Abnormal diagnosis method of diesel engine exhaust temperature based on neural network expert system

Abstract

The purpose of the present invention is to provide a method for diagnosing abnormal exhaust temperature of a diesel engine based on a neural network expert system, comprising the following steps: in view of the problem of abnormal exhaust temperature of a diesel engine, relevant operating data is obtained, and the data set is divided into two parts, one of which is used as training data and the other is used as test data; both the training data and the test data are described as a knowledge system; the fault feature matrix is ​​normalized to obtain a normalized fault feature matrix; a BP neural network is established; an expert knowledge base is established based on the output results of the BP neural network; and the model accuracy is verified. The present invention can accurately diagnose abnormal exhaust temperature of a diesel engine; solves the problem of long knowledge acquisition cycle of the expert system and difficulty in understanding the reasoning process and results of the neural network; the present invention uses the particle swarm algorithm to optimize two key parameters in the BP neural network model: connection weight and output threshold, which improves the classification accuracy and calculation time of the model.

Description

Technical Field

[0001] The invention relates to a diesel engine fault diagnosis method, in particular to a diesel engine exhaust temperature diagnosis method. Background Art

[0002] Diesel engines are the most widely used type of power machinery at present. The development of diesel engine technology has an important impact on my country's industry, agriculture, transportation, national defense construction and other aspects. Abnormal exhaust temperature of diesel engines is a specific sign of internal faults in diesel engines. It may involve single-system single fault, single-system multiple faults and multiple-system multiple faults, with complex coupling relationships. If diagnosis and maintenance are not carried out in time, it will lead to reduced mechanical efficiency and power of diesel engines, affecting the safe operation of equipment and systems, and even affecting the personal safety of operators. Therefore, the research on abnormal exhaust temperature diagnosis is of great significance to ensure the safe and efficient operation of diesel engines.

[0003] The document "Exhaust Temperature Abnormal Fault Case of Wärtsilä Auxiliary Engine" ("Navigation Technology", 2020) diagnosed the abnormal exhaust temperature of Wärtsilä diesel engine, checked the parameters such as scavenging temperature and turbine speed, disassembled and checked the relevant components between the governor and the high-pressure oil pump rack, and consulted the maintenance record book, and finally determined that the fault was the cylinder throttle reduction. The diagnostic method used in this document is in the post-maintenance stage. Its shortcomings are that the diagnosis cycle is long, the steps are cumbersome, the diagnostic means are backward, and it may destroy the optimal working state of the diesel engine during the troubleshooting and maintenance process. The document "Analysis and Troubleshooting of Large Exhaust Temperature Difference of MTU396 Diesel Engine" ("Internal Combustion Engine", 2021) disassembled the intake pipe cover, exhaust connection pipe, 8 injectors of A1~A4 and B1~B4 and other related components to address the problem of large exhaust temperature difference between A and B columns. After adjusting the injection advance angle and resetting the relevant components, the fault still exists. After disassembling the injection pump and injector for testing again, it was determined that the fault point was the stuck valve core of the return oil pipeline check valve. The diagnostic method used in this document repeatedly disassembles fault-related components and conducts multiple experiments when the fault exists, which causes great damage to the diesel engine. The fault troubleshooting process is complicated and tedious, and the technical means are backward. Summary of the invention

[0004] The purpose of the present invention is to provide a method for diagnosing abnormal exhaust temperature of a diesel engine based on a neural network expert system, which can overcome the shortcomings of a long expert system knowledge acquisition cycle and a neural network reasoning process and results that are difficult to understand during the diagnosis process.

[0005] The object of the present invention is achieved in that:

[0006] The present invention provides a method for diagnosing abnormal exhaust temperature of a diesel engine based on a neural network expert system, which is characterized by:

[0007] (1) To solve the problem of abnormal exhaust temperature of diesel engines, relevant operating data is obtained and the data set is divided into two parts, one of which is used as training data T r , and the other one is used as test data T e ;

[0008] (2) The training data T in step (1) r With test data T e are described as a knowledge system ∑ = (F, S), where the output result set F = {f 1 , f 2 , …, f n} is the typical state of the diesel engine, and the input parameter set S = {s 1 ,s 2 ,…,s k} is the actual operating parameter of the diesel engine, according to the training data T r Establish training fault feature matrix R r (∑)=(r ij ) n×k , in order to train the neural network; according to the test data T e Establish the test fault feature matrix R e (∑)=(r ij ) n×k , in order to test the neural network expert system after it is established, where r ij is the relevant parameter s in the actual operation of the diesel engine i The value of , i ranges from {i∈Z|1≤i≤n}, and j ranges from {j∈Z|1≤j≤k};

[0009] (3) For the fault feature matrix R in step (2) r (∑) and R e (∑) is normalized to obtain the normalized fault feature matrix R′ r (∑)=(r′ ij ) n×k and R′ e (∑)=(r′ ij ) n×k , where r′ ij is the relevant parameter s in the actual operation of the diesel engine i The normalized value has a value range of [0, 1], and training examples and test examples are established based on the fault feature matrix;

[0010] (4) Establish a BP neural network and use the training examples in step (3) to train the BP neural network, and use the particle swarm algorithm to accelerate the connection weight ρ jm , With the output threshold θ m , The optimization process obtains the optimal connection weight ρ jm , With the output threshold θ m , Output BP neural network model;

[0011] (5) Based on the BP neural network obtained in step (4), the connection between it and the expert system is established, that is, the expert knowledge base is established based on the output results of the BP neural network, and the inference engine and interpreter expert system structure is established, and the neural network expert system model is output;

[0012] (6) Use the neural network expert system model obtained in step (5) and the test case in step (3) to test and verify the accuracy of the model.

[0013] The present invention may also include:

[0014] 1. The step (1) of obtaining relevant operating data includes obtaining operating data of seven groups of relevant measuring points, which are respectively the data when the diesel engine is operating normally and the operating data when one of the following faults occurs: the fault type is a supercharger damage fault, an abnormal fuel injection fault, an intake pipe blockage fault, an exhaust valve damage fault, a fresh water system abnormality, and a high-pressure oil pipe leakage fault.

[0015] 2. The output result set F in step (2) includes: normal operation status f 1 and fault conditions, including supercharger damage 2 、Abnormal fuel injection 3 、Intake pipe blockage 4 , exhaust valve damage 5 、Freshwater system abnormality 6 , High-pressure oil pipe leakage 7 , that is, F = {f 1 , f 2 , f 3 , f 4 , f 5 , f 6 , f 7};

[0016] The input parameter set S includes: exhaust temperature s 1 , turbocharger speed s 2 , injection status signal 3 , air cooler outlet air pressure s 4 , exhaust valve status signal s 5 , Fresh water inlet pressure s 6 , High-pressure fuel pipe leakage level 7 , that is, S = {s 1 ,s 2 ,s 3 ,s 4,s 5 ,s 6 ,s 7};

[0017] Fault feature matrix R r (∑) and R e The form of (∑) is:

[0018]

[0019] The rows represent seven typical states that may occur in the diesel engine system, and the columns represent seven relevant operating parameters of the diesel engine that can be monitored.

[0020] 3. In step (3), the fault feature matrix R r (∑) and R e (∑) is normalized, that is, the fault feature matrix R r (∑) and R e Each column of (∑) is normalized using the maximum and minimum method:

[0021]

[0022] Among them, i (i = 1, ..., 7) is the state type identifier, j (j = 1, ..., 7) is the parameter type identifier, and the normalized value r′ ij In [0, 1], injection status signal s 3 With exhaust valve status signal s 5 The value is 1 in normal state and 0 in abnormal state;

[0023] Establish training examples or test examples, that is, take a row of the normalized fault feature matrix as the first dimension of the training example or test example, that is, the input vector, set the state corresponding to the row to 1, and set the other states to 0 to form the second dimension of the training example or test example, that is, the expected output vector. A total of 7 training examples are established; take the intake pipe blockage fault f 4 For example, the training examples and test examples are as follows:

[0024] (x 4 ,y 4 )=((r′ 41 , r′ 42 , r′ 43 , r′ 44 , r′ 45 , r′ 46 , r′ 47 ) T , (0,0,0,1,0,0,0) T ).

[0025] 4. In step (4), using the particle swarm algorithm to optimize the BP neural network includes the following steps:

[0026] (a) Initialize the BP neural network model:

[0027] A neural network structure with one hidden layer is adopted. The number of neurons in the input layer is k, the number of neurons in the output layer is n, and the number of neurons in the hidden layer is q, which is determined by the following formula:

[0028]

[0029] Where t is a constant, t∈[0,10];

[0030] Set the input of the mth (1≤m≤q) neuron node in the hidden layer to:

[0031]

[0032] where ρ jm is the connection weight between the jth neuron node in the input layer and the mth neuron node in the hidden layer, r′ ij is the normalized value of the jth input parameter in state i;

[0033] Set the output of the mth neuron node in the hidden layer to:

[0034]

[0035] in is the output threshold of the mth neuron node in the hidden layer;

[0036] Set the input of the i-th neuron node in the output layer, that is, the i-th component vector in the output vector, to:

[0037]

[0038] in is the connection weight between the mth neuron node in the hidden layer and the ith neuron node in the output layer, b h is the output of the mth hidden layer neuron node;

[0039] Set the value of the i-th neuron node in the output layer, that is, the value of the i-th component vector in the output vector is:

[0040]

[0041] where θ i is the output threshold of the i-th neuron node in the output layer;

[0042] (b) Initialize the particle swarm.

[0043] Set two particle swarm stop iteration conditions, the first is the upper limit of the number of iterations, and the second is that the BP neural network confidence reaches 95%:

[0044] Set the speed and position update formula of each particle as follows:

[0045]

[0046]

[0047] in is the component of the velocity of the k+1th generation particle u in the dth dimension, is the component of the individual extreme value of the k-th generation particle u in the d-th dimension, is the component of the global extreme value of the k-th generation particle in the d-th dimension, is the component of the position of the kth generation particle u in the dth dimension, c 1 With c 2 are all learning factors, ω is the inertia weight, ζ and η are random numbers on (0, 1);

[0048] Set the inertia weight to:

[0049]

[0050] k is the number of iterations, k max is the maximum number of iterations, and in the initial state ω=ω max =1,ω min =0.4;

[0051] Set the particle's fitness function to:

[0052]

[0053] Among them, X is the position of a particle in the particle swarm, n is the number of training examples, represents the output result of the particle under the i-th training example, y i The expected output result;

[0054] Initialize the particle velocity, the value is [v min , v max ], which is taken as [-0.05, 0.05] in the present invention. ij , With the output threshold θ m , As the position component of each particle in the particle swarm, the position of the uth particle in the particle swarm is:

[0055]

[0056] Among them, x us is the sth position component of the uth particle, and its value range is [xmin , x max ], when a position component exceeds the upper or lower bound, the component is approximated as the boundary value, and the value range of s is {s∈Z|1≤s≤q(k+n+1)+n}, that is, the value range of s is {s∈Z|1≤s≤67};

[0057] (c) If the current fitness function value of the particle is better than the individual historical optimal value, the current position of the particle is updated to the individual historical optimal position; if the individual historical optimal position of the particle is better than the global optimal position, the global optimal position of the particle is updated to the current individual historical optimal position;

[0058] (d) Update the position and velocity of each particle according to the particle swarm algorithm;

[0059] (e) If any of the stopping iteration conditions is met, the iteration is stopped, the optimal initial weights and thresholds of the BP network are obtained, and the trained BP neural network is output.

[0060] 5. In step (5), the method for establishing the connection between the BP neural network model and the expert system is as follows:

[0061] Investigate the hidden layer neurons pointing to the output results of the BP neural network and find the positive weighted input values ​​of the output layer, which have the following characteristics:

[0062]

[0063] At the same time, examine the input layer neurons pointing to this hidden layer neuron, and also find out their positive weighted input values ​​to form a rule:

[0064] IF expression 1 AND expression 2 AND…AND expression n THEN conclusion

[0065] The specific form of the expression n in the rule is that the output of a neuron in a certain input layer is positive, that is, it has a certain sign. The specific form of the conclusion in the rule is the fault type. The following rules are obtained for the abnormal exhaust temperature of the diesel engine:

[0066] IF exhaust temperature>550℃AND turbocharger speed decreases THEN turbocharger is damaged

[0067] IF exhaust temperature>550℃AND air cooler outlet pressure decreasesTHEN intake pipe is blocked

[0068] IF exhaust temperature>550℃AND fuel injection abnormality flag=1 THEN fuel injection abnormality

[0069] IF exhaust temperature>550℃AND exhaust valve damage indicator=1 THEN exhaust valve is damaged

[0070] IF exhaust temperature>550℃AND fresh water inlet pressure decreasesTHEN fresh water system abnormality

[0071] IF the exhaust temperature is low AND the high pressure fuel pipe leaks and the oil level rises THEN the high pressure fuel pipe leaks

[0072] Such rules are stored in the knowledge base to form the knowledge base of the neural network expert system, and the expert system is established based on this knowledge base.

[0073] 6. The process of testing the neural network expert system in step (6) is as follows:

[0074] (i) receiving an abnormal exhaust temperature alarm signal from a diesel engine;

[0075] (ii) The inference engine of the expert system refers to the knowledge base, performs fault diagnosis on the abnormal data of the associated measurement points, and calls the neural network module;

[0076] (iii) If the confidence level of the neural network diagnosis result is greater than 0.95, it will be used as the final diagnosis result of the neural network expert system, and the expert system will interpret the diagnosis result;

[0077] (iv) If the confidence of the output of the neural network is not greater than 0.95, the diagnosis result of the expert system is output;

[0078] (v) If both the expert system and the neural network are unable to diagnose the fault, that is, if a fault occurs for which the correspondence between the symptom and the cause cannot be established at the current stage, the data is packaged, the learning mode of the neural network expert system is started, the expert names the fault, adjusts the neural network topology, adds output nodes, and trains the neural network with the packaged offline data, and the trained neural network is updated to the neural network expert system.

[0079] The advantages of the present invention are:

[0080] 1. The present invention can accurately diagnose abnormal exhaust temperature of diesel engines;

[0081] 2. Solved the problem of long knowledge acquisition cycle of expert system and difficult to understand the reasoning process and results of neural network;

[0082] 3. The present invention uses particle swarm algorithm to optimize two key parameters in the BP neural network model: connection weight ρ jm , With the output threshold θ m , This improves the classification accuracy and computation time of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 It is a structural schematic diagram of the present invention;

[0084] Figure 2 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0085] The present invention is described in more detail below with reference to the accompanying drawings:

[0086] Combination Figure 1-2 The present invention provides a neural network expert system fault diagnosis method for abnormal diesel engine exhaust temperature, comprising the following steps:

[0087] Step 1: To solve the problem of abnormal exhaust temperature of diesel engine, relevant operation data is obtained and the data set is divided into two parts, one of which is used as training data T r , and the other one is used as test data T e ;

[0088] Step 2: Transform the training data T in step 1 r With test data T e are described as a knowledge system ∑ = (F, S), where the output result set F = {f 1 , f 2 , …, f n} is the typical state of the diesel engine, and the input parameter set S = {s 1 ,s 2 ,…,s k} are the relevant parameters in the actual operation of the diesel engine. r Establish training fault feature matrix R r (∑)=(r ij ) n×k , in order to train the neural network; according to the test data T e Establish the test fault feature matrix R e (∑)=(r ij ) n×k , in order to test the neural network expert system after it is established, where r ij is the relevant parameter s in the actual operation of the diesel engine i The value of , i ranges from {i∈Z|1≤i≤n}, and j ranges from {j∈Z|1≤j≤k};

[0089] Step 3: For the fault feature matrix R in step 2 r (∑) and R e (∑) is normalized to obtain the normalized fault feature matrix R′ r (∑)=(r′ ij ) n×k and R′ e (∑)=(r′ ij ) n×k , where r′ij is the relevant parameter s in the actual operation of the diesel engine i The normalized value has a value range of [0, 1], and training examples and test examples are established based on the fault feature matrix;

[0090] Step 4: Establish a BP neural network and use the training examples in step 3 to train the BP neural network, and use the particle swarm algorithm to accelerate the connection weight ρ jm , With the output threshold θ m , The optimization process obtains the optimal connection weight ρ jm , With the output threshold θ m , Output BP neural network model;

[0091] Step 5: Based on the BP neural network obtained in step 4, establish the connection between it and the expert system, that is, establish an expert knowledge base based on the output results of the BP neural network, and establish the expert system structure such as the inference engine and the interpreter, and output the neural network expert system model;

[0092] Step 6: Use the neural network expert system model obtained in step 5 and the test case in step 3 to test and verify the accuracy of the model.

[0093] To establish the data set in step 1, it is necessary to obtain the operating data of 7 groups of related measuring points, which are the data when the diesel engine is operating normally and the operating data when one of the following faults occurs. The fault types are supercharger damage fault, fuel injection abnormality, intake pipe blockage fault, exhaust valve damage fault, fresh water system abnormality and high-pressure oil pipe leakage fault;

[0094] The output result set F in step 2 includes: normal operation status f 1 , and fault conditions, including supercharger damage f 2 , fuel injection abnormality 3 , the intake pipe is blocked 4 , exhaust valve damage 5 , freshwater system abnormality 6 , high pressure oil pipe leakage 7 , that is, F = {f 1 , f 2 , f 3 , f 4 , f 5 , f 6 , f 7};

[0095] The input parameter set S includes: exhaust temperature s 1 , turbocharger speed s 2 , injection status signal s3 , air cooler outlet air pressure s 4 , exhaust valve status signal s 5 , fresh water inlet pressure s 6 , high pressure fuel pipe leakage level s 7 , that is, S = {s 1 ,s 2 ,s 3 ,s 4 ,s 5 ,s 6 ,s 7}.

[0096] Fault feature matrix R r (∑) and R e The form of (∑) is:

[0097]

[0098] The rows represent seven typical states that may occur in the diesel engine system, and the columns represent seven relevant operating parameters of the diesel engine that can be monitored.

[0099] Multi-parameter selection is conducive to all-round judgment of the diesel engine's operating conditions, making the diagnosis results more accurate. Due to the diversity of reference parameters, the fault diagnosis model can also comprehensively analyze the diesel engine's multi-parameter information and improve the diagnosis accuracy and reliability.

[0100] In step 3, the fault feature matrix R r (∑) and R e (∑) is normalized, that is, the fault feature matrix R r (∑) and R e Each column of (∑) is normalized using the maximum and minimum method:

[0101]

[0102] Among them, i (i = 1, ..., 7) is the state type identifier, j (j = 1, ..., 7) is the parameter type identifier, and the normalized value r′ ij In [0, 1], injection status signal s 3 With exhaust valve status signal s 5 The value is 1 in normal state and 0 in abnormal state. Normalization can eliminate the influence of dimension and magnitude differences between indicators.

[0103] Establish training examples or test examples, that is, take a row of the normalized fault feature matrix as the first dimension of the training example or test example, that is, the input vector, set the state corresponding to the row to 1, and set the other states to 0 to form the second dimension of the training example or test example, that is, the expected output vector. A total of 7 training examples are established. 4For example, the training examples and test examples are as follows:

[0104] (x 4 ,y 4 )=((r′ 41 , r′ 42 , r′ 43 , r′ 44 , r′ 45 , r′ 46 , r′ 47 ) T , (0,0,0,1,0,0,0) T )

[0105] In step 4, using particle swarm algorithm to optimize BP neural network includes the following steps:

[0106] (1) Initialize the BP neural network model.

[0107] The present invention adopts a neural network structure containing one hidden layer, the number of neurons in the input layer k, the number of neurons in the output layer n, and the number of neurons in the hidden layer q are determined by the following formula:

[0108]

[0109] Wherein, t is a constant, t∈[0,10], in the present invention, n=7, k=7, and the adjustment number t=0, so the hidden layer neuron format q is approximately taken as 4.

[0110] Set the input of the mth (1≤m≤q) neuron node in the hidden layer to:

[0111]

[0112] where ρ jm is the connection weight between the jth neuron node in the input layer and the mth neuron node in the hidden layer, r′ ij is the normalized value of the jth input parameter in state i.

[0113] Set the output of the mth neuron node in the hidden layer to:

[0114]

[0115] in is the output threshold of the mth neuron node in the hidden layer.

[0116] Set the input of the i-th neuron node in the output layer, that is, the i-th component vector in the output vector, to:

[0117]

[0118] in is the connection weight between the mth neuron node in the hidden layer and the ith neuron node in the output layer, b h is the output of the mth hidden layer neuron node.

[0119] Set the value of the i-th neuron node in the output layer, that is, the value of the i-th component vector in the output vector is:

[0120]

[0121] where θ i is the output threshold of the i-th neuron node in the output layer.

[0122] (2) Initialize the particle swarm.

[0123] Two particle swarm stop iteration conditions are set. The first is the upper limit of the number of iterations, and the second is that the confidence of the BP neural network reaches 95%.

[0124] Set the speed and position update formula of each particle as follows:

[0125]

[0126]

[0127] in is the component of the velocity of the k+1th generation particle u in the dth dimension, is the component of the individual extreme value of the k-th generation particle u in the d-th dimension, is the component of the global extreme value of the k-th generation particle in the d-th dimension, is the component of the position of the kth generation particle u in the dth dimension, c 1 With c 2 are learning factors, ω is the inertia weight, ζ and η are random numbers on (0, 1).

[0128] Set the inertia weight to:

[0129]

[0130] k is the number of iterations, k max is the maximum number of iterations, and in the initial state ω=ω max =1,ω min =0.4.

[0131] Set the particle's fitness function to:

[0132]

[0133] Among them, X is the position of a particle in the particle swarm, n is the number of training examples, represents the output result of the particle under the i-th training example, yi The expected output result.

[0134] Initialize the particle velocity, the value is [v min , v max ], which is taken as [-0.05, 0.05] in the present invention. ij , With the output threshold θ m , As the position component of each particle in the particle swarm, the position of the uth particle in the particle swarm is:

[0135]

[0136] Among them, x us is the sth position component of the uth particle, and its value range is [x min , x max ], the present invention takes it as [-1, 1]. When a position component exceeds the upper or lower limit value, the component needs to be approximated as the boundary value. The value range of s is {s∈Z|1≤s≤q(k+n+1)+n}, that is, the value range of s is {s∈Z|1≤s≤67}.

[0137] (3) If the current fitness function value of the particle is better than the individual historical optimal value, the current position of the particle is updated to the individual historical optimal position; if the individual historical optimal position of the particle is better than the global optimal position, the global optimal position of the particle is updated to the current individual historical optimal position.

[0138] (4) Update the position and velocity of each particle according to the particle swarm algorithm.

[0139] (5) If any of the stop iteration conditions is met, the iteration is stopped, the optimal initial weights and thresholds of the BP network are obtained, and the trained BP neural network is output.

[0140] In step 5, the method of establishing the connection between the BP neural network model and the expert system is as follows:

[0141] Investigate the hidden layer neurons pointing to the output results of the BP neural network and find the positive weighted input values ​​of the output layer, which have the following characteristics:

[0142]

[0143] At the same time, examine the input layer neurons pointing to this hidden layer neuron, and also find out their positive weighted input values ​​to form a rule:

[0144] IF expression 1 AND expression 2 AND…AND expression n THEN conclusion

[0145] The specific form of the expression n in the rule is that the output of a neuron in a certain input layer is positive, which means that there is a certain sign. The specific form of the conclusion in the rule is the fault type. The following rules can be obtained for the abnormal exhaust temperature of diesel engines:

[0146] IF exhaust temperature>550℃AND turbocharger speed decreases THEN turbocharger is damaged

[0147] IF exhaust temperature>550℃AND air cooler outlet pressure decreasesTHEN intake pipe is blocked

[0148] IF exhaust temperature>550℃AND fuel injection abnormality flag=1 THEN fuel injection abnormality

[0149] IF exhaust temperature>550℃AND exhaust valve damage indicator=1 THEN exhaust valve is damaged

[0150] IF exhaust temperature>550℃AND fresh water inlet pressure decreasesTHEN fresh water system abnormality

[0151] IF the exhaust temperature is low AND the high pressure fuel pipe leaks and the oil level rises THEN the high pressure fuel pipe leaks

[0152] Such rules are stored in the knowledge base to form the knowledge base of the neural network expert system, and the expert system is established based on this knowledge base.

[0153] The process of testing the neural network expert system in step 6 is:

[0154] (1) Receive the abnormal exhaust temperature alarm signal of the diesel engine;

[0155] (2) The inference engine of the expert system refers to the knowledge base, performs fault diagnosis on the abnormal data of the associated measuring points, and calls the neural network module;

[0156] (3) If the confidence level of the neural network diagnosis result is greater than 0.95, it will be used as the final diagnosis result of the neural network expert system, and the expert system will interpret the diagnosis result;

[0157] (4) If the confidence of the output of the neural network is not greater than 0.95, the diagnosis result of the expert system is output;

[0158] (5) If neither the expert system nor the neural network can diagnose the fault, that is, if a fault occurs for which the relationship between symptoms and causes cannot be established at the current stage, the data will be packaged, the learning mode of the neural network expert system will be started, the expert will name the fault, adjust the neural network topology, add output nodes, and use the packaged offline data to train the neural network. The trained neural network will then be updated to the neural network expert system.

Claims

1. Abnormal diagnosis method of diesel engine exhaust temperature based on neural network expert system, Its characteristics are: (1) To solve the problem of abnormal exhaust temperature of diesel engines, relevant operating data is obtained and the data set is divided into two parts, one of which is used as training data T r , and the other one is used as test data T e ; (2) The training data T in step (1) r With test data T e are described as a knowledge system ∑ = (F, S), where the output result set F = {f 1 , f 2 , …, f n } is the typical state of the diesel engine, and the input parameter set S = {s 1 ,s 2 ,…,s k } is the actual operating parameter of the diesel engine, according to the training data T r Establish training fault feature matrix R r (∑)=(r ij ) n×k , in order to train the neural network; according to the test data T e Establish the test fault feature matrix R e (∑)=(r ij ) n×k , in order to test the neural network expert system after it is established, where r ij is the relevant parameter s in the actual operation of the diesel engine i The value of , the value range of i is {i∈Z|z≤i≤n}, and the value range of j is {j∈Z|1≤j≤k}; (3) For the fault feature matrix R in step (2) r (∑) and R e (∑) is normalized to obtain the normalized fault feature matrix R′ r (∑) = (r′ ij ) n×k and R′ e (∑) = (r′ ij ) n×k , where r′ ij is the relevant parameter s in the actual operation of the diesel engine i The normalized value has a value range of [0, 1], and training examples and test examples are established based on the fault feature matrix; (4) Establish a BP neural network and use the training examples in step (3) to train the BP neural network, and use the particle swarm algorithm to accelerate the connection weight ρ jm , With the output threshold θ m , The optimization process obtains the optimal connection weight ρ jm , With the output threshold θ m , Output BP neural network model; (5) Based on the BP neural network obtained in step (4), the connection between it and the expert system is established, that is, the expert knowledge base is established based on the output results of the BP neural network, and the inference engine and interpreter expert system structure is established, and the neural network expert system model is output; (6) Use the neural network expert system model obtained in step (5) and the test case in step (3) to test and verify the accuracy of the model.

2. The method for diagnosing abnormal exhaust temperature of a diesel engine based on a neural network expert system according to claim 1, Its characteristics are: The step (1) of obtaining relevant operating data includes obtaining operating data of seven groups of relevant measuring points, which are respectively the data when the diesel engine is operating normally and the operating data when one of the following faults occurs: the fault type is a supercharger damage fault, an abnormal fuel injection fault, an intake pipe blockage fault, an exhaust valve damage fault, a fresh water system abnormality, and a high-pressure oil pipe leakage fault.

3. The method for diagnosing abnormal exhaust temperature of a diesel engine based on a neural network expert system according to claim 1, Its characteristics are: The output result set F in step (2) includes: normal operation status f 1 and fault conditions, including supercharger damage 2 、Abnormal fuel injection 3 、Intake pipe blockage 4 , exhaust valve damage 5 、Freshwater system abnormality 6 , High-pressure oil pipe leakage 7 , that is, F = {f 1 , f 2 , f 3 , f 4 , f 5 , f 6 , f 7 }; The input parameter set S includes: exhaust temperature s 1 , turbocharger speed s 2 , injection status signal 3 , air cooler outlet air pressure s 4 , exhaust valve status signal s 5 , Fresh water inlet pressure s 6 , High-pressure fuel pipe leakage level 7 , that is, S = {s 1 ,s 2 ,s 3 ,s 4 ,s 5 ,s 6 ,s 7 }; Fault feature matrix R r (∑) and R e The form of (∑) is: The rows represent seven typical states that may occur in the diesel engine system, and the columns represent seven relevant operating parameters of the diesel engine that can be monitored.

4. The method for diagnosing abnormal exhaust temperature of a diesel engine based on a neural network expert system according to claim 1, Its characteristics are: In step (3), the fault feature matrix R r (∑) and R e (∑) is normalized, that is, the fault feature matrix R r (∑) and R e Each column of (∑) is normalized using the maximum and minimum method: Among them, i (i = 1, ..., 7) is the state type identifier, j (j = 1, ..., 7) is the parameter type identifier, and the normalized value r′ ij In [0, 1], injection status signal s 3 With exhaust valve status signal s 5 The value is 1 in normal state and 0 in abnormal state; Establish training examples or test examples, that is, take a row of the normalized fault feature matrix as the first dimension of the training example or test example, that is, the input vector, set the state corresponding to the row to 1, and set the other states to 0 to form the second dimension of the training example or test example, that is, the expected output vector. A total of 7 training examples are established; take the intake pipe blockage fault f 4 For example, the training examples and test examples are as follows: (x 4 ,y 4 )=((r′ 41 ,r′ 42 ,r′ 43 ,r′ 44 ,r′ 45 ,r′ 46 ,r′ 47 ) T ,(0,0,0,1,0,0,0) T )。 5. The method for diagnosing abnormal exhaust temperature of a diesel engine based on a neural network expert system according to claim 1, Its characteristics are: In step (4), using the particle swarm algorithm to optimize the BP neural network includes the following steps: (a) Initialize the BP neural network model: A neural network structure with one hidden layer is adopted. The number of neurons in the input layer is k, the number of neurons in the output layer is n, and the number of neurons in the hidden layer is q, which is determined by the following formula: Where t is a constant, t∈[0,10]; Set the input of the mth (1≤m≤q) neuron node in the hidden layer to: where ρ jm is the connection weight between the jth neuron node in the input layer and the mth neuron node in the hidden layer, r′ ij is the normalized value of the jth input parameter in state i; Set the output of the mth neuron node in the hidden layer to: in is the output threshold of the mth neuron node in the hidden layer; Set the input of the i-th neuron node in the output layer, that is, the i-th component vector in the output vector, to: in is the connection weight between the mth neuron node in the hidden layer and the ith neuron node in the output layer, b h is the output of the mth hidden layer neuron node; Set the value of the i-th neuron node in the output layer, that is, the value of the i-th component vector in the output vector is: where θ i is the output threshold of the i-th neuron node in the output layer; (b) Initialize the particle swarm; Set two particle swarm stop iteration conditions, the first is the upper limit of the number of iterations, and the second is that the BP neural network confidence reaches 95%: Set the speed and position update formula of each particle as follows: in is the component of the velocity of the k+1th generation particle u in the dth dimension, is the component of the individual extreme value of the k-th generation particle u in the d-th dimension, is the component of the global extreme value of the k-th generation particle in the d-th dimension, is the component of the position of the kth generation particle u in the dth dimension, c 1 With c 2 are all learning factors, ω is the inertia weight, ζ and η are random numbers on (0, 1); Set the inertia weight to: k is the number of iterations, k max is the maximum number of iterations, and in the initial state ω=ω max =1,ω min =0.4; Set the particle's fitness function to: Among them, X is the position of a particle in the particle swarm, n is the number of training examples, represents the output result of the particle under the training example, y i The expected output result; Initialize the particle velocity, the value is [v min , v max ] is a uniformly distributed random number between [-0.05, 0.05]; the connection weight ρ in the BP neural network is ij , With the output threshold θ m , As the position component of each particle in the particle swarm, the position of the uth particle in the particle swarm is: Among them, x us is the sth position component of the uth particle, and its value range is [x min , x max ], when a position component exceeds the upper or lower bound, the component is approximated as the boundary value, and the value range of s is {s∈Z|1≤s≤q(k+n+1)+n}, that is, the value range of s is {s∈Z|1≤s≤67}; (c) If the current fitness function value of the particle is better than the individual historical optimal value, the current position of the particle is updated to the individual historical optimal position; if the individual historical optimal position of the particle is better than the global optimal position, the global optimal position of the particle is updated to the current individual historical optimal position; (d) Update the position and velocity of each particle according to the particle swarm algorithm; (e) If any of the stopping iteration conditions is met, the iteration is stopped, the optimal initial weights and thresholds of the BP network are obtained, and the trained BP neural network is output.

6. The method for diagnosing abnormal exhaust temperature of a diesel engine based on a neural network expert system according to claim 1, Its characteristics are: In step (5), the method for establishing the connection between the BP neural network model and the expert system is as follows: Investigate the hidden layer neurons pointing to the output results of the BP neural network and find the positive weighted input values ​​of the output layer, which have the following characteristics: At the same time, examine the input layer neurons pointing to this hidden layer neuron, and also find out their positive weighted input values ​​to form a rule: IF expression 1 AND expression 2 AND…AND expression n THEN conclusion The specific form of the expression n in the rule is that the output of a neuron in a certain input layer is positive, that is, it has a certain sign. The specific form of the conclusion in the rule is the fault type. The following rules are obtained for the abnormal exhaust temperature of the diesel engine: IF exhaust temperature>550℃AND turbocharger speed decreases THEN turbocharger is damaged IF exhaust temperature>550℃AND air cooler outlet pressure decreasesTHEN intake pipe is blocked IF exhaust temperature>550℃AND fuel injection abnormality flag=1 THEN fuel injection abnormality IF exhaust temperature>550℃AND exhaust valve damage indicator=1 THEN exhaust valve is damaged IF exhaust temperature>550℃AND fresh water inlet pressure decreasesTHEN fresh water system abnormality IF the exhaust temperature is low AND the high pressure fuel pipe leaks and the oil level rises THEN the high pressure fuel pipe leaks Such rules are stored in the knowledge base to form the knowledge base of the neural network expert system, and the expert system is established based on this knowledge base.

7. The method for diagnosing abnormal exhaust temperature of a diesel engine based on a neural network expert system according to claim 1, Its characteristics are: The process of testing the neural network expert system in step (6) is as follows: (i) receiving an abnormal exhaust temperature alarm signal from a diesel engine; (ii) The inference engine of the expert system refers to the knowledge base, performs fault diagnosis on the abnormal data of the associated measurement points, and calls the neural network module; (iii) If the confidence level of the neural network diagnosis result is greater than 0.95, it will be used as the final diagnosis result of the neural network expert system, and the expert system will interpret the diagnosis result; (iv) If the confidence of the output of the neural network is not greater than 0.95, the diagnosis result of the expert system is output; (v) If both the expert system and the neural network are unable to diagnose the fault, that is, if a fault occurs for which the correspondence between the symptom and the cause cannot be established at the current stage, the data is packaged, the learning mode of the neural network expert system is started, the expert names the fault, adjusts the neural network topology, adds output nodes, and trains the neural network with the packaged offline data, and the trained neural network is updated to the neural network expert system.

Images