Fault detection method and system for series diesel-electric hybrid power system

Abstract

The invention provides a fault detection method and system for a series diesel-electric hybrid power system. The fault detection method comprises the steps that a rotating speed signal is collected; decomposing the rotating speed signal once to obtain a plurality of IMF signals, performing data processing on the IMF signals to obtain corresponding Hurst indexes, and dividing the plurality of Hurst indexes to obtain IMF signals with noise as a main part and IMF signals with effective signals as a main part; deep decomposition processing is carried out on the IMF signals with noise as the main part, secondary reconstruction is carried out on the IMF signals after deep decomposition processing and the IMF signals with effective signals as the main part, and rotating speed signals after quality improvement are obtained; and inputting the rotating speed signal into a pre-constructed xLSTM deep neural network model to detect whether the current power system has a fault or not. The diesel engine rotating speed data of the system can be analyzed, the system problem can be found accurately and timely, and the safety of the series diesel-electric hybrid power system is guaranteed.

Description

Technical Field

[0001] The present application belongs to the technical field of diesel engine fault diagnosis, and in particular, relates to a series diesel-electric hybrid power system fault detection method and a detection system thereof. Background Art

[0002] In a series diesel-electric hybrid system, the data types involved usually include data from various sensors of the diesel engine, starter motor and power battery, covering temperature, pressure, speed, voltage, current and other parameters. Among them, the speed sensor and speed signal are very important in the series hybrid system, and are the key components to achieve coordinated operation of the internal combustion engine and the electric motor.

[0003] The speed sensor is used to monitor the speed of the internal combustion engine and the starter motor, and the speed signal is the key to ensure the coordinated operation between the two. For the internal combustion engine, the speed signal can help control the ignition timing and the amount of fuel injection to ensure that the engine runs in the best condition. For the electric motor, the speed signal can help control the output power of the starter motor so that it can work in coordination with the internal combustion engine to achieve the best power generation efficiency and fuel efficiency. Therefore, the speed signal has a strong internal connection and explainability with the power system operation failure, and fault detection based on the speed signal is of great significance. At the same time, in the series diesel-electric hybrid system, due to the involvement of a variety of complex mechanical and electrical components, the sensor data is often subject to various interferences, including mechanical vibration, electromagnetic interference, etc.

[0004] At present, the common methods for improving signal quality by removing noise are mostly interval threshold method based on empirical mode decomposition (EMD-IT), soft threshold method based on variational mode decomposition (VMD-SIT) and wavelet transform (WD). These methods often have some problems, such as the problem of mode mixing after the original signal is decomposed, the problem of difficulty in determining the signal decomposition parameters and the problem of missing valuable signals in the original signal after denoising. At present, such methods not only limit the quality of signal denoising and the degree of signal quality improvement, but also cause difficulties for subsequent fault diagnosis methods based on this signal. Moreover, the common fault detection and diagnosis methods are mostly single-process methods with weak parallel capabilities, poor interpretation of the previous and next relationships of data, and are easily interfered by noise, which affects the efficiency of the algorithm and the accuracy of diagnosis. For the series diesel-electric hybrid system, the system composition is more complex, and the noise components introduced into the sensor signal are more difficult to distinguish. Summary of the invention

[0005] In view of this, the present application aims to propose a series diesel-electric hybrid power system fault detection method and detection system thereof to solve at least one of the above problems.

[0006] To achieve the above purpose, the technical solution of this application is implemented as follows:

[0007] In a first aspect, the present application provides a series diesel-electric hybrid system fault detection method, comprising:

[0008] Collect the speed signal of the series diesel-electric hybrid system during operation under different working conditions;

[0009] Decomposing the speed signal once by an empirical mode decomposition algorithm to obtain a plurality of IMF signals, performing data processing on each of the IMF signals based on a detrending fluctuation analysis algorithm to obtain a corresponding Hurst index, and dividing the plurality of Hurst indexes according to a preset Hurst index threshold to obtain an IMF signal dominated by noise and an IMF signal dominated by a valid signal;

[0010] The IMF signal dominated by noise is deeply decomposed by a hard threshold wavelet denoising method, and the IMF signal after the deep decomposition processing and the IMF signal dominated by the valid signal are reconstructed twice to obtain a speed signal with improved quality, wherein the deep decomposition processing includes secondary decomposition processing, hard threshold denoising processing and primary reconstruction processing;

[0011] The speed signal after quality improvement is input into a pre-built xLSTM deep neural network model to detect whether there is a fault in the current power system, wherein the xLSTM deep neural network model is trained based on the speed signal after quality improvement.

[0012] In the second aspect, based on the same inventive concept, the present application also provides a series diesel-electric hybrid system fault detection system, comprising:

[0013] A signal acquisition module is configured to acquire a speed signal of the series diesel-electric hybrid power system during operation under different operating conditions;

[0014] a primary processing module, configured to perform a primary decomposition on the speed signal through an empirical mode decomposition algorithm to obtain a plurality of IMF signals, perform data processing on each of the IMF signals based on a detrending fluctuation analysis algorithm to obtain a corresponding Hurst index, and divide the plurality of Hurst indexes according to a preset Hurst index threshold to obtain an IMF signal dominated by noise and an IMF signal dominated by a valid signal;

[0015] A secondary processing module is configured to perform deep decomposition processing on the IMF signal dominated by noise through a hard threshold wavelet denoising method, and perform secondary reconstruction on the IMF signal after the deep decomposition processing and the IMF signal dominated by the valid signal to obtain a speed signal with improved quality, wherein the deep decomposition processing includes secondary decomposition processing, hard threshold denoising processing and primary reconstruction processing;

[0016] The fault detection module inputs the speed signal after quality improvement into a pre-built xLSTM deep neural network model to detect whether there is a fault in the current power system, wherein the xLSTM deep neural network model is trained based on the speed signal after quality improvement.

[0017] Compared with the prior art, the series diesel-electric hybrid system fault detection method and detection system described in this application have the following beneficial effects:

[0018] The present application describes a fault detection method and a detection system for a series diesel-electric hybrid power system. By analyzing the importance and characteristics of the speed signal in the series diesel-electric hybrid power system and integrating the empirical mode decomposition technology, the speed signal quality obtained by the speed sensor is improved. By building an xLSTM speed fault detection model, the diesel engine speed data of the series diesel-electric hybrid power system is analyzed. This can accurately and timely discover system problems and avoid larger faults, further ensuring the safety of the series diesel-electric hybrid power system. While effectively improving the speed sensor signal quality, the series diesel-electric hybrid power system can be fault detected and diagnosed through a deep neural network, providing efficient and accurate data support for power system maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0020] Figure 1 A flow chart of a fault detection method for a series diesel-electric hybrid power system according to an embodiment of the present application;

[0021] Figure 2 This is a flow chart of the speed signal data processing described in the embodiment of the present application;

[0022] Figure 3 This is a schematic diagram of the xLSTM deep neural network model structure described in the embodiment of the present application;

[0023] Figure 4 This is a structural schematic diagram of a series diesel-electric hybrid power system fault detection system described in an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0025] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be the usual meanings understood by people with ordinary skills in the field to which the present application belongs. The "first", "second" and similar words used in the embodiments of the present application do not represent any order, quantity or importance, but are only used to distinguish different components. "Including" or "comprising" and similar words mean that the elements or objects appearing in front of the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0026] The core of this application is to provide a fault detection scheme for a series diesel-electric hybrid system based on a speed sensor signal. In this scheme, firstly, a new signal quality improvement method is designed for the importance of the speed signal collected by the speed sensor and the characteristics of the series diesel-electric hybrid system, because it involves many complex mechanical and electrical components, and the sensor data is often subject to various interferences, including mechanical vibration, electromagnetic interference, etc. Then, the problems existing in the current common neural network methods for fault diagnosis are analyzed, and a new improved xLSTM method is designed to perform fault diagnosis based on the processed speed sensor signal, so as to further improve the real-time and accuracy of fault diagnosis.

[0027] The speed sensor is used to collect the speed data signal of the APU subsystem during operation under different working conditions. After two signal decompositions, the noise part and the effective part are separated more finely, and a more effective denoising effect will be obtained for the mixed noise in the effective frequency band.

[0028] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0029] See also Figures 1 to 3 As shown, this embodiment provides a fault detection method for a series diesel-electric hybrid system, comprising:

[0030] Step S101: collecting speed signals of the series diesel-electric hybrid power system during operation under different working conditions.

[0031] Specifically, in this embodiment, based on bench tests, the speed signal of the APU in the series diesel-electric hybrid power system under different working conditions (steady-state operation and fault state) is obtained through sensors, and based on the obtained parameter information, it is determined whether there is a fault in the current power system.

[0032] Step S102, decomposing the speed signal once by an empirical mode decomposition algorithm to obtain multiple IMF signals, performing data processing on each IMF signal based on a detrended fluctuation analysis algorithm to obtain a corresponding Hurst index, and dividing the multiple Hurst indexes according to a preset Hurst index threshold to obtain an IMF signal dominated by noise and an IMF signal dominated by a valid signal.

[0033] Specifically, in this embodiment, Figure 2 As shown, the speed signal is decomposed based on the empirical mode decomposition (ICEEMDAN) to obtain multiple intrinsic mode functions (IMFs). Based on the DFA (Detrended fluctuation analysis) method, the Hurst Exponent H of each IMF is calculated to determine the size of the Hurst index. The speed signal has a clear regularity, so the Hurst index can clearly determine the part with strong mixed noise in the signal, and then wavelet hard threshold denoising is used to clean it. The specific method is as follows:

[0034] 1) Perform the first decomposition on the original speed signal to obtain the residual component r1 and the first intrinsic mode function IMF1, where LM() represents the local mean operator.

[0035] r1=LM(x (1) ) (1)

[0036] IMF1=x-r1 (2)

[0037] First, the original signal x is processed by adding white noise to obtain x( 1) The specific formula is x (i) =x+α0EMD1(ω (i) ), α k is the amplitude coefficient of the added noise, that is, the signal-to-noise ratio.

[0038] 2) After executing step 1), further th The kth residual is calculated and the kth th An intrinsic modal function.

[0039] r k =LM(r k-1 +α k-1 EMD k (ω (i)) (3)

[0040] IMF k =r k-1 -r k (4)

[0041] 3) Repeat step 2) until the generated r k When the signal is stable, stop the iterative calculation and obtain all IMFs, where the total number of IMFs is K. The original signal x can be expressed as:

[0042]

[0043] In the above steps, EMD k () represents the kth th The operator of the IMF components, LM() represents the local mean operator, α k-1 is at the kth th The amplitude coefficient of the noise added to the components, r represents the residual component, ω (i) is the i-th Gaussian white noise, and K represents the total number of IMF signals.

[0044] The design uses Hurst Exponent to further divide the IMF signal into an effective signal part and a noise signal part. For time series, the Hurst index reflects the long-term dependence and self-similarity of the sequence. The value range of this embodiment is set to 0 to 1. The smaller the H value, the worse the self-similarity of the data. The threshold is set to 0.5. When H<0.5, it is considered that the tested time series is mainly noise. When 0.5≤H≤1, it is considered that the signal is mainly composed of effective signal components, and the proportion of noise is extremely small. In the method of calculating the Hurst index, because the differential time series analysis (DFA, Detrended Fluctuation Analysis) can effectively remove the trend influence of non-stationarity, it is very effective for detecting long-range correlations and is suitable for complex time series with noise pollution, so DFA is used to calculate the Hurst index. The algorithm steps and formulas of DFA are specifically as follows:

[0045] 1) Consider the time series IMF k According to the average After calculating the cumulative deviation, we can get:

[0046]

[0047] 2) Then y(k) is decomposed into multiple sub-intervals of the same length Δn. In each segment, N is the total number of data points, and the integrated data is locally fitted to a polynomial yΔ n (k), and the root mean square error F(Δn) is defined as:

[0048]

[0049] 3) The DFA method mainly tests the self-similarity of the signal, which can be reflected in the double logarithmic (log-log) diagram of F(Δn) versus Δn. Specifically:

[0050] F(Δn)=C(Δn) α (8)

[0051] 4) In the above formula, C is a constant and α is a scaling exponent, which can be estimated by least squares fitting. This scaling exponent α is a measure of the correlation in the noise and is an estimate of the Hurst exponent H. The above formula is transformed into the following form to obtain the final estimate of the Hurst exponent H:

[0052] ln(F(Δn))=αln(Δn)+ln(C) (9)

[0053]

[0054] For the IMF of the noise signal part, that is, the IMF of the part with H<0.5, hard threshold wavelet denoising is used to perform deeper decomposition and processing.

[0055] Step S103, deeply decompose the IMF signal dominated by noise through the hard threshold wavelet denoising method, and reconstruct the IMF signal after the deep decomposition processing and the IMF signal dominated by valid signals for the second time (superimpose and reconstruct the IMF signal dominated by valid signals obtained after the deep decomposition and the IMF signal dominated by valid signals obtained after the primary decomposition) to obtain a speed signal with improved quality, wherein the deep decomposition processing includes secondary decomposition processing, hard threshold denoising processing and primary reconstruction processing.

[0056] Specifically, in this embodiment, wavelet threshold denoising can be divided into hard threshold denoising and soft threshold denoising due to different selections of threshold functions. From the denoising effect point of view, the signal after soft threshold denoising is smoother, but it is easy to remove some valid signals. The signal after hard threshold denoising will produce oscillation, but the error after denoising is smaller than that of the soft threshold method, and does not affect the approximation between the denoised signal and the original signal. Therefore, from the perspective of ensuring that the signal has a high degree of similarity, this application adopts the hard threshold denoising method.

[0057] The original signal that needs to be decomposed by wavelet (this application uses db2 wavelet and 4-level decomposition) is an IMF signal mainly composed of noise signals that need to be denoised. After decomposition, the approximate coefficient and detail coefficient are obtained. These two parts contain different characteristics and information of the signal. After the first decomposition, the approximate coefficient cA1 and the detail coefficient cD1 are obtained. Subsequently, cA1 is further decomposed to obtain the approximate coefficient cA2 and the detail coefficient cD2, and so on. Hard threshold denoising is when the detail coefficient is greater than the threshold, it is judged as a valid signal and retained. When the detail coefficient is less than the threshold, it is judged as noise, and the signal is set to zero for subsequent operations, as shown in formula (11).

[0058]

[0059] Among them, cD iλ represents the new detail coefficient after the i-th detail coefficient is updated, and λ represents the threshold.

[0060] After the above-mentioned deep decomposition and removal of mixed noise in the signal, the signal quality is improved. After two signal reconstructions, namely wavelet reconstruction and empirical mode decomposition reconstruction, the speed signal time series with improved quality is obtained.

[0061] This application is based on the principle that the speed sensor signal during engine operation is measured by a periodic sine wave, so that the effective part of the signal is concentrated on the frequency corresponding to the speed. A signal quality improvement method combining ICEEMDAN-DFA-WD is designed and used. This method is based on the speed signal measurement mechanism and uses two different signal decomposition methods to achieve deeper signal mining and obtain better signal quality improvement effect. Compared with the several traditional methods mentioned in the background technology, it has a more reliable quality improvement effect. This also effectively solves the problem of analysis result errors caused by the sensitivity of neural networks to data to a certain extent.

[0062] Step S104: input the speed signal after quality improvement into a pre-built xLSTM deep neural network model to detect whether there is a fault in the current power system, wherein the xLSTM deep neural network model is trained based on the speed signal after quality improvement.

[0063] Specifically, in this embodiment, the processed speed signal is divided by labels, the speed signal under a healthy condition is marked as a positive sample, and the speed signal under a fault condition is marked as a negative sample.

[0064] The positive and negative samples are randomly divided into 10 parts respectively, and then divided into 10 groups based on the ten-fold cross-validation scheme. In each group, 9 positive samples and 9 negative samples are used as training sets, and 1 positive sample and 1 negative sample are used as test sets.

[0065] Based on the integration of two new variants, sLSTM and mLSTM, into the residual module, the xLSTM network structure is designed, and the established xLSTM deep neural network is trained using the training set. A ten-fold cross validation is performed to illustrate the feasibility and effectiveness of the established xLSTM network in fault detection of series diesel-electric hybrid power systems.

[0066] like Figure 3 Shown is the flowchart of the proposed improved xLSTM method for series hybrid system fault detection.

[0067] In order to solve the limitations of LSTM, the ability of LSTM is enhanced by introducing exponential gating and modifying the memory structure in xLSTM: exponential gating introduces appropriate normalization and stabilization techniques, improves the gating mechanism of LSTM, and modifies the memory structure of LSTM; then two new memory units are introduced by decomposition, sLSTM has scalar memory, scalar update and new memory hybrid technology, and mLSTM has matrix memory and covariance update rules, which can be fully parallelized. Then the sLSTM and mLSTM modules are connected in sequence to form the basic architecture unit of xLSTM.

[0068] Figure 3 The left side shows the basic xLSTM architecture unit, which is the integration of sLSTM and mLSTM memory units. These blocks are the basic building blocks of the xLSTM architecture, and the complete xLSTM architecture is formed through residual stacking. Traditional LSTM performs poorly when dealing with occasional faults and low-frequency faults because of its limited storage capacity, but xLSTM solves this problem through a matrix memory mechanism. This architecture can expand the parameter scale several times when processing long-sequence speed signal data collected by the speed sensor during APU operation, and can more effectively handle rare situations in time series signals and make correct judgments. In the APU speed detection process, it can process speed signals for longer periods of time with higher accuracy.

[0069] Figure 3The right side shows the basic structure of sLSTM and mLSTM. In sLSTM, G1 is the input gate, G2 is the forget gate, G3 is the output gate sigmod function, and u1 and u2 are memory units. The main difference from LSTM is that in order to enable LSTM to revise its storage decision, sLSTM introduces a gating mechanism of exponential activation function. The input gate and forget gate used are changed from the original sigmod function to exponential gating. Unlike traditional sigmoid gating, exponential gating can produce larger values, allowing the model to update its internal state more flexibly.

[0070] The forward propagation equation is the core update rule of sLSTM, including the calculation of cell state, normalizer state and hidden state. The calculation of the above steps directly defines how sLSTM updates its internal state at time t. The specific formula is as follows:

[0071] c t =f t c t-1 +i t z t (12)

[0072] n t =f t n t-1 +i t (13)

[0073]

[0074] Among them, formula (12) is the cell state, formula (13) is the normalizer state, and formula (14) is the hidden state.

[0075] In the formula, z t is the cell input, c t-1 is the current cell state, c t is the updated cell state, n t-1 is the normalized state of the cell, n t is the updated normalized cell state, i t is the input gate parameter, f t is the forget gate parameter, o t is the output gate parameter, which controls the update of the storage unit, h t is the hidden layer.

[0076] Since the exponential activation function may cause numerical overflow, sLSTM introduces a normalizer state, which sums the product of the input gate and all future forget gates, as shown in formula (13). In addition, in order to stabilize the gate, sLSTM uses an additional stabilizer state to control the value of the gate. Formulas (15-17) are the formulas for cell input, input gate, forget gate, and output gate, respectively. Where the weight vector ω z ,ω i ,ω f and ω o They are the network input x t The weights of the inputs between the cell input, input gate, forget gate, and output gate. z , r i , r f and r o Corresponding to the hidden state h t-1 The recurrent weights between the cell input, input gate, forget gate, and output gate. z , b i , b f and b o is the corresponding bias term. is tanh, which is the activation function for the unit input and hidden state. σ indicates that the gate activation function is sigmoid, and exp indicates the exponential activation function.

[0077]

[0078] sLSTM has a multi-head structure. The traditional LSTM network has only one head, while sLSTM has two heads. The two heads perform network operations simultaneously and allow memory mixing through loop connections, that is, crossing and mixing memory states in two parallel networks. This new memory mixing technology allows sLSTM to share information between different memory units, so that each head has its own memory mixing, but there is no cross-head memory mixing between heads. This multi-head structure combined with the memory mixing strategy provides sLSTM with a new memory mixing method, which can improve the utilization of effective information.

[0079] The input gate and forget gate of sLSTM use exponential activation functions, while the output gate still uses the sigmoid function. In order to stabilize these gates, sLSTM uses a broadcast mechanism to extend the original LSTM gating technology and bias terms to the new architecture. During the sLSTM parameter calculation process, each gate of sLSTM is associated with a weight vector and a bias term, which are learned during the model training process. The introduction of activation functions and stabilized states increases the computational complexity of the model, but also improves its expressiveness.

[0080] The core update rule of mLSTM is reflected in its forward propagation equation, which includes the calculation of matrix memory state (cellstate), normalized state (normalizer state) and hidden state (hidden state), as shown in formulas (19)-(21). These equations define how mLSTM updates its internal state at time step t.

[0081] c t =f t c t-1 +i t v t k′ t (19)

[0082] n t =f t n t-1 +i t k t (20)

[0083]

[0084] Among them, k t and v t are the key and value stored in the corresponding matrix input at time t. t-1 is the current cell state, c t is the updated cell state, n t-1 is the normalized state of the cell, n t is the updated normalized cell state, i t is the input gate parameter, f t is the forget gate parameter, o t is the output gate parameter, which controls the update of the storage unit, h t is the hidden layer information.

[0085] In order to improve the signal-to-noise ratio, sLSTM uses the covariance update rule to store and retrieve information. t (key) and value vt The (value) pairs are stored as rows or columns of a matrix, as shown in formula (22) and formula (23), thereby improving the separation and signal-to-noise ratio of the retrieval. t is the network input, W k and W v is the weight, b k and b v is the bias term, and d is the matrix dimension.

[0086]

[0087] v t =W v x t +b v (twenty three)

[0088] The basic structure of mLSTM is the same as that of sLSTM. The input gate G1 and the forget gate G2 both use exponential activation functions, and the output gate G3 is a sigmoid function. It also has a multi-head structure, as shown in formulas (24)-(26). The main difference is that there is no memory mixing structure in the two parallel network structures. mLSTM abandons the loop connection (memory mixing) between the hidden layers and allows full parallelization (Parallelization), which makes mLSTM more efficient in training and reasoning on modern hardware. And the memory units in the structure are all in matrix form. The design of mLSTM is inspired by the bidirectional associative memory model (Bidirectional Associative Memories BAMs), which uses matrix multiplication to realize information storage and retrieval. By expanding the traditional scalar memory unit into a matrix memory unit, more information is stored in the form of a matrix, thereby improving the storage capacity of the model.

[0089]

[0090] The weight vector ω i ,ω f and ω o They are the network input x t The input weights between the input gate, forget gate, and output gate, b i , b f and b o is the corresponding bias term, σ indicates that the gate activation function is sigmoid, and exp indicates the exponential activation function.

[0091] The signal after the proposed signal quality improvement method is divided into a training set and a test set. The training set data is used as input, and the operation and fault conditions of the APU system are used as labels to train the xLSTM architecture neural network. The trained xLSTM neural network is then tested using the test set. The proposed method randomly divides the data into ten parts to facilitate the ten-fold cross-validation of the proposed method. After the verified network parameters are solidified, real-time fault detection of the APU system based on the speed sensor signal is realized.

[0092] In the fault diagnosis of hybrid power system, this application is mainly aimed at the fault diagnosis of power auxiliary unit (APU). Compared with traditional internal combustion power, the internal combustion engine and the starter-in-one motor of the hybrid power system are coaxially connected. Therefore, during operation, the detection of the shaft speed can not only monitor the status of the internal combustion engine but also monitor the generator at the same time. LSTM, as a neural network method with memory ability, has certain advantages in processing time series, but its operation complexity is high, lacks parallelism, has insufficient generalization ability and limited storage capacity.

[0093] xLSTM has better performance. It uses residual connection and normalization layer technology, combined with a new exponential gating mechanism, to solve the limitations of storage, parallelism, and long-term dependency modeling, and improves the performance and scalability of the model in processing time series data. In this application, xLSTM has better performance and is compatible with the speed signal mechanism of the hybrid system. It is used for the first time in fault diagnosis based only on speed signals.

[0094] By obtaining speed signal samples of the series diesel-electric hybrid system under various steady-state conditions, normal operation and fault conditions, based on the importance, periodicity and stability characteristics of the speed signal in the series diesel-electric hybrid system, the ICEEMDAN-DFA-WD signal cleaning method combining integrated empirical mode decomposition with adaptive noise and Hurst index is designed, and the xLSTM analysis method with higher parallel capability is more suitable for such speed signals.

[0095] First, the speed signal is cleaned to remove the noise signal introduced by vibration and other reasons. Compared with the traditional signal cleaning method based on empirical mode decomposition, the signal cleaning method designed in this application is very suitable for the speed sensor signal, which can obtain a higher signal-to-noise ratio and provide a basis for subsequent data analysis. Then, the positive samples and fault samples are randomly divided into ten training sets and test sets, respectively, to facilitate the subsequent ten-fold cross validation. By extending the LSTM method (xLSTM), the enhanced memory structure can improve the parallel capability. It is worth mentioning that the improved exponential gating and residual network modules are more suitable for the analysis and detection of the speed signal sequence during the operation of the engine. An xLSTM deep neural network model is established, with the model input as the training set and the output as the speed of the sample in the fault state, and a ten-fold cross validation is performed.

[0096] By analyzing the importance and characteristics of the speed signal in the series diesel-electric hybrid system, combined with the integrated empirical mode decomposition technology, the speed signal quality obtained by the speed sensor is improved, and the xLSTM speed fault detection model is further built. The diesel engine speed data of the series diesel-electric hybrid system is analyzed, which can accurately and timely discover system problems, avoid larger faults, and further ensure the safety of the series diesel-electric hybrid system.

[0097] It should be noted that the above describes some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0098] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, an embodiment of the present application further provides a series diesel-electric hybrid power system fault detection system.

[0099] like Figure 4 As shown, the series diesel-electric hybrid power system fault detection system comprises:

[0100] The signal acquisition module 11 is configured to acquire the speed signal of the series diesel-electric hybrid power system during operation under different working conditions;

[0101] The primary processing module 12 is configured to perform a primary decomposition on the speed signal through an empirical mode decomposition algorithm to obtain a plurality of IMF signals, perform data processing on each of the IMF signals based on a detrending fluctuation analysis algorithm to obtain a corresponding Hurst index, and divide the plurality of Hurst indexes according to a preset Hurst index threshold to obtain an IMF signal dominated by noise and an IMF signal dominated by a valid signal;

[0102] The secondary processing module 13 is configured to perform deep decomposition processing on the IMF signal dominated by noise through a hard threshold wavelet denoising method, and perform secondary reconstruction on the IMF signal after the deep decomposition processing and the IMF signal dominated by the valid signal to obtain a speed signal with improved quality, wherein the deep decomposition processing includes secondary decomposition processing, hard threshold denoising processing and primary reconstruction processing;

[0103] The fault detection module 14 inputs the speed signal after quality improvement into the pre-built xLSTM deep neural network model to detect whether there is a fault in the current power system, wherein the xLSTM deep neural network model is trained based on the speed signal after quality improvement. For the convenience of description, the above device is described in various modules according to their functions. Of course, when implementing the embodiments of the present application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0104] The device of the above embodiment is used to implement the corresponding method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0105] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present application (including the claims) is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0106] The embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of the present application.

Claims

1. A fault detection method for a series diesel-electric hybrid system, characterized in that: include: Collect the speed signal of the series diesel-electric hybrid system during operation under different working conditions; Decomposing the speed signal once by an empirical mode decomposition algorithm to obtain a plurality of IMF signals, performing data processing on each of the IMF signals based on a detrending fluctuation analysis algorithm to obtain a corresponding Hurst index, and dividing the plurality of Hurst indexes according to a preset Hurst index threshold to obtain an IMF signal dominated by noise and an IMF signal dominated by a valid signal; The IMF signal dominated by noise is deeply decomposed by a hard threshold wavelet denoising method, and the IMF signal after the deep decomposition processing and the IMF signal dominated by the valid signal are reconstructed twice to obtain a speed signal with improved quality, wherein the deep decomposition processing includes secondary decomposition processing, hard threshold denoising processing and primary reconstruction processing; The speed signal after quality improvement is input into a pre-built xLSTM deep neural network model to detect whether there is a fault in the current power system, wherein the xLSTM deep neural network model is trained based on the speed signal after quality improvement.

2. The method according to claim 1, characterized in that The empirical mode decomposition algorithm includes: in, r k =LM(r k-1 +a k-1 EMD k (oh (i) ); IMF k =r k-1 -r k ; Where, EMD k () represents the kth th The operator of the IMF components, LM() represents the local mean operator, α k-1 is at the kth th The amplitude coefficient of the noise added to the components, r represents the residual component, ω (i) is the i-th Gaussian white noise, and K represents the total number of IMF signals.

3. The method according to claim 2, characterized in that The detrended fluctuation analysis algorithm includes: IMF for time series k According to the average After calculating the cumulative deviation, we can get: Decompose y(k) into multiple sub-intervals of the same length Δn. In each segment, N represents the total number of data points. The integrated data is locally fitted to a polynomial y Δn (k), and the root mean square error F(Δn) is defined as: The double logarithm of F(Δn) versus Δn is expressed as: F(Δn)=C(Δn) α ; Transform the above formula into the following form to obtain the final estimated value of the Hurst index H: ln(F(Δn))=αln(Δn)+ln(C); Where C is a constant, α is a scaling exponent, and is an estimated value of the Hurst exponent H.

4. The method according to claim 3, characterized in that: The preset Hurst index threshold is 0.5; In response to the Hurst exponent being less than 0.5, it is determined to be an IMF signal dominated by noise; In response to the Hurst index being greater than or equal to 0.5, it is determined to be an IMF signal dominated by valid signals.

5. The method according to claim 1, characterized in that The IMF signal dominated by noise is decomposed twice by hard threshold wavelet denoising method, including: The IMF signal dominated by noise signal obtained after the primary decomposition is decomposed twice by hard threshold wavelet denoising method, and the approximate coefficient and detail coefficient are obtained after decomposition. In response to the detail coefficient being greater than a set threshold, it is determined to be an IMF signal mainly composed of valid signals, and a retention process is performed; In response to the detail coefficient being less than a set threshold, it is determined to be an IMF signal dominated by noise, and the IMF signal is set to zero.

6. The method according to claim 1, characterized in that In the formula, cD iλ represents the new detail coefficient after the i-th detail coefficient is updated, and λ represents the set threshold.

7. The method according to claim 1, characterized in that The method for constructing the xLSTM deep neural network model includes: The speed signal after quality improvement is divided by labels. The speed signal under healthy working condition is marked as positive sample, and the speed signal under fault condition is marked as negative sample. The positive and negative samples are randomly divided into training sets and test sets respectively; The sLSTM and mLSTM are integrated into the residual module to obtain the xLSTM network structure. The established xLSTM deep neural network model is trained with the training set, and the trained xLSTM neural network model is tested with the test set through ten-fold cross validation.

8. A series diesel-electric hybrid power system fault detection system implemented according to any one of claims 1 to 7, characterized in that: include: A signal acquisition module is configured to acquire a speed signal of the series diesel-electric hybrid power system during operation under different operating conditions; a primary processing module, configured to perform a primary decomposition on the speed signal through an empirical mode decomposition algorithm to obtain a plurality of IMF signals, perform data processing on each of the IMF signals based on a detrending fluctuation analysis algorithm to obtain a corresponding Hurst index, and divide the plurality of Hurst indexes according to a preset Hurst index threshold to obtain an IMF signal dominated by noise and an IMF signal dominated by a valid signal; A secondary processing module is configured to perform deep decomposition processing on the IMF signal dominated by noise through a hard threshold wavelet denoising method, and perform secondary reconstruction on the IMF signal after the deep decomposition processing and the IMF signal dominated by the valid signal to obtain a speed signal with improved quality, wherein the deep decomposition processing includes secondary decomposition processing, hard threshold denoising processing and primary reconstruction processing; The fault detection module inputs the speed signal after quality improvement into a pre-built xLSTM deep neural network model to detect whether there is a fault in the current power system, wherein the xLSTM deep neural network model is trained based on the speed signal after quality improvement.

Images