A method for ranking diesel engine fault causes by using analytic hierarchy process combined with neural network

Abstract

A method for ranking diesel engine fault causes using the Analytic Hierarchy Process (AHP) combined with a neural network includes the following steps: acquiring the fault phenomena of the diesel engine, identifying internal and external factors, and establishing a fault database; using the AHP, constructing a hierarchical structure model based on internal factors, building a judgment matrix, calculating the weights of each internal factor, and obtaining an internal factor weight vector; using an AlexNet neural network to process the external factors, outputting the corresponding influence factors of each external factor; using the neural network in conjunction with external factors to optimize the weights of each internal factor obtained by the AHP, obtaining the final weights; based on the final weights, calculating the score for each fault cause in conjunction with the fault database, and prioritizing the causes according to the scores to obtain the fault cause ranking result. This method can effectively optimize the diesel engine fault detection steps, quickly locate fault points, shorten repair time, improve diagnostic efficiency, and greatly reduce the dependence on the technical level of maintenance personnel.

Description

Technical Field

[0001] This application relates to the field of diesel engine repair and testing technology, and in particular to a method for ranking the causes of diesel engine failures using the analytic hierarchy process combined with a neural network. Background Technology

[0002] Diesel engines, as complex thermodynamic machines, are widely used in vehicles, ships, generator sets, and other fields. During operation, various malfunctions inevitably occur due to factors such as wear and tear on internal components and changes in the external environment. When a malfunction occurs, quickly and accurately locating the cause is crucial to ensuring normal equipment operation and minimizing downtime losses.

[0003] Currently, diesel engine fault diagnosis and handling generally relies heavily on manual experience, lacking a scientific and systematic set of criteria. Repair effectiveness is highly dependent on the technical skill level of the repair personnel, which often varies greatly, resulting in inconsistent repair quality and an inability to quickly and accurately locate faults in emergency field situations. More advanced methods are based on expert systems, which collect operational status data through various sensors and use the expert system to make judgments based on anomaly parameter association rules. However, expert system-based diagnosis heavily relies on the establishment of expert databases and judgment rules, which has certain limitations and is difficult to adapt to dynamically changing external environments. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method for ranking the causes of diesel engine failures using the analytic hierarchy process combined with neural networks. This method addresses the technical problems in the prior art where diesel engine failure diagnosis relies on human experience, lacks scientific and systematic criteria, and has significant limitations in expert systems, making it impossible to quickly and accurately locate faults.

[0005] The technical solution of the present invention is as follows:

[0006] A method for ranking the causes of diesel engine failures using the analytic hierarchy process (AHP) combined with a neural network includes the following steps: S1. Obtain the fault phenomena of the diesel engine, determine the internal and external factors affecting the ranking of fault causes, and establish a fault database; S2. Using the analytic hierarchy process (AHP), a hierarchical structure model is constructed based on the internal factors, and a judgment matrix is ​​built to calculate the weight of each internal factor and obtain the internal factor weight vector. S3. The external factors are processed using an improved AlexNet neural network, which is a one-dimensional convolutional neural network that takes the external factors as input and outputs the influence factors corresponding to each external factor. S4. Using the neural network combined with external factors, optimize the weights of each internal factor obtained by the analytic hierarchy process to obtain the final weight vector. S5. Based on the final weight vector, calculate the score for each fault cause in conjunction with the fault database, and sort them by priority according to the scores to obtain the fault cause ranking result.

[0007] A further improvement of the technical solution of the present invention is that the internal factors include historical failure probability, maintenance convenience, safety risk and maintenance cost, and the external factors include temperature, altitude, friction time and mileage.

[0008] A further improvement of the technical solution of the present invention is that: the hierarchical structure model in S2 includes a target layer, a criterion layer and a solution layer, wherein the target layer is the priority ranking of fault cause detection, the criterion layer includes the internal factors, and the solution layer includes various fault causes.

[0009] A further improvement to the technical solution of this invention lies in the fact that: in step S2, the analytic hierarchy process (AHP) is used to construct a judgment matrix and calculate the weights of each internal factor, specifically including: 2.1 Based on the relative importance of each internal factor in the criterion layer, the judgment matrix A is constructed using the 1-9 scale method as follows:

[0010] A=(a ij ) n×n ;

[0011] Where n is the number of internal factors in the criterion layer, a ij This represents a scale indicating the importance of the i-th factor relative to the j-th factor; 2.2. Calculate the largest eigenvalue λ of the judgment matrix A using the sum-product method or the square root method for approximate calculation. max and its corresponding eigenvector ω; 2.3. Normalize the eigenvector ω to obtain the initial weight vector W for each internal factor. internal =[w hist ,w conv ,w safe ,w cost ]; 2.4 Perform a consistency check on the judgment matrix and calculate the consistency index CI=(λ) max -n) / (n-1) and the consistency ratio CR=CI / RI. When CR<0.10, the consistency is considered acceptable; otherwise, the judgment matrix is ​​adjusted. Wherein, RI is the average random consistency index.

[0012] A further improvement to the technical solution of this invention lies in the following: based on expert experience, the judgment matrix A is constructed as follows:

[0013] Historical failure probability 1 2 3 5 Ease of maintenance 1 / 2 1 2 3 Security risks 1 / 3 1 / 2 1 2 Maintenance costs 1 / 5 1 / 3 1 / 2 1

[0014] A further improvement to the technical solution of this invention lies in that: the improved AlexNet neural network outputs four influencing factors, denoted as δ. T δ H δ R δ M The values ​​correspond to the degree of influence of temperature, altitude, friction time, and mileage on the ranking of fault causes, respectively. The value range of each influencing factor is [-1, 1], with positive values ​​indicating positive influence, negative values ​​indicating negative influence, and zero indicating no influence.

[0015] A further improvement of the technical solution of the present invention is that the structure of the improved AlexNet neural network includes 6 one-dimensional convolutional layers and 4 fully connected layers. Each convolutional layer is followed by the ReLU activation function. Overlapping pooling layers are added after some convolutional layers. Dropout regularization is introduced in the first two fully connected layers, and the random deactivation ratio is 0.5.

[0016] A further improvement to the technical solution of the present invention is that: S4 specifically includes: 4.1 Calculate the algebraic sum Δ=δ of the four influencing factors. T +δ H +δ R +δ M ;

[0017] 4.2 For each internal factor i, i=1,2,3,4 and corresponding to historical failure probability, maintenance convenience, safety risk, and maintenance cost respectively, its weight after being comprehensively affected by external factors is: W′ final,i =W internal,i (1+Δ); 4.3. The calculated weights W′ final Normalization is performed so that the sum of all components is 1, resulting in the final weight vector W. final .

[0018] A further improvement to the technical solution of this invention lies in: S5 calculates the score for each fault cause based on the final weight vector and the fault database, specifically including: 5.1 Suppose there are K possible causes of failure in the fault database, and the k-th cause of failure corresponds to the set values ​​of four internal factors:

[0019] Historical failure probability calculation value P k : Calculate the historical occurrence probability of the cause of the fault in the fault database, divide the probability value by 10 and keep one decimal place, the value range is 0~10;

[0020] Maintenance convenience coefficient M k : Fault database preset value, ranging from 1 to 10;

[0021] Safety Risk Level S k : Fault database preset value, ranging from 1 to 10;

[0022] Maintenance cost level C k : Fault database preset value, ranging from 1 to 10;

[0023] 5.2 The score for the kth fault cause is:

[0024] Score k = w hist ×P k +w conv ×M k +w safe ×S k +w cost ×C k ;

[0025] 5.3. According to the aforementioned score k Sort all K fault causes from highest to lowest and output the sorted fault cause list.

[0026] A further improvement of the technical solution of the present invention is that the method further includes a growth closed-loop feedback step, that is, after the maintenance personnel perform maintenance according to the fault cause ranking result, they feed back the actual confirmed fault cause to the system; the system updates the historical fault probability value of the corresponding fault cause in the fault database according to the result, and at the same time uses the external factor feature vector of this maintenance and the actual fault cause label as new training samples for incremental training or retraining of the improved AlexNet neural network.

[0027] The beneficial effects of this invention are:

[0028] 1. This method is the first to combine the hierarchical analysis method and neural network, and takes into account both internal and external factors that affect the ranking of fault causes. Based on the feedback of maintenance results, the method can be optimized in real time, making the invention scalable. It can effectively optimize the fault detection steps of diesel engines, quickly locate the fault point, shorten the maintenance time, improve the diagnostic efficiency, and greatly reduce the dependence on the technical level of maintenance personnel.

[0029] 2. This method can run completely offline without relying on a network connection, and is suitable for rapid diagnosis of diesel engine faults in harsh field environments. Attached Figure Description

[0030] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0031] The present invention will now be described in further detail with reference to specific embodiments and the accompanying drawings.

[0032] like Figure 1 As shown, the present invention relates to a method for ranking the causes of diesel engine failures using the analytic hierarchy process (AHP) combined with a neural network, comprising the following steps:

[0033] S1. Obtain fault phenomena of diesel engines and establish a fault database; and determine the internal and external factors that affect the ranking of fault causes; the internal factors include historical fault probability, maintenance convenience, safety risks and maintenance costs, and the external factors include temperature, altitude, friction time and mileage.

[0034] S2. Using the Analytic Hierarchy Process (AHP), a hierarchical structure model is constructed based on the internal factors, and a judgment matrix is ​​built to calculate the weight of each internal factor, resulting in an internal factor weight vector. The hierarchical structure model includes a target layer, a criterion layer, and a solution layer, where the target layer is the priority ranking of fault cause detection, the criterion layer includes the internal factors, and the solution layer includes each fault cause.

[0035] The process of constructing a judgment matrix and calculating the weights of each internal factor using the analytic hierarchy process includes:

[0036] 2.1 Based on the relative importance of each internal factor in the criterion layer, a judgment matrix A is constructed using the 1-9 scale method. The judgment matrix A is as follows:

[0037] A=(a ij ) n×n ;

[0038] Where n is the number of criteria layer factors, in this embodiment n=4, namely historical failure probability, maintenance convenience, safety risk, and maintenance cost; a ij This represents a scale indicating the importance of the i-th factor relative to the j-th factor. The scales have the following meanings: scale 1 indicates that the two factors are equally important; scale 3 indicates that the former is slightly more important than the latter; scale 5 indicates that the former is significantly more important than the latter; scale 7 indicates that the former is strongly more important than the latter; scale 9 indicates that the former is extremely more important than the latter; scales 2, 4, 6, and 8 represent the intermediate values ​​of the above adjacent judgments, indicating different degrees of importance.

[0039] If the importance ratio of factors i to j is a ij Then the ratio of the importance of factor j to i is a. ji =1 / a ij For example, if the ratio of the importance of safety risk to the probability of failure is 3, then the ratio of the importance of the probability of failure to the importance of safety risk is 1 / 3.

[0040] Specifically, based on the experience of experts in the field of diesel engine repair, this embodiment constructs the judgment matrix A as follows:

[0041] Historical failure probability 1 2 3 5 Ease of maintenance 1 / 2 1 2 3 Security risks 1 / 3 1 / 2 1 2 Maintenance costs 1 / 5 1 / 3 1 / 2 1

[0042] 2.2. Use the root method or sum-product method to perform approximate calculations and calculate the largest eigenvalue λ of the judgment matrix A. max And its corresponding eigenvector ω. This embodiment takes the sum-product method as an example, and the specific sub-steps are as follows: 2.2.1 Normalize each column of the judgment matrix A, i.e., calculate... ; 2.2.2 Summing the normalized matrix row by row yields the following result. ; 2.2.3. The summed vector Further normalization, i.e. This yields the eigenvector ω = [ω1,ω2,ω3,ω4]. T That is, the initial weights of each internal factor; 2.2.4 Calculate the largest eigenvalue Where A is the judgment matrix, ω is the eigenvector, and Aω represents the product of matrix A and vector ω, the result of which is a column vector; (Aω) i Let i represent the i-th component of vector Aω.

[0043] 2.3. Normalize the eigenvector ω to obtain the initial weight vector W for each internal factor. internal =[w hist ,w conv ,w safe ,w cost ];

[0044] 2.4 Perform a consistency check on the judgment matrix. Calculate the consistency index CI = (λ) max The formulas are: -n) / (n-1) and the consistency ratio CR = CI / RI, where RI is the average random consistency index. According to the table below, when n=4, RI=0.89. If CR < 0.10 after calculation, the consistency of the judgment matrix is ​​considered acceptable; otherwise, the scale values ​​in the judgment matrix need to be adjusted until the consistency requirements are met.

[0045] RI 0 0 0.58 0.89 1.12 1.24 1.32 1.41 1.45

[0046] S3. The external factors are processed using an improved AlexNet neural network. The improved AlexNet neural network is a one-dimensional convolutional neural network that takes the external factors as input and outputs the influence factors corresponding to each external factor. Specifically, this includes the following steps:

[0047] 3.1 Collect external factor data of the current diesel engine operation to form a one-dimensional feature vector X. ext=[T, H, M,R], where T is the ambient temperature, H is the altitude, M is the cumulative mileage, and R is the running time (hours) since the last maintenance. This feature vector is preprocessed by normalization so that all dimensions fall within the [0,1] interval.

[0048] 3.2 Constructing an improved one-dimensional AlexNet neural network. This network structure includes 6 one-dimensional convolutional layers (Conv1~Conv6) and 4 fully connected layers (FC1~FC4). The specific parameter configurations are as follows: Conv1 contains 4 convolutional kernels of size 3×1 with a stride of 1; Conv2 contains 8 convolutional kernels of size 3×1 with a stride of 1; Conv3 contains 16 convolutional kernels of size 3×1 with a stride of 1; Conv4 contains 32 convolutional kernels of size 3×1 with a stride of 1; Conv5 contains 64 convolutional kernels of size 3×1 with a stride of 1; and Conv6 contains 128 convolutional kernels of size 3×1 with a stride of 1.

[0049] Each convolutional layer is followed by a ReLU activation function, f(x) = max(0,x), to introduce non-linearity and mitigate the vanishing gradient problem. Overlapping pooling layers are added after Conv1, Conv2, and Conv6, with a pooling window size of 1×1 and a stride of 1. Dropout regularization is introduced in the first two fully connected layers, with a random deactivation ratio of 0.5 to prevent overfitting. The output dimension of the last fully connected layer, FC4, is set to 4, consistent with the number of external factors.

[0050] 3.3. The normalized external factor feature vector X ext The input is fed into the improved AlexNet neural network, and after forward propagation calculation, four influencing factors are output, denoted as δ. T δ H δ R δ M The values ​​correspond to the degree of influence of temperature, altitude, friction time, and mileage on the ranking of fault causes, respectively. The value range of each influencing factor is [-1, 1], with positive values ​​indicating positive influence, negative values ​​indicating negative influence, and zero indicating no influence.

[0051] S4. Using the neural network in conjunction with external factors, optimize the weights of each internal factor obtained by the analytic hierarchy process (AHP) to obtain the final weight vector. The specific steps are as follows:

[0052] 4.1 Calculate the algebraic sum Δ=δ of the four influencing factors. T +δ H +δ R +δ M ;

[0053] 4.2 For each internal factor i, i=1,2,3,4 and corresponding to historical failure probability, maintenance convenience, safety risk, and maintenance cost respectively, its weight after being comprehensively affected by external factors is: W′ final,i =W internal,i (1+Δ); When Δ>0, the weights of all internal factors are proportionally increased; when Δ<0, they are proportionally decreased; when Δ=0, the weights remain unchanged. This formula reflects the overall adjustment of the weights of internal factors after the combined effect of four external factors: temperature, altitude, friction time, and mileage.

[0054] 4.3. The calculated weights W′ final =(W′ final,i The weight vector W is obtained by normalizing the components i=1,2,3,4 to make the sum of all components equal to 1. final =[w hist ,w conv ,w safe ,w cost ].

[0055] S5, based on the final weight vector W final By combining the baseline feature vector corresponding to each fault cause stored in the fault database, a score is calculated for each fault cause, and the causes are prioritized according to their scores to obtain the fault cause ranking result. The specific steps are as follows: 5.1 Suppose there are K possible causes of failure in the fault database, and the k-th cause of failure corresponds to the set values ​​of four internal factors:

[0056] Historical failure probability calculation value P k The probability of a fault occurring in the fault database is calculated by dividing the historical occurrence rate of that fault cause by 10 and rounding to one decimal place. The value ranges from 0 to 10. For example, if the historical occurrence rate of a fault cause is 50%, then Pf = 10. k = 50 / 10 = 5.0; if the probability is 33%, then P k = 3.3.

[0057] Maintenance convenience coefficient M k This is a preset value in the fault database, ranging from 1 to 10; the larger the value, the easier the repair.

[0058] Safety Risk Level S k , which is a preset value in the fault database, ranging from 1 to 10; the larger the value, the higher the safety risk.

[0059] Maintenance cost level C k , which is a preset value in the fault database, ranging from 1 to 10; the larger the value, the higher the maintenance cost.

[0060] 5.2 The formula for calculating the score of the kth fault cause is as follows:

[0061] Score k = w hist ×P k +w conv ×M k +w safe ×S k +w cost ×C k ;

[0062] Among them, w hist、 w conv、 w safe and w cost For the final weight vector W final The components correspond to the weights of historical failure probability, ease of maintenance, safety risk, and maintenance cost, respectively.

[0063] 5.3. According to the aforementioned score k Sort all K fault causes from highest to lowest and output a sorted list of fault causes. Higher scores indicate a greater likelihood of the fault occurring, and maintenance personnel should prioritize checking that cause.

[0064] The method also includes a closed-loop feedback step, whereby maintenance personnel perform maintenance based on the fault cause ranking results and then feed back the confirmed fault causes to the system. The system updates the historical fault probability value corresponding to that fault cause in the fault database based on this result (i.e., increases the count of that fault cause and recalculates the normalized value), and simultaneously updates the external factor feature vector X of this maintenance. ext Using the actual fault cause labels as new training samples, the improved AlexNet neural network is incrementally trained or retrained periodically, enabling the neural network to adapt to new external environmental changes and continuously improve the accuracy of weight optimization.

[0065] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for ranking the causes of diesel engine failures using the analytic hierarchy process (AHP) combined with neural networks, characterized in that, The steps are as follows: S1. Obtain the fault phenomena of the diesel engine, determine the internal and external factors affecting the ranking of fault causes, and establish a fault database; S2. Using the analytic hierarchy process (AHP), a hierarchical structure model is constructed based on the internal factors, and a judgment matrix is ​​built to calculate the weight of each internal factor and obtain the internal factor weight vector. S3. The external factors are processed using an improved AlexNet neural network, which is a one-dimensional convolutional neural network that takes the external factors as input and outputs the influence factors corresponding to each external factor. S4. Using the neural network combined with external factors, optimize the weights of each internal factor obtained by the analytic hierarchy process to obtain the final weight vector. S5. Based on the final weight vector, calculate the score for each fault cause in conjunction with the fault database, and sort them by priority according to the scores to obtain the fault cause ranking result.

2. The method according to claim 1, characterized in that: The internal factors include historical failure probability, ease of maintenance, safety risks, and maintenance costs, while the external factors include temperature, altitude, friction time, and mileage.

3. The method according to claim 2, characterized in that: The hierarchical model in S2 includes a target layer, a criterion layer, and a solution layer. The target layer is the priority ranking of fault cause detection, the criterion layer includes the internal factors, and the solution layer includes various fault causes.

4. The method according to claim 3, characterized in that, In step S2, the analytic hierarchy process (AHP) is used to construct the judgment matrix and calculate the weights of each internal factor, specifically including: 2.1 Based on the relative importance of each internal factor in the criterion layer, the judgment matrix A is constructed using the 1-9 scale method as follows: A = (a ij ) n×n ; where n is the number of internal factors of the criterion layer, a ij represents the importance degree scale of the ith factor relative to the jth factor; 2.

2. Calculate the largest eigenvalue λ of the judgment matrix A using the sum-product method or the square root method. max and its corresponding eigenvector ω; 2.

3. Normalize the eigenvector ω to obtain the initial weight vector W for each internal factor. internal =[w hist ,w conv ,w safe ,w cost ]; 2.4 Perform a consistency check on the judgment matrix and calculate the consistency index CI=(λ) max -n) / (n-1) and the consistency ratio CR=CI / RI. When CR<0.10, the consistency is considered acceptable; otherwise, the judgment matrix is ​​adjusted. Wherein, RI is the average random consistency index.

5. The method according to claim 4, characterized in that: Based on expert experience, the judgment matrix A is constructed as follows:

6. The method according to claim 1, characterized in that: The improved AlexNet neural network outputs four influencing factors, denoted as δ. T δ H δ R δ M The values ​​correspond to the degree of influence of temperature, altitude, friction time, and mileage on the ranking of fault causes, respectively. The value range of each influencing factor is [-1, 1], with positive values ​​indicating positive influence, negative values ​​indicating negative influence, and zero indicating no influence.

7. The method according to claim 6, characterized in that, The improved AlexNet neural network structure includes six one-dimensional convolutional layers and four fully connected layers. Each convolutional layer is followed by a ReLU activation function. Overlapping pooling layers are added after some convolutional layers. Dropout regularization is introduced in the first two fully connected layers, and the random deactivation ratio is 0.

5.

8. The method according to claim 6, characterized in that: S4 specifically includes: 4.1 Calculate the algebraic sum Δ=δ of the four influencing factors. T +δ H +δ R +δ M ; 4.2 For each internal factor i, i=1,2,3,4 and corresponding to historical failure probability, maintenance convenience, safety risk, and maintenance cost respectively, its weight after being comprehensively affected by external factors is: W′ final,i =W internal,i (1+Δ); 4.

3. The calculated weights W′ final Normalization is performed so that the sum of all components is 1, resulting in the final weight vector W. final .

9. The method according to claim 1, characterized in that: S5 calculates the score for each fault cause based on the final weight vector and the fault database, specifically including: 5.1 Suppose there are K possible causes of failure in the fault database, and the k-th cause of failure corresponds to the set values ​​of four internal factors: Historical failure probability calculation value P k : Calculate the historical occurrence probability of the cause of the fault in the fault database, divide the probability value by 10 and keep one decimal place, the value range is 0~10; Maintenance convenience coefficient M k : Fault database preset value, ranging from 1 to 10; Safety Risk Level S k : Fault database preset value, ranging from 1 to 10; Maintenance cost level C k : Fault database preset value, ranging from 1 to 10; 5.2 The score for the kth fault cause is: Score k = w hist ×P k +w conv ×M k +w safe ×S k +w cost ×C k ; 5.

3. According to the aforementioned score k Sort all K fault causes from highest to lowest and output the sorted fault cause list.

10. The method according to any one of claims 1-9, characterized in that, The method also includes a growth closed-loop feedback step, whereby maintenance personnel perform maintenance based on the fault cause ranking results and feed back the actual confirmed fault causes to the system; the system updates the historical fault probability values ​​of the corresponding fault causes in the fault database based on the results, and uses the external factor feature vector and actual fault cause label of this maintenance as new training samples for incremental training or retraining of the improved AlexNet neural network.

Images