Railway locomotive diesel engine adaptive health diagnosis system based on human-in-the-loop

Abstract

The present invention relates to the technical field of railway locomotive diesel engine adaptive health diagnosis systems, in particular to a railway locomotive diesel engine adaptive health diagnosis system based on human-in-the-loop. The system performs algorithm classification on time sequence signals collected by sensors deployed on a locomotive diesel engine, and comprises four modules a working condition segmentation module, a deduplication detection module, a model training module and an anomaly detection model module, and two steps an offline training process and an online inference process, so as to complete identification and classification of features of anomaly signals, thereby satisfying the requirements of operation and maintenance engineers for health diagnosis and monitoring of health states of subsystems. In addition, the present invention can associate anomaly signals with original sensor data to conveniently check fault reasons, and for faults with various influence factors and unknown reasons, the accuracy of health diagnosis after model construction is completed can be higher. The present invention has the possibility of the health diagnosis model evolving from a fault identification function to a fault classification function, thereby reserving a direction of development for higher requirements of followup intelligent operation and maintenance.

Description

An adaptive health diagnosis system for railway locomotive diesel engines based on human-in-the-loop Technical Field

[0001] The present invention relates to the technical field of railway locomotive diesel engine adaptive health diagnosis system, in particular to a railway locomotive diesel engine adaptive health diagnosis system based on man-in-the-loop. Background Art

[0002] With the increasing demand for digitalization in China's rail transit industry and the gradual popularization and rapid development of big data technology, the demand for intelligent operation and maintenance technology for mainline locomotives and their major components has also increased significantly. Therefore, the use of big data algorithms to replace manual fault detection of product operation data and then build a locomotive diesel engine health diagnosis model improves the efficiency and scientificity of fault diagnosis while reducing the manpower and time costs required for detection.

[0003] There are several problems with building intelligent health diagnosis for locomotive diesel engines: First, the influencing factors of various types of diesel engine faults are complex, and the attribution of each fault sample may be different. When cleaning and classifying the sampled data, the dimension is easily too high, resulting in the algorithm being unable to converge; second, my country has a vast territory, and the application environment of mainline locomotive diesel engines is complex and changeable. It is difficult for a single algorithm to achieve the evolution of adaptive self-learning functions; third, although only fault identification functions are being developed and researched at the current stage, with the continuous improvement of railway customers' demand for intelligence, the possibility of evolving fault classification / cause-finding functions must also be considered in the future.

[0004] Predicting or monitoring early signs of failure through time series data collected by sensors is an important foundation for the control management and predictive maintenance of complex systems. Failure precursors generally refer to early symptoms of impending failure in complex systems. Failure and fault precursor signals are essentially abnormal compared to the signals of a healthy complex system. The main difference lies in the extent to which they affect the "normal" operation of the system. When a fault precursor appears, although the performance of the component or subsystem has declined and the time series signal involved has "small" changes, the system will generally still operate normally. Over time, the abnormality of the component or subsystem will propagate to more related upstream and downstream components and subsystems, triggering anomalies that are observable from the time series data.

[0005] In the existing technology, although it is possible to achieve early warning of the status. However, in the technology mentioned above, when issuing an early warning, it is necessary to first simulate and predict the digital twin model based on the characteristics of the state quantity time series, and then update the digital twin model based on the simulation prediction data, and then compare the characteristics of the updated digital twin model with the characteristics of the initial digital twin model to issue an early warning. Therefore, the accuracy of the early warning is often determined by the accuracy of the updated digital twin model, and the updated digital twin model is predicted based on the data of the initial digital twin model. In other words, the accuracy of the early warning will depend on the data initially collected. Whether it is a digital twin model based on a mechanism model, a digital drive model or a hybrid model, it has the following shortcomings:

[0006] 1. Learning multi-dimensional heterogeneous time series data with temporal dynamics and intercorrelations is very difficult;

[0007] 2. The dataset lacks fault samples (currently there are no fault samples), not to mention accurate annotation of fault precursors;

[0008] 3. Purely data-driven models require a large amount of data, but the dataset only contains sample data of the prototype vehicle under limited operating conditions;

[0009] 4. There are certain differences between the engine operating condition distribution and actual operation.

[0010] Summary of the Invention

[0011] In order to overcome the existing deficiencies, the present invention provides a railway locomotive diesel engine adaptive health diagnosis system based on man-in-the-loop.

[0012] The technical solution adopted by the present invention to solve its technical problems is: a railway locomotive diesel engine adaptive health diagnosis system based on human in the loop, including four modules: working condition segmentation module, deduplication detection module, model training module, and anomaly detection model module, and two steps: offline training process and online reasoning process; the offline training process includes parallel pipeline 1 and pipeline 2; the pipeline 1 divides the time series data into different time series segments through the working condition segmentation module, eliminates duplicate data segments from the above time series segments through the deduplication detection module, generates digital signatures for the new samples after deduplication detection, writes them into the Redis database, and generates new samples and writes them into the MYSQL database; the pipeline 2 imports the new samples from the MYSQL database of pipeline 1, drives the model training module to cycle, and writes the model files into the Redis database. The folder is saved in the local path of the model and the model version information is written to the MYSQL database; the online reasoning process includes parallel pipelines three and four; the pipeline three divides the time series data into different time series segments through the working condition segmentation module, calculates the anomaly score of the above time series segments through the anomaly detection model module, stores the generated anomaly score in the MYSQL database and compares it with the artificial pre-defined threshold in the MYSQL database to determine whether there is an anomaly, and writes the anomaly judgment into the MYSQL database; the pipeline four divides the working condition segmentation module in pipeline three into different time series segments and eliminates duplicate data segments through the deduplication detection module, generates a digital signature for the new sample after deduplication detection, writes it into the Redis database, and generates a new sample and writes it into the MYSQL database.

[0013] According to another embodiment of the present invention, the operating condition segmentation module further includes dividing the diesel engine into different handle gears in different operating modes, and segmenting the time series data into data under each gear according to the above handle gear data; for the time series data in the current gear interval, a statistical sliding window is applied to the actual power of the diesel engine starting from the leftmost side of the interval; if the power mean value in the current sliding window falls within the statistical interval of ±95% of the mean value of the statistical window after a specific time T, it is considered that the engine has entered a steady-state interval, and the sliding window at the current moment is moved backward by time T, and the above process is repeated until the power mean value in the sliding window at the current moment exceeds the statistical interval of the sliding window after time T or the current sliding window has reached the end of the current gear interval; if the power mean value in the current sliding window does not fall within the statistical interval of ±95% of the mean value of the statistical window after the specific time T, it is considered that the engine has not yet entered a steady-state interval, and the sliding window at the current moment is moved backward by one time, and the above process is repeated until the end of the current gear interval.

[0014] According to another embodiment of the present invention, the anomaly detection model module further includes algorithm processing, anomaly score and anomaly judgment; the algorithm processing is to call the anomaly detection model to obtain the time series segment X after the working condition is segmented. i Extract feature Y i, and use the autoencoder in the model to reconstruct the data fragments to generate algorithm features After splicing, we get a vector signal The anomaly score maps the vector signal q to a sphere with a radius of 1, calculates its distance from the normal signal mapping value in the MYSQL database, and writes the generated anomaly score to the MYSQL database; the anomaly judgment compares the generated anomaly score with a manually pre-defined threshold in the MYSQL database to determine whether there is an anomaly.

[0015] According to another embodiment of the present invention, the anomaly detection model in the anomaly detection model module and the model in the model training module are both composed of a feature extractor, a time series encoder, a time series decoder and a projection layer; the feature extractor extracts the time series segment X through a one-dimensional convolutional network. i Feature Y in i The temporal encoder and temporal decoder extract the spatiotemporal correlation feature Z through a recurrent neural network i , the reconstructed signal is The Y i and After being mapped to the high-dimensional sphere surface of the projection layer, they are respectively vector v i and vector Then the vector The goal of the loss function is to maximize the vector v i ,vector and vector c e The cosine similarity of

[0016] According to another embodiment of the present invention, the anomaly score is further included. The result is compared with the pre-defined threshold in the MYSQL database. Let the threshold be c. If the value is greater than c, it is abnormal, and if it is less than c, it is normal.

[0017] According to another embodiment of the present invention, the deduplication detection module divides the working condition into time series segments, maps each channel of the multidimensional time series data into a string using the SAX algorithm, and then reconstructs it into equal-length cross-dimensional phrases, compares the similarity with the historical data set and excludes duplicate data segments according to a manually preset threshold, generates a digital signature for the new sample after deduplication detection through a hash algorithm, writes it into the Redis database, and generates a new sample and writes it into the MYSQL database.

[0018] The beneficial effect of the present invention is that a human-in-the-loop mode is introduced in the process of developing the health diagnosis algorithm, and the identification and classification of abnormal signal features are completed through the combination of big data algorithms and manual calibration of the results. It mainly relies on human intelligence to help machines become more intelligent. It takes the form of sampling to label representative or confusing samples and then feeds them back to the model so that it can adjust the iteration according to the feedback. Although this method is more difficult to build a model than previous methods, and the manual workload and completion cycle of the later model training are greatly increased, the accuracy of health diagnosis is higher for faults with many influencing factors and unknown causes after the model is built; the most important thing is that this method has the possibility of evolving the health diagnosis model fault identification to the fault classification function, which reserves a development direction for the higher requirements of subsequent intelligent operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below with reference to the accompanying drawings and examples.

[0020] FIG1 is a schematic flow diagram of the present invention;

[0021] Figure 2: Distribution of actual engine power and number of handle gears at different engine gears;

[0022] Figure 3 is a curve showing the change of the handle gear position and the actual power of the diesel engine over time;

[0023] Figure 4 shows the curve changes of the handle gear position and the actual power of the diesel engine during upshifting and downshifting;

[0024] FIG5 is a schematic diagram of the diesel engine steady-state data segmentation logic;

[0025] Figure 6 is a diagram of the sample deduplication and new sample labeling pipeline;

[0026] Figure 7 is a diagram of the model structure. DETAILED DESCRIPTION

[0027] Figure 1 is a structural diagram of the present invention, a railway locomotive diesel engine adaptive health diagnosis system based on human-in-the-loop, including four modules: working condition segmentation module, deduplication detection module, model training module, and anomaly detection model module, as well as two steps: offline training process and online reasoning process; the offline training process includes parallel pipelines 1 and 2; the pipeline 1 divides the time series data into different time series segments through the working condition segmentation module, eliminates duplicate data segments from the above time series segments through the deduplication detection module, generates digital signatures for the new samples after deduplication detection, writes them into the Redis database, and generates new samples and writes them into the MYSQL database; the pipeline 2 imports the new samples of the pipeline 1 MYSQL database, drives the model training module to loop, and saves the model folder There is a local path for the model and the model version information is written to the MYSQL database; the online reasoning process includes parallel pipelines three and four; the pipeline three divides the time series data into different time series segments through the working condition segmentation module, calculates the anomaly score of the above time series segments through the anomaly detection model module, stores the generated anomaly score in the MYSQL database and compares it with the artificial pre-defined threshold in the MYSQL database to determine whether there is an anomaly, and writes the anomaly judgment into the MYSQL database; the pipeline four divides the working condition segmentation module in pipeline three into different time series segments and eliminates duplicate data segments through the deduplication detection module, generates a digital signature for the new sample after deduplication detection, writes it into the Redis database, and generates a new sample and writes it into the MYSQL database.

[0028] According to another embodiment of the present invention, the operating condition segmentation module further includes dividing the diesel engine into different handle gears in different operating modes, and segmenting the time series data into data under each gear according to the above handle gear data; for the time series data in the current gear interval, a statistical sliding window is applied to the actual power of the diesel engine starting from the leftmost side of the interval; if the power mean value in the current sliding window falls within the statistical interval of ±95% of the mean value of the statistical window after a specific time T, it is considered that the engine has entered a steady-state interval, and the sliding window at the current moment is moved backward by time T, and the above process is repeated until the power mean value in the sliding window at the current moment exceeds the statistical interval of the sliding window after time T or the current sliding window has reached the end of the current gear interval; if the power mean value in the current sliding window does not fall within the statistical interval of ±95% of the mean value of the statistical window after the specific time T, it is considered that the engine has not yet entered a steady-state interval, and the sliding window at the current moment is moved backward by one time, and the above process is repeated until the end of the current gear interval.

[0029] Specifically, as the energy source for electric-driven internal combustion locomotives, diesel engines output mechanical energy, which in turn drives the main generator to generate electricity and excite the main generator. This electrical energy supplies energy to the traction and auxiliary systems. Therefore, engine power can be considered equal to the sum of traction power, auxiliary system power, and main generator excitation power. The engine's key states, which this project focuses on, are closely related to factors such as the engine's operating load, operating mode, and external environment. Therefore, it is necessary to compile the engine's operating data under different operating conditions as input for subsequent algorithm design. The most basic operating mode division of a diesel engine is the different handle gear positions under different locomotive conditions, which correspond to different engine operating gears and output powers.

[0030] Table 1 shows the possible gears and working modes of the engine under different locomotive conditions:

[0031] It can be seen that the engine gear responds differently to the locomotive control handle position under different locomotive states. The diesel engine's operating state varies with the locomotive state and handle position. Even under the same locomotive state and handle position, such as traction gear 12, the diesel engine's actual output power is affected by factors such as auxiliary system power fluctuations and traction power fluctuations caused by the line. Therefore, under certain operating conditions, when the diesel engine output power enters a "steady state" (here, steady state refers to the situation where the diesel engine's actual power fluctuations are relatively small, as opposed to the situation where the power fluctuates significantly during engine shifting), the sliding mean of its actual output power follows the central limit theorem, and its distribution is close to a normal distribution, as shown in Figure 2. The core concern of this invention is the quantitative assessment of the degree of abnormality in the operation of the diesel engine and its related subsystems. The diesel engine power directly reflects its own operating state. However, under non-steady-state operating conditions such as gear switching and power boosting, the time series data related to each item is also in a non-steady state, making it difficult to extract effective data features from it to reflect the current health status of the diesel engine. Therefore, the subsequent focus is on the diesel engine's time series data samples in the steady state as input for evaluating its health status.

[0032] Figure 3 depicts the temporal relationship between locomotive handle position and diesel engine actual power. It can be seen that actual power is positively correlated with handle position under traction conditions. Under resistance braking conditions, the diesel engine actual power is very low, providing only excitation power, consistent with the operating condition table described above. During each gear position, in most cases, the diesel engine actual power remains essentially stable, fluctuating only within a relatively small range. As shown in Figure 4, during the initial period of upshifting or downshifting into a specific gear, the diesel engine power rapidly increases or decreases to the actual power range for the corresponding gear and stabilizes. This steady state persists until the end of the gear position. During this period, the diesel engine actual power may fluctuate (the source of this fluctuation has been described previously), but in most cases, this fluctuation is negligible relative to the engine's actual power (as seen in the 10-gear cycle on the left side of Figure 3). However, in some cases, the fluctuation is significant, as seen in the power fluctuations during the 12th gear position in the middle of Figure 3. In contrast, the power during the 10th gear position on the right side of Figure 4 never reaches a steady state. To identify and isolate steady-state diesel engine operating samples, we designed a steady-state sample segmentation logic based on the engine's actual power and the gear position, as shown in Figure 5. When an upshift occurs, the power curve rises rapidly and then gradually stabilizes, entering the first steady-state sample interval. After a period of time, the actual power increases slightly and stabilizes again until the next upshift. Based on the above and Figure 6, the main logic for steady-state sample segmentation is divided into the following three steps:

[0033] 1. Divide the time series data into data under each gear according to the handle gear data;

[0034] 2. For the time series data within the current gear range, a statistical sliding window is applied to the actual diesel engine power starting from the leftmost side of the range. If the power mean value within the current sliding window falls within the ±95% statistical interval of the mean value of the statistical window after a specific time T, it is considered to have entered the steady-state range. The current sliding window is then moved back by time T, and the above process is repeated until the power mean value within the current sliding window exceeds the sliding window statistical interval after time T or the current sliding window has reached the end of the current gear range.

[0035] 3. If the power mean in the current sliding window does not fall within the ±95% statistical interval of the mean of the statistical window after the specific time T, it is considered that the steady-state interval has not yet been entered. The current sliding window is moved back by one time and the above process is repeated until the end of the current gear interval.

[0036] Through the above steps, the division of steady-state samples is based on local statistical information. Compared with global information as the division threshold, its advantage is better stability and the ability to fully collect steady-state data samples in different forms. During the same gear period, the actual power of the diesel engine may be affected by multiple factors and produce power fluctuations. The above method can exclude the fluctuating part and only identify the steady-state section. Therefore, multiple steady-state samples may be identified during one gear period. A minimum duration limit for steady-state samples can be imposed to improve the quality of sample division. Each sample has corresponding statistical information such as gear, power mean and variance as sample meta-information. This information can be used as supplementary input besides time series data.

[0037] According to another embodiment of the present invention, the anomaly detection model module further includes algorithm processing, anomaly score and anomaly judgment; the algorithm processing is to call the anomaly detection model to obtain the time series segment X after the working condition is segmented. i Extract feature Y i , and use the autoencoder in the model to reconstruct the data fragments to generate algorithm features After splicing, we get a vector signal The anomaly score maps the vector signal q to a sphere with a radius of 1, calculates its distance from the normal signal mapping value in the MYSQL database, and writes the generated anomaly score to the MYSQL database; the anomaly judgment compares the generated anomaly score with a manually pre-defined threshold in the MYSQL database to determine whether there is an anomaly.

[0038] Specifically, in the field of data-driven anomaly detection, with the development of deep learning technology, network-based representation learning is often used to extract features of a system's "normal" state. These representation learning methods uniformly rely on many assumptions: automatic prediction (AE) and generative adversarial networks (GANs) assume that normal samples are easier to reconstruct or generate from the latent space than abnormal samples; single-category classification assumes that normal samples can be described by a single distribution. However, these overly strong assumptions make it difficult to cope with the various anomalies in multidimensional time series data. Data augmentation and negative samples are the most commonly used methods in contrastive learning and representation learning. However, time series augmentation easily introduces noise, and our dataset only contains positive samples, not negative ones. To address these issues: a feature extractor is used to smooth out small differences between samples; a temporal encoder and a temporal decoder are used to generate positive samples; representation learning and the previous two modules are jointly trained for anomaly detection; and if negative samples are present, adding them to the subsequent adaptive training should further improve model performance.

[0039] According to another embodiment of the present invention, the anomaly detection model in the anomaly detection model module and the model in the model training module are both composed of a feature extractor, a time series encoder, a time series decoder, and a projection layer; as shown in FIG7 , the feature extractor extracts the time series segment X through a one-dimensional convolutional network. i Feature Y in i The temporal encoder and temporal decoder extract the spatiotemporal correlation feature Z through a recurrent neural network i , the reconstructed signal is The Y i and After being mapped to the high-dimensional sphere surface of the projection layer, they are respectively vector v i and vector Then the vector The goal of the loss function is to maximize the vector v i ,vector and vector c e The cosine similarity of

[0040] According to another embodiment of the present invention, the anomaly score is further included. The result is compared with the pre-defined threshold in the MYSQL database. Let the threshold be c. If the value is greater than c, it is abnormal, and if it is less than c, it is normal.

[0041] According to another embodiment of the present invention, the deduplication detection module divides the working condition into time series segments, maps each channel of the multidimensional time series data into a string using the SAX algorithm, and then reconstructs it into equal-length cross-dimensional phrases, compares the similarity with the historical data set and excludes duplicate data segments according to a manually preset threshold, generates a digital signature for the new sample after deduplication detection through a hash algorithm, writes it into the Redis database, and generates a new sample and writes it into the MYSQL database.

[0042] Specifically, as shown in Figure 2, the engine power distribution between 0 and 3500 kW is uneven. There are numerous samples at low power levels, but only a few at high power levels. This uneven distribution of samples can easily lead the algorithm to learn relevant features for samples in the low-power cluster, weakening the feature extraction efficiency for samples in the high-power cluster. Since the actual engine power consists of traction power and auxiliary system power, the samples between power clusters are relatively sparse. During real-world locomotive operation, the power cluster distribution may shift to the right (due to more carriages or increased power demand from auxiliary systems). This can lead to concept drift between the input data samples and the training sample distribution after deployment in a real-world vehicle operating environment. In this case, the output performance of data-driven algorithms cannot be guaranteed. Algorithms such as DTW, which calculate time series similarity, have high computational complexity and are difficult to perform for sample deduplication using similarity calculations on massive multi-dimensional time series data. To this end, we convert time series samples into strings and use the Simhash and Minhash algorithms to generate digital signatures. Finally, we use the Hamming distance to measure the distance between samples.

[0043] As shown in Figure 6, the process of deduplication and new sample marking is as follows:

[0044] First, set up a time series dataset where X i is a time series X of length L i ={x1, x2, ..., xL}, X j =R d is a d-dimensional vector. The essence of sample deduplication is to compare and i X j Similarity, Sim(X i ,X j )>r, where r is the similarity threshold.

[0045] Step 1: Use SAX algorithm to transform multidimensional time series data X i Each channel of is mapped to a string;

[0046] Step 2: Assuming a sliding window length of W, define the cross-channel bag of words within the sliding window. If there are three signals, the box represents the first word cad in the first sliding window. If the sliding window length is 3 and the stride is 3, then sliding window 1 is compressed into the sentence "cad bbb ccb."

[0047] Step 3, X i Can be compressed into one text based on multiple sliding windows;

[0048] Step 4: Use Simhash or Minhash algorithm to generate a digital signature and write it into the Redis database;

[0049] Step 5: Calculate the distance between samples and cluster them.

[0050] Specifically, during the offline training process, during a cold start, pipeline one splits the steady-state data, removes duplicate samples, and writes the processed training samples to a MySQL database. Pipeline two trains the model offline, saves the model file to a local path, and writes the model version information to the MySQL database. After deployment, only pipeline two needs to be executed to train the model. During the online inference process, pipelines three and four run in parallel. Pipeline three splits the steady-state data, calculates the anomaly score, and compares it to a predefined anomaly threshold to determine if it is abnormal. Pipeline four determines whether the split stream data contains samples that are significantly different from the existing training set. If so, the samples are entered into the MySQL database.

[0051] The above description is only illustrative of the present invention and not restrictive. Those skilled in the art will understand that many modifications, changes or equivalents may be made without departing from the spirit and scope defined by the appended claims, but all of them will fall within the scope of protection of the present invention.

Claims

1. An adaptive health diagnosis system for railway locomotive diesel engines based on human-in-the-loop, Characterized in that, It includes four modules: a working condition segmentation module, a duplicate removal detection module, a model training module, and an anomaly detection model module, as well as two steps: an offline training process and an online inference process; The offline training process includes parallel pipeline one and pipeline two; Pipeline one segments the time series data into different time series segments through the working condition segmentation module, excludes duplicate data segments from the above time series segments through the duplicate removal detection module, generates digital signatures for the new samples after duplicate removal detection and writes them into the Redis database, and generates new samples and writes them into the MYSQL database; Pipeline two imports the new samples in the MYSQL database of pipeline one, drives the model training module to loop, saves the model folder in the local model path, and writes the model version information into the MYSQL database; The online inference process includes parallel pipeline three and pipeline four; Pipeline three segments the time series data into different time series segments through the working condition segmentation module, calculates the anomaly score for the above time series segments through the anomaly detection model module, stores the generated anomaly score in the MYSQL database, and compares it with the threshold predefined manually in the MYSQL database to determine whether there is an anomaly, and writes the anomaly judgment into the MYSQL database; Pipeline four excludes duplicate data segments from the different time series segments segmented by the working condition segmentation module in pipeline three through the duplicate removal detection module, generates digital signatures for the new samples after duplicate removal detection and writes them into the Redis database, and generates new samples and writes them into the MYSQL database.

2. The adaptive health diagnosis system for railway locomotive diesel engines based on human-in-the-loop according to claim 1, Characterized in that, The working condition segmentation module divides the diesel engine into different handle positions in different working modes, and segments the time series data according to the above handle position data into data under each position; for the time series data within the current position range, a statistical sliding window is applied to the actual power of the diesel engine starting from the leftmost side of the range. If the power mean within the current sliding window falls within the mean ± 95% statistical interval of the statistical window after a specific time T, it is considered to enter the steady state interval, and the sliding window at the current moment is moved backward by T moments, and the above process is repeated until the power mean within the current moment sliding window exceeds the statistical interval of the sliding window after T moments or the current sliding window has reached the end of the current position range; if the power mean within the current sliding window does not fall within the mean ± 95% statistical interval of the statistical window after a specific time T, it is considered not to have entered the steady state interval, then the sliding window at the current moment is moved backward by 1 moment, and the above process is repeated until the end of the current position range.

3. The adaptive health diagnosis system for railway locomotive diesel engines based on human-in-the-loop according to claim 1, Characterized in that, The abnormal detection model module includes algorithm processing, abnormal score, and abnormal judgment; the algorithm processing extracts features Y from the time series segment X after working condition segmentation by calling the abnormal detection model i and uses the autoencoder in the model to reconstruct and generate algorithm features for the data segment i , and after splicing, a vector signal is obtained The abnormal score maps the vector signal q into a sphere with a radius of 1, and calculates the distance between it and the mapped value of the normal signal in the MYSQL database as the abnormal score of the segment data, and writes the generated abnormal score into the MYSQL database; the abnormal judgment compares the generated abnormal score with the threshold predefined manually in the MYSQL database to judge whether there is an abnormality.

4. The adaptive health diagnosis system for railway locomotive diesel engines based on human-in-the-loop according to claim 3, Characterized in that, The anomaly detection model in the anomaly detection model module and the model in the model training module are both composed of a feature extractor, a temporal encoder, a temporal decoder, and a projection layer; the feature extractor extracts features Y from the time series segment X through a one-dimensional convolutional network i in i ; the temporal encoder and the temporal decoder extract spatio-temporal related features Z i and the reconstructed signal is the said Y i and after being mapped to the surface of the high-dimensional sphere of the projection layer, are the vector v i and the vector Then the vector The objective of the loss function is to maximize the cosine similarity of the vector v i , the vector and the vector c e , that is, the anomaly score 5. The adaptive health diagnosis system for railway locomotive diesel engines based on human-in-the-loop according to claim 4, Characterized in that, The abnormal score Compare and judge with the threshold value predefined manually in the MYSQL database. Let the threshold value be c. If it is greater than c, it is abnormal; if it is less than c, it is normal.

6. The human-in-the-loop based adaptive health diagnosis system for railway locomotive diesel engines according to claim 1, characterized in that, the duplicate removal detection module divides the time series segments after working condition segmentation, maps each channel of the multi-dimensional time series data into a string by using the SAX algorithm, then reconstructs it into equal-length cross-dimensional words and sentences, compares the similarity with the historical data set, and excludes duplicate data segments according to the artificially preset threshold. The new samples after duplicate removal detection generate digital signatures through the hash algorithm, write them into the Redis database, and generate new samples to write into the MYSQL database.

Images