A real-time accurate early warning method for subway turnout overload failure

Abstract

A real-time, accurate early warning method for subway turnout overload failures is proposed. First, the root mean square value of the characteristic points of the turnout power curve is continuously calculated in real time to obtain a load characteristic sequence, which serves as the prediction input. A multi-description complementary prediction mechanism is then designed, and the base predictor parameters are optimized. A machine learning model for multi-description complementary prediction is established to perform multi-step, single-point range prediction of short-term load characteristics, forming a prediction feature matrix. Using the elements of the prediction feature matrix as samples, the probability density function and overall confidence interval of the turnout load characteristics during the prediction period are calculated using an adaptive kernel density estimation method. Combined with the confidence intervals of historical load characteristics, the overload degree of the turnout during the prediction period is calculated as a warning indicator. Finally, the relationship between the overload degree and the warning threshold is compared to determine whether a subsequent failure is likely and to infer the time of failure occurrence. Furthermore, the warning threshold is optimized based on user needs to balance the warning success rate and false alarm rate.

Description

Technical Field

[0001] The present invention relates to the field of turnout mechanical fault prediction and early warning methods, and in particular to a real-time accurate early warning method for overload action faults of subway turnouts. Background Art

[0002] Turnouts are currently the most critical infrastructure for rail transit. Failures can cause delays at best, or even derailments at worst. Subway switchbacks frequently operate, and due to wear and tear combined with environmental influences, they often enter an overloaded state for short periods, leading to mechanical failures such as jamming. Compared to sudden electrical failures such as wire breaks and short circuits, the degradation process of mechanical failures is more easily observed and captured, offering valuable early warning capabilities. Predicting risks before a failure occurs, and deploying repair resources or rerouting, can minimize the impact of a failure or even prevent it from occurring. Therefore, real-time warning of switch overload is crucial for ensuring subway safety.

[0003] There are three main types of technical methods currently used for turnout fault prediction and condition warning. The first type of method uses a state degradation model based on multiple dynamic and static state parameters to predict the remaining useful life of the equipment. This method is suitable for modeling the evolution of equipment failures, and the model is generally based on specific wear tests. However, the operating environment of turnouts is complex and uncertain, and many static parameters cannot be obtained in real time. Therefore, this method is not suitable for sensing short-term state changes. The second approach extracts the time-domain features of the power monitoring curve generated by turnout movements and learns to obtain state thresholds for real-time fault prediction and diagnosis. Although this approach can achieve real-time state prediction and early warning, it has some important drawbacks. Fault prediction based solely on anomaly identification of the power curve generated by a single turnout movement is rather one-sided. Field experience shows that normal monitoring data is generally stable but fluctuates. A single anomaly in the turnout movement curve only indicates a potential problem with that turnout movement. However, a fault is often the result of multiple anomalies leading to qualitative changes. Therefore, using single-turnout characteristics for fault early warning is somewhat accidental and can lead to increased misjudgment. An improvement has been made. First, a state threshold is adaptively calculated as a reference for real-time turnout anomaly identification. Then, a statistical strategy is used to determine the criticality of a fault. This mitigates the one-sidedness of fault early warning based solely on single-turnout anomaly identification. However, the problem is that the warning still relies on existing movement characteristics. When a warning is issued, it often indicates that a fault is approaching. For short-distance trains, it may be too late to take emergency measures at this time, weakening the significance of early warning.

[0004] The existing third type of technology has improved upon the second type of technology. The main idea is to adopt a two-step process of "feature prediction-fault warning / prediction", that is, to make real-time predictions on the accumulated action features, and then use the predicted values ​​for fault warning and prediction. This can to a certain extent overcome the one-sidedness of the second type of technology, which only uses the action features generated at the current moment to identify anomalies to predict faults, and is conducive to earlier perception of fault signs.

[0005] Existing technologies all use multi-step single-point deterministic feature prediction, which has two drawbacks:

[0006] The standard grey prediction model used is a direct prediction model that ignores external patterns and only considers internal connections. Its prediction results depend on the initial and final values ​​of the input sequence, resulting in poor generalization. Based on single-point deterministic predictions, large errors in predicted values ​​or significant prediction lags can lead to continuous deviations in subsequent judgments. Short-term deterministic predictions have a limited number of steps, lacking sufficient prediction samples to characterize the switch's operational state in subsequent time periods.

[0007] Existing technologies directly use predicted values ​​as features of subsequent turnout states for fault prediction or early warning, a flawed approach. Any prediction inevitably involves errors. As noted in the first flaw description, single-point deterministic predictions with a limited number of steps are subject to multiple negative impacts, including prediction error, prediction lag, and insufficient prediction samples. Existing technologies often employ a two-step process: "feature prediction-fault prediction / early warning." This process remains at the "fuzzy early warning" stage, determining whether a fault has occurred, and cannot achieve effective "precise early warning." The concept of precise early warning proposed in this invention refers to a method that not only predicts whether a subsequent fault will occur, but also infers the operating period and even the specific moment of the fault. Clearly, this so-called "precise early warning" better meets the needs of field applications, facilitating users to take targeted emergency measures based on the advance warning time. This represents the future development trend of all fault early warning methods. Existing technologies are subject to several limitations, and currently no method clearly demonstrates the ability to achieve relatively accurate early warning judgments. Regarding the final fault prediction or early warning judgment, existing technologies also suffer from unreasonable early warning threshold settings and insufficiently balanced early warning performance. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed to provide a real-time accurate early warning method for subway turnout overload operation failure that overcomes the above problems or at least partially solves the above problems.

[0009] In order to solve the above technical problems, the embodiments of the present application disclose the following technical solutions:

[0010] A real-time accurate early warning method for subway turnout overload failure, comprising:

[0011] S100 obtains the cumulative continuous turnout power curve when the i-th turnout overload action is performed, and calculates the real-time action load characteristic sequence {R};

[0012] S200. Using a multi-description complementary prediction engine, the real-time action load feature sequence {R} is multi-step value range predicted to construct a subsequent short-term action load feature prediction sample matrix G;

[0013] S300. Using the sample matrix G of the action load characteristics of the subsequent time period as a sample, based on the adaptive kernel density estimation method, the probability density function and the overall confidence interval of the load characteristics of the forecast period are calculated;

[0014] S400. Calculate the overload degree S of the switch action during the forecast period and use it as an early warning evaluation indicator;

[0015] S500. Using overload fault and non-fault data to perform a preference search, determine the warning threshold t, and make a warning judgment based on the warning threshold t and the overload degree S of the switch action during the forecast period;

[0016] S600. Infer the fault occurrence time based on the current driving interval and the warning judgment result.

[0017] Furthermore, in S100, the method for calculating the real-time action load characteristic sequence {R} includes: taking the current i-th action moment as a reference, taking the previous i-2L turnout action power curves in the same direction, calculating the root mean square value of the sampling point as the action load characteristic, and obtaining the turnout action load characteristic sequence {R i |i=1,2,3...,2L}, 2L represents the sequence length.

[0018] Furthermore, in S200, the method for predicting the action load characteristic sample matrix G of the subsequent time period is as follows: performing odd and even sampling on the characteristic sequence {R} respectively to obtain two groups of subsequences, using the basis predictor to predict the original sequence to obtain {r(i)|i=1,2,...,2l}, and then using the basis predictor to predict the two groups of subsequences to obtain {r′(i)|i=1,2,...,l}, {r″(i)|i=1,2,...,l}, and a given weight sequence {k1,k2,...,k 2l}, the final prediction sequence {r f The single-point prediction value at each step can be obtained by interweaving and weighting r, r′, and r″; then the weight ki of each prediction step is optimized to obtain n changing values ​​to construct the optimal compensation weight matrix Final use The original prediction sequence {r f} is expanded into an n×2l prediction sample matrix G.

[0019] Furthermore, the prediction sequence {r f The calculation formula for} is:

[0020]

[0021] Among them, {r(i)|i=1,2,...,2l} is obtained by using the basis predictor to predict the original sequence, {r′(i)|i=1,2,...,l} and {r″(i)|i=1,2,...,l} are obtained by using the basis predictor to predict the two groups of sampling sequences, and {k1,k2,...,k 2l} is a given weight sequence.

[0022] Further, it is used to predict the sequence {r f} is expanded into the prediction sample matrix G The way to obtain is:

[0023]

[0024] Among them, Q is the number of learning samples covered, loss(K n×2l ) is the prediction feature in K n×2l The loss function below.

[0025] Furthermore, loss(K n×2l ) is calculated as follows: the weighted prediction sequence {r f} is input, and a set of weight sequences {k i ,k i +b,...,k i +(n-1)b}, and get the compensation weight matrix K n×2l It is used to construct a prediction feature sequence with a changing value at each step to form a prediction feature matrix, which is expected to be enveloping or as close to the true value as possible within a small range; that is, given a weight domain B, the weighted single point prediction value is r, the single point real value is V, and the prediction feature is constructed in K n×2l The loss function is:

[0026]

[0027] In the formula Based on the KL divergence modification, F T and F P are the discrete probability distributions of the true sequence and the predicted feature under the same sample value interval distribution Ω, and λ is the conditional parameter.

[0028] Furthermore, the base predictor is selected and optimized. The specific method is as follows:

[0029] The standard grey model DGM is used as the base predictor, and its standard structure is improved and optimized, including:

[0030] The weight parameter w and translation parameter c are introduced to weight and translate the original sequence in pairs to obtain the intermediate sequence, and then accumulate to generate the prediction input sequence, which improves the prediction accuracy and generalization ability from the two perspectives of sequence smoothing and adjustment level ratio. Assume w∈W, c∈C, use each group (w, c) in the value space to pre-model and predict the original sequence, and calculate the root mean square error (RMSE) between the original sequence and the corresponding simulated prediction sequence. (w,c) , Fréchet distance FD (w,c) , establish the binary consistency constraint equation:

[0031]

[0032] Where e r 、e d The root mean square error and Fréchet distance between the simulated prediction sequence and the original sequence obtained by standard DGM modeling are respectively, and the parameter group that minimizes the target E(w,c) by traversing (w,c) values ​​is:

[0033]

[0034] Then, the pre-optimized parameters (w0, c0) are substituted into the prediction model to achieve optimized prediction.

[0035] Furthermore, the method for calculating the probability density function and the overall confidence interval of the load characteristics during the forecast period is as follows:

[0036] The predicted feature matrix element is recorded as {g i}, the number of elements 2nl = M, the one-dimensional kernel density estimation function is as follows:

[0037]

[0038] Where, is the kernel function, and the Gaussian kernel is selected as the representation below; h represents the bandwidth, and the variable value h is determined adaptively according to the sample position. i Instead; therefore, the adaptive Gaussian kernel density estimation function can be obtained:

[0039]

[0040] where h i Use the asymptotically integrated mean square error method to obtain the optimal solution:

[0041]

[0042] In the formula Represents the optimal fixed window width dominated by the sample standard deviation σ; ω is a parameter; further, the overall confidence interval [c1, c2] of the prediction space under a given confidence level P can be obtained:

[0043]

[0044] Furthermore, the method for calculating the overload degree S of the turnout operation during the prediction period is:

[0045] Collect a large number of historical motion load feature samples and use the method in step 3 to calculate their confidence interval [z1, z2] at the same confidence level as the representation of the empirical level of motion load features;

[0046] Combining the probability density function and confidence interval of the prediction feature matrix, the action overload degree S during the prediction period is defined and calculated as follows:

[0047]

[0048] The closer S is to 1, the greater the probability that the overall predicted level exceeds the upper limit of the historical experience level, the greater the exceedance, and the greater the possibility that the turnout will fail due to overload operation in the short term.

[0049] Furthermore, the specific method of S500 includes: making a warning judgment based on the warning threshold value obtained by optimizing the two-aspect preference, first judging whether to issue a warning according to the warning threshold value t and the overload degree S of the switch action in the predicted period, if the overload degree S calculated in the previous step is greater than or equal to t, it is considered that the switch may have an overload action failure in the subsequent period and a warning is required; if a warning is issued, the specific action period and action sequence of the subsequent failure are further inferred.

[0050] The beneficial effects of the above technical solutions provided by the embodiments of the present invention include at least:

[0051] The present invention discloses a real-time accurate early warning method for overload action failure of subway turnouts. First, the characteristic points of the turnout power / tension curve are collected in real time and continuously, and the root mean square value is calculated to obtain the turnout action load characteristic sequence as a prediction input; then a multi-description complementary prediction mechanism is designed, parameter learning optimization is performed on the selected base predictor, and a machine learning model of the multi-description complementary prediction is established to perform multi-step single-point range prediction of short-term action load characteristics to form a prediction feature matrix; using the elements of the prediction feature matrix as samples, the probability density function and the overall value confidence interval of the turnout action load characteristics during the prediction period are calculated based on an adaptive kernel density estimation method, and the action overload degree of the prediction period is defined as a warning indicator in combination with the historical load characteristic value confidence interval; finally, by comparing the size relationship between the overload degree and the warning threshold, it is judged whether a subsequent failure is likely to occur, and the warning threshold is optimized with preference so that the method balances the warning success rate and the false alarm rate according to user needs, thereby effectively achieving accurate warning.

[0052] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0054] Figure 1 This is a flow chart of a real-time accurate early warning method for subway turnout overload operation failure in embodiment 1 of the present invention. DETAILED DESCRIPTION

[0055] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0056] In order to solve the problems existing in the prior art, an embodiment of the present invention provides a real-time accurate early warning method for subway turnout overload operation failure.

[0057] Example 1

[0058] A real-time accurate early warning method for subway turnout overload failure, characterized by comprising:

[0059] S100. Obtain the continuous turnout power curve accumulated during the i-th turnout overload operation, and calculate the real-time operation load characteristic sequence {R}; Specifically, the turnout operation load characteristic index: Referring to the definition of mechanical load, the turnout operation load can refer to the total power drawn from the switch machine to overcome resistance at a certain moment of operation. Therefore, current technology often uses the power / tension curve generated in real time as the turnout operates to characterize its mechanical operation state. Since a turnout operation lasts for approximately 6 seconds, the present invention calculates the root mean square value of the characteristic points of the power / tension curve during the turnout operation (unlocking to locking) as the load characteristic. The sampling points of the characteristic points can be determined by the user according to actual conditions, and generally include the curve entry value, peak value, exit value, etc.

[0060] In this embodiment S100, the method for calculating the real-time action load characteristic sequence {R} includes: taking the current i-th action moment as a reference, taking the previous i-2L turnout action power curves in the same direction, calculating the root mean square value of the sampling point as the action load characteristic, and obtaining the turnout action load characteristic sequence {R} i |i=1,2,3...,2L}, 2L represents the sequence length.

[0061] S200. Using a multi-description complementary prediction engine, the real-time action load feature sequence {R} is multi-step value range predicted to construct a subsequent short-term action load feature prediction sample matrix G;

[0062] Specifically, the original input sequence is sampled to obtain two sets of sequences {R1, R3, ..., R 2L-1}, {R2,R4,...,R 2L}, which is regarded as the information description of the original sequence with 0.5 sparsity. Then, the base predictor is used to predict the original sequence to obtain {r(i)|i=1,2,...,2l}, and then the two sets of sample sequences are predicted to obtain {r′(i)|i=1,2,...,l}, {r″(i)|i=1,2,...,l}. Given the weight sequence {k1,k2,...,k 2l}, the single-point prediction value of each step in the final prediction sequence can be obtained by interweaving and weighting r, r′, and r″:

[0063]

[0064] Since the operation intervals of subway turnouts are short and the overall changes are not large, the operation load characteristics can be regarded as a time series. The equally spaced prediction values ​​of the sampling sequence can be used as a supplementary description of the continuous prediction values ​​of the original sequence. The prediction results with different descriptions of the information span are intermittently complementary, which helps to regulate the prediction trend.

[0065] The weighted prediction sequence {r f} is input, and a set of weight sequences {k i ,k i +b,...,k i +(n-1)b}, and get the compensation weight matrix K n×2l It is used to construct a prediction feature sequence with a changing value at each step to form a prediction feature matrix, which is expected to be within a small range or as close to the true value as possible. That is, given a weight domain B, the weighted single point prediction value is r, the single point real value is V, and the prediction feature is constructed in K n×2l The loss function is:

[0066]

[0067] In the formula Based on the KL divergence modification, F T and F P are the discrete probability distributions of the true sequence and the predicted feature under the same sample value interval distribution Ω, and λ is the conditional parameter. Covering Q samples, the learning goal is to obtain the compensation weight matrix that minimizes the total loss

[0068]

[0069] In summary, the multi-description complementary prediction machine proposed in this embodiment is characterized by transforming the existing sequence single-point deterministic prediction into a reliable sequence single-point value range prediction, thereby improving the adaptability of the prediction; at the same time, it can expand the prediction features that vary within the expected range to estimate the switch operation load state in subsequent time periods.

[0070] In this embodiment, the base forecast period is selected and optimized by the following method:

[0071] The standard grey model DGM is used as the base predictor, and its standard structure is improved and optimized, including:

[0072] The weight parameter w and translation parameter c are introduced to weight and translate the original sequence in pairs to obtain the intermediate sequence, and then accumulate to generate the prediction input sequence, which improves the prediction accuracy and generalization ability from the two perspectives of sequence smoothing and adjustment level ratio. Assume w∈W, c∈C, use each group (w, c) in the value space to pre-model and predict the original sequence, and calculate the root mean square error (RMSE) between the original sequence and the corresponding simulated prediction sequence. (w,c) , Fréchet distance FD (w,c) , establish the binary consistency constraint equation:

[0073]

[0074] Where e r 、e dThe root mean square error and Fréchet distance between the simulated prediction sequence and the original sequence obtained by standard DGM modeling are respectively, and the parameter group that minimizes the target E(w,c) by traversing (w,c) values ​​is:

[0075]

[0076] Then, the pre-optimized parameters (w0, c0) are substituted into the prediction model to achieve optimized prediction.

[0077] S300. Using the sample matrix G of the action load characteristics of the subsequent time period as a sample, based on the adaptive kernel density estimation method, the probability density function and the overall value confidence interval of the load characteristics of the forecast period are calculated; in this embodiment, the method for calculating the probability density function and the overall value confidence interval of the load characteristics of the forecast period is:

[0078] The predicted feature matrix element is recorded as {g i}, the number of elements 2nl = M, the one-dimensional kernel density estimation function is as follows:

[0079]

[0080] Where, is the kernel function, and the Gaussian kernel is selected as the representation below; h represents the bandwidth, and the variable value h is determined adaptively according to the sample position. i Instead; therefore, the adaptive Gaussian kernel density estimation function can be obtained:

[0081]

[0082] where h i Use the asymptotically integrated mean square error method to obtain the optimal solution:

[0083]

[0084] In the formula Represents the optimal fixed window width dominated by the sample standard deviation σ; ω is a parameter; further, the overall confidence interval [c1, c2] of the prediction space under a given confidence level P can be obtained:

[0085]

[0086] S400. Calculate the overload degree S of the switch action during the forecast period and use it as an early warning evaluation indicator. In this embodiment, the method for calculating the overload degree S of the switch action during the forecast period is:

[0087] Collect a large number of historical motion load feature samples and use the method in step 3 to calculate their confidence interval [z1, z2] at the same confidence level as the representation of the empirical level of motion load features;

[0088] Combining the probability density function and confidence interval of the prediction feature matrix, the action overload degree S during the prediction period is defined and calculated as follows:

[0089]

[0090] The closer S is to 1, the greater the probability that the overall predicted level exceeds the upper limit of the historical experience level, the greater the exceedance, and the greater the possibility that the turnout will fail due to overload operation in the short term.

[0091] S500. Using overload fault and non-fault cases to perform a preference search, determine the warning threshold t, and make a warning judgment based on the warning threshold t and the overload degree S of the switch action during the predicted period;

[0092] In S500 of this embodiment, the given warning threshold is t (0<t≤1). This embodiment performs two aspects of warning judgment:

[0093] 1) Determine whether to issue an early warning. If the overload degree S calculated in the previous step is ≥ t, it is considered that the turnout may experience an overload operation failure in the subsequent period and an early warning is required.

[0094] 2) If an early warning is issued, the specific action period and action sequence of the subsequent failure are further inferred (i.e., accurate early warning). The overload degree defined and calculated by the method of the present invention can be regarded as the probability of the switch overload action occurring within the predicted action steps. Therefore, the early warning threshold t means that within the predicted period, when the switch overload action rate is above t, it is easy to cause a failure. Assuming that the predicted action steps are T, the action sequence T of the failure can be inferred. f It should be between the tTth and T-1st turnout actions after the first warning is issued (the action sequence of the first warning is considered as the first possible overload action and is included). Since the action sequence is an integer, according to probability knowledge, we can get:

[0095]

[0096] Where, Indicates rounding down, S f It indicates the overload degree when the first warning is issued. Obviously, it is positive only when tT ≥ 2. The value of tT can be used to test the rationality of the prediction step setting and warning threshold.

[0097] In the judgment of S500, the size of the warning threshold has an important impact on the performance of the method. Based on the actual application, the present invention comprehensively considers the following three aspects of warning performance: warning success rate F s (ratio of warnings issued on fault days), false alarm rate F f (ratio of warnings issued on non-fault days), accurate warning rate F a(The ratio of the actual fault action time being within the inferred fault action time period). s and F f They can be considered together as a description of the accuracy of early warning and taken into consideration.

[0098] In order to better balance the above-mentioned warning performance and avoid the uncertainty caused by subjective setting of empirical thresholds, the present invention optimizes the warning threshold based on known fault cases and non-fault case data. The method is as follows:

[0099] Take N1 data samples from the day of an overload action failure and N2 data samples from a non-fault day. Perform a simulated early warning test on each sample using the method of the present invention. Stop once an early warning is issued. The number of missed alarms obtained from the test is n1, the number of false alarms is n2, and the number of accurate early warnings is n3. The following biased optimization objective function is established:

[0100]

[0101] The optimization objective is to achieve both correct and precise warnings. α is a weight, and the goal is to find the warning threshold that maximizes the objective function. Users can adjust α based on their needs, choosing whether to prioritize correct or precise warnings.

[0102] Further focus on the "correct warning part", the optimization result of this part is expected to be F s As large as possible and F f As small as possible. Then let the total number of warning errors E = n1 + n2, the sample ratio β = N1 / N2, and substitute the "correct warning part" to get:

[0103]

[0104] It is easy to analyze the monotonicity of the above formula. Assuming that the number of samples is fixed, different sample ratios β can make the optimization process have preferences, that is:

[0105] When β<1, the optimization goal is to minimize the total error and minimize the missed alarms, while a certain amount of false alarms is allowed.

[0106] When β>1, the optimization goal is to minimize the total error and minimize the false alarm, while a certain amount of missed alarms is allowed.

[0107] When β=1, the optimization goal is to minimize the total error and have a neutral preference.

[0108] In summary, the weight parameter α and the ratio of faulty / non-faulty samples fed into the system, β, are defined as two preference factors of the objective function. Users can adjust these preference factors to optimize the warning threshold, ensuring that warning performance is biased and meets actual application requirements. From a safety perspective, users tend to prioritize warning accuracy over accuracy. Regarding warning accuracy, sites prefer to maximize the warning success rate (i.e., reduce missed warnings), allowing for a certain amount of false alarms.

[0109] S600. Infer the fault occurrence time based on the current driving interval and the warning judgment result. According to the warning judgment result of S500, two types of warning feedback information can be finally formed:

[0110] 1. There is no risk of overload failure during the current forecast period. The system does not need to issue an early warning and continues to perform real-time early warning calculations with the next turnout operation. Only the overload level is fed back to the user, without inferring the time of failure.

[0111] 2. There is a risk of overload failure in the turnout during the current forecast period. The system needs to issue a fault warning report as soon as possible. Based on the content of the previous step and the current driving interval H, the specific time range for the fault to occur is inferred to be The occurrence time is approximately Users can take emergency measures in advance based on early warning information and on-site conditions.

[0112] Compared to existing technologies, this embodiment primarily adopts a two-step approach: "feature prediction - fault warning / prediction judgment." This embodiment follows a four-stage approach: "feature prediction - state assessment - fault prediction - precise warning." This approach uses the currently continuously generated turnout load characteristics as input to predict the load characteristics for a certain number of steps in the future. The predicted load characteristics are then used as samples to establish a model to assess the turnout state for subsequent short-term periods. Finally, quantitative warning judgments are made based on state assessment indicators and reference thresholds. If a warning is issued, the time of subsequent fault occurrence is further inferred. This method establishes a machine learning process based on data samples for multiple stages of the main line. It uses innovative prediction methods to obtain more and more accurate prediction features, designs a quantitative assessment model to calculate more appropriate state expressions, and employs a biased optimization of warning thresholds to balance multiple aspects of warning performance.

[0113] It should be understood that the specific order or hierarchy of steps in the disclosed processes is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process can be rearranged without departing from the scope of the present disclosure. The accompanying method claims present elements of the various steps in an exemplary order and are not intended to be limited to the specific order or hierarchy described.

[0114] In the foregoing detailed description, various features are grouped together in a single embodiment to simplify the disclosure. This method of disclosure should not be interpreted as reflecting an intention that embodiments of the claimed subject matter require more features than are expressly recited in each claim. On the contrary, as reflected in the appended claims, the invention comprises less than all the features of any individual disclosed embodiment. The appended claims are therefore hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

[0115] Those skilled in the art will also appreciate that the various illustrative logic blocks, modules, circuits, and algorithmic steps described in conjunction with the embodiments herein may be implemented as electronic hardware, computer software, or a combination thereof. In order to clearly illustrate the interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps described above are generally described around their functions. Whether such functions are implemented as hardware or software depends on the specific application and the design constraints imposed on the entire system. A skilled person may implement the described functions in an adaptable manner for each specific application, but such implementation decisions should not be interpreted as departing from the scope of protection of this disclosure.

[0116] The steps of the methods or algorithms described in conjunction with the embodiments herein may be directly embodied as hardware, software modules executed by a processor, or a combination thereof. The software module may be located in a RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, register, hard disk, removable disk, CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is connected to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium may also be an integral part of the processor. The processor and storage medium may be located in an ASIC. The ASIC may be located in a user terminal. Of course, the processor and storage medium may also be present in a user terminal as discrete components.

[0117] For software implementation, the techniques described in this application can be implemented using modules (e.g., procedures, functions, etc.) that perform the functions described in this application. These software codes can be stored in a memory unit and executed by a processor. The memory unit can be implemented within the processor or external to the processor. In the latter case, it is communicatively coupled to the processor via various means, which are well known in the art.

[0118] The foregoing description includes examples of one or more embodiments. Of course, it is not possible to describe all possible combinations of components or methods for the purposes of describing the above embodiments, but one of ordinary skill in the art will recognize that the various embodiments may be further combined and arranged. Therefore, the embodiments described herein are intended to encompass all such changes, modifications and variations that fall within the scope of the appended claims. Furthermore, to the extent the term "comprising" is used in the specification or claims, the term is intended to be encompassed in a manner similar to the term "including," as explained in terms of "including," used as a transitional word in the claims. Furthermore, any use of the term "or" in the specification of the claims is intended to mean a "non-exclusive or."

Claims

1. A real-time accurate early warning method for subway turnout overload failure, characterized by: include: S100 obtains the cumulative continuous turnout power curve when the i-th turnout overload action is performed, and calculates the real-time action load characteristic sequence {R}; S200. Using a multi-description complementary prediction engine, the real-time action load feature sequence {R} is multi-step value range predicted to construct a subsequent short-term action load feature prediction sample matrix G; In S200, the method for predicting the motion load feature prediction sample matrix G for the subsequent time period is as follows: perform odd and even sampling on the feature sequence {R} respectively to obtain two groups of subsequences, use the basis predictor to predict the original sequence to obtain {r(i)|i=1,2,...,2l}, and then use the basis predictor to predict the two groups of subsequences to obtain {r′(i)|i=1,2,...,l}, {r″(i)|i=1,2,...,l}, and give the weight sequence {k1,k2,...,k 2l }, the final prediction sequence {r f The single point prediction value at each step can be obtained by interweaving and weighting r, r′, and r″; then the weight k of each step is i All learn to optimize and obtain n changing values ​​to construct the optimal compensation weight matrix Final use Expand the original prediction sequence {rf} of length 2l into an n×2l action load feature prediction sample matrix G; For the prediction sequence {r f } is expanded into the action load feature prediction sample matrix G The way to obtain is: Among them, Q is the number of learning samples covered, loss(K n×2l ) is the prediction feature in K n×2l The loss function under loss(K n×2l The calculation method is as follows: take the weighted prediction sequence {rf} as input, use the known sequence sample set to train and learn a set of weight sequences {k} with a change step size of b for each single point prediction value. i ,k i +b,...,k i +(n-1)b}, and get the compensation weight matrix K n×2l It is used to construct a prediction feature sequence with a changing value at each step to form a prediction sample matrix, which is expected to be within a small range or as close to the true value as possible; that is, given a weight domain B, the weighted single point prediction value is r, the single point real value is V, and the prediction feature is constructed in K n×2l The loss function is: In the formula Based on the KL divergence modification, F T and F P are the discrete probability distributions of the true sequence and the predicted feature under the same sample value interval distribution Ω, and λ is the conditional parameter; S300. Taking the subsequent time period action load feature prediction sample matrix G as the sample, based on the adaptive kernel density estimation method, calculate the probability density function and the overall value confidence interval of the load feature of the prediction period; The method for calculating the probability density function and overall confidence interval of the load characteristics during the forecast period is: The predicted sample matrix elements are recorded as {g i }, the number of elements 2nl = M, the one-dimensional kernel density estimation function is as follows: Where, is the kernel function, and the Gaussian kernel is selected as the representation below; h represents the bandwidth, and the variable value h is determined adaptively according to the sample position. i Instead; therefore, the adaptive Gaussian kernel density estimation function can be obtained: where h i Use the asymptotically integrated mean square error method to obtain the optimal solution: In the formula represents the optimal fixed window width dominated by the sample standard deviation σ; ω is a parameter; Furthermore, we can obtain the overall confidence interval [c1, c2] of the prediction space at a given confidence level P: S400. Calculate the overload degree S of the switch action during the forecast period and use it as an early warning evaluation indicator; The method for calculating the overload degree S of the turnout operation during the forecast period is: Collect a large number of historical motion load feature samples and use S300 to calculate their value confidence interval [z1, z2] at the same confidence level as the representation of the experience level of motion load features; Combining the probability density function and confidence interval of the prediction sample matrix, the action overload degree S of the prediction period is defined and calculated by the following formula: The closer S is to 1, the greater the probability that the overall prediction level exceeds the upper limit of the historical experience level, the greater the excess amount, and the greater the possibility that the turnout will fail due to overload operation in the short term; S500. Using overload fault and non-fault data to perform a preference search, determine the warning threshold t, and make a warning judgment based on the warning threshold t and the overload degree S of the switch action during the forecast period; S600. Infer the fault occurrence time based on the current driving interval and the warning judgment result.

2. A real-time accurate early warning method for subway turnout overload failure according to claim 1, characterized in that: In S100, the method for calculating the real-time action load characteristic sequence {R} includes: taking the current i-th action moment as a reference, taking the previous i-2L turnout action power curves in the same direction, calculating the root mean square value of the sampling point as the action load characteristic, and obtaining the turnout action load characteristic sequence {R i |i=1,2,3...,2L}, 2L represents the sequence length.

3. A real-time accurate early warning method for subway turnout overload failure according to claim 1, characterized in that: Prediction sequence {r f The calculation formula for} is: Among them, {r(i)|i=1,2,...,2l} is obtained by using the basis predictor to predict the original sequence, {r′(i)|i=1,2,...,l} and {r″(i)|i=1,2,...,l} are obtained by using the basis predictor to predict the two groups of sampling sequences, and {k1,k2,...,k 2l } is a given weight sequence.

4. A real-time accurate early warning method for subway turnout overload failure according to claim 1, characterized in that: The base predictor is selected and optimized by: The standard grey model DGM is used as the base predictor, and its standard structure is improved and optimized, including: The weight parameter w and translation parameter c are introduced to weight and translate the original sequence in pairs to obtain the intermediate sequence, and then accumulate to generate the prediction input sequence, which improves the prediction accuracy and generalization ability from the two perspectives of sequence smoothing and adjustment level ratio. Assume w∈W, c∈C, use each group (w, c) in the value space to pre-model and predict the original sequence, and calculate the root mean square error (RMSE) between the original sequence and the corresponding simulated prediction sequence. (w,c) , Fréchet distance FD (w,c) , establish the binary consistency constraint equation: Where e r 、e d The root mean square error and Fréchet distance between the simulated prediction sequence and the original sequence obtained by standard DGM modeling are respectively, and the parameter group that minimizes the target E(w,c) by traversing (w,c) values ​​is: Then, the pre-optimized parameters (w0, c0) are substituted into the prediction model to achieve optimized prediction.

5. A real-time accurate early warning method for subway turnout overload failure according to claim 1, characterized in that: The specific method of S500 includes: making a warning judgment based on the warning threshold value obtained by optimizing the two-aspect preference, first judging whether to issue a warning based on the warning threshold value t and the overload degree S of the switch action in the predicted period, if the overload degree S calculated in the previous step is greater than or equal to t, it is considered that the switch may have an overload action failure in the subsequent period and a warning is required; if a warning is issued, the specific action period and action sequence of the subsequent failure are further inferred.

Images