Turnout state monitoring and alarming method based on multi-dimensional time sequence curve correlation analysis

Abstract

The invention discloses a turnout state monitoring and alarming method based on multi-dimensional time sequence curve correlation analysis, relates to the technical field of railway facility monitoring, fundamentally solves the problem that a traditional supervised learning algorithm cannot be applied due to extremely sparse fault samples, can complete training of a core model without any fault data, and improves the training efficiency. The qualitative leap from numerical alarm to behavior mode alarm is realized. According to the method, by analyzing the correlation behavior mode of the current and the displacement, early hidden faults which cannot be found only by single data, such as switch rail jitter caused by normal current but slow action and current fluctuation can be accurately identified, and the sensitivity and the early warning capability of fault detection are remarkably improved. And finally, as the model deeply understands the influence of normal working condition changes such as temperature and load on curve relevance, the model can effectively distinguish normal working condition fluctuation and abnormal fault symptoms, so that the false alarm rate in the prior art is greatly reduced, and the alarm information is more credible.

Description

Technical Field

[0001] The present invention belongs to the technical field of railway facility monitoring, and in particular relates to a turnout state monitoring and alarm method based on multi-dimensional time series curve correlation analysis. Background Art

[0002] Railway turnouts are critical equipment for controlling train direction, and the reliability of their operating status is directly related to railway transportation safety. Currently, turnout condition monitoring primarily relies on threshold alarms based on operating current. This method determines whether the equipment is abnormal by setting fixed upper and lower current limits. However, it ignores the morphological changes of the current curve over a complete operating cycle and is highly susceptible to interference from normal operating conditions such as ambient temperature and load fluctuations. This results in high rates of false alarms and missed alarms, making it difficult to effectively identify potential faults in their early stages. To improve accuracy, some technical solutions have attempted to introduce supervised learning algorithms to train classification models to identify fault modes. However, a fundamental technical bottleneck of this approach is its heavy reliance on a large and balanced set of fault samples for training. In high-security sectors like railways, real equipment failures are extremely rare events, and fault samples are not only sparse but also expensive to obtain. This makes the models used in these methods ineffective for training, leaving those skilled in the art with the technical dilemma of being unable to build models without fault data. Therefore, there is an urgent need for a new method for high-precision and robust condition monitoring and alarming that does not rely on massive fault samples to address these existing challenges. Summary of the Invention

[0003] In view of the defects in the prior art, the present invention provides a turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis to solve the above technical problems.

[0004] The turnout condition monitoring and alarm method based on multi-dimensional time series curve correlation analysis includes the following steps: Acquire a multi-dimensional timing curve data set generated when the turnout machine is running, wherein the data set includes at least one operating current timing curve and an analog quantity timing curve collected synchronously with the operating current timing curve; Based on the historical normal data in the data set, a baseline model is trained to characterize the intrinsic correlation between the action current timing curve and the analog quantity timing curve; Inputting the currently collected curve data into the trained baseline model in real time, and calculating a deviation value representing the difference between the current curve data and the baseline model; The deviation value is compared with an alarm threshold, and an alarm signal is generated when the deviation value is greater than the alarm threshold.

[0005] Preferably, the sampling frequency of the analog timing curve is not less than 1 kHz, and is time-aligned with the action current curve. The analog timing curve specifically includes: Displacement curve collected by laser displacement sensor; Pressure value curve collected by strain gauge pressure sensor; Vibration curve collected by vibration sensor.

[0006] Preferably, when training a baseline model for characterizing the intrinsic correlation between the action current timing curve and the analog quantity timing curve, the following steps are specifically further included: The historical normal data is amplified by data augmentation technology to generate a training data set for training; The data enhancement technology includes one or more of noise injection, time series distortion and using a generative adversarial network to synthesize simulated fault data.

[0007] Preferably, the construction process of the baseline model is specifically as follows: A variational autoencoder is used to extract features from the input curve data to generate a potential feature vector; A long short-term memory network is used to perform temporal modeling on the latent feature vectors to learn the dynamic associations between curves.

[0008] Preferably, the method further includes a model updating step, wherein the model updating step adopts a semi-supervised learning mechanism, which specifically includes the following contents: Using unlabeled normal operation data, the feature encoding capability of the baseline model is continuously optimized through contrastive learning; A small amount of manually verified labeled fault data is used to fine-tune the alarm threshold or the decision boundary of the model.

[0009] Preferably, the alarm threshold is an adaptive threshold, and the method for generating the adaptive threshold is specifically as follows: Calculate the mean μ and standard deviation σ of historical deviation values ​​based on the sliding time window; Set the adaptive threshold to a, where a=μ+kσ; The value of the coefficient k is associated with the collected ambient temperature or the historical frequency of use of the turnout.

[0010] Preferably, the alarm signal is a graded alarm signal, and the method for generating the graded alarm signal is specifically as follows: If the deviation value is continuously greater than the alarm threshold, and the extracted deviation curve features match the patterns in the preset fault feature library, a level 1 alarm signal is generated; If the deviation value is continuously greater than the alarm threshold, but no matching pattern is found in the fault feature library, a secondary alarm signal is generated; If the instantaneous rate of change of the deviation value exceeds the preset emergency threshold, a level 3 alarm signal is generated.

[0011] As preferably, the following steps are also included: When generating the first-level alarm signal, the top N fault types with the highest matching degree with the current deviation curve characteristics are extracted from the fault feature library as recommended diagnosis conclusions, where N≥1.

[0012] Preferably, the method is implemented through an edge-cloud collaborative architecture, and the method steps under the architecture include: Execute the steps on the edge computing unit deployed at the turnout site and perform real-time deviation calculation; When the deviation value is greater than a locally preset threshold, uploading the complete multi-dimensional timing curve to a cloud server; The model training and updating in the steps are performed on the cloud server, and the updated model is compressed by knowledge distillation and sent to the edge computing unit.

[0013] Preferably, when the deviation value is compared with an alarm threshold and an alarm signal is generated when the deviation value is greater than the alarm threshold, the method further includes the following steps: Calculating the distance between the currently acquired multidimensional time series curve and the reference curve output by the baseline model using a dynamic time warping algorithm, and using the distance as the deviation value; When the deviation values ​​of M consecutive sampling points are all greater than the alarm threshold, the alarm signal is generated, where M≥3.

[0014] The beneficial effects of the present invention are: the present invention fundamentally solves the problem that traditional supervised learning algorithms cannot be applied due to the extreme sparsity of fault samples, and the core model training can be completed without any fault data, achieving a qualitative leap from numerical alarm to behavioral pattern alarm. The traditional threshold method can only determine whether the current is out of limit, while this method can accurately identify early hidden faults such as normal current but slow action, current fluctuation causing point rail jitter, etc., which cannot be discovered by a single data, by analyzing the correlation behavior pattern of current and displacement. It significantly improves the sensitivity of fault detection and early warning capabilities. Finally, because the model has a deep understanding of the impact of normal operating conditions such as temperature and load on the correlation of curves, it can effectively distinguish between normal operating fluctuations and abnormal fault signs, thereby greatly reducing the false alarm rate of existing technologies, making alarm information more credible, and improving operation and maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0016] Figure 1 A schematic flow chart of a turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis provided by the present invention; Figure 2 This is a diagram showing the baseline model structure and deviation calculation principle of the turnout condition monitoring and alarm method based on multi-dimensional time series curve correlation analysis provided by the present invention. DETAILED DESCRIPTION

[0017] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0018] The disclosure below provides many different embodiments or examples to realize different structures of the present invention. In order to simplify the disclosure of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the present invention.

[0019] The embodiments of the invention are described in detail below with reference to the accompanying drawings.

[0020] like Figure 1 、 Figure 2 As shown, the turnout state monitoring and alarm method based on multi-dimensional time series curve correlation analysis includes the following steps: Acquire a multi-dimensional timing curve data set generated when the turnout machine is running, wherein the data set includes at least one operating current timing curve and an analog quantity timing curve collected synchronously with the operating current timing curve; Based on the historical normal data in the data set, a baseline model is trained to characterize the intrinsic correlation between the action current timing curve and the analog quantity timing curve; Inputting the currently collected curve data into the trained baseline model in real time, and calculating a deviation value representing the difference between the current curve data and the baseline model; The deviation value is compared with an alarm threshold, and an alarm signal is generated when the deviation value is greater than the alarm threshold.

[0021] In its implementation, the present invention first synchronously captures the switch machine's operating current and switch rail displacement during each full switching cycle within the switch machine's control cabinet. This periodic data is converted into a time-series curve and assigned a uniform timestamp, forming a multidimensional curve data pair that represents a one-to-one correspondence between current and displacement. Subsequently, during the data preprocessing phase, the collected raw curves are filtered, de-noised, and normalized. Next, during the model training phase, thousands of healthy multidimensional curve data pairs collected under normal operating conditions in different seasons and loads are selected as the training set. This data is then trained using a sequence-to-sequence autoencoder model based on a long short-term memory network. The encoder portion of the model reads the complete current curve and compresses it into a thought vector containing temporal dynamic information. The decoder portion, based on this thought vector, attempts to reconstruct the corresponding displacement curve point by point. The training objective is to minimize the root mean square error between the reconstructed displacement curve and the actual captured normal displacement curve. Ultimately, the model becomes a system that represents the inherent physical relationship between current and displacement under normal conditions, namely a baseline model. During the real-time monitoring phase, when a new data pair is collected, its current curve is immediately input into the trained baseline model, which quickly outputs a predicted displacement curve, representing the expected displacement of the device based on the current current. The dynamic time warping algorithm then calculates the distance between the predicted displacement curve and the actual measured displacement curve. This distance represents the deviation between the ideal predicted data and the actual measured data. Ultimately, the system compares this deviation with an adaptive threshold that is dynamically adjusted based on historical deviation statistics and external environmental factors. Once the deviation exceeds this threshold, an alarm signal is immediately generated and pushed to the operation and maintenance platform.

[0022] Compared with the existing technology, the present invention fundamentally solves the problem that traditional supervised learning algorithms cannot be applied due to the extreme sparsity of fault samples. This method can complete the training of the core model without any fault data, and cleverly transforms the problem from learning what is a fault to mastering what is absolutely normal. Secondly, it achieves a qualitative leap from numerical alarm to behavioral pattern alarm. The traditional threshold method can only determine whether the current is out of limit, while this method can accurately identify early hidden faults such as normal current but slow action, current fluctuations causing point rail jitter, etc. by analyzing the correlation behavior pattern of current and displacement, which cannot be discovered by a single data alone. It significantly improves the sensitivity of fault detection and early warning capabilities. Finally, because the model has a deep understanding of the impact of normal operating conditions such as temperature and load on the correlation of curves, it can effectively distinguish between normal operating condition fluctuations and abnormal fault signs, thereby greatly reducing the false alarm rate of existing technologies, making alarm information more credible, and improving operation and maintenance efficiency.

[0023] More specifically, the sampling frequency of the analog timing curve is not less than 1 kHz, and is time-aligned with the action current curve. The analog timing curve specifically includes: Displacement curve collected by laser displacement sensor; Pressure value curve collected by strain gauge pressure sensor; Vibration curve collected by vibration sensor.

[0024] First, three types of high-precision sensors are deployed in parallel within the switch machine or on key transmission components: a laser displacement sensor, whose beam is aimed at the non-working surface of the switch rail, measures the precise millisecond position of the rail throughout the entire switching process without contact or wear. A strain gauge pressure sensor, integrated into the locking rod connecting the motor and the locking mechanism, quantifies in real time the tensile and compressive stresses experienced by the rod during the push-pull and locking processes. An industrial-grade triaxial MEMS vibration accelerometer, mounted on the switch machine's reduction gearbox housing, captures the high-frequency vibration signals generated by the equipment during operation. These three analog signals, along with the operating current signal, are synchronously connected to a multi-channel high-speed data acquisition card, sampling at a minimum of 1kHz. The rising edge of the motor's starting current serves as a unified hardware trigger signal, ensuring the precise alignment of the four curves on a time basis.

[0025] More specifically, when training a baseline model for characterizing the intrinsic correlation between the action current timing curve and the analog quantity timing curve, the following steps are further included: The historical normal data is amplified by data augmentation technology to generate a training data set for training; The data enhancement technology includes one or more of noise injection, time series distortion and using a generative adversarial network to synthesize simulated fault data.

[0026] In practice, noise is injected into each pair of collected multidimensional curve data. This involves adding subtle Gaussian white noise, tailored to the characteristics of the on-site electrical environment, to the original curves. This simulates the inevitable random interference of sensors and electrical circuits, forcing the model to learn to ignore these meaningless fluctuations and instead identify the core correlations. Time warping techniques are then applied to smoothly stretch or compress the time axis of the curves nonlinearly. For example, this can simulate a normal scenario where the overall operating time is slightly longer due to the increased viscosity of lubricating oil in winter. A generative adversarial network trained on a large amount of normal data is then fine-tuned at its input latent vectors to synthesize a large amount of "simulated sub-healthy" data that lies on the fuzzy boundary between "normal" and "abnormal." For example, a pressure curve with slight jitter at the end of the curve can be synthesized, but the overall curve remains within the normal range. These three techniques, combined, can amplify thousands of original normal samples into hundreds of thousands of "pan-normal" training datasets encompassing a variety of subtle variations, covering different environmental influences, and spanning different wear stages, for training the final baseline model.

[0027] More specifically, the construction process of the baseline model is as follows: A variational autoencoder is used to extract features from the input curve data to generate a potential feature vector; A long short-term memory network is used to perform temporal modeling on the latent feature vectors to learn the dynamic associations between curves.

[0028] In implementation, the input 4D time series curve is fed into the encoder portion of a variational autoencoder at each time step. Through convolution and pooling operations, the encoder extracts and maps the high-dimensional, complex curve snapshot at the current moment into a low-dimensional, Gaussian-distributed latent feature vector. Next, the sequence of these time-series-generated latent feature vectors is fed into a long-short-term memory (LSTM) network. The LSTM network memorizes the feature vector at the start of the action, how to smoothly transition in the middle, and the pattern during the locking phase, thereby constructing a dynamic correlation model of the entire turnout operation at different stages. Finally, the time series information modeled by the LSTM network is reconstructed by the variational autoencoder to restore the original 4D curve. The training goal of this model is to minimize the difference between the reconstructed curve and the input curve.

[0029] More specifically, the method further includes a model updating step, wherein the model updating step adopts a semi-supervised learning mechanism, which specifically includes the following contents: Using unlabeled normal operation data, the feature encoding capability of the baseline model is continuously optimized through contrastive learning; A small amount of manually verified labeled fault data is used to fine-tune the alarm threshold or the decision boundary of the model.

[0030] First, the system continuously performs contrastive learning in the background, leveraging massive amounts of normal operating data that requires no manual annotation. For a newly acquired normal curve, a positive example is created through data augmentation, while other unrelated normal curves are used as negative examples. Next, using a specific contrastive loss function, such as SimCLR, the variational autoencoder in the baseline model is continuously optimized, aiming to keep the latent feature vector generated for the positive example as close as possible in feature space to that of the positive example and as far away as possible from the feature vectors of the negative examples. This continuous process forces the baseline model to learn a more robust representation of the concept of normality. Simultaneously, during maintenance, if a very small number of real faults are manually confirmed and labeled, this labeled data triggers the second stage of fine-tuning. This data is fed into the model, and its deviation is calculated. If the deviation exceeds the alarm threshold, no adjustment is required. However, if it falls within the fuzzy area of ​​the alarm boundary, the system fine-tunes the weights of a few network layers of the baseline model close to the decision boundary at a very small learning rate, based on the severity of the fault, to slightly shift the decision boundary away from the fault.

[0031] More specifically, the alarm threshold is an adaptive threshold, and the method for generating the adaptive threshold is as follows: Calculate the mean μ and standard deviation σ of historical deviation values ​​based on the sliding time window; Set the adaptive threshold to a, where a=μ+kσ; The value of the coefficient k is associated with the collected ambient temperature or the historical frequency of use of the turnout.

[0032] The system maintains a historical deviation value database based on a sliding time window. This database stores deviation values ​​from hundreds or even thousands of past switch switching operations that were deemed normal. Each time a new normal switch is completed, its deviation value is stored in the database, while the oldest deviation value is discarded, ensuring that the database always reflects the most recent and up-to-date health of the equipment. Based on this database, the system calculates the mean μ and standard deviation σ of all historical deviation values ​​in real time and on a rolling basis. μ represents the average level of recent normal equipment performance, while σ quantifies the acceptable range of fluctuation within this normal level.

[0033] The k-factor is a key factor in the alarm system. A larger k-value increases the tolerance range, making the system less sensitive and less likely to generate an alarm. A smaller k-value narrows the tolerance range, making the system more sensitive and more likely to generate an alarm. Extreme low temperatures are a primary cause of turnout failures. Low temperatures generally present the following potential risks: Lubricating oil in slide plates, rollers, and other areas becomes extremely viscous, causing a sharp increase in frictional resistance; metal components shrink, potentially reducing clearances and increasing the risk of seizure; and low temperatures reduce motor output efficiency. Based on the real-time ambient temperature, the first k-factor adjustment is determined using the following temperature-k-factor relationship table.

[0034] Temperature-k value relationship table

[0035] Frequently used turnouts also wear out their mechanical components faster and accumulate fatigue damage at a higher rate. High-frequency use often exposes turnouts to the following risks: The degree of wear on components like the switch rail, slide plate, and locking mechanism is positively correlated with the number of switches; and high-frequency shock and vibration accelerate the growth of fatigue cracks in the material. The system counts the number of turnout actuations over a period of time, such as 24 hours, and uses the following table to determine the second adjustment factor for the k value.

[0036] Action frequency-k value relationship table

[0037] In actual applications, the system will synchronously access real-time ambient temperature sensor data and historical usage frequency data recorded by the switch controller. The final k value is the result of the combined effect of the first adjustment coefficient and the second adjustment coefficient, and the final k=3×min(f1,f2).

[0038] More specifically, the alarm signal is a graded alarm signal, and the method for generating the graded alarm signal is specifically as follows: If the deviation value is continuously greater than the alarm threshold, and the extracted deviation curve features match the patterns in the preset fault feature library, a level 1 alarm signal is generated; If the deviation value is continuously greater than the alarm threshold, but no matching pattern is found in the fault feature library, a secondary alarm signal is generated; If the instantaneous rate of change of the deviation value exceeds the preset emergency threshold, a level 3 alarm signal is generated.

[0039] When a deviation value is first confirmed to be persistently greater than an adaptive threshold, the system immediately initiates a hierarchical diagnostic process. First, the system extracts morphological features from the time-series curve of the abnormal deviation value, such as peak height, kurtosis, duration, curve area, and spectral distribution derived through Fourier transform. These features are known as the deviation curve signature. The system then rapidly compares the deviation curve signature with a pre-established fault signature library. This library contains standard fault signatures corresponding to various known faults, such as foreign object jamming in turnouts, loose locks, and motor brush wear. If the current fingerprint successfully matches a pattern in the library, for example, by meeting criteria such as a cosine similarity greater than 0.95, the system generates a clear first-level alarm. If no matching pattern is found in the fault signature library, this indicates that the system has discovered a new phenomenon, a confirmed anomaly but with an unknown cause, and a second-level alarm is generated. Throughout this process, the instantaneous rate of change monitoring module concurrently calculates the incremental deviation value in real time within a set time window. Once this incremental value exceeds the emergency threshold, it will instantly trigger a level 3 alarm signal, which will bypass all upper-layer applications and directly send the highest instruction representing the level 3 alarm signal to the hardware driver layer.

[0040] More specifically, the following steps are also included: When generating the first-level alarm signal, the top N fault types with the highest matching degree with the current deviation curve characteristics are extracted from the fault feature library as recommended diagnosis conclusions, where N≥1.

[0041] When the system determines that the deviation value is consistently above the threshold and decides to generate a Level 1 alarm, it uses the cosine similarity algorithm to calculate the similarity score between the current deviation curve feature vector and all fault feature vectors in the library. This calculation method allows the system to focus only on the directional proximity of two vectors, ignoring their absolute values. This allows it to accurately capture the morphological similarity of fault patterns without being affected by the severity of individual faults.

[0042] More specifically, the method is implemented through an edge-cloud collaborative architecture, and the method steps under the architecture include: Execute the steps on the edge computing unit deployed at the turnout site and perform real-time deviation calculation; When the deviation value is greater than a locally preset threshold, uploading the complete multi-dimensional timing curve to a cloud server; The model training and updating in the steps are performed on the cloud server, and the updated model is compressed by knowledge distillation and sent to the edge computing unit.

[0043] An industrial-grade edge computing unit is deployed next to the control cabinet of each turnout. This unit is directly connected to four sensors. Its core task is to continuously run the baseline model during each turnout movement. In most normal movements, the deviation value is below a threshold, and the edge unit records and archives the data without generating any uplink communication. However, if the deviation value of a particular movement exceeds this threshold, the edge unit immediately uploads the complete multi-dimensional timing curve data package for that movement to a remote cloud server via the 4G / 5G network. After receiving the data, the cloud server invokes the baseline model for refined diagnosis, graded alarms, and fault tracing. The cloud server also aggregates both abnormal and normal data uploaded by all edge nodes, using this data for continuous model retraining and semi-supervised learning updates. Once a model with improved performance is trained, it uses knowledge distillation to compress its core knowledge into a new, lightweight model, which is then distributed to edge computing units across the network.

[0044] More specifically, when the deviation value is compared with an alarm threshold and an alarm signal is generated when the deviation value is greater than the alarm threshold, the method further includes the following steps: Calculating the distance between the currently acquired multidimensional time series curve and the reference curve output by the baseline model using a dynamic time warping algorithm, and using the distance as the deviation value; When the deviation values ​​of M consecutive sampling points are all greater than the alarm threshold, the alarm signal is generated, where M≥3.

[0045] After a turnout completes a movement, the system invokes a dynamic time warping algorithm to nonlinearly align the multidimensional time series curves currently being collected in real time with the reference curve output by the baseline model. It intelligently stretches or compresses the time axis to find the optimal morphological match between the two curves, thereby calculating a distance value that truly reflects the essential difference in shape between the two curves. This distance value is a more robust deviation value. The system then checks whether the deviation values ​​of M consecutive sampling points, starting from the current sampling point, are greater than the dynamically adaptive alarm threshold. If so, it is considered a persistent exceedance. Here, M is generally set to an integer between 3 and 5. This setting is intended to effectively filter out instantaneous random interference and ensure that only persistent anomalies are confirmed as alarm events. Only when this persistent exceedance condition is met will the system finally confirm that this is an abnormal event with real physical significance and formally generate an alarm signal.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis is characterized in that: The following steps are involved: Acquire a multi-dimensional timing curve data set generated when the turnout machine is running, wherein the data set includes at least one operating current timing curve and an analog quantity timing curve collected synchronously with the operating current timing curve; Based on the historical normal data in the data set, a baseline model is trained to characterize the intrinsic correlation between the action current timing curve and the analog quantity timing curve; Inputting the currently collected curve data into the trained baseline model in real time, and calculating a deviation value representing the difference between the current curve data and the baseline model; The deviation value is compared with an alarm threshold, and an alarm signal is generated when the deviation value is greater than the alarm threshold.

2. The turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis according to claim 1 is characterized in that: The sampling frequency of the analog timing curve is not less than 1kHz, and is time-aligned with the action current curve. The analog timing curve specifically includes: Displacement curve collected by laser displacement sensor; Pressure value curve collected by strain gauge pressure sensor; Vibration curve collected by vibration sensor.

3. The turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis according to claim 1 is characterized in that: When training a baseline model for characterizing the intrinsic correlation between the action current timing curve and the analog quantity timing curve, the following steps are specifically included: The historical normal data is amplified by data augmentation technology to generate a training data set for training; The data enhancement technology includes one or more of noise injection, time series distortion and using a generative adversarial network to synthesize simulated fault data.

4. The turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis according to claim 3 is characterized in that: The construction process of the baseline model is specifically as follows: A variational autoencoder is used to extract features from the input curve data to generate a potential feature vector; A long short-term memory network is used to perform temporal modeling on the latent feature vectors to learn the dynamic associations between curves.

5. The turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis according to claim 4 is characterized in that: The model updating step is also included. The model updating step adopts a semi-supervised learning mechanism, which specifically includes the following contents: Using unlabeled normal operation data, the feature encoding capability of the baseline model is continuously optimized through contrastive learning; A small amount of manually verified labeled fault data is used to fine-tune the alarm threshold or the decision boundary of the model.

6. The turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis according to claim 1 is characterized in that: The alarm threshold is an adaptive threshold, and the method for generating the adaptive threshold is specifically as follows: Calculate the mean μ and standard deviation σ of historical deviation values ​​based on the sliding time window; Set the adaptive threshold to a, where a=μ+kσ; The value of the coefficient k is associated with the collected ambient temperature or the historical frequency of use of the turnout.

7. The turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis according to claim 1 is characterized in that: The alarm signal is a graded alarm signal, and the generation method of the graded alarm signal is specifically as follows: If the deviation value is continuously greater than the alarm threshold, and the extracted deviation curve features match the patterns in the preset fault feature library, a level 1 alarm signal is generated; If the deviation value is continuously greater than the alarm threshold, but no matching pattern is found in the fault feature library, a secondary alarm signal is generated; If the instantaneous rate of change of the deviation value exceeds the preset emergency threshold, a level 3 alarm signal is generated.

8. The turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis according to claim 7 is characterized in that: The following steps are also included: When generating the first-level alarm signal, the top N fault types with the highest matching degree with the current deviation curve characteristics are extracted from the fault feature library as recommended diagnosis conclusions, where N≥1.

9. The turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis according to claim 1 is characterized in that: The method is implemented through an edge-cloud collaborative architecture, and the method steps under the architecture include: Execute the steps on the edge computing unit deployed at the turnout site and perform real-time deviation calculation; When the deviation value is greater than a locally preset threshold, uploading the complete multi-dimensional timing curve to a cloud server; The model training and updating in the steps are performed on the cloud server, and the updated model is compressed by knowledge distillation and sent to the edge computing unit.

10. The turnout status monitoring and alarm method based on multi-dimensional time series curve correlation analysis according to claim 1 is characterized in that: The deviation value is compared with an alarm threshold, and when the deviation value is greater than the alarm threshold, an alarm signal is generated, specifically further comprising the following steps: Calculating the distance between the currently acquired multidimensional time series curve and the reference curve output by the baseline model using a dynamic time warping algorithm, and using the distance as the deviation value; When the deviation values ​​of M consecutive sampling points are all greater than the alarm threshold, the alarm signal is generated, where M≥3.

Images