiency ratio; X6 indicates the total amount of data whose computing efficiency exceeds the preset value; s n represents the score of each peak data; N represents the total amount of data in the data set.

Images