Railway dc relay operating state evaluation method based on distance metric

Abstract

The present application relates to the field of railway relay state evaluation, especially to a railway DC relay operating state evaluation method based on distance measurement. It can visually represent the operating state of the relay, provide a theoretical basis for relay state evaluation, and provide a theoretical basis for equipment maintenance decision-making, making the operation of the railway train more safe and reliable. It includes: 1) extracting multiple characteristic parameters for analysis; 2) data denoising and trend enhancement processing; 3) quantitative analysis; 4) degradation distance representation; 5) degradation distance representation; 6) providing distance sample set for subsequent analysis; 7) clustering analysis; 8) feedback verification; 9) constructing an improved Elman neural network prediction model; 10) using the harmony search algorithm to optimize the related parameters of the improved Elman model; 11) constructing an improved HS-Elman neural network for relay state evaluation and prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of railway relay state evaluation, in particular to a railway DC relay operating state evaluation method based on distance measurement and improved Elman. BACKGROUND

[0002] With the continuous development of China's rail transportation industry, the intelligent construction of railway transportation system is vigorously promoted, and has become an important foundation of China's transportation industry and equipment manufacturing industry. Railway DC relay is widely used in train control system, which is an important basic switching device, and requires a high operating frequency, which requires the relay to have better stability and service life to ensure the normal operation of the train control system. The service life of the relay is divided into mechanical life and electrical life, and the electrical life is often several orders of magnitude lower than the mechanical life. Therefore, when studying the performance degradation process of the relay, the electrical life is often the main research. With the increase of the number of relay actions, the breaking performance of the contact system will gradually degrade, thereby causing the overall state performance of the relay to deteriorate. Under this background, the operating state recognition of the railway DC relay is studied, the true life state of the relay contact is reflected, which can provide a scientific basis for the maintenance decision of the relay and provide protection for the stable and safe operation of the train, and has significant theoretical and practical significance.

[0003] At present, most of the researches on railway relays at home and abroad are life prediction and failure mode discrimination, and few involve relay operating state division and recognition problems. Among the few state evaluation researches, the long-time sequence degradation in the relay operating process and the selection of state intervals in state recognition classification have not been analyzed in depth, and the division of different state intervals according to the degradation degree still needs further research. Therefore, the present application will propose a railway DC relay operating state evaluation method based on distance measurement and improved Elman to provide a more perfect theoretical basis for state recognition evaluation research. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a railway DC relay operating state evaluation method based on distance measurement and improved Elman.

[0005] The technical scheme adopted by the present application is: the purpose of the embodiment of the present application is to provide a railway DC relay operating state evaluation method based on distance measurement and improved Elman, by building a railway signal relay electrical life test system platform, the voltage and current data of the relay coil and the contact are accurately collected. According to the degradation characteristics of the relay, a variety of performance degradation characteristic parameters are extracted from the test data, and a feature parameter processing model combining singular spectrum analysis (SSA) and Spearman rank correlation coefficient (SRCC) is constructed, and the feature parameters are subjected to data enhancement and quantization processing. The Mahalanobis distance and cosine similarity distance are used to convert the enhanced multi-dimensional features into one-dimensional state indicators, and the two are combined for spatial visualization representation. Then, the improved K-means algorithm is used for clustering analysis, combined with the visual degradation representation curve for supervised feedback, the relay operating state is reasonably divided into intervals, and finally an improved HS-Elman prediction model is constructed to evaluate and predict the relay operating state. The method of the present application can realize effective evaluation and prediction of the operating state of the railway DC relay, solve the problems that a single characteristic parameter cannot comprehensively represent the degradation process, and the state interval of degradation cannot be selected according to prior knowledge, and can visually represent the operating state of the relay, provide a theoretical basis for relay state evaluation work, and provide a theoretical basis for equipment maintenance decision-making, so that the operation of the railway train is more safe and reliable.

[0006] The railway DC relay operating state evaluation method based on distance measurement and improved Elman of the present application comprises the following steps:

[0007] Step 1, collecting test data by relay electrical life test, extracting multi-dimensional feature parameters therefrom and performing data enhancement processing on the feature parameters. Wherein:

[0008] Step 1.1, performing electrical life test, obtaining initial test data, and extracting multi-dimensional feature parameters capable of representing relay performance degradation from the initial test data.

[0009] The present application is built and tested according to the requirements of GB14048.4-2016 and GB14048.5-2016. The Siemens 3RH2122-2KF40 type railway DC relay commonly used in railway trains is taken as the test object.

[0010] According to the performance degradation characteristics of the relay, the calculation methods of various feature parameters are as follows:

[0011] 1) Attraction time

[0012] t x =t2-t1.

[0013] In the formula, t1 is the coil power-on time, and t2 is the first contact time of the moving and static contacts.

[0014] 2) bounce time

[0015] t t = t3 - t2

[0016] where t3 is the end time of voltage and current fluctuation.

[0017] 3) overtravel time

[0018] t c = t4 - t2

[0019] where t4 is the stable attraction time.

[0020] 4) contact resistance

[0021]

[0022] where N is the number of collected points, u n , i n are the contact voltage and current.

[0023] 5) release time

[0024] t s = t6 - t5

[0025] where t5 is the coil power-off time, and t6 is the arcing time.

[0026] 6) arc energy

[0027]

[0028] where Δt is the time interval between adjacent sampling points, and f s is the sampling rate.

[0029] Step 1.2, SSA is used to denoise and enhance the initial feature parameters.

[0030] SSA is a method for processing nonlinear time series data. By decomposing and reconstructing the trajectory matrix of the time series to be studied, different component sequences (long-term trend, noise, etc.) in the time series are extracted, and the time series is analyzed or denoised. It mainly includes four steps: embedding, decomposition, grouping, and reconstruction. SSA decomposes one-dimensional signals into two-dimensional trajectory matrices, and extracts the main trend components in the time series by decomposing and reconstructing the trajectory matrix, achieving data enhancement. The calculation steps are as follows.

[0031] 1) Embedding: assume that the time series with length N is Y N = (y1, y2, …, y N), N > 2, set the embedding dimension as L, L is an integer and 1 < L ≤ N / 3, K = N - L + 1, define the delay vector y i = [y i ,y i+1 ,…,y i+L-1 ] T , the trajectory matrix X of Y can be expressed as:

[0032]

[0033] 2) Singular value decomposition: singular value decomposition is performed on the trajectory matrix X. First, eigenvalue decomposition is performed on XX T , L non-negative eigenvalues in descending order are obtained, and corresponding orthogonal eigenvectors U1, U2, … U L , let d be the number of non-zero eigenvalues, then X can be expressed in the following form:

[0034] X = E1 + E2 + … + E d .

[0035] where is the singular value.

[0036] 3) Grouping: the first r E i approximates X i , and is divided into p groups, the subscript {1, 2, …, r} is divided into consecutive subsets I1, I2, … I p , and I j = {j1, j2, … j m} is the subscript corresponding to the I j th group, and the matrices in each group are added to obtain a new matrix The trajectory matrix X can be expressed as:

[0037]

[0038] In the formula, the variance contribution rate of

[0039] 4) Diagonal average: restore the matrix to a time series. Let the elements of x ij , then can be restored to the corresponding time series {z0, z1, …, z n-1} by the following formula:

[0040]

[0041] Where, L* = min(L, K), K* = max(L, K).

[0042] Step 1.3: The Spearman rank correlation coefficient (SRCC) was used to quantitatively analyze the data enhancement effect.

[0043] SRCC studies the correlation between two variables based on ordinal data. This form of representation does not have the same strict requirements on variable data as the Pearson product-moment correlation coefficient. It does not require the variables to be normally distributed or linearly correlated. Therefore, this algorithm is suitable for studying nonlinear characteristic variables.

[0044] Assume two variables are X and Y, and the i-th (1≤i≤N) value of each variable is represented by X. i Y i This means that sorting X and Y (either in descending or ascending order) yields two ranked sets x and y, where element x is a set of elements in the set x. i y i X i Ranking in X and Y i The ranking in Y. Subtracting the corresponding elements in sets x and y yields a ranking difference set d. The Spearman rank correlation coefficient r between variables X and Y can be calculated from x, y, and d, as follows:

[0045]

[0046] Where d i =x i -y i , 1≤i≤N. The calculated value of SRCC is between -1 and 1. The closer its absolute value is to 1, the greater the correlation between the two.

[0047] Step 2: Construct a three-dimensional spatial visualization representation model of relay degradation based on distance metric.

[0048] Step 2.1: Use cosine similarity distance to measure the degradation of multidimensional feature parameters from a spatial perspective.

[0049] Converting a multidimensional feature vector reflecting the health status of electrical appliances into a one-dimensional health index can be achieved based on the principle of distance measurement. For an n-dimensional feature vector x, the cosine similarity cos(x,u) to the initial state u is:

[0050]

[0051] In the above formula, dot(·) is the inner product operation, and ||·|| is the 2-norm of the vector. The calculated cosine similarity value is between 0 and 1; the closer the value is to 1, the more similar x and u are. However, this calculation method easily amplifies the impact of large amounts of data on similarity. Therefore, when calculating cosine similarity, it is improved to a standardized cosine similarity.

[0052] The feature parameter x is normalized to a dimensionless value:

[0053]

[0054] In the formula, max() and min() are the maximum and minimum value functions, respectively.

[0055] Standardized cosine similarity cos norm (x,u) is:

[0056]

[0057] The measurement of similarity between vectors should be standardized so that the smaller the value, the higher the similarity. Therefore, the cosine similarity distance d is defined. cos for:

[0058] d cos =1-cos norm (x,u).

[0059] Step 2.2: Use Mahalanobis distance to measure the degradation of multidimensional feature parameters from the spatial numerical direction.

[0060] For similarity calculations in multidimensional space, in addition to considering angular distance, the influence of absolute numerical distance should also be taken into account. Mahalanobis distance, when calculating distance similarity, can not only ignore the influence of dimensions but also consider the similarity between overall samples. Therefore, Mahalanobis distance is used to calculate the numerical similarity of multidimensional vectors. Let d be the Mahalanobis distance d between an n-dimensional state vector x and its initial state u. ma for:

[0061]

[0062] In the formula, A is the covariance matrix.

[0063]

[0064] Step 2.3: Combine cosine similarity distance and Mahalanobis distance to perform three-dimensional spatial visualization representation of relay degradation.

[0065] A distance criterion is introduced for similarity measurement. Taking the initial contact state of a railway DC relay as the healthy ground state, the similarity distance between the remaining samples and the healthy ground state is calculated. The distance is used as a degradation index to establish a relay contact degradation model. Mahalanobis distance is used to measure the spatial numerical changes of multidimensional feature parameters, and cosine similarity distance is used to measure the spatial angular changes. This provides a three-dimensional spatial visualization of relay performance degradation, offering a visual basis and degradation spatial state samples for subsequent state interval division.

[0066] Step 3: Construct an improved K-means clustering model to divide the relay state intervals.

[0067] Step 3.1: Construct an improved K-means clustering model for cluster analysis.

[0068] The K-means algorithm is a partition-based clustering algorithm in data mining. This algorithm aggregates unlabeled data according to similar attributes, and the clustering process consists of two parts: cluster assignment and cluster center movement. Its computation is simple and has low time complexity, making it suitable for partitioning relay state intervals. However, the clustering effect of this algorithm depends on the initialization of cluster centers; inappropriate initial value selection can lead to increased computation time and the existence of local optima in the clustering results. The specific implementation steps of the improved K-means algorithm to address these issues are as follows:

[0069] 1) To address the issue of the algorithm being sensitive to initial values, K samples are extracted from the state space sample set as initial cluster centers.

[0070] 2) To avoid the impact of large amounts of data on the calculation of the distance between cluster centers and samples, the sample data is normalized, and the standard Euclidean distance d between each sample and each cluster center is calculated. ou Based on the principle of proximity, the sample is assigned to the nearest cluster.

[0071]

[0072] In the formula, s is x i Standard deviation.

[0073] 3) Calculate the mean of each cluster:

[0074]

[0075] Where: n j C represents the number of data contained in the j-th cluster. j This represents the j-th cluster.

[0076] The formula for calculating termination condition E is:

[0077]

[0078] 4) Repeat steps 2 and 3, iterating until the value of E remains basically unchanged. At this point, the centers of each cluster no longer change, satisfying the loop termination condition of the clustering algorithm.

[0079] Step 3.2: Verify the state interval division results by combining the average amplitude and standard deviation of each stage of the spatial characterization curve.

[0080] The K-means algorithm is an unsupervised learning algorithm; therefore, based on cluster analysis, spatial visualization curve data is used to back-validate the partitioning results. The average amplitude of the degenerate spatial curve is used. It reflects the central tendency of the sample data, and the standard deviation s reflects the dispersion of the sample data, thus enabling supervised feedback verification of the clustering algorithm analysis.

[0081]

[0082]

[0083] In the formula, Here, N is the mean, and N is the number of samples.

[0084] Step 4: Construct an improved HS-Elman model to evaluate and predict the relay's operating state range.

[0085] Step 4.1: Construct an improved Elman neural network prediction model.

[0086] The Elman neural network is a dynamic feedback network consisting of an input layer, hidden layers (intermediate layers), a support layer, and an output layer. Compared to the traditional BP network, its network structure incorporates delay and storage in the support layer, enhancing its local feedback capability. Due to the presence of delay units, the Elman network exhibits strong memory and temporal stability. The improved Elman network adds an output feedback loop, feeding information from the output layer back to the support layer. This allows the support layer to not only remember the information from the hidden layer at the current k-th time and past k-1, k-2, ..., kn-th times, but also the information from the output layer at the current k-th time and some past times. Let the number of nodes in the input, hidden, and output layers of the Elman network be M, L, and N, respectively, and the number of nodes in the support layer be equal to the number of nodes in the hidden layer. The improved Elman neural network model is as follows:

[0087]

[0088] In the formula: x i (k) represents the input of node i in the input layer at time k; y q (k) represents the output of the output layer node q at time k; These are the input and output of hidden layer node j at time k, respectively; These are the input and output of the receiving layer node z at time k, respectively; The connection weights from input layer node i to hidden layer node j; Let be the connection weight from hidden layer node j to output layer node q; f(*) represents the connection weight from node z in the receiving layer to node j in the hidden layer; f(*) is the transfer function of the hidden layer neuron, usually the Sigmoid function.

[0089] Let Y q(k) represents the desired output of the network at time k. Therefore, the objective function, or approximation error function, of the network in the time interval [0, T] is:

[0090]

[0091] The gradient descent principle is used to calculate the adjustment amount Δω for the connection weights of the Elman neural network.

[0092]

[0093] In the formula: η is the learning rate.

[0094] In summary, the improved Elman neural network algorithm can be obtained as follows:

[0095]

[0096] For weights The calculations are performed using the chain rule for differentiation:

[0097]

[0098]

[0099]

[0100] In the formula: η1, η2, η3, and η4 are respectively The learning step size.

[0101] Step 4.2: Optimize and improve the relevant parameters of the Elman model using the Harmony Search (HS) algorithm.

[0102] When improving the Elman neural network algorithm by training the network parameters using gradient descent, it is prone to getting trapped in local optima. To overcome this drawback, this paper utilizes the HS algorithm to improve the relevant parameters of the Elman model.

[0103] First, set the relevant parameters of the HS-Elman model. Define the number of parameters N for each solution vector and the four parameters of the Elman model used: Harmony Library Size (HMS), Memory Library Probability (HMCR), Fine-tuning Probability (PAR), and Pitch Fine-tuning Bandwidth BW.

[0104] Secondly, a set of relevant parameters ω generated during the training of the Elman neural network algorithm will be used. 1 ,ω 2 ,ω 3 ,ω 4 As a set of harmonic ω in the HS algorithm i ={ω 1 ,ω 2 ,ω3 ,ω 4 Then, HMS ω values ​​are randomly selected from the parameter set generated during the Elman neural network algorithm training process. i and these ω i Perform real number encoding. Encode these SHM ω values. i As the initial HM, i.e. {ω1,ω1,…,ω SHM HM is the HMS tuple of harmony ω.

[0105] Subsequently, a new solution is searched in the harmony memory using the probability value HMCR. The new solution is generated as follows:

[0106]

[0107] If the new harmony comes from HM, then the new variable ω needs to be adjusted. i Fine-tuning will be performed, specifically the following methods:

[0108]

[0109] In the formula: rand is a random number uniformly distributed on [0,1]; Z i For the range of variables; Let be the j-th solution of HM obtained through the i-th calculation, where j = 1, 2, ..., SHM.

[0110] ω i After adjustment, then ω i Update the HM and finally estimate the fitness of the new solution; if the new solution is better than the worst fitness value in the HM, then add the new solution to the HM. The method for adding the new solution is shown in the formula:

[0111]

[0112] Repeat the above steps until T does not exceed T. max Up to this point, the optimal solution in HM is the optimal value of the relevant parameters of the Elman algorithm.

[0113] Step 4.3: Construct an improved HS-Elman neural network for relay state assessment and prediction.

[0114] From the degenerate space samples, 70% of the samples are selected as the training set and 30% as the test set in a 7:3 ratio. The training sample data is normalized to construct the Elman neural network prediction model to be optimized.

[0115] Set the relevant parameters for the improved Elman model. Define the number of parameters N for each solution vector and the four parameters of the Elman model used: Harmony Memory Size (HMS), Harmony Memory Probability (HMCR), Pitch Adjustment Probability (PAR), and Pitch Adjustment Bandwidth BW. Then, use the probability value HMCR in the harmony memory to search for new solutions. If the new harmony comes from HM, then the new variable ω needs to be adjusted. i Make fine adjustments, and then adjust ω. i Update the HM and finally estimate the fitness of the new solution; if the new solution is better than the worst fitness value in the HM, add the new solution to the HM and repeat the above steps until T does not exceed T. max Until then, the optimal solution in HM is the optimal value of the relevant parameters of the Elman algorithm, satisfying the algorithm's iteration termination condition.

[0116] The optimal solution is selected as the input parameter of the improved Elman model to determine the improved HS-Elman prediction model. The test sample set is then input to obtain the final state evaluation result. Attached Figure Description

[0117] Figure 1 This is the SSA flowchart.

[0118] Figure 2 It is a data-enhanced image.

[0119] Figure 3 This is a graph showing the correlation coefficient of SRCC.

[0120] Figure 4 It is a degenerate representation of cosine similarity distance.

[0121] Figure 5 This is a characterization diagram of Mahalanobis distance degradation.

[0122] Figure 6 It is a three-dimensional spatial visualization representation.

[0123] Figure 7 This is a comparison chart of the average amplitude across different state intervals.

[0124] Figure 8 This is a comparison chart of the standard deviations of each state interval.

[0125] Figure 9 This is a schematic diagram of the improved Elman model structure.

[0126] Figure 10 This is a flowchart of the improved HS-Elman prediction model.

[0127] Figure 11 This is a graph showing the prediction results of the improved HS-Elman model. Detailed Implementation

[0128] To make the objectives, technical solutions, and advantages of the present invention clearer, the following detailed description of a railway DC relay operation status evaluation method based on distance measurement and improved Elman measurement, in conjunction with specific embodiments and accompanying drawings, is provided.

[0129] (1) Collect test data through relay electrical life test, extract multidimensional feature parameters from it and perform data enhancement processing on the feature parameters.

[0130] The setup and testing were conducted according to the requirements of GB14048.4-2016 and GB14048.5-2016. The Siemens 3RH2122-2KF40 railway DC relay, commonly used in railway trains, was used as the test object. The feature parameters extracted from the electrical life test data contained environmental noise and fluctuated drastically, thus masking their own time-series trend information. Therefore, the SSA algorithm was used to denoise and enhance the trend of the extracted feature parameters, and SRCC was used to quantitatively analyze the data enhancement effect and the rationality of feature selection. See step 1 for details.

[0131] (2) Construct a three-dimensional spatial visualization representation model of relay degradation based on distance metric.

[0132] The multidimensional feature vectors reflecting the health status of electrical equipment are transformed into a one-dimensional health index using distance metrics. Cosine similarity distance is used for spatial angular measurement, while Mahalanobis distance is used for spatial numerical measurement. Taking the initial contact state of the relay as the healthy ground state, the similarity distance between the remaining samples and the healthy ground state is calculated, and the distance magnitude is used as an indicator of degradation. Finally, the two methods are combined to achieve a visual representation from the multidimensional feature space to the three-dimensional state space. See step 2 for details.

[0133] (3) Construct an improved K-means clustering model to divide the relay state interval.

[0134] This patent constructs a state partitioning model based on an improved K-means clustering algorithm, as detailed in step 3. By selecting initial values ​​from the degenerate sample set and conducting comparative analysis with different cluster numbers K, the problem of local optima arising from the model's reliance on cluster center initialization is resolved. Based on the clustering analysis, spatial visualization curve data is used to back-validate the partitioning results. The average amplitude of the degenerate spatial curve is employed. The standard deviation s reflects the central tendency of the sample data and the dispersion of the sample data, enabling supervised feedback verification of the clustering algorithm analysis and improving the reliability of the model partitioning.

[0135] (4) Construct an improved HS-Elman model to evaluate and predict the relay operating state range.

[0136] The improved HS-Elman evaluation and prediction model constructed in this patent is detailed in step 4. The improved Elman model is optimized using the Harmony Search (HS) algorithm, which solves the problem of the Elman model easily getting trapped in local optima when training network parameters using gradient descent, thus improving the accuracy of the Elman model's recognition and prediction.

[0137] The following is a specific implementation example:

[0138] This invention refers to the requirements of GB14048.4-2016 and GB14048.5-2016 standards to construct a railway DC relay electrical life measurement test platform to collect voltage and current signals of the contacts and coils during contact operation. Through relay electrical life testing, the contact engagement and disengagement processes are observed, and various performance degradation parameters are extracted. The SSA algorithm is used to perform data denoising and trend enhancement processing on the extracted feature parameters, such as... Figure 1 As shown, the original signal of the 249,349 interruption times of the extracted 6 feature parameters and its trend signal after SSA enhancement processing are as follows: Figure 2 As shown, SRCC was used to quantitatively analyze the enhanced characteristic parameters, verifying the effectiveness of the enhancement and feature selection. After SSA data enhancement, the correlation values ​​between each characteristic parameter and its lifespan significantly increased, indicating that the characteristic parameters enhanced by SSA can better characterize the performance degradation trend of the relay. Figure 3 As shown, specific quantitative indicators were used to verify the effectiveness of SSA in data augmentation, and the rationality of selecting six feature parameters as performance indicators was also verified.

[0139] Cosine similarity distance is used to measure spatial angular change, and the distance calculation results are as follows: Figure 4 As shown, Mahalanobis distance is used to measure the spatial numerical variation of multidimensional feature parameters, such as... Figure 5 As shown, to combine the overall trend and local differences of the metric, both are used for a three-dimensional spatial degradation visualization representation, and to provide a sample set for subsequent state partitioning, such as... Figure 6 As shown.

[0140] An improved K-means clustering algorithm was used to perform state clustering analysis on distance samples representing spatial representations. However, this algorithm can lead to the decomposition of identical states by creating too many clusters, resulting in complex and meaningless state divisions. To address the algorithm's sensitivity to initial values, initial cluster centers were selected from the samples, with cluster numbers set to 2, 3, 4, 5, and 6 respectively. The clustering statistics are shown in Table 1.

[0141] Table 1. Statistics on the proportion of clustered samples

[0142]

[0143]

[0144] When the number of clusters is 2 or 3, the number of samples at the end of the life cycle accounts for about 30%, which does not reflect the actual operating conditions of the equipment. When the number of clusters is 5 or 6, the state intervals are over-subdivided due to the increase in the number of clusters. From the visualized state curves and cluster statistics, it can be seen that stages III and IV with a cluster number of 5 are derived from stage III with a cluster number of 4, and stages II and III with a cluster number of 6 are derived from stage II with a cluster number of 5. This over-subdivision of state intervals will increase the complexity of subsequent state identification and has no obvious practical significance. Therefore, stage clustering with a cluster number of 4 is the most reasonable. Based on the feedback verification of the characterization curves, it is concluded that the degradation of the relay can be divided into 4 stages, as shown in Table 2. Figure 7 , Figure 8 As shown.

[0145] Table 2 Relay Status Level Classification

[0146]

[0147] like Figure 9 , Figure 10 A study was conducted on the operation status assessment and prediction of railway DC relays using an improved Elman neural network optimized by a harmony search algorithm. The results are as follows: Figure 11 As shown, compared with other recognition algorithms, the improved HS-Elman prediction model constructed in this invention can accurately and effectively identify the operating state of the relay, and is more suitable for research on the operating state identification of relays.

[0148] Table 3 Comparison of test results for different models

[0149]

[0150] Traditional BP neural network models suffer from low accuracy in state recognition due to their computational simplicity. While Elman and improved Elman models offer better accuracy than BP, they still fall short of requirements. The improved HS-Elman model, with its HS optimization algorithm, compensates for the deficiencies in weight and threshold selection found in the improved Elman model and addresses the problem of local optima. Although it slightly increases computational cost, it further enhances state recognition accuracy, meeting the research requirements.

[0151] The railway DC relay operation status assessment method based on distance metric and improved Elman's algorithm, constructed in this invention, can compare and divide relay degradation state intervals without prior conditions, and achieve assessment and prediction analysis of corresponding state stages. It has advantages such as fast data convergence, stable prediction results, and high accuracy. In test experiments, the model achieved a relay state identification accuracy of 98.12%, making it suitable for assessing and identifying the operation status of railway DC relays.

[0152] The purpose of this invention is to propose a method for evaluating the operating status of railway DC relays based on distance metrics and an improved Elman algorithm. This method involves constructing an electrical life test platform to collect test data and extracting various feature parameters for subsequent analysis. SSA (Single-Signal Analysis) is used to denoise and enhance the feature data, extracting key trend information from time-series features. SRCC (Signal-Rated Conversion) is used to quantitatively calculate the data enhancement effect and the rationality of parameter selection. Distance metrics are applied to the feature parameters, considering both absolute numerical distance and angular distance in the multi-dimensional feature space. Mahalanobis distance and cosine similarity distance are used to visualize the degradation state of the relay in the degradation sample space. The spatial representation results are more three-dimensional and comprehensive than planar representation curves, providing state samples and visualization basis for the next step of state interval clustering. An improved K-means algorithm is used to cluster the spatial representation samples. Supervised feedback is performed based on the visualized state representation results to determine that the operating status of this type of relay can be divided into four stages, with specific stage divisions provided for subsequent operating status evaluation and identification. The Elman neural network model was optimized and improved using the harmony search algorithm, and an improved HS-Elman model was constructed to identify and predict the relay status, providing a theoretical basis for maintenance decisions of railway DC relays and providing a favorable guarantee for the safe and reliable operation of trains.

Claims

1. A method for evaluating the operating status of railway DC relays based on distance metrics, characterized in that: Includes the following steps: 1) Through relay electrical life tests, extract various characteristic parameters: pull-in time, release time, contact resistance, bounce time, overtravel time, and arc energy for analysis; 2) Singular spectral analysis (SSA) was used to denoise and enhance the trend of the initial feature parameters; 3) The Spearman rank correlation coefficient (SRCC) was used to quantitatively analyze the data augmentation effect and verify the rationality of the parameter selection; 4) Cosine similarity distance is used to represent the degradation distance from the perspective of the feature space; 5) The degenerate distance is represented from the numerical direction of the feature space using Mahalanobis distance; 6) Combining both angular and numerical aspects, it achieves a visual representation of the degenerate state from multi-dimensional space to three-dimensional space, and provides a distance sample set for subsequent analysis; 7) An improved K-means clustering algorithm was used to perform cluster analysis on the distance samples representing spatial characteristics; 8) The state interval division results are verified by combining the average amplitude and standard deviation of each stage of the spatial characterization curve; 9) Construct an improved Elman neural network prediction model; 10) Optimize and improve the relevant parameters of the Elman model using the Harmony Search (HS) algorithm; 11) Construct an improved HS-Elman neural network for relay state assessment and prediction; Step 9) involves constructing an improved Elman neural network prediction model, including: The improved Elman network adds an output feedback loop, which feeds the information of the output layer back to the receiving layer. This allows the receiving layer to not only remember the information of the hidden layer at the current k time and the past k-1, k-2, ..., kn times, but also the information of the output layer at the current k time and some past times. Let the number of nodes in the input layer, hidden layer, and output layer of the Elman network be M, L, and N, respectively, and the number of nodes in the connecting layer be equal to the number of nodes in the hidden layer. The improved Elman neural network model is as follows: In the formula: x i (k) represents the input of node i in the input layer at time k; y q (k) represents the output of the output layer node q at time k; These are the input and output of hidden layer node j at time k, respectively; These are the input and output of the receiving layer node z at time k, respectively; The connection weights from input layer node i to hidden layer node j; Let be the connection weight from hidden layer node j to output layer node q; f(*) represents the connection weights from node z in the receiving layer to node j in the hidden layer; f(*) is the transfer function of the hidden layer neuron, which is the Sigmoid function. Let Y q (k) represents the desired output of the network at time k. Therefore, the objective function, or approximation error function, of the network in the time interval [0, T] is: The gradient descent principle is used to calculate the adjustment amount Δω for the connection weights of the Elman neural network. In the formula: η is the learning rate; In summary, the improved Elman neural network algorithm is as follows: For weights The calculations are performed using the chain rule for differentiation: In the formula: η1, η2, η3, and η4 are respectively The learning step size.

2. The railway DC relay operation status evaluation method based on distance metric as described in claim 1, characterized in that: Step 1) describes the extraction and analysis of various characteristic parameters through relay electrical life testing, including: pull-in time, release time, contact resistance, bounce time, overtravel time, and arc energy. Based on the relay performance degradation characteristics, the calculation methods for various characteristic parameters are as follows: 1) Adsorption time t x =t2-t1 In the formula, t1 is the moment when the coil is energized, and t2 is the moment when the moving and stationary contacts first make contact; 2) Jump time t t t3-t2 In the formula, t3 is the time when the voltage and current fluctuations end; 3) Overrun time t c =t4-t2 In the formula, t4 is the stable adsorption time; 4) Contact resistance In the formula, N is the number of data collection points, u n i n Contact voltage and current; 5) Release time t s =t6-t5 In the formula, t5 is the moment when the coil is de-energized, and t6 is the moment when the arc begins to ignite; 6) Arc energy In the formula, Δt is the time interval between adjacent sampling points, and f s The sampling rate.

3. The method for evaluating the operating status of railway DC relays based on distance metrics according to claim 1, characterized in that: Step 2) describes using Singular Spectrum Analysis (SSA) to denoise and enhance the trends of the initial feature parameters, which includes four steps: embedding, decomposition, grouping, and reconstruction. 1) Embedding: Assume a time series of length N is Y N =(y1,y2,…,y N Given N > 2, set the embedding dimension to L, where L is an integer and 1 < L ≤ N / 3, K = N - L + 1, and define the delay vector y. i =[y i ,y i+1 ,…,y i+L-1 ] T The trajectory matrix X of Y can be represented as: 2) Singular Value Decomposition: Perform singular value decomposition on the trajectory matrix X; first, decompose XX... T Perform eigenvalue decomposition to obtain L non-negative eigenvalues ​​in descending order, and corresponding orthogonal eigenvectors U1, U2, ... U1. L Let d be the number of non-zero eigenvalues. (i = 1, 2, ..., d), then X can be represented in the following form: X=E1+E2+…+E d in (i = 1, 2, ..., d), It is a singular value; 3) Grouping: Take the first r E i Approximate X i They are then divided into p groups, and the subscripts {1,2,…,r} are further divided into continuous subsets I1,I2,…I p , note I j ={j1,j2,…j m } is the I j The subscripts of the groups are used to sum the matrices within each group to obtain a new matrix. The trajectory matrix X can be represented as: In the formula, The variance contribution rate 4) Diagonal Mean: Restores the matrix to a time series; let... The element is x ij ,but The corresponding time series {z0, z1, ..., z} can be restored by the following formula. n-1 }: Where L* = min(L,K) and K* = max(L,K).

4. The method for evaluating the operating status of railway DC relays based on distance metrics according to claim 1, characterized in that: Step 3) describes using the Spearman Rank Correlation Coefficient (SRCC) to quantitatively analyze the data augmentation effect and verify the rationality of parameter selection, including: Assume two variables are X and Y, and the i-th value of each variable is represented by X. i Y i This means that sorting X and Y results in two ranked sets x and y, where element x is the set of elements in the set x. i y i X i Ranking in X and Y i The ranking in Y; subtracting the corresponding elements in sets x and y yields a ranking difference set d. The Spearman rank correlation coefficient r between variables X and Y can be calculated from x, y, and d, as follows: Where d i =x i -y i ,1≤i≤N; The calculated value of SRCC is between -1 and 1. The closer the absolute value is to 1, the greater the correlation between the two; the closer the absolute value is to 0, the smaller the correlation between the two.

5. The method for evaluating the operating status of railway DC relays based on distance metrics according to claim 1, characterized in that: In step 5), the degradation distance representation from the numerical direction of the feature space using Mahalanobis distance includes: calculating the numerical similarity of multidimensional vectors using Mahalanobis distance, where the n-dimensional state vector x has a Mahalanobis distance d from the initial state u. ma for: In the formula, A is the covariance matrix; 6. The method for evaluating the operating status of railway DC relays based on distance metrics according to claim 1, characterized in that: Step 6) combines both angular and numerical aspects to achieve a degenerate visualization representation from multi-dimensional space to three-dimensional space, and provides a distance sample set for subsequent analysis, including: A distance criterion is introduced for similarity measurement. Taking the initial contact state of the railway DC relay as the healthy ground state, the similarity distance between the remaining samples and the healthy ground state is calculated. The distance is used as the degradation index, and then a relay contact degradation model is established. Mahalanobis distance is used to measure the spatial numerical changes of multidimensional feature parameters, and cosine similarity distance is used to measure the spatial angular changes. This allows for a three-dimensional spatial visualization of relay performance degradation, providing a visual basis and degradation spatial state samples for subsequent state interval division.

7. The method for evaluating the operating status of railway DC relays based on distance metrics according to claim 1, characterized in that: Step 8) The feedback verification of the state interval division results by combining the average amplitude and standard deviation of each stage of the spatial characterization curve includes: on the basis of cluster analysis, using spatial visualization characterization curve data to perform reverse verification of the division results; Average amplitude of the degradation space curve It reflects the central tendency of the sample data, and the standard deviation s reflects the dispersion of the sample data, thus enabling supervised feedback verification of the clustering algorithm analysis; In the formula, is the mean, and N is the number of samples.

8. The method for evaluating the operating status of railway DC relays based on distance metrics according to claim 1, characterized in that: Step 10) describes optimizing and improving the Elman model using the Harmony Search (HS) algorithm, which includes the following parameters: First, set the relevant parameters of the HS-Elman model; the number of parameters N for each solution vector and the four parameters of the Elman model used: harmony library size HMS, memory library value probability HMCR, fine-tuning probability PAR, and pitch fine-tuning bandwidth BW; Secondly, a set of relevant parameters ω generated during the training of the Elman neural network algorithm will be used. 1 ,ω 2 ,ω 3 ,ω 4 As a set of harmonic ω in the HS algorithm i ={ω 1 ,ω 2 ,ω 3 ,ω 4 Then, HMS ω values ​​are randomly selected from the parameter set generated during the Elman neural network algorithm training process. i and these ω i Perform real number encoding; convert these SHM ω i As the initial HM, i.e. {ω1,ω1,…,ω SHM }; HM is the HMS tuple of harmony ω; Subsequently, a new solution is searched in the harmony memory using the probability value HMCR. The new solution is generated as follows: If the new harmony comes from HM, then the new variable ω needs to be adjusted. i Fine-tuning will be performed, specifically the following methods: In the formula: rand is a random number uniformly distributed on [0,1]; Z i For the range of variables; Let HM be the j-th solution obtained through the i-th calculation, where j = 1, 2, ..., SHM; ω i After adjusting, then ω i Update the HM and finally estimate the fitness of the new solution; if the new solution is better than the worst fitness value in the HM, then add the new solution to the HM. The method for adding the new solution is shown in the formula: Repeat the above steps until T does not exceed T. max Up to this point; the optimal solution in HM is the optimal value of the relevant parameters of the Elman algorithm; Step 11) involves constructing an improved HS-Elman neural network for relay state assessment and prediction, which includes: From the degenerate space samples, 70% of the samples are selected as the training set and 30% of the samples are selected as the test set in a 7:3 ratio. The training sample data is normalized to construct the Elman neural network prediction model to be optimized. Set the relevant parameters for the improved Elman model; specify the number of parameters N for each solution vector and the four parameters of the Elman model used: harmony library size HMS, memory value probability HMCR, fine-tuning probability PAR, and pitch fine-tuning bandwidth BW; then, use the probability value HMCR in the harmony memory to search for new solutions. If the new harmony comes from HM, then the new variable ω needs to be adjusted. i Make fine adjustments, and then adjust ω. i Update the HM and finally estimate the fitness of the new solution; if the new solution is better than the worst fitness value in the HM, add the new solution to the HM and repeat the above steps until T does not exceed T. max Until then; at this point, the optimal solution in HM is the optimal value of the relevant parameters of the Elman algorithm, satisfying the algorithm iteration termination condition; The optimal solution is selected as the input parameter of the improved Elman model to determine the improved HS-Elman prediction model. The test sample set is then input to obtain the final state evaluation result.

Images