Brake system solenoid valve fault diagnosis method based on wavelet packet decomposition and BP neural network

Abstract

The invention discloses a braking system solenoid valve fault diagnosis method based on wavelet packet analysis and a BP neural network. The method is used for rapidly positioning a fault solenoid valve and diagnosing a fault reason. According to the method, firstly, the working current of the electromagnetic valve of the braking system is collected in real time through vehicle-mounted equipment, then a fault diagnosis model is operated through a calculation processing unit of the vehicle-mounted equipment, and the fault diagnosis model adopts wavelet packet analysis to perform three-layer decomposition on original current signals collected by the vehicle-mounted equipment; energy characteristics of different frequency bands are extracted based on an energy entropy mode, and finally, the extracted energy characteristics are classified by adopting a BP neural network algorithm. The effectiveness of the method is verified in simulation data and real object fault data, and the method is used for fault diagnosis of the electromagnetic valve of the rail vehicle braking system, so that the handling efficiency of rail traffic maintenance personnel is greatly improved.

Description

Technical Field

[0001] The invention relates to a rail vehicle braking system, in particular to a braking system electromagnetic valve fault diagnosis method based on wavelet packet decomposition and BP neural network. Background Art

[0002] As one of the most important sub-components in the rail vehicle braking control system, the solenoid valve plays an important role in applying, relieving and preventing air braking force. Once the solenoid valve fails, it may cause problems such as brake failure and wheel rubbing, which in turn leads to train delays, passenger clearance and rescue. According to after-sales operation and maintenance feedback, there are many faults reported by trains during self-inspection and operation due to solenoid valve failure. When the fault occurs, maintenance personnel need to check step by step according to the maintenance manual, and then locate the solenoid valve component by means of disassembly and replacement, and finally the technical department will disassemble and analyze the faulty valve. This relatively passive fault diagnosis solution not only consumes a lot of manpower and material resources, but also fails to detect the problem in a timely manner, resulting in low troubleshooting efficiency. Summary of the invention

[0003] In view of the problems existing in the prior art, it is considered to analyze the solenoid valve coil current to determine whether the solenoid valve is faulty. When the solenoid valve is stuck or other faults occur, the coil current will change accordingly, and the characteristics of this change may be difficult to extract in the time domain. Therefore, it is considered to extract the characteristics of the coil current in the frequency domain.

[0004] The present invention provides a method for fault diagnosis of electromagnetic valves in brake systems based on wavelet packet analysis and BP neural network, which is used to quickly locate faulty electromagnetic valves and diagnose the causes of faults. The effectiveness of the method has been verified in simulation data and real fault data, and its application in fault diagnosis of electromagnetic valves in rail vehicle brake systems has greatly improved the handling efficiency of rail transit maintenance personnel.

[0005] The purpose of the present invention is achieved through the following technical solutions.

[0006] A brake system solenoid valve fault diagnosis method based on wavelet packet decomposition and BP neural network, comprising the following steps:

[0007] Step 1: Collect the coil current of the solenoid valve under normal and fault conditions;

[0008] Step 2: Use wavelet packet decomposition to decompose the current signal under various working conditions into three layers to obtain the decomposed wavelet packet coefficients; Step 3: Extract the energy entropy characteristics of the third layer wavelet packet coefficients.

[0009] Step 4: Use the energy features extracted in step 3 to train the BP neural network model;

[0010] Step 5: For the sample to be tested, extract the energy entropy characteristics of the sample according to steps 2 and 3;

[0011] Step 6: Bring the features extracted in step 5 into the trained neural network model for prediction output to determine whether the current solenoid valve is faulty.

[0012] The step 3 is specifically as follows:

[0013] First, calculate the wavelet packet energy of 8 nodes in the third layer. The energy calculation formula is as follows:

[0014]

[0015] Where: S 3,i (k) is the kth wavelet packet coefficient of the i-th node in the third layer, and M represents the number of wavelet packet coefficients of the node;

[0016] Then the energy ratio of each node in the third layer is used as the final feature, namely:

[0017]

[0018] The mathematical description of the BP neural network model in step 4 is: in is the output value of the jth neuron in the n-1th layer, represents the weight from neuron j to neuron i in the n-1th layer to the nth layer, is the bias of neuron i in the nth layer, and function f is the activation function.

[0019] The solution steps of the BP neural network model in step 4 include forward propagation and back propagation. Forward propagation is to gradually solve the output layer according to the known connection weights and biases, and back propagation is to update the weights and biases of the model by reverse derivation according to the solved errors.

[0020] Compared with the prior art, the advantages of the present invention are: the present invention is used to quickly locate the faulty solenoid valve and diagnose the cause of the fault. The method first collects the working current of the brake system solenoid valve in real time through the on-board equipment, and then runs the fault diagnosis model through the on-board equipment computing processing unit, wherein the fault diagnosis model uses wavelet packet analysis to perform a three-layer decomposition of the original current signal collected by the on-board equipment, and then uses an energy entropy-based method to extract energy features of different frequency bands, and finally uses the BP neural network algorithm to classify the extracted energy features. The effectiveness of the proposed method has been verified in simulation data and actual fault data, and the use of the method for fault diagnosis of the solenoid valve of the rail vehicle brake system has greatly improved the handling efficiency of rail transit maintenance personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1This is the solenoid valve structure diagram.

[0022] Figure 2 Schematic diagram of wavelet packet decomposition.

[0023] Figure 3 This is the numerical simulation signal waveform.

[0024] Figure 4 It is the 3-layer wavelet packet decomposition coefficient of the numerical simulation signal.

[0025] Figure 5 Energy entropy eigenvalue distribution of numerical simulation signal based on wavelet packet decomposition coefficients.

[0026] Figure 6 It is a neural network model.

[0027] Figure 7 The figure is a fault diagnosis flowchart.

[0028] Figure 8 is the solenoid valve coil current under three working conditions.

[0029] Fig. 9 is the energy ratio of each frequency band under three working conditions.

[0030] Fig.10 Prediction results for the simulation test set based on BP neural network.

[0031] Fig.11 It is the working current acquisition circuit of the solenoid valve.

[0032] Fig.12 This is a comparison chart of the waveforms of the working circuits of three types of solenoid valves.

[0033] Fig.13 The energy ratio of each frequency band of the measured current of three types of solenoid valves.

[0034] Fig.14 This is the prediction result of the measured data test set based on the BP neural network. DETAILED DESCRIPTION

[0035] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] There are many types of solenoid valves in the brake control system, including emergency solenoid valves, high-speed solenoid valves for common braking, forced release solenoid valves, etc. According to the power supply range, they can be divided into DC24V and DC110V, according to the channel, they can be divided into 2-position 3-way and 2-position 2-way, and according to the closing mode, they can be divided into normally open and normally closed solenoid valves. The following is a detailed description of the high-speed solenoid valve for common braking as an example.

[0037] The high-speed solenoid valve used in the rail vehicle braking system is mainly composed of valve body, coil, static iron core, moving iron core, valve core, reset spring and other main components. When the coil is energized or de-energized, the operation of the magnetic core will cause the fluid to pass through the valve body or be cut off to achieve the purpose of changing the direction of the fluid. Figure 1 As shown: when the solenoid valve is energized, the electromagnetic force overcomes the spring force of the moving iron core and pushes the moving iron core to move. Under the action of the valve core spring, the plug inside the valve core is sealed with the upper valve port. At this time, port 1 and port 2 are connected.

[0038] When the solenoid valve loses power, the moving iron core moves under the action of the reset spring, pushing the valve core and compressing the valve core spring. The valve core plug contacts and seals the lower valve port, port 1 and port 2 are cut off, and port 2 and port 3 are connected.

[0039] The normally closed (NC) solenoid valve takes in air from port 1 and port 3 is connected to the atmosphere; the normally open (NO) solenoid valve takes in air from port 3 and port 1 is connected to the atmosphere.

[0040] There are many reasons for solenoid valve failure, such as spring failure, valve core sticking, air leakage, coil degradation, etc.

[0041] (1) Spring failure: The essence of spring failure is that the magnitude of the spring force changes during the operation of the solenoid valve. When the spring is under high temperature and high pressure conditions, it may cause the spring to fail.

[0042] (2) Valve core stuck: The essence of the stuck fault is that the friction force has changed. The clearance between the solenoid valve sleeve and the valve core is very small, and it is usually assembled in one piece. When mechanical impurities are introduced or there is too little lubricating oil, it is easy to get stuck.

[0043] (3) Air leakage: When the sealing gasket is damaged or the sliding valve is worn, the solenoid valve will leak air, which will lead to insufficient air pressure and make it difficult to open and close the pneumatic valve.

[0044] (4) Coil degradation or burnout: The causes of coil degradation and burnout include the coil being damp, which causes poor insulation and magnetic leakage, resulting in excessive current in the coil and burning. In addition, the spring is too hard, the reaction force is too large, the number of coil turns is too small, and the suction force is insufficient, which can also cause the coil to burn out.

[0045] Energy entropy eigenvalue extraction based on wavelet packet decomposition:

[0046] Wavelet packet is based on wavelet analysis theory. Through the expansion and translation of wavelet basis function parameters, it can subdivide time at high frequencies and subdivide frequency at low frequencies, thus realizing localized analysis of signals in time and frequency domains. In wavelet analysis theory, continuous wavelet transform, discrete wavelet transform and discrete wavelet transform decomposition and reconstruction are its core theories.

[0047] Wavelet packet definition: In multi-resolution analysis, and ψ(t) satisfy the two-scale equation. Let W1(t)=ψ(t), which can be written as a recursive formula:

[0048]

[0049] In the formula, h(k) and g(k) satisfy ∑h(n-2k)h(n-2l)=δ kl , g(k)=(-1)kh(ik), n, k, l, i are all integers. The function system {W n (t)} n∈N called by The determined orthogonal wavelet packets.

[0050] In the process of multi-layer signal decomposition, the signal is generally decomposed into a wavelet packet tree. The three-layer wavelet packet tree decomposition process is as follows: Figure 2 shown.

[0051] It can be seen from the above decomposition that wavelet packet decomposition further refines the detail features, that is, the high-frequency part of the signal, on the basis of wavelet decomposition. The signal is decomposed into different frequency bands through high-pass and low-pass filter groups. Based on the wavelet packet coefficients obtained by decomposition, the energy entropy method is further used to quantitatively analyze the wavelet packet coefficients. The following takes the three-layer wavelet packet decomposition as an example, combined with formulas (3)-(5), to describe in detail the energy entropy eigenvalue extraction process based on the wavelet packet decomposition coefficients:

[0052] (1) Perform three-layer wavelet packet decomposition on the original signal to obtain the decomposed wavelet packet coefficients;

[0053] (2) Extract the energy entropy features of the third layer wavelet packet coefficients. First, calculate the energy entropy of the wavelet packet of the 8 nodes in the third layer.

[0054] The energy calculation formula is as follows:

[0055]

[0056] Where: S 3,i (k) is the kth wavelet packet coefficient of the i-th node in the third layer, and M represents the number of wavelet packet coefficients of this node.

[0057] (3) Then the energy ratio of each node in the third layer is used as the final feature, that is:

[0058]

[0059] The following is a specific numerical simulation example to illustrate the process of wavelet packet decomposition and energy entropy eigenvalue extraction.

[0060] a. Assume that the signal x is composed of a 100Hz sine wave and a 300Hz sine wave superimposed, that is, x = sin(2π*100t)+sin(2π*300t), the sampling frequency is 1024Hz, and the sampling time is 0.5s. The original signal is as follows: Figure 3 As shown, the horizontal axis represents the sampling point, and the vertical axis represents the signal amplitude;

[0061] b. Perform 3-layer wavelet packet decomposition on the signal x to obtain the wavelet packet coefficients of the 8 nodes in the 3rd layer. The coefficient waveforms of all nodes in the 3rd layer are as follows: Figure 4 As shown in the figure, the horizontal axis and the vertical axis represent the number and amplitude of wavelet packet coefficients respectively. It can be seen from the figure that the wavelet packet coefficients are mainly concentrated in the second and fifth frequency bands, namely 64-128Hz and 256-320Hz, corresponding to the frequencies of the two sine wave signals of the source signal x, 100Hz and 300Hz.

[0062] c. Based on the obtained wavelet packet coefficients, the wavelet packet energy of the 8 nodes in the third layer is calculated using formula (3), and then the proportion of wavelet energy of each node is calculated using formulas (4) and (5), that is, the final energy entropy feature is extracted. The energy entropy of the sinusoidal superposition signal is as follows: Figure 5 As shown; it can be quantitatively seen from the figure that the source signal energy is mainly concentrated in the 2nd frequency band and the 5th frequency band, namely 64~128Hz and 256~320Hz, corresponding to the frequencies of the two sinusoidal wave signals of the source signal x.

[0063] BP neural network is the most widely used neural network model, which includes input layer, hidden layer and output layer. Its model structure is as follows Figure 6 As shown, the mathematical description is as follows:

[0064]

[0065] in is the output value of the jth neuron in the n-1th layer, represents the weight from neuron j to neuron i in the n-1th layer to the nth layer, is the bias of neuron i in the nth layer, and function f is the activation function.

[0066] The neural network solution steps include forward propagation and back propagation. Forward propagation is to gradually solve the output layer based on the known connection weights and biases, and back propagation is to update the weights and biases of the model by reverse derivative based on the solved error.

[0067] like Figure 7 As shown, the steps include:

[0068] Step 1: Collect the coil current of the solenoid valve under normal and fault conditions;

[0069] Step 2: Use wavelet packet decomposition to perform three-layer decomposition on the current signal under various working conditions to obtain the decomposed wavelet packet coefficients;

[0070] Step 3: Extract the energy entropy features of the third layer wavelet packet coefficients according to the method described in formulas (3)-(5).

[0071] Step 4: Use the energy features extracted in step 3 to train the neural network model;

[0072] Step 5: For the sample to be tested, extract the energy entropy characteristics of the sample according to steps 2 and 3;

[0073] Step 6: Bring the features extracted in step 5 into the trained neural network model for prediction output to determine whether the current solenoid valve is faulty.

[0074] Combined with the frequency of solenoid valve failures on the actual line, the simulink simulation platform is used to simulate the three working conditions of the solenoid valve: normal, stuck, and spring failure. 40 sets of current data are generated under each working condition, and the data collection frequency is 12000Hz. The current data under the three working conditions are as follows Figure 8 As shown in the figure, the wavelet packet tree decomposition in Section 2.1 is used to decompose the current simulation data of each working condition into three layers to obtain the wavelet packet coefficients of each layer after decomposition. Then, the decomposed wavelet packet coefficients are extracted using formulas (3), (4), and (5). The energy entropy characteristics under the three working conditions are as follows: Fig. 9 shown.

[0075] Fig. 9 The horizontal axis represents the 8 frequency bands of the third layer after wavelet packet decomposition, and the vertical axis is the energy percentage in the frequency band. According to the Nyquist sampling theorem, the sampling frequency of 12000Hz can only retain the frequency components within 6000Hz of the signal, so each frequency band contains a frequency of 750Hz, and the frequencies increase from 1 to 8.

[0076] from Figure 8 It can be seen that due to the complete stagnation of the valve core, the solenoid valve current has been rising and maintained at the rated current of the solenoid valve, so its frequency domain characteristics are relatively concentrated, all concentrated in the first frequency band, that is, 0-749Hz. However, the current waveforms under normal working conditions and spring failure conditions are relatively similar. The difference is that under the spring failure condition, the solenoid valve current responds faster. The reason is that under the spring failure condition, the valve core overcomes the spring force during movement, so it will reach the pull-in process faster. In the frequency domain characteristics, that is, under the spring failure condition, the characteristic energy of the third frequency band accounts for less than the energy under normal conditions. The discrimination of this part is mainly based on the results of feature extraction of simulation data and actual data.

[0077] After the above feature extraction, the 120 sets of current data generated by the simulink simulation platform are divided into training set and test set at a ratio of 3:1. That is, 90 sets of samples are brought into the BP neural network for training, and the remaining 30 sets are brought into the trained model as test sets. The prediction results of the model are as follows: Fig.10 As shown;

[0078] Fig.10 The horizontal and vertical axes represent the test samples and sample categories, where categories 0, 1, and 2 represent normal, spring failure, and valve core stuck conditions, respectively. The red mark in the figure is the actual label of the sample, and the green mark is the predicted result of the sample. It can be seen from the figure that only one prediction error occurred in 30 test samples, and the accuracy rate reached 96.7%.

[0079] In order to further verify the effectiveness of this method from the engineering implementation aspect, physical fault verification was carried out. A total of three types of solenoid valves, normal solenoid valve, valve core stuck solenoid valve, and return spring failure solenoid valve, were used for physical fault verification experiments. Fig.11 The acquisition circuit shown in the figure collects the working current of these three types of solenoid valves, such as Fig.12 shown.

[0080] Similarly, the wavelet packet tree decomposition in Section 2.1 is used to perform three-layer decomposition of the measured current data of three types of solenoid valves: normal, valve core stuck, and return spring failure, and the wavelet packet coefficients of each layer after decomposition are obtained. Then, the decomposed wavelet packet coefficients are extracted using formulas (4), (5), and (6). The energy entropy characteristics of the three different types of solenoid valves: normal, valve core stuck, and return spring failure are as follows Fig.13 shown.

[0081] From this figure, it can be seen that the energy entropy characteristic distribution extracted after wavelet packet decomposition of the working current of the three types of solenoid valves, namely normal, valve core stuck, and return spring failure, is consistent with the simulation data analysis results, which further verifies the effectiveness of the method proposed in this paper from the aspect of engineering implementation.

[0082] Consistent with the simulation data classification test process, in order to verify the accuracy of physical fault classification based on the features extracted from the measured current data, the 120 sets of measured current data collected were extracted with the above features and divided into training set and test set at a ratio of 3:1. That is, 90 sets of samples were brought into the BP neural network for training, and the remaining 30 sets were brought into the trained model as test sets. The prediction results of the model are shown in Figure 2. Fig.14 shown.

[0083] Similarly, this figure shows the test samples and sample categories, where categories 0, 1, and 2 are normal, return spring failure, and valve core stuck conditions, respectively. The red mark in the figure is the actual label of the sample, and the green mark is the predicted result of the sample. It can be seen from the figure that the classification results of the 30 test samples are accurate, with an accuracy rate of 100%. The results show that this method can achieve accurate detection of solenoid valve faults and rapid diagnosis of fault types.

[0084] In order to solve the problem of difficulty in detecting solenoid valve faults during operation and maintenance, a solenoid valve fault diagnosis method based on wavelet packet analysis and BP neural network is proposed, and simulation research and experimental verification are carried out. The method is feasible. This method uses wavelet packet decomposition to extract the frequency domain signal of the solenoid valve coil current, and then diagnoses the characteristic signal through the neural network algorithm. This method can not only efficiently handle the mainline operation fault, but also provide further guidance for the analysis of the return of faulty solenoid valves to the factory. The solenoid valve fault diagnosis and fault warning method proposed in this paper needs to add a solenoid valve current sampling circuit on the basis of the existing system, which has certain engineering implementation difficulties. It is necessary to consider the needs of solenoid valve current collection and fault warning in the new train design process.

Claims

1. A brake system solenoid valve fault diagnosis method based on wavelet packet decomposition and BP neural network, characterized in that the steps include: Step 1: Collect the coil current of the solenoid valve under normal and fault conditions; Step 2: Use wavelet packet decomposition to perform three-layer decomposition on the current signal under various working conditions to obtain the decomposed wavelet packet coefficients; Step 3: Extract energy entropy features from the third layer wavelet packet coefficients. Step 4: Use the energy features extracted in step 3 to train the BP neural network model; Step 5: For the sample to be tested, extract the energy entropy characteristics of the sample according to steps 2 and 3; Step 6: Bring the features extracted in step 5 into the trained neural network model for prediction output to determine whether the current solenoid valve is faulty.

2. The method for diagnosing faults of electromagnetic valves in a brake system based on wavelet packet decomposition and BP neural network according to claim 1 is characterized in that The step 3 is specifically as follows: First, calculate the wavelet packet energy of 8 nodes in the third layer. The energy calculation formula is as follows: Where: S 3,i (k) is the kth wavelet packet coefficient of the i-th node in the third layer, and M represents the number of wavelet packet coefficients of the node; Then the energy ratio of each node in the third layer is used as the final feature, namely:

3. The method for diagnosing faults of electromagnetic valves in a brake system based on wavelet packet decomposition and BP neural network according to claim 1 is characterized in that The mathematical description of the BP neural network model in step 4 is: in is the output value of the jth neuron in the n-1th layer, represents the weight from neuron j to neuron i in the n-1th layer to the nth layer, is the bias of neuron i in the nth layer, and function f is the activation function.

4. The method for diagnosing faults of electromagnetic valves in a brake system based on wavelet packet decomposition and BP neural network according to claim 3 is characterized in that The solution steps of the BP neural network model in step 4 include forward propagation and back propagation. Forward propagation is to gradually solve the output layer according to the known connection weights and biases, and back propagation is to update the weights and biases of the model by reverse derivation according to the solved errors.

Images