A data-driven method for locating and identifying fault information in diesel engine vibration signals

Abstract

The application discloses a data-driven fault information positioning and identification method for diesel engine vibration signals, and belongs to the technical field of equipment state monitoring and diagnosis. First, diesel engine fault change experimental data are collected, training set, verification set and test set samples are divided, and corresponding label sets are established. Second, a fault positioning model with a sparse attention module, a label prediction module and a prior distribution module is established, and a model training method is established. Third, multiple fault positioning models are obtained through multiple training to form a model set. Finally, a model evaluation index is established, a fault positioning model with the optimal index is obtained from the model set, a sparse weight vector of the fault positioning model is obtained, and the position of the sparse vector with the value of 1 is analyzed to position the fault information, thereby providing data support for the inspection and maintenance of the unit.

Description

Technical Field

[0001] This invention relates to a data-driven method for locating and identifying fault information in diesel engine vibration signals, applicable to the field of diesel engine vibration monitoring and fault diagnosis technology. Background Technology

[0002] Diesel engines, as one of the core power equipment in large-scale industrial fields, have a wide range of applications. Condition monitoring and fault diagnosis of diesel engines can effectively diagnose abnormal or faulty conditions, improving the reliability and safety of equipment operation. Since the vibration signal of diesel engine equipment is a typical angular domain signal, there is a correspondence between the vibration signal and the motion process of key moving parts of the diesel engine. Faults in moving parts will cause abnormal symptoms in the corresponding angular domain signal segment of the vibration signal. Diesel engine vibration signals exhibit complex coupling characteristics of multi-source impacts. Traditional methods of locating fault information based on fault mechanisms are limited by the degree of understanding of these mechanisms. Data-driven methods seek the intrinsic differences in signals caused by faults from the perspective of the data itself. By combining sparse attention mechanisms and deep learning to construct a fault location model, fault information can be located and identified in the signal, providing guidance for diesel engine fault diagnosis and improving the efficiency and accuracy of fault analysis.

[0003] Based on the structure and working principle of diesel engines, the fault information of typical faults is mostly concentrated in one or more narrow angular domain signal segments, that is, the fault information presents a fragmented form in the angular domain of the vibration signal. This invention fully considers the above characteristics of typical diesel engine fault information, employing a sparse attention mechanism and weight distribution constraints to obtain adaptive segmented sparse weights for the signal, achieving autonomous segmented localization and identification of fault information. This makes the identification results of the data-driven method more consistent with the angular domain distribution characteristics of typical diesel engine fault information, providing effective technical support for fault-sensitive feature extraction and fault diagnosis. Summary of the Invention

[0004] The purpose of this invention is to provide a data-driven method for locating and identifying fault information in vibration signals in the field of diesel engine fault diagnosis.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] First, establish a sample dataset, collect vibration data of diesel engines under normal and fault conditions, and perform normalization processing; map a label with a unique heat-coded label to each sample and establish a label set; divide the data into training set, validation set, and test set samples.

[0007] Second, a fault location model is established, which consists of three modules. The first module is a sparse attention module, which is the basic module for realizing the ability to locate fault information. The second module is a label prediction module, which is based on its function to ensure the accuracy of fault information location. The third module is a prior distribution module, which is based on its function to ensure that the result of fault information location conforms to the mechanism that diesel engine fault impact has short-term continuity.

[0008] Third, establish a training mechanism to train the fault localization model. This step mainly involves specifying the training methods for each parameter in the model, including the sparse weight vector in the sparse attention module and all trainable weights in the label prediction module. Finally, multiple model results are obtained and combined into a model set.

[0009] Fourth, establish a model evaluation method to select the optimal model from the model set and locate the fault information; that is, obtain the sparse weight vector of the model with the best evaluation index in the model set, read its value, and if the value is 1, it is considered that the fault information exists in this signal segment; otherwise, it is considered that there is no fault information.

[0010] A data-driven method for locating fault information in diesel engines, characterized by comprising the following steps:

[0011] Step 1: Create a sample dataset

[0012] 1.1 Collect diesel engine experimental data according to the fault distribution, and organize them to obtain the total sample set D, as shown in formulas (1), (2), and (3):

[0013]

[0014] Where D is the total sample set. This refers to a three-dimensional array with T×N×K elements defined in the real number field. Where T, N, and K are all positive integers. T represents the set of positive integers; in fault diagnosis, T represents the number of fault categories, N represents the number of samples under each fault category, and K represents the number of sequence points contained in each sample; F t Refers to the set of fault samples of type t;

[0015]

[0016] Where F refers to a set of fault samples, and X is any sample; X∈D, Let F be any set of fault samples of any type in the total sample set D, with the number of elements being N×K and defined in the real number field; let X be any sample, with the number of elements being K and defined in the real number field; let F contain any set of fault samples of type N; and let X contain any sequence points of type K.

[0017] 1.2 The collected signals are normalized, that is, each sample X is updated in the following way, as shown in formula (4):

[0018]

[0019] Where, x new For the normalized sample, x min Let x be the minimum value in the sequence of samples x. max Let x be the maximum value in the sequence of samples x. mean The average value in the sample x sequence;

[0020] 1.3 Divide the total sample set D into training set, validation set, and test set according to the number of samples, with a ratio of 5:1:1, denoted as D0. a D b D c And for all samples, standard one-hot encoding is performed according to the fault as the label of the sample, denoted as l for the label of X. From standard one-hot encoding, it can be seen that the number of elements in any label l is T; therefore, the total set of labels L can be obtained from the total set of samples D. A three-dimensional array with a number of elements of T×N×T, defined in the real number field;

[0021] Step 2: Establish a fault location model based on prior mechanisms

[0022] The purpose of the fault location model is to locate fault information in sample X. The model consists of three modules: the first module is the sparse attention module, which is the basic module for realizing the ability to locate fault information; the second module is the label prediction module, which is based on its function to ensure the accuracy of fault information location; and the third module is the prior distribution module, which is based on its function to ensure that the result of fault information location conforms to the mechanism that diesel engine fault impact has short-term continuity.

[0023] 2.1 The sparse attention module is established as follows;

[0024] Let the number of sequence points of sample X be K, as shown in formula (5):

[0025]

[0026] Establish a trainable attention weight vector w s Note that w s This process is shared across all samples, as shown in formula (6):

[0027]

[0028] Establish a sparse weight vector m. Note that m is shared by all samples, as shown in formula (7):

[0029] m = [m1, m2, ..., m i ,…,m K ] T ∈{0,1} K (7)

[0030] in, A one-dimensional vector with K elements, defined in the real number field; {0,1} K A one-dimensional vector of number K elements, defined only on the positive integers 0 and 1; x i For the i-th segment of sample x, For x i In the attention weight vector w s The corresponding value in, m i For x i The corresponding value w in the sparse weight vector m s All elements in m are initialized using standard He initialization, and all elements in m are initialized to 1.

[0031] The sample X update method in the sparse attention mechanism is established as shown in formula (8):

[0032]

[0033] in, For sparse samples; Sigmoid is the Sigmoid activation function; ⊙ is the multiplication flag for corresponding positions in the vector; X is the sample, m is the sparse weight vector, and w s This represents the trainable attention weight vector within the mechanism; This refers to the process after performing the calculations using the formula (8) above. The number of elements in it is still K, and its domain is still in the real number field;

[0034] 2.2 Establish the label prediction module, the process is shown in formula (9):

[0035]

[0036] Here, 1DCNN refers to a diagnostic model that uses any one-dimensional convolutional neural network, due to the sparse attention mechanism in the output. Without changing the number of elements and the domain of sample X, any mature one-dimensional convolutional neural network diagnostic model can use the sparse attention mechanism as a prerequisite. , where W represents sparse samples; W represents all trainable parameters in the one-dimensional convolutional neural network diagnostic model. For predicting labels, This refers to an array with T elements defined in the real number field;

[0037] 2.3 Establish the prior distribution module, as shown in formula (10):

[0038]

[0039] Where p is the prior distribution module; K is the number of sequence points in sample X; β is the loss balance coefficient, which is related to the number of digits in K. When K is two digits or less, β = 0.01; when K is more than two digits, ... The number of bits in the K value; w s′ For w s Find the first derivative; |·| is for finding the absolute value; ReLU refers to the standard ReLU function;

[0040] The fault location model described in 2.4 needs to define a loss function, training method, and number of iterations, as shown below;

[0041] The loss function is established as shown in formula (11):

[0042]

[0043] Among them, c c For categorical crossentropy The standard multi-class cross-entropy loss function requires setting the predicted label and the true label; Here, 'l' represents the predicted label, 'l' represents the true label, and 'Loss' represents the predicted label. The loss function calculated with respect to the true label l;

[0044] The training method is established as shown in formula (12):

[0045] Adam(0.001),minimize(Loss) (12)

[0046] Where Qdam(0.001) is the standard Adam training algorithm with an initial learning rate of 0.001; minimize(Loss) means that the learning objective is to minimize the loss.

[0047] Establish the number of iterations E P E P It is a positive integer whose value is related to the training target. During a single training cycle, Loss increases with E. P As it increases, it continuously decreases, requiring E to be continuously increased. P Until the loss stops decreasing;

[0048] Step 3: Establish a training mechanism to train the fault localization model.

[0049] 3.1 The training set D a Import the fault location model for training, and then use it on the validation set D. bVerification is performed on the test set D; after training, the data is taken from the test set D. c Find the model with the smallest loss and derive its w. s W, m and the diagnostic accuracy acc of the test set; this process is represented by the following equation (13):

[0050]

[0051] Wherein, U refers to the training process. This refers to the result after the i-th training iteration on the test set D. c Find the model with the minimum loss; derive the trainable attention weight vector w in this model. s In a one-dimensional convolutional neural network diagnostic model, all trainable parameters W, sparse weight vector m, and diagnostic accuracy acc on the test set are denoted as follows:

[0052] 3.2 Establish η, i.e., the number of training iterations, which is a positive integer; define the sparse threshold ε and its training method, and the training method for the sparse weight vector m;

[0053] The training method for ε is shown in formula (14):

[0054]

[0055] Where ε is the sparsity threshold; η refers to the number of training iterations set by the engineer, and its value is related to the diagnostic accuracy acc on the test set. After a certain training iteration, the fault location model results... At this point, η needs to be further increased, that is, the next round of training should be carried out, until... The diagnostic accuracy on the test set after the (η-1)th training iteration;

[0056] The training method for the sparse weight vector m is shown in Equation (15):

[0057]

[0058] Where ⊙ is the symbol for multiplying corresponding positions of vectors. For the sparse weight vector m and attention weight vector w after the i-th training iteration... s The value of ; ε is the sparsity threshold; η is the number of training iterations; for The j-th sequence point of the vector obtained after multiplying corresponding positions; K is the number of sequence points of sample X, which is also numerically equal to The number of sequence points; The sparse weight vector is the input required for the (i+1)th training of the fault location model;

[0059] 3.3 Set other weight training methods between adjacent training iterations, namely:

[0060]

[0061] Among them, ⊙ is the symbol for multiplying corresponding positions of vectors, which are respectively the value of the attention weight vector w s after the i-th training and the value of all trainable parameters W in the one-dimensional convolutional neural network diagnosis model; which are respectively the value of the sparse weight vector m required for the (i + 1)-th training, the value of the attention weight vector w s and the value of all trainable parameters W in the one-dimensional convolutional neural network diagnosis model;

[0062] 3.4 Traverse all training times to obtain η fault location models and form a set V, that is

[0063]

[0064] Step 4: Establish an evaluation method, select the optimal model and locate the fault information

[0065] 4.1 Establish a model evaluation index based on the sparse threshold ε and the diagnostic accuracy acc of the test set, as shown in formula (18):

[0066] J = θ × acc + (1 - θ) × ε, θ ∈ (0, 1) (18)

[0067] Among them, θ is the evaluation measurement coefficient, used to adjust the proportion of the diagnostic accuracy acc of the test set and the sparse threshold ε in the evaluation index; θ is related to the number of fault categories T. It is recommended that when 2 < T < 10, θ = 0.1 × T; when T ≥ 10, ε is the sparse threshold;

[0068] 4.2 In the set V, obtain the data-driven model when J is the largest, export the sparse weight vector m under this model, and read the positions where the value is 1, that is, the signal segments where the fault information occurs. Description of the Drawings

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0070] Figure 1 is a flowchart of a data-driven diesel engine fault information location method provided according to an example of the present application;

[0071] Figure 2This is a schematic diagram of a fault signal and a tag, provided based on an example of this application;

[0072] Figure 3 This is a sparse attention structure diagram provided based on an example from this application;

[0073] Figure 4 This is a label prediction model structure diagram provided based on an example of this application;

[0074] Figure 5 This is a schematic diagram of a data-driven model structure for multiple training iterations, based on an example from this application.

[0075] Figure 6 This is a schematic diagram of a sparse vector result provided based on an example of this application;

[0076] Figure 7 This is a comparison diagram of a worn bearing bush and a normal bearing bush, based on an example provided in this application. Detailed Implementation

[0077] To better understand the technical solution of this invention, the cylinder head vibration signal of a TBD234V12 diesel engine is used as a specific implementation object to calculate and implement the data-driven diesel engine fault information localization method proposed in this invention.

[0078] Figure 1 This is a flowchart of a data-driven diesel engine fault information localization method provided in this application. See also... Figure 1 The implementation process and results of the present invention are as follows:

[0079] A vibration acceleration sensor is placed at the crankshaft main bearing housing of the engine to collect the main bearing housing signal. The sampling frequency is 51200Hz, and the signal is divided into angular domain signals with a crankshaft turning angle of 720° as one cycle.

[0080] Step 1: Create a sample dataset

[0081] 1.1 Collect diesel engine experimental data according to the fault distribution, and organize them to obtain the total sample set D, as shown in formulas (1), (2), and (3):

[0082]

[0083] Where D is the total sample set. This refers to a three-dimensional array with T×N×K elements defined in the real number field. Where T, N, and K are all positive integers. T represents the set of positive integers; in fault diagnosis, T represents the number of fault categories, N represents the number of samples under each fault category, and K represents the number of sequence points contained in each sample; F tRefers to the set of fault samples of type t;

[0084]

[0085] Where F refers to a set of fault samples, and X is any sample; X∈D, Let F be any set of fault samples of any type in the total sample set D, with the number of elements being N×K and defined in the real number field; let X be any sample, with the number of elements being K and defined in the real number field; let F contain any set of fault samples of type N; and let X contain any sequence points of type K.

[0086] In this example, the lubricating oil supply was reduced for an extended period of time, and the collected signals were divided into four fault states according to the time sequence: Fault I, Fault II, Fault III, and Fault IV, with T=4, N=100, and K=6144.

[0087] 1.2 The collected signals are normalized. Let X refer to a sample. Each sample is updated as follows, as shown in formula (4):

[0088]

[0089] Where, x new For the normalized sample, x min Let x be the minimum value in the sequence of samples x. max Let x be the maximum value in the sequence of samples x. mean This is the average value in the sample x sequence.

[0090] 1.3 Divide the total sample set D into training set, validation set, and test set according to sample category, with a ratio of 5:1:1, denoted as D0. a D b D c And all samples are tagged with standard one-hot codes based on their faults, X (t,i) The label is written as L (t,i) ;

[0091] In this example, the sample label for fault I is (1,0,0,0), the sample label for fault II is (0,1,0,0), the sample label for fault III is (0,0,1,0), and the sample label for fault IV is (0,0,0,1).

[0092] Step 2: Establish a fault location model based on prior mechanisms

[0093] The purpose of the fault location model is to locate fault information in sample X. The model consists of three modules: the first module is the sparse attention module, which is the basic module for realizing the ability to locate fault information; the second module is the label prediction module, which is based on its function to ensure the accuracy of fault information location; and the third module is the prior distribution module, which is based on its function to ensure that the result of fault information location conforms to the mechanism that diesel engine fault impact has short-term continuity.

[0094] 2.1 The sparse attention module is established as follows;

[0095] Let the number of sequence points of sample X be K, as shown in formula (5):

[0096]

[0097] Establish a trainable attention weight vector w s Note that w s This process is shared across all samples, as shown in formula (6):

[0098]

[0099] Establish a sparse weight vector m. Note that m is shared by all samples, as shown in formula (7):

[0100] m = [m1, m2, ..., m i ,…,m K ] T ∈{0,1} K (7)

[0101] in, A one-dimensional vector with K elements, defined in the real number field; {0,1} K A one-dimensional vector of number K elements, defined only on the positive integers 0 and 1; x i For the i-th segment of sample x, For x i In the attention weight vector w s The corresponding value in, m i For x i The corresponding value w in the sparse weight vector m s All elements in m are initialized using standard He initialization, and all elements in m are initialized to 1.

[0102] The sample X update method in the sparse attention mechanism is established as shown in formula (8):

[0103]

[0104] in, For sparse samples; Sigmoid is the Sigmoid activation function; ⊙ is the multiplication flag for corresponding positions in the vector; X is the sample, m is the sparse weight vector, and w s This represents the trainable attention weight vector within the mechanism; This refers to the process after performing the calculations using the formula (8) above. The number of elements in it is still K, and its domain is still in the real number field;

[0105] In this example, Figure 3 The calculation process for this step is described;

[0106] 2.2 Establish the label prediction module, the process is shown in formula (9):

[0107]

[0108] Here, 1DCNN refers to a diagnostic model that uses any one-dimensional convolutional neural network, due to the sparse attention mechanism in the output. Without changing the number of elements and the domain of sample X, any mature one-dimensional convolutional neural network diagnostic model can use the sparse attention mechanism as a prerequisite. , where W represents sparse samples; W represents all trainable parameters in the one-dimensional convolutional neural network diagnostic model. For predicting labels, This refers to an array with T elements defined in the real number field;

[0109] The process is as follows: Figure 4 As shown.

[0110] 2.3 Establish the prior distribution module, as shown in formula (10):

[0111]

[0112] Where p is the prior distribution module; K is the number of sequence points of sample X; β is the loss balance coefficient, which is related to the number of digits in k. When K is two digits or less, β = 0.01; when K is more than two digits, ... The number of bits in the K value; w s′ For w s Find the first derivative; |·| is for finding the absolute value; ReLU refers to the standard ReLU function;

[0113] The fault location model described in 2.4 needs to define a loss function, training method, and number of iterations, as shown below;

[0114] The loss function is established as shown in formula (11):

[0115]

[0116] Among them, c c For categorical crossentropy The standard multi-class cross-entropy loss function requires setting the predicted label and the true label; Here, 'l' represents the predicted label, 'l' represents the true label, and 'Loss' represents the predicted label. The loss function calculated with respect to the true label l;

[0117] The training method is established as shown in formula (12):

[0118] Adam(0.001),minimize(Loss) (12)

[0119] Here, Adam(0.001) is the Adam training algorithm with a standard initial learning rate of 0.001; minimize(Loss) means that the learning objective is to minimize the loss.

[0120] Establish the number of iterations E P E P It is a positive integer whose value is related to the training target. During a single training cycle, Loss increases with E. P As it increases, it continuously decreases, requiring E to be continuously increased. P Until the loss stops decreasing;

[0121] In this example, β = 0.5;

[0122] Step 3: Establish a training mechanism to train the fault localization model.

[0123] 3.1 The training set D a Import the fault location model for training, and then use it on the validation set D. b Verification is performed on the test set D; after training, the data is taken from the test set D. c Find the model with the smallest loss and derive its w. s W, m and the diagnostic accuracy acc of the test set; this process is represented by the following equation (13):

[0124]

[0125] Wherein, U refers to the training process. This refers to the result after the i-th training iteration on the test set D. c Find the model with the minimum loss; derive the trainable attention weight vector w in this model. s In a one-dimensional convolutional neural network diagnostic model, all trainable parameters W, sparse weight vector m, and diagnostic accuracy acc on the test set are denoted as follows:

[0126] 3.2 Establish η, i.e., the number of training iterations, which is a positive integer; define the sparse threshold ε and its training method, and the training method for the sparse weight vector m;

[0127] The training method for ε is shown in formula (14):

[0128]

[0129] Where ε is the sparsity threshold; η refers to the number of training iterations set by the engineer, and its value is related to the diagnostic accuracy acc on the test set. After a certain training iteration, the fault location model results... At this point, η needs to be further increased, that is, the next round of training should be carried out, until... The diagnostic accuracy on the test set after the (η-1)th training iteration;

[0130] The training method for the sparse weight vector m is shown in Equation (15):

[0131]

[0132] Where ⊙ is the symbol for multiplying corresponding positions of vectors. For the sparse weight vector m and attention weight vector w after the i-th training iteration... s The value of ; ε is the sparsity threshold; η is the number of training iterations; for The j-th sequence point of the vector obtained after multiplying corresponding positions; K is the number of sequence points of sample X, which is also numerically equal to The number of sequence points; The sparse weight vector is the input required for the (i+1)th training of the fault location model;

[0133] 3.3 Set other weight training methods between adjacent training iterations, namely:

[0134]

[0135] Where ⊙ is the symbol for multiplying corresponding positions of vectors. The attention weight vector w after the i-th training iteration is... s The values ​​of and the values ​​of all trainable parameters W in the one-dimensional convolutional neural network diagnostic model; These represent the values ​​of the sparse weight vector m and the attention weight vector w, respectively, which are the input values ​​required for the (i+1)th training iteration. s The values ​​of and the values ​​of all trainable parameters W in the one-dimensional convolutional neural network diagnostic model;

[0136] 3.4 Iterate through all training iterations to obtain η fault location models, forming a set V, i.e.

[0137]

[0138] In this example, E P = 100, η = 20, ε = Ψ(i) = (1×(1 - (η - i) 2 / η 2 )) 0.5 Figure 5 The calculation process of this step is introduced;

[0139] Step 4: Establish an evaluation method, select the optimal model and locate the fault information

[0140] 4.1 Establish a model evaluation index based on the sparse threshold ε and the diagnostic accuracy acc of the test set, as shown in formula (18):

[0141] J = θ×acc + (1 - θ)×ε, θ∈(0,1) (18)

[0142] Where, θ is the evaluation measurement coefficient, used to adjust the proportion of the diagnostic accuracy acc of the test set and the sparse threshold ε in the evaluation index; θ is related to the number of fault categories T. It is recommended that when 2 < T < 10, θ = 0.1×T; when T ≥ 10, ε is the sparse threshold;

[0143] 4.2 In the set V, obtain the data-driven model when J is the largest, derive the sparse weight vector m under this model, and read the positions with a value of 1, that is, the signal segments where the fault information occurs.

[0144] In this example, θ = 1, μ = 0.5, and the final result is selected as Figure 6 shown. The part inside the black box is the part with a sparse value of 1, and the rest is 0. It can be seen that the signal indicates that the fault occurs around 0, 180, 360, 540, 720 degrees, that is, the top and bottom dead points of the piston movement, indicating that there is a certain rubbing between the connecting rod, bearing bush and crankshaft. Disassembly and inspection found that the bearing bush was worn, as Figure 7 shown.

[0145] In summary, this example proves the effectiveness of the proposed data-driven diesel engine fault information location method.

Claims

1. A data-driven method for locating fault information in a diesel engine, characterized by comprising the following steps: Step 1: Create a sample dataset 1.1 Collect diesel engine experimental data according to the fault distribution, and organize them to obtain the total sample set D, as shown in formulas (1), (2), and (3): in, D is the total sample set. This refers to a three-dimensional array with T×N×K elements defined in the real number field. Where T, N, and K are all positive integers. Represents the set of positive integers; In fault diagnosis, T represents the number of fault categories, N represents the number of samples under each fault category, and K represents the number of sequence points contained in each sample; F t Refers to the set of fault samples of type t; Where F refers to a set of fault samples, and X is any sample; Let F be any set of fault samples of any type in the total sample set D, with the number of elements being N×K and defined in the real number field; let X be any sample, with the number of elements being K and defined in the real number field; let F contain any set of fault samples of type N; and let X contain any sequence points of type K. 1.2 The collected signals are normalized, that is, each sample X is updated in the following way, as shown in formula (4): Where, x new For the normalized sample, x min Let x be the minimum value in the sequence of samples x. max Let x be the maximum value in the sequence of sample x. mean The average value in the sample x sequence; 1.3 Divide the total sample set D into training set, validation set, and test set according to the number of samples, with a ratio of 5:1:1, denoted as Di. a D b D c And for all samples, standard one-hot encoding is performed according to the fault as the label of the sample, denoted as l for the label of X. From standard one-hot encoding, it can be seen that the number of elements in any label l is T; therefore, the total set of labels L can be obtained from the total set of samples D. A three-dimensional array with T×N×T elements, defined in the real number field; Step 2: Establish a fault location model based on prior mechanisms The purpose of the fault location model is to locate fault information in sample X. The model consists of three modules: the first module is the sparse attention module, which is the basic module for realizing the ability to locate fault information; the second module is the label prediction module, which is based on its function to ensure the accuracy of fault information location; and the third module is the prior distribution module, which is based on its function to ensure that the result of fault information location conforms to the mechanism that diesel engine fault impact has short-term continuity. 2.1 The sparse attention module is established as follows; Let the number of sequence points of sample X be K, as shown in formula (5): Establish a trainable attention weight vector w s Note that w s This process is shared across all samples, as shown in formula (6): Establish a sparse weight vector m. Note that m is shared by all samples, as shown in formula (7): m=[m1,m2,…,m i ,…,m K ] T ∈{0,1} K (7) in, A one-dimensional vector with K elements, defined in the real number field; {0,1} K A one-dimensional vector of number K elements, defined only on the positive integers 0 and 1; x i For the i-th segment of sample x, For x i In the attention weight vector w s The corresponding value in, m i For x i The corresponding value w in the sparse weight vector m s All elements in m are initialized using standard He initialization, and all elements in m are initialized to 1. The sample X update method in the sparse attention mechanism is established as shown in formula (8): in, For sparse samples; Sigmoid is the Sigmoid activation function; ⊙ is the multiplication flag for corresponding positions in the vector; X is the sample, m is the sparse weight vector, and w s This represents the trainable attention weight vector within the mechanism; This refers to the process after performing the calculations using the formula (8) above. The number of elements in it is still K, and its domain is still in the real number field; 2.2 Establish the label prediction module, the process is shown in formula (9): Here, 1DCNN refers to a diagnostic model that uses any one-dimensional convolutional neural network, due to the sparse attention mechanism in the output. Without changing the number of elements and the domain of sample X, any mature one-dimensional convolutional neural network diagnostic model can use the sparse attention mechanism as a prerequisite. , where W represents sparse samples; W represents all trainable parameters in the one-dimensional convolutional neural network diagnostic model. For predicting labels, This refers to an array with T elements defined in the real number field; 2.3 Establish the prior distribution module, as shown in formula (10): Where p is the prior distribution module; K is the number of sequence points in sample X; β is the loss balance coefficient, which is related to the number of digits in K. When K is two digits or less, β = 0.01; when K is more than two digits, ... The number of bits in the K value; w s′ For w s Find the first derivative; |·| is for finding the absolute value; ReLU refers to the standard ReLU function; The fault location model described in 2.4 needs to define a loss function, training method, and number of iterations, as shown below; The loss function is established as shown in formula (11): Among them, c c For categorical crossentropy The standard multi-class cross-entropy loss function requires setting the predicted label and the true label; Here, 'l' represents the predicted label, 'l' represents the true label, and 'Loss' represents the predicted label. The loss function calculated with respect to the true label l; The training method is established as shown in formula (12): Adam(0.001), minimize(Loss) (12) where Adam(0.001) is the Adam training algorithm with a standard initial learning rate of 0.001; minimize(Loss) means that the learning objective is to minimize the loss. Establish the number of iterations E P E P It is a positive integer whose value is related to the training target. During a single training cycle, Loss increases with E. P As it increases, it continuously decreases, requiring E to be continuously increased. P Until the loss stops decreasing; Step 3: Establish a training mechanism to train the fault localization model. 3.1 The training set D a Import the fault location model for training, and then use it on the validation set D. b Verification is performed on the test set D; after training, the data is taken from the test set D. c Find the model with the smallest loss and derive its w. s W, m and the diagnostic accuracy acc of the test set; this process is represented by the following equation (13): Wherein, U refers to the training process. This refers to the result after the i-th training iteration on the test set D. c Find the model with the minimum loss; derive the trainable attention weight vector w in this model. s In a one-dimensional convolutional neural network diagnostic model, all trainable parameters W, sparse weight vector m, and diagnostic accuracy acc on the test set are denoted as follows: 3.2 Establish η, i.e., the number of training iterations, which is a positive integer; define the sparse threshold ε and its training method, and the training method for the sparse weight vector m; The training method for ε is shown in formula (14): in, ε is the sparsity threshold; η refers to the number of training iterations set by the engineer. Its value is related to the diagnostic accuracy (acc) on the test set. After a certain training iteration, the fault location model results... At this point, η needs to be further increased, that is, the next round of training should be carried out, until... The diagnostic accuracy on the test set after the (η-1)th training iteration; The training method for the sparse weight vector m is shown in Equation (15): Where ⊙ is the symbol for multiplying corresponding positions of vectors. For the sparse weight vector m and attention weight vector w after the i-th training iteration... s The value of ; ε is the sparsity threshold; η is the number of training iterations; for The j-th sequence point of the vector obtained after multiplying corresponding positions; K is the number of sequence points of sample X, which is also numerically equal to The number of sequence points; The sparse weight vector is the input required for the (i+1)th training of the fault location model; 3.3 Set other weight training methods between adjacent training iterations, namely: Where ⊙ is the symbol for multiplying corresponding positions of vectors. The attention weight vector w after the i-th training iteration is... s The values ​​of and the values ​​of all trainable parameters W in the one-dimensional convolutional neural network diagnostic model; These represent the values ​​of the sparse weight vector m and the attention weight vector w, respectively, which are the input values ​​required for the (i+1)th training iteration. s The values ​​of and the values ​​of all trainable parameters W in the one-dimensional convolutional neural network diagnostic model; 3.4 Iterate through all training iterations to obtain η fault location models, forming a set V, i.e. Step 4: Establish evaluation methods, select the optimal model, and locate fault information. 4.1 Establish a model evaluation index based on sparse threshold ε and test set diagnostic accuracy acc, as shown in formula (18): J=θ×acc+(1-θ)×ε,θ∈(0,1) (18) in, θ is the evaluation metric coefficient used to adjust the proportion of test set diagnostic accuracy acc and sparse threshold ε in the evaluation metrics. θ is related to the number of fault categories T. It is recommended that when 2 < T < 10, θ = 0.1×T; when T ≥ 10, ε is the sparse threshold; 4.2 In set V, obtain the fault location model when J is maximum, derive the sparse weight vector m under this model, and read the position where the value is 1, that is, the signal segment where the fault information occurs.

Images