Diesel engine multi-sensor multi-scale fusion feature extraction and fault early warning method

Abstract

A diesel engine multi-sensor multi-scale fusion feature extraction and fault early warning method is suitable for the technical field of diesel engine state monitoring and fault diagnosis. According to the method, firstly, collected cylinder cover vibration signals are preprocessed, and the preprocessing comprises multi-period superposition averaging, filtering noise reduction, envelope line extraction and phase alignment; secondly, a multi-sensor signal multi-scale similarity matrix is constructed, the similarity degree of multi-cylinder signals is calculated through sliding of different windows, eigenvalue decomposition is carried out, and the multi-scale similarity matrix serves as input of a multi-branch residual convolutional neural network fault recognition model; and finally, constructing a multi-branch residual convolutional neural network model, and obtaining a fault recognition result through a full connection layer and a softmax classifier by adopting an Adam optimization algorithm and a cross entropy loss function through a multi-channel input fusion layer, a residual module and a multi-branch output fusion layer. The method has high fault identification accuracy and robustness under a single working condition and a complex variable working condition.

Description

Technical Field

[0001] The present invention relates to a method for identifying equipment faults, and in particular to a method for extracting multi-sensor multi-scale fusion features and performing fault warning, which is applicable to the technical field of condition monitoring and fault diagnosis of diesel engines. Background Art

[0002] Medium- and high-speed, high-power diesel engines are the driving force behind crucial equipment in many key sectors, including shipping and electric power. Their complex structures and systems are becoming increasingly integrated, comprehensive, and intelligent, featuring numerous cylinders, high explosion pressure, high power density, high levels of integration, and compact structures. Furthermore, they often operate under complex and variable operating conditions, subject to strong impacts from multiple sources and subject to harsh operating conditions, resulting in a high failure rate. For medium- and high-speed, high-power diesel engines, effective monitoring using advanced sensor technology, coupled with signal processing and analysis, is being implemented, gradually enabling intelligent early warning and diagnosis of diesel engine failures. This enables timely detection of faults and prevents them from worsening, playing a crucial role in improving the operational safety and reliability of these equipment systems.

[0003] Currently, common methods for diesel engine condition monitoring include vibration monitoring, instantaneous speed monitoring, acoustic emission monitoring, and oil monitoring. Because vibration signal monitoring is a non-invasive condition monitoring method and diesel engine cylinder head vibration signals contain rich operating status information, diesel engine fault identification currently relies on vibration signal feature extraction and the selection of fault-sensitive characteristic parameters. However, due to the typical non-stationary nature of diesel engine vibration signals, vibration-based fault identification methods face numerous challenges. Current diesel engine fault identification methods based on vibration signals typically only consider fault characteristic parameters under a single operating condition. However, in actual engineering applications, diesel engine operating conditions frequently vary. Furthermore, the types and magnitudes of diesel engine fault-sensitive characteristic parameters under different operating conditions are often not identical, making it difficult for fault identification methods to maintain high accuracy under varying operating conditions. Furthermore, because diesel engine faults are often manifested in short-term impacts in the angular domain, and the impact energy and duration angle are both related to the fault type and severity, multi-sensor fusion analysis requires special attention to the multi-scale nature of the signal. Therefore, the present invention discloses a variable operating condition diesel engine fault identification method based on multi-sensor signal multi-scale fusion. Summary of the Invention

[0004] In order to solve the problem that the relevant information between cylinders in the multi-cylinder diesel engine fault identification method is not fully and effectively utilized, and taking into account the characteristics that the mechanical structure and working status of each cylinder of a multi-cylinder diesel engine are basically consistent and the multi-scale characteristics of fault information, the present invention proposes a multi-sensor multi-scale fusion feature extraction and fault warning method.

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

[0006] First, the raw signal is preprocessed using methods such as multi-cycle superposition averaging, filtering and noise reduction, and envelope extraction. Phase alignment of the signals for each cylinder is performed in the angular domain according to the diesel engine's ignition order, ensuring clearer and more reliable feature extraction from the raw vibration signal. Second, sliding windows of varying sizes are used to calculate the similarity of multi-cylinder signals at different scales, generating a multi-scale similarity matrix and extracting multi-scale fusion features from the multi-sensor signals. Finally, a multi-branch residual convolutional neural network model is constructed, with each branch processing the similarity matrix at a single scale. Finally, the similarity matrices are fused at the output layer to produce the classification results.

[0007] 1. A diesel engine sensor multi-scale fusion feature extraction and fault warning method, characterized by comprising the following steps:

[0008] Step 1: Preprocess the collected vibration signal

[0009] 1.1 Noise Reduction Preprocessing

[0010] The original signal is processed by multi-cycle superposition and averaging. The specific process is expressed as follows:

[0011]

[0012] Where i = 1, 2, ..., N d , y is the signal after multi-period superposition and averaging, V acc (i, j) represents the vibration value of the i-th data point in the j-th engine working cycle, N a Indicates the number of superposition cycles, N a =60, N d Represents the number of data points in a single engine operating cycle;

[0013] For the input signal y after multi-period averaging, the expression for 5kHz high-pass filtering using a Butterworth high-pass filter is as follows:

[0014] y f =HPF(y,f c ,f s ,order)

[0015] Among them, y f is the output signal after filtering, HPF(·) is the high-pass filter function, f c is the cutoff frequency of the filter, f c =5kHz, f s is the sampling frequency of the signal, which is initially set to 51200 Hz, and order is the filter order, which ranges from 3 to 8;

[0016] 1.2 Envelope Extraction First, find the signal y f All local maximum points in are first-order extreme points;

[0017] Then, find all the maximum points from the first-order extreme points as the connection envelope points, and remove the mean of the extracted envelope points; suppose the sequence of a certain order envelope feature point is n d is the total number of data points, then the formula for removing the mean is expressed as:

[0018]

[0019] Finally, the series obtained after removing the mean Perform Hermite interpolation to obtain the envelope y e ;

[0020] 1.3 Phase alignment signal y e The signal collected by the diesel engine in one working cycle, that is, 720° crankshaft angle, is N. d ; The signal y after phase alignment according to the different phase angles of different cylinders p for:

[0021] y p =stack[y e (angle / 720×N d :N d ),y e (1:angle / 720×N d )]

[0022] Among them, stack[a1,a2] represents the concatenation of signals a1 and a2, y e (n1:n2) represents signal y e The n1th to n2th elements of , angle represents the corresponding phase angle;

[0023] Step 2: Construct the Multi-Sensor signal Multi-Scale Similarity Matrix MSMSSM

[0024] For two discrete signals x1 and x2, the cross-correlation function between them is expressed as:

[0025]

[0026] Where τ = 1-N, 2-N, …, -1, 0, 1, …, N-2, N-1; m is the index of the signal sample point, N is the length of the signal, and by default, the signal lengths are consistent. If they are inconsistent, they are aligned through resampling, truncation, or zero padding. τ is the delay.

[0027] The obtained cross-correlation function is normalized so that the maximum value of the autocorrelation function is 1 at the time delay τ = 0:

[0028]

[0029] The maximum value of the normalized cross-correlation function is taken as a measure of the similarity between the two signals. The similarity r of signals x1 and x2 is expressed as:

[0030]

[0031] The similarity r of signals x1 and x2 ranges from [0,1], where a larger value indicates a greater similarity. When r = 0, the signals are completely dissimilar. When r = 1, the signals are exactly the same at the time delay τ.

[0032] The similarity matrix R of multi-sensor signals is constructed as follows:

[0033]

[0034] Among them, for a multi-cylinder diesel engine, the number of cylinders is n, and the corresponding number of cylinder head vibration signal sensors is also n; represents a set of n×n real matrices, r pq represents the similarity between the cylinder head vibration signals collected by the pth sensor and the qth sensor after preprocessing, p,q∈(1,2,…,n); in addition, the diagonal elements of the matrix represent the similarity of the sensor signals themselves, so the values ​​of the diagonal elements are all 1;

[0035] Then, the multi-sensor signal similarity matrix R is subjected to eigenvalue decomposition;

[0036] In the formula, since the correlation of sensor signals is not zero, all entries of the matrix R are positive. According to the Perron-Frobenius theorem, the matrix R has a unique maximum real eigenvalue, which is expressed as λ max , the corresponding eigenvector is expressed as v = [v1, v2, ..., v p ,…,v n ] T , with a strictly positive component v p , p=1,2,…,n; at the same time, Rv=λ max v relationship holds, which is expressed as:

[0037]

[0038] Further expanding the above formula, we can get:

[0039] r p1 v1+r p2 v2+…+(1-λ max )v p +…+r pn v n =λ max v p

[0040] Where λ max v p Comprehensive r p1 ,r p2 ,…,r pn Therefore, λ max v p It is regarded as a comprehensive correlation measure between all sensors and the p-th sensor;

[0041] Define the indicator Ind based on the PF vector p To detect and locate faulty sensor signals, the specific formula is as follows:

[0042]

[0043] Ind p is the fault detection index, which is the comprehensive support degree of all sensors to the pth sensor, v q is the index of the characteristic component in the characteristic vector; if there is no fault in the measurement data, the signals of all sensors are similar, and the index Ind p is 1 / n; on the contrary, the component corresponding to the fault data will be smaller than the non-fault data, and the index Ind p will be less than 1 / n;

[0044] Then, the signal is segmented at multiple scales to construct the multi-sensor signal similarity matrix R at different scales. c According to the characteristics of the impact duration of diesel engine vibration signals, the sample size of multi-cylinder vibration signals in one working cycle is n×N, where n represents the number of sensors and N represents the signal length. Four different window sizes c are used, namely 10°, 30°, 60°, and 720°, with a step size ranging from 1° to 5°. With a step size of 1°, the number of similarity matrices at each scale is 711, 691, 661, and 1, respectively. For each scale c, the signal is divided into sub-signals of length c / 720×N, with a number of (720-c) / 1+1.

[0045] definition is the kth sub-signal of sensor p at scale c; calculate the sub-signal and Similarities between The multi-sensor similarity matrix at scale c is expressed as:

[0046]

[0047] The multi-scale similarity matrix of multi-sensor signals is further expressed as Where l = ∑((720-c) / 1+1) is the number of channels of the matrix;

[0048] Finally, the constructed multi-scale similarity matrix of multi-sensor signals is used as the input of the multi-branch residual convolutional neural network;

[0049] Step 3: Build a multi-branch residual convolutional neural network

[0050] The multi-branch residual convolutional neural network includes a multi-channel input fusion module, a residual module, and a multi-branch output fusion module. It uses a residual learning mechanism and a multi-branch fusion mechanism to improve classification accuracy and robustness.

[0051] 3.1 Multi-channel Input Fusion Layer Design

[0052] The multi-channel fusion layer fuses the multi-channel matrices into a single-channel matrix through convolution operation; the input feature set is I E ={I1,I2,I3,I4}, where is the input of each branch, E represents the index of the four scales, n represents the number of sensors, l E Indicates the number of similarity matrices at each scale; the multi-channel fusion layer uses the Conv2D type, with a size of 1×1, a step size of 1, and 16 convolution kernels to perform multi-channel fusion on the input of each branch; the output calculation formula of each branch after multi-channel fusion is as follows:

[0053]

[0054] Among them, X E (B t ,D t ) indicates that the output feature is at position (B t ,D t ) value, I E (B t ,D t ,k) indicates the input feature at position (B t ,D t ) and the value at channel k, and K(1,1,k) represents the weight of convolution kernel K at position (1,1,k);

[0055] 3.2 Residual Module Design After the multi-channel input fusion layer, each network branch is mainly composed of a stack of residual modules. Residual module 1 contains two consecutive Conv2D layers. The first Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 32. The second Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 32. Residual module 2 also contains two Conv2D layers. The first Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 64. The second Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 64.

[0056] The convolution operation in the residual module is expressed as:

[0057] x r (a) = g r [w r (a)×x r (a-1)+b r (a)]

[0058] Among them, x r (a) is the output feature map of layer a, x r (a-1) is the output feature map of the a-1th layer, w r (a) is the convolution kernel, b r (a) is the network bias term; g r (·) represents the nonlinear activation function, using equation g r (x) = ReLU (x) = max (x, 0) to express it; In addition, BN is the batch normalization function, which is also used in the convolution layer and is expressed as BN (·); When the input and output sizes do not match, a 1×1 convolution operation is used for alignment; The convolution operation formula of the residual block is:

[0059] x r (a) = w r (a)×ReLu{BN[x r (a-1)]}+b r (a)+x r (a-2)

[0060] 3.3 Multi-branch Output Fusion Layer Design A multi-branch output fusion layer is designed at the end of the multi-branch network. For the information learned by each branch at different scales, the feature map extracted at the end of each branch is first expanded into a one-dimensional feature vector, namely the Flatten layer. Then, the output of each branch is weighted and fused through the multi-branch fusion layer.

[0061] The specific process is as follows: By expressing the feature vector extracted from each branch as in The dimensions are equal, n b is the number of branches; the obtained fusion feature vector z f Expressed as:

[0062]

[0063] in, are the feature vectors learned during training. The fusion weight of ; the dimension of the fusion weight is the same as the dimension of the feature vector;

[0064] After the multi-branch output fusion layer, the fault identification result is obtained through the fully connected layer;

[0065] The number of output nodes in the fully connected layer is 2, corresponding to the two categories of normal state and fault state. At the same time, the softmax activation function is used to convert the output value of the fully connected layer into a probability distribution, so that the value of each output node represents the probability of the sample belonging to the corresponding category, that is, to determine whether the sample belongs to the normal state or the fault state. The softmax function is defined as follows:

[0066]

[0067] Among them, y u (h) represents the output value of the u-th sample, e is a natural constant, h is the category index from 1 to C, and C is the number of categories;

[0068] The loss function uses the Adam optimization algorithm and the cross entropy loss function. The training goal is to minimize the loss function through the optimization algorithm. The cross entropy loss function formula is as follows:

[0069]

[0070] Among them, N t is the number of training samples, γ uh is the class h label in the actual label of the u-th sample, where γ uh The value is 0 or 1, indicating whether the sample belongs to category h, ρ uh It uses the softmax function to predict the probability that the u-th sample belongs to category h. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flow chart of a diesel engine multi-sensor multi-scale fusion feature extraction and fault warning method provided according to an example of this application;

[0072] Figure 2 This is a waveform diagram of the angular signal impulse before and after multi-cycle superposition and averaging processing, provided according to an example of this application;

[0073] Figure 3 This is a waveform diagram of an angular signal impulse after filtering and noise reduction processing provided according to an example of this application;

[0074] Figure 4 This is a waveform diagram of the angular signal impact before and after envelope extraction provided according to an example of this application;

[0075] Figure 5 This is a waveform diagram of the angular domain signal impulse before and after phase alignment processing provided according to an example of the present application;

[0076] Figure 6 This is a sensor indicator diagram under different health states (from left to right: normal, B3 misfire, A3 bearing wear) provided according to the example of this application;

[0077] Figure 7 This is a multi-sensor signal multi-scale similarity matrix construction process provided according to the example of this application;

[0078] Figure 8 This is a similarity matrix diagram under different health states (from left to right: normal, B3 misfire, A3 bearing wear) provided according to the example of this application;

[0079] Figure 9 This is a multi-branch residual convolutional neural network model structure diagram provided according to the example of this application;

[0080] Figure 10 It is the variable speed fault identification result of different sample inputs provided according to the example of this application;

[0081] Figure 11 It is the variable load fault identification result of different sample inputs provided according to the example of this application;

[0082] Figure 12 This is the variable speed + variable load fault identification result provided by different sample inputs based on the examples of this application. DETAILED DESCRIPTION

[0083] To more clearly illustrate the technical solution of the present invention, the following will be described clearly and completely with reference to the embodiments and accompanying drawings. It is obvious that the embodiments described are only some of the embodiments of the present invention, rather than all of them. For those skilled in the art, all other embodiments obtained without creative work are within the scope of protection of the present invention.

[0084] Figure 1 This is a flowchart of a diesel engine multi-sensor multi-scale fusion feature extraction and fault warning method provided by this application. Figure 1 , the implementation process and results of the present invention are as follows:

[0085] To verify the effectiveness of the proposed method, a dataset acquired on a medium-speed, high-power diesel engine fault simulation test bench was used for validation. The test was conducted on a Hechai TBD234V12 diesel engine. The sensors corresponding to the 12 cylinders were numbered A1, A2, A3, A4, A5, A6, B1, B2, B3, B4, B5, and B6. Three engine health states were set: normal, A3 cylinder bearing wear, and B3 cylinder misfire. To prevent severe faults such as shaft seizure, the bearing wear test was conducted at only 1000 rpm and without load. Therefore, a dataset with three health states (normal, A3 cylinder bearing wear, and B3 cylinder misfire) and two health states (normal and B3 cylinder misfire) were constructed under the same operating conditions, as shown in Tables 1 and 2. The sample set is divided into training samples, validation samples, and test samples with proportions of 0.6, 0.2, and 0.2, respectively. The numbers in the table represent the corresponding sample sizes.

[0086] Table 1 Datasets with three health states under the same working conditions

[0087]

[0088] Table 2 Datasets with two health states under the same working conditions

[0089]

[0090] Step 1: Preprocess the collected vibration signal

[0091] 1.1 Noise Reduction Preprocessing

[0092] The original signal is processed by multi-cycle superposition and averaging. The specific process is expressed as follows:

[0093]

[0094] Where i = 1, 2, ..., N d , y is the signal after multi-period superposition and averaging, V acc (i, j) represents the vibration value of the i-th data point in the j-th engine working cycle, N a Indicates the number of superposition cycles, N a =60, N d Represents the number of data points in a single engine operating cycle;

[0095] In this example, the waveform of the angular domain signal impulse before and after multi-cycle superposition and averaging is as follows: Figure 2 shown.

[0096] For the input signal y after multi-period averaging, the expression for 5kHz high-pass filtering using a Butterworth high-pass filter is as follows:

[0097] y f =HPF(y,f c ,f s ,order)

[0098] Among them, y f is the output signal after filtering, HPF(·) is the high-pass filter function, f c is the cutoff frequency of the filter, f c =5kHz, f s is the sampling frequency of the signal, which is initially set to 51200 Hz, and order is the filter order, which ranges from 3 to 8;

[0099] In this example, the impact waveform of the angular domain signal after filtering and noise reduction is as follows: Figure 3 shown.

[0100] 1.2 Envelope Extraction First, find the signal y f All local maximum points in are first-order extreme points;

[0101] Then, find all the maximum points from the first-order extreme points as the connection envelope points, and remove the mean of the extracted envelope points; suppose the sequence of a certain order envelope feature point is n d is the total number of data points, then the formula for removing the mean is expressed as:

[0102]

[0103] Finally, the series obtained after removing the mean Perform Hermite interpolation to obtain the envelope y e ;

[0104] In this example, the waveform of the angular domain signal impulse before and after envelope extraction is as follows: Figure 4 shown.

[0105] 1.3 Phase alignment signal y e The signal collected by the diesel engine in one working cycle, that is, 720° crankshaft angle, is N. d ; The signal y after phase alignment according to the different phase angles of different cylinders p for:

[0106] y p =stack[y e (angle / 720×N d :N d ),y e(1:angle / 720×N d )]

[0107] Among them, stack[a1,a2] represents the concatenation of signals a1 and a2, y e (n1:n2) represents signal y e The n1th to n2th elements of , angle represents the corresponding phase angle;

[0108] In this example, the waveforms of the angular domain signal impulse before and after phase alignment are as follows: Figure 5 shown.

[0109] Step 2: Construct the Multi-Sensor signal Multi-Scale Similarity Matrix MSMSSM

[0110] For two discrete signals x1 and x2, the cross-correlation function between them is expressed as:

[0111]

[0112] Where τ = 1-N, 2-N, …, -1, 0, 1, …, N-2, N-1; m is the index of the signal sample point, N is the length of the signal, and by default, the signal lengths are consistent. If they are inconsistent, they are aligned through resampling, truncation, or zero padding. τ is the delay.

[0113] The obtained cross-correlation function is normalized so that the maximum value of the autocorrelation function is 1 at the time delay τ = 0:

[0114]

[0115] The maximum value of the normalized cross-correlation function is taken as a measure of the similarity between the two signals. The similarity r of signals x1 and x2 is expressed as:

[0116]

[0117] The similarity r of signals x1 and x2 ranges from [0,1], where a larger value indicates a greater similarity. When r = 0, the signals are completely dissimilar. When r = 1, the signals are exactly the same at the time delay τ.

[0118] The similarity matrix R of multi-sensor signals is constructed as follows:

[0119]

[0120] Among them, for a multi-cylinder diesel engine, the number of cylinders is n, and the corresponding number of cylinder head vibration signal sensors is also n; represents a set of n×n real matrices, r pq represents the similarity between the cylinder head vibration signals collected by the pth sensor and the qth sensor after preprocessing, p,q∈(1,2,...,n); in addition, the diagonal elements of the matrix represent the similarity of the sensor signals themselves, so the values ​​of the diagonal elements are all 1;

[0121] Then, the multi-sensor signal similarity matrix R is subjected to eigenvalue decomposition;

[0122] In the formula, since the correlation of sensor signals is not zero, all entries of the matrix R are positive. According to the Perron-Frobenius theorem, the matrix R has a unique maximum real eigenvalue, which is expressed as λ max , the corresponding eigenvector is expressed as v = [v1, v2, ..., v p ,…,v n ] T , with a strictly positive component v p , p=1,2,…,n; at the same time, Rv=λ max v relationship holds, which is expressed as:

[0123]

[0124] Further expanding the above formula, we can get:

[0125] r p1 v1+r p2 v2+…+(1-λ max )v p +…+r pn v n =λ max v p

[0126] Where λ max v p Comprehensive r p1 ,r p2 ,…,r pn Therefore, λ max v p It is regarded as a comprehensive correlation measure between all sensors and the p-th sensor;

[0127] Define the indicator Ind based on the PF vector p To detect and locate faulty sensor signals, the specific formula is as follows:

[0128]

[0129] Ind pis the fault detection index, which is the comprehensive support degree of all sensors to the pth sensor, v q is the index of the characteristic component in the characteristic vector; if there is no fault in the measurement data, the signals of all sensors are similar, and the index Ind p is 1 / n; on the contrary, the component corresponding to the fault data will be smaller than the non-fault data, and the index Ind p will be less than 1 / n; the number of sensor nodes n is equal to 12, so the ideal threshold is set to 1 / n = 0.083;

[0130] In this example, the sensor indicators in different health states are as follows: Figure 6 shown.

[0131] Then, the signal is segmented at multiple scales to construct the multi-sensor signal similarity matrix R at different scales. c According to the characteristics of the impact duration of diesel engine vibration signals, the sample size of multi-cylinder vibration signals in one working cycle is n×N, where n represents the number of sensors and N represents the signal length. Four different window sizes c are used, namely 10°, 30°, 60°, and 720°, with a step size ranging from 1° to 5°. With a step size of 1°, the number of similarity matrices at each scale is 711, 691, 661, and 1, respectively. For each scale c, the signal is divided into sub-signals of length c / 720×N, with a number of (720-c) / 1+1.

[0132] definition is the kth sub-signal of sensor p at scale c; calculate the sub-signal and Similarities between The multi-sensor similarity matrix at scale c is expressed as:

[0133]

[0134] The multi-scale similarity matrix of multi-sensor signals is further expressed as Where l = ∑((720-c) / 1+1) is the number of channels of the matrix;

[0135] In this example, the process of constructing the multi-scale similarity matrix of multi-sensor signals is as follows: Figure 7 shown.

[0136] Finally, the constructed multi-scale similarity matrix of multi-sensor signals is used as the input of the multi-branch residual convolutional neural network;

[0137] Step 3: Build a multi-branch residual convolutional neural network model MBRCNN (Multi-Branch Residual Convolutional Neural Network)

[0138] The overall structure of the multi-branch residual convolutional neural network model is as follows Figure 9 The multi-branch residual convolutional neural network mainly includes a multi-channel input fusion module, a residual module, and a multi-branch output fusion module. It uses residual learning and multi-branch fusion mechanisms to improve classification accuracy and robustness. The key parameters of the multi-branch residual convolutional neural network model are shown in Table 3.

[0139] Table 3 Key parameters of the multi-branch residual convolutional neural network model

[0140]

[0141] The specific structure is as follows:

[0142] 3.1 Design of Multi-channel Input Fusion Layer The multi-channel fusion layer fuses the multi-channel matrices into a single-channel matrix through convolution operation. The input feature set is I E ={I1,I2,I3,I4}, where is the input of each branch, E represents the index of the four scales, n represents the number of sensors, l E Indicates the number of similarity matrices at each scale; the multi-channel fusion layer uses the Conv2D type, with a size of 1×1, a step size of 1, and 16 convolution kernels to perform multi-channel fusion on the input of each branch; the output calculation formula of each branch after multi-channel fusion is as follows:

[0143]

[0144] Among them, X E (B t ,D t ) indicates that the output feature is at position (B t ,D t ) value, I E (B t ,D t ,k) indicates the input feature at position (B t ,D t ) and the value at channel k, and K(1,1,k) represents the weight of convolution kernel K at position (1,1,k);

[0145] 3.2 Residual Module Design After the multi-channel input fusion layer, each network branch is mainly composed of a stack of residual modules. Residual module 1 contains two consecutive Conv2D layers. The first Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 32. The second Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 32. Residual module 2 also contains two Conv2D layers. The first Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 64. The second Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 64.

[0146] The convolution operation in the residual module is expressed as:

[0147] x r (a) = g r [w r (a)×x r (a-1)+b r (a)]

[0148] Among them, x r (a) is the output feature map of layer a, x r (a-1) is the output feature map of the a-1th layer, w r (a) is the convolution kernel, b r (a) is the network bias term; g r (·) represents the nonlinear activation function, using equation g r (x) = ReLU (x) = max (x, 0) to express it; In addition, BN is the batch normalization function, which is also used in the convolution layer and is expressed as BN (·); When the input and output sizes do not match, a 1×1 convolution operation is used for alignment; The convolution operation formula of the residual block is:

[0149] x r (a) = w r (a)×ReLu{BN[x r (a-1)]}+b r (a)+x r (a-2)

[0150] 3.3 Multi-branch Output Fusion Layer Design A multi-branch output fusion layer is designed at the end of the multi-branch network. For the information learned by each branch at different scales, the feature map extracted at the end of each branch is first expanded into a one-dimensional feature vector, namely the Flatten layer. Then, the output of each branch is weighted and fused through the multi-branch fusion layer.

[0151] The specific process is as follows: By expressing the feature vector extracted from each branch as in The dimensions are equal, nb is the number of branches; the obtained fusion feature vector z f Expressed as:

[0152]

[0153] in, are the feature vectors learned during training. The fusion weight of ; the dimension of the fusion weight is the same as the dimension of the feature vector;

[0154] After the multi-branch output fusion layer, the fault identification result is obtained through the fully connected layer;

[0155] The number of output nodes in the fully connected layer is 2, corresponding to the two categories of normal state and fault state. At the same time, the softmax activation function is used to convert the output value of the fully connected layer into a probability distribution, so that the value of each output node represents the probability of the sample belonging to the corresponding category, that is, to determine whether the sample belongs to the normal state or the fault state. The softmax function is defined as follows:

[0156]

[0157] Among them, y u (h) represents the output value of the u-th sample, e is a natural constant, h is the category index from 1 to C, and C is the number of categories;

[0158] The loss function uses the Adam optimization algorithm and the cross entropy loss function. The training goal is to minimize the loss function through the optimization algorithm. The cross entropy loss function formula is as follows:

[0159]

[0160] Among them, N t is the number of training samples, γ uh is the class h label in the actual label of the u-th sample, where γ uh The value is 0 or 1, indicating whether the sample belongs to category h, ρ uh It uses the softmax function to predict the probability that the u-th sample belongs to category h.

[0161] Since the experimental object in this paper is a 12-cylinder diesel engine, the size of a single multi-cylinder similarity matrix is ​​12×12. Each sample contains similarity matrices at scales of 10°, 30°, 60°, and 720°. Therefore, the sample sizes corresponding to each scale are 711×12×12, 691×12×12, 661×12×12, and 1×12×12, respectively. To facilitate model input, the samples at each scale are merged, resulting in a single sample size of 2064×12×12. Details are shown in Table 4. The number of samples for each operating condition of each diesel engine state is 450. The model was trained for 100 epochs. The Adam optimizer was used with a learning rate of 0.001. The KL divergence weight coefficient was set to 1 to ensure a balance between reconstruction loss and KL divergence loss. The batch size during training was set to 64.

[0162] Table 4 Detailed information of the constructed input samples and corresponding model structures

[0163]

[0164] Experimental results analysis

[0165] First, the proposed method's fault prediction performance was verified under a single operating condition, meaning that the training, validation, and test samples were all from the same operating condition. Several different sets of model sample inputs were constructed for comparison: multi-channel raw vibration signal samples from multiple cylinders and similarity matrix samples at a single scale. Because bearing wear fault data is only available under the 1000 rpm / 0 N·m operating condition, a dataset containing three states—normal, misfire, and bearing wear faults—under the 1000 rpm speed and 0 N·m load conditions was used for training, validation, and testing. The prediction accuracy results for the test set are shown in Table 5.

[0166] Table 5 Accuracy of normal, misfire and bearing wear datasets

[0167]

[0168] The model was trained, validated, and tested using datasets containing both normal and misfire conditions under different operating conditions. Six operating conditions were selected: 1200 rpm, 400 N·m; 1200 rpm, 800 N·m; 1200 rpm, 1200 N·m; 1500 rpm, 400 N·m; 1500 rpm, 800 N·m; and 1500 rpm, 1200 N·m. The prediction accuracy results for the test set are shown in Table 6. It should be noted that to avoid random errors, the average diagnostic accuracy of 10 consecutive test sets under the same parameter settings was taken as the final result.

[0169] Table 6 Accuracy of normal and misfire datasets

[0170]

[0171] It can be seen that when the model is trained, verified, and tested using a dataset from the same operating condition, different input samples achieve relatively high fault recognition accuracy in the test set. Specifically, using multi-channel vibration signals, SSSM-30°, SSSM-60°, and MSSM as input samples, a 100% fault recognition accuracy rate was achieved in both the binary and ternary classification tasks. This demonstrates the effectiveness of the proposed method for fault recognition under a single operating condition, and also demonstrates that the quality of the constructed dataset is relatively high. However, the fault recognition accuracy in the binary and ternary classification tasks using SSSM-10° and SSSM-720° as input samples did not reach 100%, indicating that single-scale sample input has certain limitations, and multi-scale fusion is more effective.

[0172] To verify the adaptability of the proposed method to varying operating conditions, the model was trained and validated using a dataset from one operating condition. It was then tested using datasets from other operating conditions and compared with multi-channel raw vibration signal samples from multiple cylinders and similarity matrix sample inputs at a single scale. Specifically, three variable operating conditions were analyzed: constant load but varying speed, constant speed but varying load, and simultaneous speed and load variations. The datasets used were the six operating conditions listed in Table 6. For example, the number A represents a dataset with a speed of 1200 rpm and a load of 400 N·m. A→D indicates that the dataset from operating condition A was used as the training and validation set, and the dataset from operating condition D was used as the test set.

[0173] (1) Under the same load, use the dataset at one speed for training and verification, and use the dataset at another speed for testing. Figure 10 The recognition accuracy and standard deviation of the test set for different sample inputs under each variable speed task are shown, and the average recognition accuracy and standard deviation of different sample inputs under all variable speed tasks are calculated.

[0174] from Figure 10 As can be seen from the results, the proposed method achieves the best performance in each variable speed prediction task, with fault identification accuracy exceeding 90% and an average accuracy of 95.20%. This demonstrates that the proposed method has good adaptability to variable speeds. Furthermore, the proposed method has the lowest standard deviation, demonstrating its robustness in variable speed fault identification tasks.

[0175] (2) When the speed is the same, use the dataset under one load for training and validation, and use the dataset under another load for testing. Figure 11 The recognition accuracy and standard deviation of the test set for different sample inputs under each variable load task are shown, and the average recognition accuracy and standard deviation of different sample inputs under all variable load tasks are calculated.

[0176] from Figure 11 As can be seen from the results, the proposed method achieves the best performance for each variable load prediction task, with fault identification accuracy exceeding 90% and an average accuracy of 96.02%. This demonstrates the proposed method's excellent adaptability to variable loads. Furthermore, the proposed method has the lowest standard deviation, demonstrating its robustness under variable load fault identification tasks. Compared to the variable speed task, the accuracy under the variable load task is relatively high, while the standard deviation is similar, indicating that the impact of load changes on the signal may be smaller than that of speed changes.

[0177] (3) In order to further verify the fault identification ability of the model under complex and variable working conditions, a dataset in which the speed and load change simultaneously is used for training, verification, and testing. Figure 12 The test set recognition accuracy and standard deviation of different sample inputs under each variable speed + variable load task are shown, and the average recognition accuracy and standard deviation of different sample inputs under all variable speed + variable load tasks are also calculated.

[0178] from Figure 12 As can be seen from the results, the proposed method achieves the best performance in each variable speed + variable load prediction task, with fault identification accuracy exceeding 90% and an average accuracy of 94.58%. This demonstrates that the proposed method is highly adaptable even when both speed and load vary simultaneously. Furthermore, the proposed method has the lowest standard deviation, demonstrating its robustness in fault identification tasks involving simultaneous speed and load variations. Compared to tasks involving only speed variation and only load variation, the accuracy for the variable speed + variable load task is relatively low, while the standard deviation is similar, indicating that simultaneous changes in speed and load have a greater impact on the signal.

[0179] Finally, the average accuracy and standard deviation of the three tasks of variable speed, variable load, and variable speed + variable load under different sample inputs are calculated and shown in Table 7.

[0180] Table 7 Average recognition accuracy and standard deviation of comprehensive variable working conditions for different sample inputs

[0181]

[0182] As can be seen, the proposed method achieved an average accuracy of 95.28% in the comprehensive variable operating condition fault identification task, achieving the highest accuracy and lowest standard deviation. Furthermore, although the accuracy of single-scale inputs, such as SSSM-10° and SSSM-10°, was lower than that of the original multi-channel sample input, higher accuracy was achieved after fusing multiple single-scale inputs. This demonstrates that the proposed multi-branch residual convolutional neural network fault identification model, which uses a multi-scale similarity matrix as input, can effectively extract multi-sensor multi-scale fusion features, improving the accuracy and robustness of vibration anomaly detection and fault identification under complex variable operating conditions.

Claims

1. A diesel engine sensor multi-scale fusion feature extraction and fault warning method, characterized in that: The following steps are involved: Step 1: Preprocess the collected vibration signal 1.1 Noise Reduction Preprocessing The original signal is processed by multi-cycle superposition and averaging. The specific process is expressed as follows: Where i = 1, 2, ..., N d , y is the signal after multi-period superposition and averaging, V acc (i,j) represents the vibration value of the i-th data point in the j-th engine working cycle, N a Indicates the number of superposition cycles, N a =60, N d Represents the number of data points in a single engine operating cycle; For the input signal y after multi-period averaging, the expression for 5kHz high-pass filtering using a Butterworth high-pass filter is as follows: y f =HPF(y,f c ,f s ,order) Among them, y f is the output signal after filtering, HPF(·) is the high-pass filter function, f c is the filter cutoff frequency, f c =5kHz, f s is the sampling frequency of the signal, which is initially set to 51200Hz, and order is the filter order, which ranges from 3 to 8; 1.2 Envelope extraction First, find the signal y f All local maximum points in are first-order extreme points; Then, find all the maximum points from the first-order extreme points as the connection envelope points, and remove the mean of the extracted envelope points; suppose the sequence of a certain order envelope feature point is n d is the total number of data points, then the formula for removing the mean is expressed as: Finally, the sequence obtained after removing the mean Perform Hermite interpolation to obtain the envelope y e ; 1.3 Phase Alignment Signal y e is the signal collected by the diesel engine in one working cycle, that is, 720° crankshaft angle, and the number of sampling points is N d ; The signal y after phase alignment according to the different phase angles of different cylinders p for: y p =stack[y e (angle / 720×N d :N d ),y e (1:angle / 720×N d )] Among them, stack[a1,a2] represents the concatenation of signals a1 and a2, y e (n1:n2) represents signal y e The n1th to n2th elements of , angle represents the corresponding phase angle; Step 2: Construct the multi-sensor signal multi-scale similarity matrix MSMSSM For two discrete signals x1 and x2, the cross-correlation function between them is expressed as: Wherein, τ=1-N,2-N,...,-1,0,1,...,N-2,N-1; m is the index of the signal sample point, N is the length of the signal, and the signal length is consistent by default. If it is inconsistent, it is aligned by resampling, truncation or zero padding, and τ is the delay; The obtained cross-correlation function is normalized so that the maximum value of the autocorrelation function is 1 at the time delay τ = 0: The normalized cross-correlation function takes the maximum value as the measure of the similarity between the two signals, and the similarity r of the signals x1 and x2 is expressed as: The similarity r of signals x1 and x2 ranges from [0,1], and the larger the value, the greater the similarity. When r = 0, it means that the signals are completely dissimilar. When r = 1, it means that the signals are exactly the same at the time delay τ. The similarity matrix R of multi-sensor signals is constructed as follows: Among them, for a multi-cylinder diesel engine, the number of cylinders is n, and the number of corresponding cylinder head vibration signal sensors is also n; represents the set of n×n real matrices, r pq It represents the similarity between the cylinder head vibration signals collected by the pth sensor and the qth sensor after preprocessing, p,q∈(1,2,...,n); in addition, the diagonal elements of the matrix represent the similarity of the sensor signals themselves, so the values ​​of the diagonal elements are all 1; Then, the multi-sensor signal similarity matrix R is subjected to eigenvalue decomposition; In the formula, since the correlation of sensor signals is not 0, all entries of the matrix R are positive. According to the Perron-Frobenius theorem, the matrix R has a unique maximum real eigenvalue, expressed as λ max , the corresponding eigenvector is represented as v = [v1, v2, ..., v p ,...,v n ] T , with a strictly positive component v p , p=1,2,...,n; at the same time, Rv=λ max v relationship holds, which is expressed as: Further expanding the above formula, we get: r p1 v1+r p2 v2+…+(1-λ max )v p +…+r pn v n =λ max v p In the formula, λ max v p Comprehensive r p1 ,r p2 ,...,r pn Therefore, λ max v p It is regarded as a comprehensive correlation measure between all sensors and the p-th sensor; Define the indicator Ind based on the PF vector p To detect and locate faulty sensor signals, the specific formula is as follows: Ind p is the fault detection index, which is the comprehensive support degree of all sensors to the pth sensor, v q is the index of the characteristic component in the characteristic vector; if there is no fault in the measured data, the signals of all sensors are similar, and the index Ind p is 1 / n; on the contrary, the component corresponding to the fault data will be smaller than the non-fault data, and the index Ind p will be less than 1 / n; Then, the signal is segmented at multiple scales to construct the multi-sensor signal similarity matrix R at different scales. c ; According to the characteristics of the impact duration of diesel engine vibration signals, the sample size of multi-cylinder vibration signals in one working cycle is n×N, where n represents the number of sensors and N represents the signal length; four different window sizes c are used, namely 10°, 30°, 60° and 720°, with a step size range of 1° to 5°. According to the step size of 1°, the number of similarity matrices at each scale is 711, 691, 661 and 1 respectively; for each scale c, the signal is divided into sub-signals of length c / 720×N, and the number is (720-c) / 1+1; definition is the kth sub-signal of sensor p at scale c; calculate the sub-signal and Similarities between The multi-sensor similarity matrix at scale c is expressed as: The multi-scale similarity matrix of multi-sensor signals is further expressed as Where l = ∑((720-c) / 1+1) is the number of channels of the matrix; Finally, the constructed multi-scale similarity matrix of multi-sensor signals is used as the input of the multi-branch residual convolutional neural network; Step 3: Construct a multi-branch residual convolutional neural network The multi-branch residual convolutional neural network includes a multi-channel input fusion module, a residual module and a multi-branch output fusion module, and adopts a residual learning mechanism and a multi-branch fusion mechanism to improve classification accuracy and robustness; 3.1 Multi-channel input fusion layer design The multi-channel fusion layer fuses the multi-channel matrices into a single-channel matrix through convolution operations; the input feature set is I E ={I1,I2,I3,I4}, where is the input of each branch, E represents the index of the four scales, n represents the number of sensors, l E Indicates the number of similarity matrices at each scale; the multi-channel fusion layer uses the Conv2D type, with a size of 1×1, a step size of 1, and 16 convolution kernels to perform multi-channel fusion on the input of each branch; the output calculation formula of each branch after multi-channel fusion is as follows: Among them, X E (B t ,D t ) indicates that the output feature is at position (B t ,D t ), I E (B t ,D t ,k) indicates that the input feature is at position (B t ,D t ) and the value at channel k, and K(1,1,k) represents the weight of convolution kernel K at position (1,1,k); 3.2 Residual Module Design After the multi-channel input fusion layer, the branches of each network are mainly composed of stacked residual modules. Residual module 1 contains two consecutive Conv2D layers. The first Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 32. The second Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 32. Residual module 2 also contains two Conv2D layers. The first Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 64. The second Conv2D layer has a convolution kernel size of 3×3, a stride of 1, and a number of 64. The convolution operation in the residual module is expressed as: x r (a)=g r [w r (a)×x r (a-1)+b r (a)] Among them, x r (a) is the output feature map of layer a, x r (a-1) is the output feature map of the a-1th layer, w r (a) is the convolution kernel, b r (a) is the network bias term; g r (·) represents the nonlinear activation function, using equation g r (x) = ReLU (x) = max(x, 0) to represent it; in addition, BN is the batch normalization function, which is also used in the convolution layer and is represented as BN(·); when the input and output sizes do not match, a 1×1 convolution operation is used for alignment; the convolution operation formula of the residual block is: x r (a)=w r (a)×ReLu{BN[x r (a-1)]}+b r (a)+x r (a-2) 3.3 Multi-branch output fusion layer design A multi-branch output fusion layer is designed at the end of the multi-branch network. For the information learned by each branch at different scales, the feature map extracted at the end of each branch is first expanded into a one-dimensional feature vector, namely the Flatten layer. Then, the output of each branch is weighted and fused through the multi-branch fusion layer. The specific process is as follows: By expressing the feature vector extracted from each branch as in The dimensions are equal, n b is the number of branches; the obtained fusion feature vector z f It is expressed as: in, are the feature vectors learned during the training process. The fusion weight of ; the dimension of the fusion weight is the same as the dimension of the feature vector; After the multi-branch output fusion layer, the fault identification result is obtained through the fully connected layer; Among them, the number of output nodes of the fully connected layer is 2, corresponding to the two categories of normal state and fault state. At the same time, the softmax activation function is used to convert the value output by the fully connected layer into a probability distribution, so that the value of each output node represents the probability that the sample belongs to the corresponding category, that is, to judge whether the sample belongs to the normal state or the fault state; the softmax function is defined as follows: Among them, y u (h) represents the output value of the u-th sample, e is a natural constant, h is the category index from 1 to C, and C is the number of categories; The loss function uses the Adam optimization algorithm and the cross entropy loss function; the training goal is to minimize the loss function through the optimization algorithm; the cross entropy loss function formula is as follows: Among them, N t is the number of training samples, γ uh is the class h label in the actual label of the u-th sample, where γ uh The value is 0 or 1, indicating whether the sample belongs to category h, ρ uh It uses the softmax function to predict the probability that the u-th sample belongs to category h.

Images