Turnout switch machine intelligent diagnosis system and method based on small sample multi-fault mode

Abstract

The invention provides a turnout switch machine intelligent diagnosis system and method based on a small sample multi-fault mode, and the system comprises a data management module which collects equipment operation condition data in real time, carries out the data preprocessing, and stores the data in a source database in the data management module; the fault diagnosis module is used for identifying and outputting normal data and fault data G0 by taking the data in the source database as input; the data enhancement module is used for amplifying the fault data G0 identified by the fault diagnosis module into new fault data G1, and storing the fault data G0 and the new fault data G1 into an amplification database in the data management module; and the fault positioning module is used for identifying and outputting a fault type by taking the data in the amplification database as input. According to the invention, the operation process of the equipment is monitored and managed.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial equipment fault diagnosis, and in particular relates to an intelligent diagnosis system and method for turnout machines based on small sample and multiple fault modes. Background Art

[0002] With the rapid expansion of urban rail transit networks, the reliability of turnout machines, as the core equipment for controlling the direction of train operation, has become a key factor affecting the transportation efficiency and safety of rail transit systems.

[0003] Taking the subway operating system as an example, with a large-scale network serving over 4 million passengers daily and over 500 kilometers of operating mileage, its turnout equipment must perform over 10 million high-precision switching operations annually. However, current engineering practices face two challenges: first, the structural flaws of traditional maintenance models, such as rigid cycles and delayed response; second, a shortage of fault data and the complex mechanisms of equipment failures, which can be both hidden and sudden. Intelligent diagnostic systems can effectively and quickly locate equipment failure mechanisms, significantly reducing maintenance costs and avoiding resource waste, thereby optimizing operational management processes and providing strong guarantees for the safe and stable operation of rail transit systems. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent diagnosis system and method for turnout machines based on small sample and multiple fault modes, which can monitor and manage the operation of the equipment, facilitate equipment managers and enterprise managers to have a more comprehensive understanding of the health status of all equipment, and facilitate more refined management. The technical solutions adopted are: An intelligent diagnosis system for turnout machines based on small sample and multiple fault modes, comprising: The data management module collects equipment operation data in real time, performs data preprocessing, and stores the data in the source database of the data management module; Fault diagnosis module, which takes data from the source database as input, identifies and outputs normal data and fault data G0; A data enhancement module is used to amplify the fault data G0 identified by the fault diagnosis module into new fault data G1, and store the fault data G0 and the new fault data G1 in the amplification database in the data management module; The fault location module takes the data in the augmented database as input to identify and locate the fault type.

[0005] Preferably, the fault diagnosis module is a trained Hidden Markov Model HMM.

[0006] Preferably, the fault diagnosis module trains the parameters of the hidden Markov model HMM based on the dynamic Hippo optimization algorithm DHOA module and the Baum-Welch algorithm module; The dynamic Hippo optimization algorithm DHOA module converts the exploration weight in the Hippo optimization algorithm HOA ( is a random number between 0 and 1) is replaced by a dynamic exploration weight ; The M-step of the Baum-Welch algorithm module is based on dynamic exploration weights Update the mean and variance and train the Hidden Markov Model HMM.

[0007] Preferably, the dynamic exploration weight for: = ; in, -The current iteration number of the dynamic Hippo optimization algorithm DHOA module, - Total number of iterations of the DHOA module of the dynamic Hippo optimization algorithm.

[0008] Preferably, the moving flip bottleneck convolution MBConv of the fault location module introduces a sparse squeeze and excitation module, adaptive channel compression and H-Swish activation function.

[0009] Preferably, the data enhancement module is an exponential moving average vector quantization variational autoencoder EMA_VQVAE.

[0010] Preferably, the data enhancement module includes a residual block Res-Stack, and a vector quantization method with an exponential moving average mechanism is applied to the output of the first convolutional layer after the residual block Res-Stack. The sparse fault sample representation capability is enhanced through a stable code book update mechanism, effectively solving the problems of gradient oscillation and coding space collapse.

[0011] Specifically, an intelligent diagnosis system for turnout machines based on small sample and multiple fault modes includes a data management module, a fault diagnosis module, a data enhancement module and a fault location module.

[0012] A data management module, which collects equipment operation data in real time through a circuit sensing system, and stores the data in a source database after pre-processing; The fault diagnosis module uses data from the source database as input. It uses the Dynamic Hippo Optimization Algorithm (DHOA) and the Baum-Welch algorithm to collaboratively optimize the parameters of the Hidden Markov Model (HMM). The proposed DHO-BW_HMM can quickly and accurately perform binary classification between normal and fault states, thereby addressing timeliness issues.

[0013] The data enhancement module is used to take the fault data in the source database of the data management module as input and use the exponential moving average vector quantized variational autoencoder (EMA_VQVAE) to amplify the data; The fault location module uses data from the augmented database as input and locates faults through an improved and efficient network. It accurately identifies specific faults from fourteen types of faults, displays them in real time, and pushes them to users, issuing alarms for equipment failures and effectively dispatching relevant personnel for repairs.

[0014] Furthermore, the data management module includes data acquisition function, data preprocessing function and database management function.

[0015] Data acquisition function, used to collect real-time data from the circuit sensor system on the turnout machine and store it in the source database; Data preprocessing function, used to standardize the original data; The database management function is used for data transmission between the source database and the augmented database. The source database uses MySQL to store and manage real-time data, and uploads it to the database for storage at a fixed time every day. It is used for data analysis functions of modules such as fault diagnosis and data enhancement, and can also provide users with real-time data query functions; the augmented database is used to store new fault data and source fault data generated by the data enhancement module, and is used for data analysis functions of modules such as fault location, and can also provide users with visual data query functions.

[0016] The fault diagnosis module adopts the classification and recognition method, using DHO-BW_HMM to quickly and accurately perform binary classification of normal state and fault state. The corresponding method is as follows.

[0017] First define the forward variables , backward variables , and then find and .

[0018] Indicates the slave state i Transfer to state j The expected number of Indicates the slave state i The expectation that a state transition will occur.

[0019] :Given observation sequence O and model , define the forward probability as the hidden state at time t , the observation state sequence is The probability of .

[0020] The specific process is: First, initialize the forward probability of each hidden state at time 1: ; Then recursively deducing time 2 to time T+1 The forward probability ; The final calculation result is .

[0021] : Define the backward probability as the hidden state at time t , the observation state sequence is The probability of .

[0022] The specific process is: First initialize the time T The backward probability of each hidden state ; Second, the recursive time T-1 Backward probability to time 1 ; The final calculation result is .

[0023] According to the forward and backward variables, The maximum we can get is: .

[0024] : The probability that the hidden state of the device at time t is i and the hidden state at time t+1 is transferred to j.

[0025] ; :Device at time t The hidden state of i probability.

[0026] ; Then cross EM Iterate until the parameters converge.

[0027] use represents the initial state probability revaluation.

[0028] ; use represents the initial state probability revaluation.

[0029] ; In the formula Indicates the slave state i Transfer to state j The expected number of Indicates that it is in state i The expected number of times.

[0030] in, for The current optimal value.

[0031] Observation probability distribution The Gaussian model is used for fitting. In the following formula represent weight, mean and variance respectively.

[0032] use Represents the revaluation of the mean, the specific formula is as follows: ; = ; ; in, for The current optimal value. for Hidden state at the moment A collection of Similarly, calculate the variance :use Represents the revaluation of variance. The specific formula is as follows: ; Then, the trained DHO-BW_HMM is obtained: .

[0033] in The elements in ; The elements in ; The elements in are the observation probability distributions ,Right now 、 .

[0034] for The current optimal value; for Observation sequence at time O ; The final model Before and after variables 、 , and then use the forward algorithm or backward algorithm to perform accurate binary classification.

[0035] Furthermore, the data enhancement module uses a vector quantization variation autoencoder with an exponential moving average update mechanism to enhance the fault sample data, and then stores it in the augmentation database.

[0036] The vector quantization mutation autoencoder with exponential moving average update mechanism uses exponential moving average (EMA) to update the codebook, and the total loss function removes the vector quantization loss term. , the two terms represent reconstruction loss and commitment loss respectively. In the quantization process, they can be abstracted as: ; Represents the current mini-batch sample is mapped to the codebook entry The number of eigenvectors of is the smoothing attenuation coefficient. Third, the commitment loss , by constraining the geometric distance between the encoder output and the selected embedding, it prevents the latent space from diverging due to lack of regularization.

[0037] In addition, we selected Learned Perceptual Image Patch Similarity (LPIPS) and Structural Similarity (SSIM) as evaluation metrics. LPIPS measures image perceptual quality by distance in deep feature space, capturing high-level semantic differences to which human visual sensitivity is sensitive. SSIM quantifies pixel-level structural similarity across three dimensions: brightness, contrast, and structure. Combining these two metrics allows for a comprehensive assessment of the perceptual quality and structural fidelity of reconstructed data.

[0038] Furthermore, the fault location module uses an improved high-efficiency network to locate multiple faults and generate alarms. This improved high-efficiency network integrates a sparse squeeze and excitation (SE) module, adaptive channel compression, and the H-Swish activation function. This method accurately identifies fault types, visualizes them, and generates alarms for users.

[0039] An intelligent diagnosis method for a turnout machine based on small sample and multiple fault modes, characterized by comprising the following steps: Step S1: The data management module pre-processes the real-time device data collected by the sensor and stores it in the source database; Step S2: The fault diagnosis module requests device data from the source database and determines whether the device is normal or faulty using a statistical diagnosis model. Step S3: The data enhancement module requests device data from the source database, converts the original data into two dimensions, and amplifies the fault data using an exponential moving average vector quantization variational autoencoder. The original fault data and the amplified fault data are stored in the amplified database. Step S4: The fault location module inputs the data in the augmented database and uses the improved efficient network to accurately locate 14 types of faults, alarms the fault status in real time, and dispatches corresponding tools and personnel for timely maintenance.

[0040] Compared with the prior art, the advantages of the present invention are: 1. This invention identifies and manages the operational health of railway turnout machines, enabling equipment managers and enterprise managers to gain a more comprehensive understanding of the health of all equipment and facilitates more refined management, resolving two key issues raised. Specifically, fault diagnosis and fault location modules can rapidly identify, locate, and issue alerts regarding the health of industrial equipment, enabling more efficient dispatch of maintenance personnel.

[0041] 2. The present invention integrates the dynamic Hippo algorithm and the Baum-Welch algorithm and builds the DHO-BW_HMM. The processed data is input into the model, and the binary classification is performed through the forward and backward algorithm and the system is visualized to the user.

[0042] 3. This embodiment of the present invention improves efficient networks by integrating a sparse squeeze and excitation (SE) module, adaptive channel compression, and the H-Swish activation function. This allows for faster and more efficient fault location, rapid identification of fault types, and alarm generation, which can then be visually displayed to users through the system.

[0043] 4. The present invention systematizes the functions of the above-mentioned multiple modules to develop an intelligent diagnostic system, which realizes the management and diagnosis of current and power data of railway turnout machines, and improves the efficiency of scheduling management personnel and staff. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a diagram of the system implementation architecture of the present invention; Figure 2This is a flowchart of the system workflow of the present invention; Figure 3 Schematic diagram of the fault diagnosis method of the present invention; Figure 4 Schematic diagram of the data enhancement method using EMA_VQVAE in the present invention; Figure 5 This is a schematic diagram of the fault location method using IEfficientNet in the present invention; Figure 6 This is a comparative experimental result diagram corresponding to the fault diagnosis method of the present invention; Figure 7 This is a diagram of experimental results corresponding to the fault location method of the present invention; Figure 8 This is the experimental result diagram obtained based on the original EfficientNet model; Figure 9 The experimental results obtained by adding sparse squeezing and excitation modules to the original EfficientNet model. DETAILED DESCRIPTION

[0045] The following is a more detailed description of the intelligent diagnostic system and method for turnout machines based on small sample sizes and multiple fault modes, with reference to schematic diagrams. Preferred embodiments of the present invention are shown, and it should be understood that those skilled in the art may modify the present invention described herein while still achieving the beneficial effects of the present invention. Therefore, the following description should be understood as generally known to those skilled in the art and not as a limitation of the present invention.

[0046] like Figures 1 to 9 The present invention provides an intelligent diagnosis system and method for turnout machines based on small sample and multiple fault modes, aiming to perform fault diagnosis and fault location for railway turnout machines with small sample and multiple faults, including a system solution and a method solution.

[0047] In the intelligent diagnosis system solution for railway turnout machines, Figure 1 As shown, it includes data management module, fault diagnosis module, data enhancement module, fault location module, etc.

[0048] The data management module primarily consists of three functional components: data acquisition, data preprocessing, and database management. The data acquisition function collects real-time data from the circuit sensor system on the turnout machine and stores it in the database. This real-time data includes operational circuit data, including current and power changes during operation. The data sampling frequency is 400ms.

[0049] The data preprocessing function is used to standardize the raw data and generate original features. First, the collected raw data is cleaned, outliers are deleted, missing values are filled, etc., and then standardized to generate original features.

[0050] The database management function is used to transfer data between the source database and the augmented database. The source database uses MySQL to store and manage real-time data, which is uploaded to the source database at a fixed time each day for data analysis in modules such as fault diagnosis and data enhancement. The augmented database stores fault data augmented by the data enhancement module and the original fault data for data analysis in modules such as fault location, and also provides users with visual historical data query capabilities. When the source database receives a data migration request from the augmented database, it compiles an SQL statement based on the request, retrieves the corresponding real-time data from the source database, sends it to the augmented database, and saves the migration record for future queries.

[0051] The source database mainly stores the following categories: Table 1. Equipment field information table Field Name meaning type 1#Normal A-phase current Normal phase A current int 1#Normal B-phase current Normal B phase current int 1#Normal C-phase current Normal C phase current int 1#Normal Power Normal power int 2#1DQJ self-locking A-phase current 1DQJ self-closing A phase current int 2#1DQJ self-locking B-phase current 1DQJ self-closing B phase current int 2#1DQJ self-locking C-phase current 1DQJ self-closing C phase current int 2#1DQJ self-locking Power 1DQJ self-closing power int 3#2DQJ fails A-phase current 2DQJ non-reversing A-phase current int 3#2DQJ fails B-phase current 2DQJ non-rotating B-phase current int 3#2DQJ fails C-phase current 2DQJ non-reversing C-phase current int 3#2DQJ fails Power 2DQJ non-rotating pole power int 4#Indicates fault A-phase current Indicates the circuit fault A phase current int 4#Indicates fault B-phase current Indicates the circuit fault B phase current int 4#Indicates fault C-phase current Indicates the circuit fault C phase current int 4#Indicates fault Power Indicates circuit fault power int 5#Disconnect X1 A-phase current Cut off X1 A phase current int 5#Disconnect X1 B-phase current Cut off X1 B phase current int 5#Disconnect X1 C-phase current Cut off X1 C phase current int 5#Disconnect X1 Power Cut off X1 power int …… …… int 13#Unlocking A-phase current Unlock the blocking current of phase A int 13#Unlocking B-phase current Unlock the blocked B phase current int 13#Unlocking C-phase current Unlock the blocked C phase current int 13#Unlocking Power Unlocking blocking power int 14#Locking A-phase current Locking block A phase current int 14#Locking B-phase current Locking and blocking B phase current int 14#Locking C-phase current Locking and blocking C phase current int 14#Locking Power Locking resistance power int 15#Conversion A-phase current Converter resistance A phase current int 15#Conversion B-phase current Converter resistance B phase current int 15#Conversion C-phase current Conversion card resistance C phase current int 15#Conversion Power Conversion card resistance power int The expansion database mainly stores the following categories: Table 2. Field information table Field Name meaning type 1#1DQJ self-locking A-phase current 1DQJ self-closing A phase current int 1#1DQJ self-locking B-phase current 1DQJ self-closing B phase current int 1#1DQJ self-locking C-phase current 1DQJ self-closing C phase current int 1#1DQJ self-locking Power 1DQJ self-closing power int 2#2DQJ fails A-phase current 2DQJ non-reversing A-phase current int 2#2DQJ fails B-phase current 2DQJ non-polarity B phase current int 2#2DQJ fails C-phase current 2DQJ non-reversing C-phase current int 2#2DQJ fails Power 2DQJ non-rotating pole power int 3#Indicates fault A-phase current Indicates the circuit fault A phase current int 3#Indicates fault B-phase current Indicates the circuit fault B phase current int 3#Indicates fault C-phase current Indicates the circuit fault C phase current int 3#Indicates fault Power Indicates circuit fault power int 4#Disconnect X1 A-phase current Cut off X1 A phase current int 4#Disconnect X1 B-phase current Cut off X1 B phase current int 4#Disconnect X1 C-phase current Cut off X1 C phase current int 4#Disconnect X1 Power Cut off X1 power int 5#Disconnect X2 A-phase current Cut off X2 A phase current int 5#Disconnect X2 B-phase current Cut off X2 B phase current int 5#Disconnect X2 C-phase current Cut off X2 C phase current int 5#Disconnect X2 Power Cut off X2 power int 6#Disconnect X3 A-phase current Cut off X3 A phase current int 6#Disconnect X3 B-phase current Cut off X3 B phase current int 6#Disconnect X3 C-phase current Cut off X3 C phase current int 6#Disconnect X3 Power Cut off X3 power int 7#Disconnect X4 A-phase current Cut off X4 A phase current int 7#Disconnect X4 B-phase current Cut off X4 B phase current int 7#Disconnect X4 C-phase current Cut off X4 C phase current int 7#Disconnect X4 Power Cut off X4 power int 8#Disconnect X5 A-phase current Cut off X5 A phase current int 8#Disconnect X5 B-phase current Cut off X5 B phase current int 8#Disconnect X5 C-phase current Cut off X5 C phase current int 8#Disconnect X5 Power Cut off X5 power int 9#The diode broken down A-phase current Indicates normal, diode breakdown A phase current int 9#The diode broken down B-phase current Indicates normal, diode breakdown B phase current int 9#The diode broken down C-phase current Indicates normal, diode breakdown C phase current int 9#The diode broke down Power Indicates normal, diode breakdown power int 10#The diode is open A-phase current The action circuit is normal, the diode is open and the A phase current int 10#The diode is open B-phase current The action circuit is normal, the diode is open and the B phase current int 10#The diode is open C-phase current The action circuit is normal, the diode is open and the C phase current int 10#The diode is open Power The action circuit is normal, the diode open circuit power int 11#Disconnect the circuit breaker A-phase current Phase A current of the circuit breaker int 11#Disconnect the circuit breaker B-phase current Breaker B phase current int 11#Disconnect the circuit breaker C-phase current Breaker C phase current int 11#Disconnect the circuit breaker Power Breaker power int 12#Unlocking A-phase current Unlock the blocking current of phase A int 12#Unlocking B-phase current Unlock the blocked B phase current int 12#Unlocking C-phase current Unlock the blocked C phase current int 12#Unlocking Power Unlocking blocking power int 13#Locking A-phase current Locking block A phase current int 13#Locking B-phase current Locking and blocking B phase current int 13#Locking C-phase current Locking and blocking C phase current int 13#Locking Power Locking resistance power int 14#Conversion A-phase current Converter resistance A phase current int 14#Conversion B-phase current Conversion card resistance B phase current int 14#Conversion C-phase current Conversion card resistance C phase current int 14#Conversion Power Conversion card resistance power int Fault diagnosis module, which uses the pre-processed data collected by the data management module as input to perform fault diagnosis. Condition monitoring uses the DHO-BW_HMM method: As most of the data types in this project are current, the data from the source database is used as input. First, the samples are balanced and stratified, and then input into the model for classification. The process is as follows: Figure 3 As shown, the details are as follows: The first step is to initialize the initial probability ; Step 2: E step according to and , solve the time corresponding to each sample t The device is hidden i , the hidden state is transferred to j The expected value of probability and time t The hidden state of the device is i Probability .

[0052] Step 3, M Step 1, use E The expected value obtained by the step is used to update the model parameters .

[0053] Step 4: Loop EM Step until parameter When converged, the trained DHO-BW_HMM is obtained: .

[0054] in, E The expected value obtained by the step is 、 .

[0055] The final model Before and after variables 、 , and then use the forward algorithm or backward algorithm to perform accurate binary classification.

[0056] The data enhancement module is used to enhance the fault data classified by the fault diagnosis module. It will first convert the data into two dimensions and then input it into the model. The encoder extracts the features of the input data through a convolutional neural network and maps these features to a discrete latent space. In order to achieve discrete latent representation, an exponential moving average vector quantization mechanism is introduced to map the continuous latent space to a finite vector dictionary, thereby ensuring that the generated latent representation is a discrete symbol. The decoder is responsible for restoring the discrete latent representation output by the encoder back to an approximation of the original data. The decoder also uses a convolutional neural network, takes the discrete latent representation as input, and generates reconstructed data as close to the original data as possible through a learning restoration process, and then stores the enhanced fault data and the source fault data in the augmented database. The specific process is as follows: Figure 4 shown.

[0057] Figure 4 shows the augmentation of the fault data in this paper using EMA_VQVAE (data augmentation module). This model is trained and reconstructed using the fault data image in this paper, producing an image that is highly similar to the original image.

[0058] The improved efficient network (Imporved EfficientNet, IEfficientNet) is used to locate and identify multiple faults. The specific process is as follows: Figure 5 The improved efficient network model is based on the efficient network model and incorporates a sparse squeeze and excitation (SE) module, an adaptive channel compression mechanism, and an H-Swish activation function with lower computational complexity, achieving efficient reconstruction of the computational graph.

[0059] In Figure 5, IEfficientNet (fault location module) is used to diagnose and locate the 14 fault conditions in this article.

[0060] After 250 rounds of training using images of 14 different fault states as input, the model was trained to produce the final model. Then, given any image of a fault state, the system can calculate the probability of a particular fault state based on the final model.

[0061] For example: if you input a picture of fault state 2, the system can output: the probability of this picture being fault state 0 is 0.1%, the probability of being fault state 1 is 0.1%, the probability of being fault state 2 is 99%, the probability of being fault state 3 is 0.1%, the probability of being fault state 4 is 0.1%, the probability of being fault state 5 is 0.1%, the probability of being fault state 6 is 0.1%, the probability of being fault state 7 is 0.1%, the probability of being fault state 8 is 0.1%, the probability of being fault state 9 is 0.%, the probability of being fault state 10 is 0.1%, the probability of being fault state 11 is 0%, the probability of being fault state 12 is 0%, and the probability of being fault state 13 is 0%.

[0062] The model is divided into width ( ),depth( ) and resolution ( ) Consider the problem in three dimensions at the same time, and abstract the entire network as follows: ; Convolutional network layer i It can be defined as a function: , represents an operator, Indicates that the stage i Executed Second-rate, X Indicates input stage i The characteristic matrix of express X In order to explore the height, width, and number of channels. The impact of these three factors on the final accuracy will be Adding this to the formula, we can get the abstracted optimization problem, namely: ; ; ; ; in, dUsed to scale depth , r Used for image resolution and , w To scale the number of channels of the feature matrix , To limit memory , is limited by FLOPs. Then a blending factor is used φ Go to unified scaling , the specific calculation is as follows: ; ; ; ; ; ; So the total FLOPs rate can be approximated by To indicate that when the restriction ≈ 2, for any FLOPs has increased considerably times. Then fix , and search based on the formula given above to get the optimal parameters Then it is input into the network structure for training and testing, and the parameters are continuously updated to locate the fault.

[0063] Among them, it includes fault alarm, which is used to receive information on different health status sent by the fault location module, and push the health status of the equipment in real time according to the fault information. It can also make multi-platform alarms through mobile phone apps, etc.

[0064] The system workflow of the embodiment of the present invention is as follows: Figure 2 As shown, the following steps are included: 1. The data management module pre-processes the real-time device data collected by the sensors and stores it in the source database. The source database data is used as input for fault diagnosis and for data enhancement.

[0065] 2. The fault diagnosis module requests equipment data from the source database and makes abnormal judgments through DHO-BW_HMM.

[0066] The fault data is then used as the input of the data enhancement module for data amplification.

[0067] 3. The fault location module requests equipment data from the augmented database and uses the improved efficient network to identify and locate multiple faults, then issues an alarm in the system and pushes it to the equipment management personnel.

[0068] In Figure 6, the DHO-BW_HMM model in this paper is compared with the MCA-LSTM, COA_HSMM, GA_HSMM, DHO_RandomForest and RQA_SVM models under the same experimental conditions, and the average accuracy of each 10 rounds is calculated as follows: Figure 6 As shown in the figure, it can be seen that the model proposed in this paper has the highest average accuracy on the dataset of this paper.

[0069] Among them, Epochs represents the number of iterations, and Average Accuracy represents the average accuracy.

[0070] Figure 7 shows the experimental results of the fault localization method. The final model, obtained by training the IEfficientNet (ImporvedEfficientNet) model for 250 rounds, was tested on a test set. The IEfficientNet model was trained using a training set and a test set. The training set is used for model training, while the test set data is not used for model training. Instead, it is used to verify model performance and prevent overfitting.

[0071] Among them, 0 to 13 represent different types of faults. The vertical axis is the actual fault status, and the horizontal axis is the predicted fault status.

[0072] For example, the 0-0 position represents the fault state 0 (i.e., the state of "1DQJ self-closing"). 34 pieces of data were tested, and all 34 pieces of data were correctly identified, i.e., 100% identification.

[0073] The 11-11 position represents the fault state 11 (i.e., the "unlocked card group" state). 35 data were tested, 34 of which were correctly identified, with a recognition rate of 97.1%. One data was identified as state 13 (i.e., one data was identified as the "conversion blocked" state), so the recognition rate of this model for state 11 did not reach 100%.

[0074] Figures 8 and 9 are ablation experiments conducted to demonstrate the effectiveness of the two improved parts of this application (sparse squeezing and excitation module and H-Swish activation function).

[0075] Figure 8 shows the results of training with the original EfficientNet model without any improvement. The training conditions are the same as Figure 7 consistent.

[0076] Figure 9 shows an experiment in which the sparse squeeze and excitation modules are added to the original EfficientNet model. The training conditions are the same as those in Figure 7.

[0077] In summary, the present invention obtains real-time data on the operation of the equipment through a microcomputer circuit monitoring system, uses multi-fault diagnosis technology to establish an underlying model algorithm, and performs real-time evaluation and monitoring of the health status of the equipment based on the sensor data; the system includes a data management module that collects raw current, power and other data and stores them in a database after preprocessing; a fault diagnosis module uses a heuristic algorithm to optimize the statistical model for binary classification of normal and fault states; a data enhancement module uses an exponential moving average vector quantization variational autoencoder to amplify small sample fault data; a fault location module uses an improved high-efficiency network to locate multiple faults on the enhanced fault data; an intelligent diagnostic system performs data amplification, identification and alarm on the status of the equipment; the present invention can realize monitoring, fault diagnosis, data enhancement and fault location of the railway turnout switch machine, which helps management personnel to understand the health status of the equipment in a timely manner and achieve more refined management.

[0078] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other changes to the technical solution and technical content disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. An intelligent diagnosis system for turnout machines based on small sample and multiple fault modes, characterized by: include: The data management module collects equipment operation data in real time, performs data preprocessing, and stores the data in the source database of the data management module; Fault diagnosis module, which takes data from the source database as input, identifies and outputs normal data and fault data G0; A data enhancement module is used to amplify the fault data G0 identified by the fault diagnosis module into new fault data G1, and store the fault data G0 and the new fault data G1 in the amplification database in the data management module; a fault location module that uses data from the augmented database as input to identify and locate the fault type; The fault diagnosis module is a trained hidden Markov model HMM; The fault diagnosis module trains the parameters of the hidden Markov model HMM based on the dynamic Hippo optimization algorithm DHOA module and the Baum-Welch algorithm module; The dynamic Hippo optimization algorithm DHOA module converts the exploration weight in the Hippo optimization algorithm HOA Replaced with dynamic exploration weight ; The M-step of the Baum-Welch algorithm module is based on dynamic exploration weights Update the mean and variance and train the Hidden Markov Model HMM.

2. The intelligent diagnosis system for turnout machines based on small sample and multiple fault modes according to claim 1 is characterized in that: The dynamic exploration weight for: = ； in, -The current iteration number of the dynamic Hippo optimization algorithm DHOA module, - Total number of iterations of the DHOA module of the dynamic Hippo optimization algorithm.

3. The intelligent diagnosis system for turnout machines based on small sample and multiple fault modes according to claim 1 is characterized in that: The mobile flip bottleneck convolution MBConv of the fault localization module introduces a sparse squeezing and excitation module, adaptive channel compression and H-Swish activation function.

4. The intelligent diagnosis system for turnout machines based on small sample and multiple fault modes according to claim 1 is characterized in that: The data enhancement module is an exponential moving average vector quantization variational autoencoder EMA_VQVAE.

5. The intelligent diagnosis system for turnout machines based on small sample and multiple fault modes according to claim 4 is characterized in that: The data enhancement module includes a residual block Res-Stack, and a vector quantization method using an exponential moving average mechanism is applied to the output of the first convolutional layer after the residual block Res-Stack.

6. An intelligent diagnosis method for turnout machines based on small sample and multiple fault modes, characterized in that: The following steps are involved: Step S1: The data management module pre-processes the real-time device data collected by the sensor and stores it in the source database; Step S2: The fault diagnosis module requests device data from the source database and determines whether the device is normal or faulty using a statistical diagnosis model. Step S3: The data enhancement module requests device data from the source database, converts the original data into two dimensions, and amplifies the fault data using an exponential moving average vector quantization variational autoencoder. The original fault data and the amplified fault data are stored in the amplified database. Step S4: The fault location module inputs the data in the augmented database and uses the improved efficient network to accurately locate 14 types of faults, alarms the fault status in real time, and dispatches corresponding tools and personnel for timely maintenance.

Images