A fault diagnosis method for turnout switch machines

Abstract

This invention discloses a fault diagnosis method for a turnout switch machine, belonging to the field of fault diagnosis technology. The method includes collecting sound, vibration, and current signals from the turnout switch machine; removing local data exhibiting unstable variations from the obtained three-modal raw data; determining the error interference coefficient based on the inter-modal correlation error characteristics under ideal conditions; obtaining the error influence factor based on the correlation influence error variation characteristics of historical environmental data characteristics on the additional errors belonging to the three-modal data from historical detection data; performing error compensation on the initial processed modal data; extracting deep semantic features of the three modes after linear mapping of the obtained final processed data; conducting quality assessment based on signal-to-noise ratio, kurtosis, and spectral flatness; if the quality score to be judged is higher than a preset low quality threshold, performing cross-modal semantic interaction to obtain three sets of enhanced features, thereby determining the fault category of the turnout switch machine. This invention can improve the accuracy of fault category determination.

Description

Technical Field

[0001] This invention relates to the field of fault early warning technology, and in particular to a fault diagnosis method for turnout switching machines. Background Technology

[0002] High-speed railways are a core component of the national comprehensive transportation system. Switch machines are key equipment for ensuring the safety and efficiency of high-speed railway operations and dispatching. The accuracy, real-time performance, and intelligence level of their fault diagnosis directly determine the safety of high-speed railway operations and the quality of passenger services.

[0003] The current high-speed railway turnout switching machine operates in a complex environment with highly variable working conditions. Existing intelligent fault diagnosis technology still has three major shortcomings that restrict its practical engineering application:

[0004] 1. Limitations of single-modal information: Most existing diagnostic methods rely on signals from a single sensor, which cannot comprehensively depict the operating status of the switch machine from multiple dimensions such as mechanical, electrical, and structural aspects. The incomplete collection of fault characteristics directly results in a low diagnostic accuracy.

[0005] 2. Weak generalization ability across equipment and operating conditions: Switch machines of different models, lines and environments have large differences in structural parameters and operating conditions, and the distribution of collected data is significantly offset; the performance of traditional models declines significantly in new equipment and new scenarios, and they lack adaptive transfer capabilities.

[0006] 3. Limited Model Deployment: Existing deep learning diagnostic models have a large number of parameters and high computational complexity, making it difficult to adapt to the computing power constraints of embedded and edge computing hardware in railway sites, and thus unable to achieve real-time inference and large-scale engineering deployment. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, the present invention provides a method for diagnosing faults in turnout switching machines.

[0008] This invention provides a fault diagnosis method for a turnout switch machine, the method comprising:

[0009] Step S1: Collect the sound signal, vibration signal and current signal of the turnout switch machine to obtain three-mode raw data. Remove local data with unstable changes from the three-mode raw data to obtain the initial processed mode data.

[0010] Step S2: Obtain historical detection data. Based on the error variation characteristics of the correlation between sound signals, vibration signals, and current signals under ideal environmental conditions in the historical detection data, obtain the error interference coefficient. Based on the error variation characteristics of the correlation between historical environmental data characteristics and the additional error influence of sound signals, vibration signals, and current signals in the historical detection data, obtain the error influence factor. Based on the error interference coefficient and the error influence factor, perform error compensation on the initial processed modal data to obtain the final processed modal data.

[0011] Step S3: Linearly map the final processed modal data to obtain trimodal mapping features. Input the trimodal mapping features into the Transformer encoder for deep semantic feature representation to obtain trimodal deep semantic feature representation. If the quality score of the trimodal deep semantic feature representation obtained by quality evaluation based on signal-to-noise ratio, kurtosis, and spectral flatness is greater than a preset low quality threshold, construct a cross-attention mechanism with three pairs of modal interaction structures. Input the trimodal deep semantic feature representation into the cross-attention mechanism for cross-modal semantic interaction calculation and output three sets of modal enhancement features. Based on the three sets of modal enhancement features, obtain the fault category of the turnout switch machine.

[0012] Preferably, the variation trends of various modal signals in the three-modal raw data are statistically analyzed to obtain the variation trend characteristics to be measured;

[0013] A preset baseline determination threshold is set, and local features that exceed the baseline determination threshold in the trend features to be tested are marked to obtain the local features to be tested.

[0014] The temporal continuity of the local features of the marker to be tested is statistically analyzed to obtain the temporal continuity of the marker to be tested.

[0015] A preset continuity determination threshold is set. If the continuity of the time sequence to be tested is greater than the continuity determination threshold, the local feature of the corresponding modal signal in the local feature of the marker to be tested is determined to be an unstable variable feature and is removed to obtain the modal data to be tested.

[0016] If the continuity of the time sequence to be tested is less than or equal to the continuity determination threshold, then the local feature of the corresponding modal signal in the local feature of the label is determined to be a stable variation feature and is retained to obtain the second modal data to be tested.

[0017] The data from either the first or second modal data to be tested is selected as the initial processing modal data.

[0018] Preferably, historical detection data is acquired, and the detection error values ​​of sound signals, vibration signals, and current signals under the desired environmental conditions are extracted from the historical detection data, and a historical error feature set is obtained by combining them.

[0019] Historical three-modal data are extracted from historical detection data. Based on the correlation between any one modal signal in the historical three-modal data and the error feature values ​​corresponding to the other two modal signals in the historical error feature set, the error variation characteristics are affected. The correlation coefficient between any one modal signal in the historical three-modal data and the error feature values ​​corresponding to the other two modal signals in the historical error feature set is calculated, and the error interference coefficient is obtained by combining them.

[0020] Preferably, the environmental data characteristics of the collected trimodal raw data are detected to obtain the environmental data characteristics to be tested, which include temperature data, humidity data and electromagnetic intensity data;

[0021] The additional impact error values ​​of the sound signal, vibration signal and current signal due to environmental influence are extracted from the historical detection data after the interference between the three modes of signals has been compensated. The additional error values ​​of the historical sound, historical vibration and historical current are obtained.

[0022] Based on the characteristics of the environmental data to be tested, historical environmental data characteristics are obtained. Based on the correlation between the historical environmental data characteristics and historical additional sound error, historical additional vibration error, and historical additional current error, the error variation characteristics are affected. The correlation factors between the historical environmental data characteristics and historical additional sound error, historical additional vibration error, and historical additional current error are statistically analyzed, and the error influence factors are obtained by combining them.

[0023] Based on the error interference coefficient, error influence factor, and characteristics of the test environment data, error compensation is performed on the initial processed modal data to obtain the final processed modal data.

[0024] Preferably, the signal-to-noise ratio, kurtosis, and spectral flatness of the three-modal deep semantic feature representation are statistically analyzed to obtain pre-evaluation index features;

[0025] Based on the aforementioned pre-evaluation index characteristics, the quality of each modality feature in the three-modal deep semantic feature representation is evaluated to obtain the quality score to be judged.

[0026] A low quality threshold is preset. If the quality score to be judged is less than or equal to the low quality threshold, the sound signal, vibration signal and current signal of the turnout switch machine are re-acquired and error-compensated to obtain modal data re-detection information.

[0027] Preferably, if the quality score to be judged is greater than the low quality threshold, the feature fusion preprocessing information is output.

[0028] Based on the feature fusion preprocessing information, and based on the peak delay time estimation of the cross-correlation function, the peak delay time of the historical three-mode data in the historical detection data is estimated to obtain the delay time features;

[0029] Based on peak detection of normalized cross-correlation function, the delay direction and magnitude of historical three-mode data are statistically determined, and the delay direction and magnitude features are obtained;

[0030] Based on the aforementioned delay direction and magnitude characteristics, a directed acyclic causal graph is constructed;

[0031] Based on the aforementioned delay time characteristics and the directed acyclic causal graph, a cross-attention mechanism with three pairs of modal interaction structures is constructed.

[0032] Preferably, the three-modal deep semantic feature representations are input into a cross-attention mechanism for cross-modal semantic interaction calculation, and three sets of modality-enhanced features are output.

[0033] Based on the Sigmoid adaptive gating feature fusion mechanism, the importance weight of each modality feature in the three sets of modality enhancement features is adaptively calculated, and dynamic weighted fusion is performed to output the fused feature;

[0034] Obtain the number of known fault types of the turnout switch machine, construct a fully connected classification layer, wherein the input dimension of the fully connected classification layer is the same as the dimension of the fused feature, and the output dimension is equal to the number of known fault types;

[0035] The fused features are input into a fully connected classification layer and then mapped by a Softmax function to output the fault category of the turnout switch machine.

[0036] Compared with the prior art, the present invention has the following characteristics and beneficial effects:

[0037] By precisely removing temporally continuous unstable local data from the trimodal raw data, random instantaneous noise interference can be effectively filtered out. Compared with traditional simple noise reduction methods, this approach can improve the targeting and reliability of the original signal preprocessing while preserving effective fault characteristics, providing a high-quality data foundation for subsequent error compensation and feature extraction. Combining the error interference coefficient of intermodal correlation interference under ideal conditions and the error influence factor brought by environmental factors, error compensation is performed on the initial processed modal data. This fully considers the dual impact of intermodal coupling interference and external environmental noise such as temperature, humidity, and electromagnetic fields, solving the problem of insufficient compensation accuracy in existing technologies that only perform single error correction, significantly improving the authenticity of the modal signal data. A multi-index joint quality assessment mechanism, including signal-to-noise ratio, kurtosis, and spectral flatness, is implemented to verify the quality of deep semantic features output by the Transformer encoder. Low-quality feature data is promptly re-sampled and reprocessed to avoid the direct input of low-quality features into the model, which would lead to a decrease in fault diagnosis accuracy. By employing a Sigmoid adaptive gating feature fusion mechanism, modal importance weights are dynamically assigned to the three sets of modal enhancement features under different samples, achieving adaptive weighted fusion of multimodal features. Furthermore, a fully connected classification layer with a matching dimension is customized according to the number of fault types, and fault discrimination is completed through Softmax. This can adaptively adapt to the differences in the contribution of each modality fault under different operating conditions, effectively improving the accuracy and robustness of multimodal fault diagnosis for turnout switching machines. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the steps of a fault diagnosis method for a turnout switch machine, which is the main feature of this embodiment. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the following embodiments.

[0040] Reference Figure 1 A fault diagnosis method for a turnout switch machine, the method comprising the following steps:

[0041] Step S1: Collect the sound signal, vibration signal and current signal of the turnout switch machine to obtain three-mode raw data. Remove local data with unstable changes from the three-mode raw data to obtain the initial processed mode data.

[0042] Step S2: Obtain historical detection data. Based on the error variation characteristics of the correlation between sound signals, vibration signals, and current signals under ideal environmental conditions in the historical detection data, obtain the error interference coefficient. Based on the error variation characteristics of the correlation between historical environmental data characteristics and the additional error influence of sound signals, vibration signals, and current signals in the historical detection data, obtain the error influence factor. Based on the error interference coefficient and the error influence factor, perform error compensation on the initial processed modal data to obtain the final processed modal data.

[0043] Step S3: Linearly map the final processed modal data to obtain three-modal mapping features. Input the three-modal mapping features into the Transformer encoder for deep semantic feature representation to obtain three-modal deep semantic feature representation. If the quality score of the three-modal deep semantic feature representation obtained by quality evaluation based on signal-to-noise ratio, kurtosis and spectral flatness is greater than the preset low quality threshold, then construct a cross-attention mechanism with three pairs of modal interaction structures. Input the three-modal deep semantic feature representation into the cross-attention mechanism for cross-modal semantic interaction calculation and output three sets of modal enhancement features. Based on the three sets of modal enhancement features, obtain the fault category of the turnout switch machine.

[0044] Specifically, by precisely removing temporally continuous unstable local data from the trimodal raw data, random instantaneous noise interference can be effectively filtered out. Compared with traditional simple noise reduction methods, this approach can improve the targeting and reliability of the original signal preprocessing while preserving effective fault characteristics, providing a high-quality data foundation for subsequent error compensation and feature extraction. By combining the error interference coefficient of intermodal correlation interference under ideal conditions and the error influence factor brought by environmental factors, error compensation is performed on the initial processed modal data. This fully considers the dual impact of intermodal coupling interference and external environmental noise such as temperature, humidity, and electromagnetic fields, solving the problem of insufficient compensation accuracy in existing technologies that only perform single error correction, significantly improving the authenticity of modal signal data. By introducing a joint quality assessment mechanism based on multiple indicators such as signal-to-noise ratio, kurtosis, and spectral flatness, the deep semantic features output by the Transformer encoder are quality-verified. Low-quality feature data is promptly re-sampled and reprocessed to avoid the direct input of low-quality features into the model, which would lead to a decrease in fault diagnosis accuracy. By adopting a sigmoid adaptive gating feature fusion mechanism, modal importance weights are dynamically assigned to the three sets of modal enhancement features under different samples, realizing adaptive weighted fusion of multimodal features. Furthermore, a fully connected classification layer with a matching dimension is customized according to the number of fault types, and fault discrimination is completed through Softmax. This can adaptively adapt to the differences in the contribution of each modality fault under different working conditions, effectively improving the accuracy and robustness of multimodal fault diagnosis of turnout switching machines.

[0045] The specific step S1 includes the following sub-steps:

[0046] The variation trends of various modal signals in the three-modal raw data are statistically analyzed to obtain the variation trend characteristics to be measured.

[0047] A preset baseline determination threshold is set, and local features that exceed the baseline determination threshold in the trend features to be tested are marked to obtain the local features to be tested.

[0048] The temporal continuity of the local features of the marker to be tested is statistically analyzed to obtain the temporal continuity of the marker.

[0049] A preset continuity determination threshold is set. If the continuity of the time sequence to be tested is greater than the continuity determination threshold, the local feature of the corresponding modal signal in the local feature of the marker to be tested is determined to be an unstable variable feature and is removed to obtain the modal data to be tested.

[0050] If the continuity of the time sequence to be tested is less than or equal to the continuity determination threshold, then the local feature of the corresponding modal signal in the local feature of the marker is determined to be a stable variation feature and is retained to obtain the second modal data to be tested.

[0051] Select either the first or second modal data to be tested as the initial processing modal data.

[0052] Specifically, for example, the ZD6 type turnout switch machine actually in operation on high-speed railway lines uses hardware such as iFlytek's 6-microphone ring sound array, Witt Intelligent's three-axis vibration sensor, and Huihuang Technology's speed-up turnout current acquisition module, along with a 1DQJ contact status acquisition device to achieve multi-modal time synchronization. The sampling frequencies are uniformly set as follows: sound 25600Hz, vibration 100Hz, and current sampling rate 1000Hz. All data is transmitted to the host computer for raw storage via the acquisition module. The measured trend characteristics (such as using time-domain mean, variance, and peak factor as trend statistical indicators, and performing batch calculations based on existing signal time-domain analysis algorithms) are used. A preset change baseline judgment threshold is set (such as calculating the upper and lower limits of amplitude fluctuation using n sets of steady-state signals under standard normal operating conditions from historical data, for example, 0.15). The amplitude fluctuation range is used as a safety baseline. Signal segments exceeding this range are automatically marked as local features to be tested. The continuity of the timing to be tested is calculated (e.g., the number of continuous sampling points of abnormal signals is counted segment by segment using the number of sampling points as the statistical unit). A preset continuity judgment threshold is set (e.g., based on historical data, the number of abnormal sampling points for multiple sets of instantaneous interference (train passing, electromagnetic instantaneous pulse) does not exceed 8, so the continuity judgment threshold is fixed at 8 sampling points). If the continuity of the timing to be tested is greater than 8 sampling points, it indicates that it is not an instantaneous interference, but a continuous signal fluctuation caused by a real equipment failure. It is judged as an unstable variation feature and directly eliminated. If the continuity of the timing to be tested is less than or equal to 8 sampling points, it is judged as a stable fluctuation feature caused by instantaneous interference in the field environment. All of them are retained, and the initial processed modal data after noise interference is obtained.

[0053] The specific step S2 includes the following sub-steps:

[0054] Historical detection data is acquired, and the detection error values ​​of sound signals, vibration signals, and current signals under the desired environmental conditions are extracted from the historical detection data, and the sets are used to obtain the historical error feature set.

[0055] Historical three-mode data are extracted from historical detection data. The error variation characteristics are influenced by the correlation between any one mode signal in the historical three-mode data and the error feature values ​​corresponding to the other two mode signals in the historical error feature set. The correlation coefficients between any one mode signal in the historical three-mode data and the error feature values ​​corresponding to the other two mode signals in the historical error feature set are statistically analyzed, and the error interference coefficient is obtained by combining them.

[0056] The environmental data characteristics of the collected trimodal raw data are detected to obtain the environmental data characteristics to be measured, which include temperature data, humidity data and electromagnetic intensity data.

[0057] The additional impact error values ​​of sound signal, vibration signal and current signal due to environmental influence are extracted from historical detection data after the interference between the three modes of signal has been compensated. This yields the historical additional sound error, historical additional vibration error and historical additional current error.

[0058] Based on the characteristics of the environmental data to be tested, historical environmental data characteristics are obtained. The correlation between the historical environmental data characteristics and historical additional errors in sound, vibration, and current is used to influence error variation. The correlation factors between the historical environmental data characteristics and historical additional errors in sound, vibration, and current are statistically analyzed, and the error influence factors are obtained by combining them.

[0059] Based on the error interference coefficient, error influence factor, and characteristics of the test environment data, error compensation is performed on the initial processed modal data to obtain the final processed modal data.

[0060] Specifically, an ideal environment is artificially selected and calibrated as a standard operating condition with a temperature of 20℃, humidity of 55%, no external electromagnetic interference, and no train traffic interference. From historical detection data, i sets of fault-free steady-state data under this ideal environment are selected. The time-domain detection error values ​​of sound, vibration, and current signals are extracted, including mean error, amplitude error, and phase error. These are then summarized to form a standardized historical error feature set. Error interference coefficients are calculated (for example, from M sets of historical three-modal data, m sets of mixed operating condition samples are randomly selected. Based on existing correlation analysis algorithms, the correlation coefficients between six pairs of modes—sound and vibration, sound and current, vibration and sound, vibration and current, current and sound, and current and vibration—are calculated to quantify the degree of mutual coupling interference between modes. For example, the interference coefficient of sound due to vibration is 0.28, and the interference coefficient due to current is 0.32; the interference coefficient of vibration due to sound is 0.25, and the interference coefficient due to current is 0.30; the interference coefficient of current due to sound is 0. 0.33. Vibration interference coefficient 0.29. All coefficients were calculated in batches using the Pearson correlation coefficient algorithm. All statistically obtained correlation values ​​were integrated and summarized to form a complete and reusable error interference coefficient library. This library can accurately quantify the mutual error influence of the three modes during mechanical coupling and electrical linkage, providing quantitative numerical basis for subsequent modal data error correction and solving the problem of ignoring cross-modal coupling interference in single-modal error correction. Environmental temperature, humidity, and electromagnetic intensity signals were detected using temperature and humidity sensors and a high-precision electromagnetic intensity detector (e.g., temperature 26℃, humidity 62%, electromagnetic intensity 0.45mT). These three parameters were integrated to construct complete data characteristics of the tested environment. Additional influence error values ​​were also considered (e.g., selecting several sets of samples from historical data of the past 3 years that have completed inter-modal coupling interference compensation, covering temperatures of 10℃~35℃, humidity of 40%~80%, and electromagnetic intensity of 0.2mT~0.8mT). For all operating conditions, additional error values ​​of sound, vibration, and current affected by the environment are extracted from each sample group to establish an environment-error correspondence sample library. Error influence factors (e.g., based on the multiple linear regression algorithm, the correlation degree of the changes in the additional errors of the three modes is statistically analyzed when the temperature increases by 1℃, the humidity increases by 5%, and the electromagnetic intensity increases by 0.1mT. After batch fitting calculation, the temperature correlation factor is 0.012, the humidity correlation factor is 0.008, and the electromagnetic correlation factor is 0.025. All factors are generated by fitting historical big data). Finally, the modal data is processed (e.g., using the historical error feature set, six sets of modal error interference coefficients, and temperature, humidity, and electromagnetic correlation factors constructed above, the existing multiple linear error compensation algorithm is used as the core correction algorithm; the algorithm uniformly incorporates three types of input parameters: the initial processed modal data after unstable feature removal, the inter-modal interference coefficient calculated by the Pearson correlation coefficient algorithm, and the environmental error influence factor obtained by multiple linear regression fitting).The correction process follows a fixed sequence: first, eliminate modal coupling errors, then compensate for environmental superposition errors. First, using a single sampling point as the smallest processing unit, all sampling time data for each mode of the turnout switch machine (sound, vibration, and current) are traversed. The pairwise interference coefficients (0.28, 0.32, etc.) obtained earlier are substituted into the linear compensation formula, and the inherent errors caused by mechanical coupling and electrical linkage between modes are calculated and deducted point by point. Second, the turnout is divided into segments according to fixed time domain intervals for start-up, transmission, and locking. Within each time domain segment, the measured field temperature (26℃), humidity (62%), electromagnetic intensity (0.45mT), and corresponding correlation factors (0.012, 0.008, 0.025) are substituted to correct the additional signal offset caused by external temperature, humidity, and electromagnetic interference segment by segment. After all sampling points and time-domain segments have undergone layered correction, the standard steady-state signal under ideal environmental conditions of 20℃ and 55% humidity is used as a reference. Through dual verification of waveform similarity and amplitude fluctuation range, when the similarity of the corrected modal waveform reaches more than 95% and the amplitude fluctuation is controlled within the allowable range, the final processed modal data is generated, which simultaneously eliminates internal coupling interference and dual errors of temperature, humidity, and electromagnetic external environment.

[0061] The specific step S3 includes the following sub-steps:

[0062] The signal-to-noise ratio, kurtosis, and spectral flatness of the three-modal deep semantic feature representations are statistically analyzed to obtain the pre-evaluation index features.

[0063] Based on the characteristics of the pre-evaluation indicators, the quality of each modality feature in the trimodal deep semantic feature representation is evaluated to obtain the quality score to be judged.

[0064] If a low quality threshold is preset, and the quality score to be judged is less than or equal to the low quality threshold, the sound signal, vibration signal and current signal of the turnout switch machine are re-acquired and error-compensated to obtain modal data re-detection information.

[0065] If the quality score to be judged is greater than the low quality threshold, then output the feature fusion preprocessing information.

[0066] Based on the feature fusion preprocessing information, and using the peak delay time estimation based on the cross-correlation function, the peak delay time of historical three-mode data in the historical detection data is estimated to obtain delay time features.

[0067] Based on peak detection of the normalized cross-correlation function, the delay direction and magnitude of historical three-mode data are statistically determined, and the delay direction and magnitude features are obtained.

[0068] Based on the characteristics of delay direction and magnitude, a directed acyclic causal graph is constructed.

[0069] Based on the time delay characteristics and the directed acyclic causal graph, a cross-attention mechanism with three pairs of modal interaction structures is constructed.

[0070] The three-modal deep semantic feature representations are input into the cross-attention mechanism for cross-modal semantic interaction computation, and three sets of modality-enhanced features are output.

[0071] Based on the Sigmoid adaptive gating feature fusion mechanism, the importance weight of each modality feature in the three sets of modality enhancement features is adaptively calculated, and dynamic weighted fusion is performed to output the fused feature.

[0072] Obtain the number of known fault types of the turnout switch machine, construct a fully connected classification layer, the input dimension of the fully connected classification layer is the same as the dimension of the fused feature, and the output dimension is equal to the number of known fault types.

[0073] The fused features are input into a fully connected classification layer and then mapped by a Softmax function to output the fault category of the turnout switch machine.

[0074] Specifically, this includes pre-evaluation index characteristics (e.g., based on existing signal processing techniques such as short-time Fourier transform, higher-order statistical moments in the time domain, and FFT spectral analysis, calculating the signal-to-noise ratio (SNR), kurtosis, and spectral flatness, respectively; for example, under normal operating conditions, the SNR is 38.5 dB, kurtosis 2.1, and spectral flatness 0.86; for minor gear faults, the SNR is 29.2 dB, kurtosis 2.8, and spectral flatness 0.72; and for mechanical jamming faults, the SNR is 21.8 dB, kurtosis 3.5, and spectral flatness 0.58), preset low-quality thresholds (obtained based on historical data statistics, such as by statistically analyzing the lower limits of several groups of distorted and strongly interfered degraded samples; for example, a comprehensive calibration of SNR 25 dB, kurtosis 3.2, and spectral flatness 0.60), and quality scores to be judged (e.g., assigning 0.4, 0.3, and 0.3 to the three indicators respectively). (A weighted composite of data with fixed weights and normalized values ​​is used to obtain the quality score to be judged.) When the quality score to be judged is lower than a preset threshold, it is determined that the current modal features are severely affected by environmental interference and the fault features are blurred, making them unusable for subsequent diagnosis. This process involves readjusting the sensor installation position, collecting three-modal data again under the same operating conditions, removing duplicate data, and performing error compensation until the feature quality score meets the standard before proceeding to the feature fusion stage. Delay time features (e.g., using a cross-correlation function peak delay time estimation algorithm to calculate sound-vibration and sound-...) The peak delay times between the current, vibration, and other signals are calculated to be 12ms, 18ms, and 9ms respectively. The timing relationship of the fault signals in different modes is statistically analyzed. The delay direction and magnitude characteristics are analyzed (e.g., using a normalized cross-correlation function for peak detection, automatically identifying the direction of peak advance and lag of each signal and the specific amplitude offset, clarifying the mode timing coupling law). A directed acyclic causal graph is constructed (based on the two core parameters of delay time and delay direction, for example, using existing graph theory modeling methods, a directed acyclic causal graph of the turnout switching machine is constructed, with nodes corresponding to the three modes and directed edges representing the signal timing transmission relationship). Based on the node association logic of the directed acyclic causal graph, a cross-attention mechanism structure with pairwise interaction is constructed for sound-vibration, sound-current, and vibration-current. Three sets of modal enhancement features are also constructed (e.g., cross-modal semantic interaction is completed through Query, Key, and Value matrix mapping, fault association features between different modes are mined, and three sets of modal enhancement features are output).The fusion features (introducing the Sigmoid adaptive gating fusion mechanism (this mechanism is a general adaptive weighted algorithm for deep learning, the weights do not need to be fixed manually and can be automatically updated according to real-time working conditions), for mechanical jamming conditions, if the sound feature weight is automatically assigned 0.35, the vibration feature weight is 0.45, and the current feature weight is 0.20; for gear wear conditions, the weight is adaptively adjusted to sound 0.40, vibration 0.30, and current 0.30, and weighted fusion is completed according to the dynamic weights to generate a unified fusion feature vector with a dimension of 256), if the fault categories of the ZD6 type turnout switch machine are fixed as normal operation, gear wear, mechanical jamming, and poor electrical contact, a fully connected classification layer is built according to the number of known fault categories. If the input dimension is set to 256 dimensions of fusion features, and the output dimension matches the 4 types of faults. Finally, the fused features are input into the fully connected classification layer, and the predicted confidence probability of each type of fault is calculated by the Softmax function. If the output is converted into a 0~1 probability distribution for each category, the probability values ​​corresponding to the four types of faults are output respectively. The category with the highest probability value is the final diagnosis result, that is, the fault category of the turnout switch machine.

[0075] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for diagnosing faults in a turnout switch machine, characterized in that, Includes the following steps: Step S1: Collect the sound signal, vibration signal and current signal of the turnout switch machine to obtain three-mode raw data. Remove local data with unstable changes from the three-mode raw data to obtain the initial processed mode data. Step S2: Obtain historical detection data. Based on the error variation characteristics of the correlation between sound signals, vibration signals, and current signals under ideal environmental conditions in the historical detection data, obtain the error interference coefficient. Based on the error variation characteristics of the correlation between historical environmental data characteristics and the additional error influence of sound signals, vibration signals, and current signals in the historical detection data, obtain the error influence factor. Based on the error interference coefficient and the error influence factor, perform error compensation on the initial processed modal data to obtain the final processed modal data. Step S3: Linearly map the final processed modal data to obtain trimodal mapping features. Input the trimodal mapping features into the Transformer encoder for deep semantic feature representation to obtain trimodal deep semantic feature representation. If the quality score of the trimodal deep semantic feature representation obtained by quality evaluation based on signal-to-noise ratio, kurtosis, and spectral flatness is greater than a preset low quality threshold, construct a cross-attention mechanism with three pairs of modal interaction structures. Input the trimodal deep semantic feature representation into the cross-attention mechanism for cross-modal semantic interaction calculation and output three sets of modal enhancement features. Based on the three sets of modal enhancement features, obtain the fault category of the turnout switch machine.

2. The fault diagnosis method for a turnout switching machine according to claim 1, characterized in that, Step S1 includes: The variation trends of various modal signals in the three-modal raw data are statistically analyzed to obtain the variation trend characteristics to be measured. A preset baseline determination threshold is set, and local features that exceed the baseline determination threshold in the trend features to be tested are marked to obtain the local features to be tested. The temporal continuity of the local features of the marker to be tested is statistically analyzed to obtain the temporal continuity of the marker to be tested. A preset continuity determination threshold is set. If the continuity of the time sequence to be tested is greater than the continuity determination threshold, the local feature of the corresponding modal signal in the local feature of the marker to be tested is determined to be an unstable variable feature and is removed to obtain the modal data to be tested. If the continuity of the time sequence to be tested is less than or equal to the continuity determination threshold, then the local feature of the corresponding modal signal in the local feature of the label is determined to be a stable variation feature and is retained to obtain the second modal data to be tested. The data from either the first or second modal data to be tested is selected as the initial processing modal data.

3. The fault diagnosis method for a turnout switching machine according to claim 2, characterized in that, Step S2 includes: Historical detection data is acquired, and the detection error values ​​of sound signals, vibration signals, and current signals under the desired environmental conditions are extracted from the historical detection data, and the sets are used to obtain a historical error feature set. Historical three-modal data are extracted from historical detection data. Based on the correlation between any one modal signal in the historical three-modal data and the error feature values ​​corresponding to the other two modal signals in the historical error feature set, the error variation characteristics are affected. The correlation coefficient between any one modal signal in the historical three-modal data and the error feature values ​​corresponding to the other two modal signals in the historical error feature set is calculated, and the error interference coefficient is obtained by combining them.

4. The fault diagnosis method for a turnout switching machine according to claim 3, characterized in that, Step S2 also includes: The environmental data characteristics of the collected trimodal raw data are detected to obtain the environmental data characteristics to be tested, which include temperature data, humidity data and electromagnetic intensity data; The additional impact error values ​​of the sound signal, vibration signal and current signal due to environmental influence are extracted from the historical detection data after the interference between the three modes of signals has been compensated. The additional error values ​​of the historical sound, historical vibration and historical current are obtained. Based on the characteristics of the environmental data to be tested, historical environmental data characteristics are obtained. Based on the correlation between the historical environmental data characteristics and historical additional sound error, historical additional vibration error, and historical additional current error, the error variation characteristics are affected. The correlation factors between the historical environmental data characteristics and historical additional sound error, historical additional vibration error, and historical additional current error are statistically analyzed, and the error influence factors are obtained by combining them. Based on the error interference coefficient, error influence factor, and characteristics of the test environment data, error compensation is performed on the initial processed modal data to obtain the final processed modal data.

5. The fault diagnosis method for a turnout switching machine according to claim 4, characterized in that, Step S3 includes: The signal-to-noise ratio, kurtosis, and spectral flatness of the three-modal deep semantic feature representations are statistically analyzed to obtain pre-evaluation index features; Based on the aforementioned pre-evaluation index characteristics, the quality of each modality feature in the three-modal deep semantic feature representation is evaluated to obtain the quality score to be judged. A low quality threshold is preset. If the quality score to be judged is less than or equal to the low quality threshold, the sound signal, vibration signal and current signal of the turnout switch machine are re-acquired and error-compensated to obtain modal data re-detection information.

6. The fault diagnosis method for a turnout switching machine according to claim 5, characterized in that, Step S3 also includes: If the quality score to be judged is greater than the low quality threshold, output the feature fusion preprocessing information; Based on the feature fusion preprocessing information, and based on the peak delay time estimation of the cross-correlation function, the peak delay time of the historical three-mode data in the historical detection data is estimated to obtain the delay time features; Based on peak detection of normalized cross-correlation function, the delay direction and magnitude of historical three-mode data are statistically determined, and the delay direction and magnitude features are obtained; Based on the aforementioned delay direction and magnitude characteristics, a directed acyclic causal graph is constructed; Based on the aforementioned delay time characteristics and the directed acyclic causal graph, a cross-attention mechanism with three pairs of modal interaction structures is constructed.

7. The fault diagnosis method for a turnout switching machine according to claim 6, characterized in that, Step S3 also includes: The three-modal deep semantic feature representations are input into a cross-attention mechanism for cross-modal semantic interaction computation, and three sets of modality-enhanced features are output. Based on the Sigmoid adaptive gating feature fusion mechanism, the importance weight of each modality feature in the three sets of modality enhancement features is adaptively calculated, and dynamic weighted fusion is performed to output the fused feature; Obtain the number of known fault types of the turnout switch machine, construct a fully connected classification layer, wherein the input dimension of the fully connected classification layer is the same as the dimension of the fused feature, and the output dimension is equal to the number of known fault types; The fused features are input into a fully connected classification layer and then mapped by a Softmax function to output the fault category of the turnout switch machine.

Images