A diesel engine fault diagnosis method under variable load based on MDUNN
By adopting the multi-deep convolution structure and consolidation learning method of MDUNN in diesel engine fault diagnosis, the problem of insufficient feature extraction and multi-domain feature set under variable load of diesel engine is solved, and the accurate diagnosis and robustness of diesel engine faults are achieved.
Patent Information
- Application Number
- CN202210835032.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-15
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2042-07-15
AI Technical Summary
In the fault diagnosis of diesel engines under variable load, the multi-domain feature set is expanded and the intersection is reduced, and the intersection feature cannot meet the diagnosis needs of attention mechanisms in complex environments.
Using a multi-deep convolutional neural network based on MDUNN, a multi-deep convolutional structure is constructed to process multi-deep convolutional structure by increasing the number of first-layer convolution kernels, reducing the degree of abstraction, global average pooling, selective connections and imitating the subconscious mechanism. Domain features to avoid network degradation caused by reduction of common features.
It realizes accurate diagnosis of diesel engine failure under variable loads, fully grasps signal characteristics and forms accurate decisions, and improves the robustness and adaptability of the diagnosis.
Smart Images

Figure CN115200881B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of diesel engine fault diagnosis, and in particular to a diesel engine fault diagnosis method under variable load based on MDUNN. Background Art
[0002] Diesel engines are general-purpose power machines with high thermal efficiency, wide power range, and good mobility. They are widely used in heavy commercial vehicles, engineering and agricultural machinery, ships, power stations, etc. Most of the above fields served by diesel engines play an important role in the national economy. If timely fault detection and repair cannot be achieved, serious accidents caused by accumulated faults will cause economic losses and threaten the lives of related personnel. Therefore, timely and accurate diagnosis of internal combustion engines is very important.
[0003] Fault diagnosis technology based on deep learning has the advantages of high diagnostic accuracy, fast running speed and adaptive feature extraction. For diesel engines, the cylinder head vibration signal has significant characteristics for various diesel engine faults and is suitable as the data basis for deep learning diagnostic models. However, the working environment of diesel engines is complex, and the operating load is adjusted in real time according to actual needs. It is difficult for deep learning fault training sets and test sets to ensure power stability. The actual fault data set of diesel engines and the vibration signals collected by real-time operation and maintenance have serious non-independent and identically distributed characteristics, which makes it difficult to apply diesel engine fault diagnosis technology based on deep learning in real variable load scenarios.
[0004] In the field of fault diagnosis, some scholars have developed deep learning models to solve similar problems, which have certain robustness and generalization. Some classic robust deep convolutional networks have achieved considerable results in the fault diagnosis of relatively simple mechanical structures such as bearings, gearboxes, and wind turbines. However, the vibration signal of the diesel engine cylinder head has the following two characteristics: the transmission path is long and involves many components; the transmission path is diverse, and the same fault excitation is transmitted from multiple paths to the same receiving point. At the same time, due to its complex working environment, the diesel engine vibration signal is coupled with strong background noise. Therefore, the diesel engine needs a model with stronger signal analysis capabilities and better robustness to realize fault diagnosis under variable loads. It is of great significance to construct a diagnostic model for diesel engines under variable loads.
[0005] At present, there are the following problems in the diagnosis of diesel engine variable load faults:
[0006] 1. The first layer of convolution determines the number of features that enter CNN. The common single-channel "shuttle-shaped" CNN network architecture with a narrow first layer and a wide middle layer cannot fully extract features.
[0007] 2. Multi-domain mixing rapidly expands the total feature set, while the feature intersection rapidly decreases, forming a non-independent and identically distributed scenario. Common deep learning models are based on the independent and identically distributed assumption and cannot adapt to multi-domain environments.
[0008] 3. The commonly used attention mechanism can enable the network to focus on learning the intersection features, which can play a significant role in simple machines with highly overlapping feature sets in different domains. However, due to the huge differences in signals caused by changes in the working state of the diesel engine, the rapidly decreasing intersection features cannot meet the diagnostic needs.
[0009] Due to the above-mentioned deficiencies in the prior art, there is a need for a variable load diesel engine diagnosis method that can fully capture signal features and form accurate decisions in a multi-domain feature space.
[0010] The information disclosed in the background technology section of the present invention is only intended to deepen the understanding of the general background technology of the present invention, and should not be regarded as acknowledging or suggesting in any form that the information constitutes the prior art already known to those skilled in the art. Summary of the invention
[0011] In order to solve the technical problems mentioned in the above background technology or at least partially solve the above technical problems, the present application provides a diesel engine fault diagnosis method under variable load based on MDUNN, comprising the following steps:
[0012] S100, collecting vibration signals of different faults under different loads based on an acceleration sensor disposed on a cylinder head of a diesel engine;
[0013] S200, segmenting vibration signals of different faults under different loads according to fixed signal lengths and establishing fault labels corresponding to fault types one by one to obtain a fault data set, wherein the fault data set includes all fault labels;
[0014] S300, build the MDUNN model, and use the fault data set to train the MDUNN in stages;
[0015] S400: Input the cylinder head vibration signal collected in real time into MDUNN, and output the diesel engine fault type.
[0016] Step S300 specifically includes the following steps:
[0017] S310, build multi-depth convolutional neural networks;
[0018] S320, consolidating the feature matrix weights for training;
[0019] S330, consolidating the training of the depth weight matrix.
[0020] Step S310 includes the following steps:
[0021] S311. Construct a convolutional neural network, increase the number of kernels in the first convolution layer to 128, and keep the number of kernels in subsequent layers at 64. The first convolution layer with 128 kernels and the hidden layer with 64 kernels constitute a feature extractor. Add a convolution kernel of size 1 after the feature extractor, and adjust the number of channels to a preset number of categories. The number of categories is selected according to the fault types in the fault data set, so that the convolution output features can be directly used for decision-making.
[0022] S312. The convolution layer containing 64 convolution kernels in the above convolution structure is reduced, thereby reducing the degree of abstraction and forming convolution paths of different depths. The results of these convolution paths are directly subjected to global average pooling, which is referred to as GAP. The SoftMax classifier is used to perform gradient descent after taking cross entropy with the fault label to complete the initial training. After the training, the GAP and SoftMax classifier are discarded.
[0023] Step S320 includes the following steps:
[0024] S321. The output features of convolution embeddings with different depths are averaged within the row after dot product with matrices of the same shape, which is equivalent to adding a layer of selectively connected unbiased fully connected layer so that the output still maintains category independence. The module output is cross-entropy with the fault label through the SoftMax classifier and then gradient descent. After training, the SoftMax classifier is discarded and the output is used as the feature input of the next repeated module.
[0025] S322. After the stage training is completed, thresholds are assigned to different output features of the convolution. Features above the threshold will directly determine the output category advantage, while features below the threshold will serve to build thresholds for other features. The feature threshold is calculated using the following formula:
[0026]
[0027] Among them, the number of rows of FWM is the number of categories, and the number of columns n is the number of features in a single channel;
[0028] FWM ab is the weight of the corresponding position in row a and column b of the FWM weight matrix;
[0029] FWM ij is the weight corresponding to the position in row i and column j of the FWM weight matrix except row a and column b;
[0030] O ij The feature or score value outputted in row i and column j except row a and column b of the FWM weight matrix;
[0031] Thresold ab is the feature threshold corresponding to the position in row a and column b of the FWM weight matrix.
[0032] The FWM weight matrix is constrained by the sigmoid function, and the formula is as follows:
[0033]
[0034] Among them, x is the additional registered parameter;
[0035] λ is the value range constraint coefficient;
[0036] σ(x) is the weight of the sigmoid constraint.
[0037] Step S330 specifically includes the following steps: adding a deep weight matrix of the same form as the feature weight layer after the feature weight layer, and summing it within the row after performing dot product with the output feature of the feature weight layer, repeating step S320 to obtain the deep weight layer. The output of the deep weight layer has the same shape as the fault label and can be used as the final output used in actual fault diagnosis to complete the training of MDUNN.
[0038] The S400 specifically includes:
[0039] S100, obtaining a cylinder head vibration signal based on an acceleration sensor disposed on a cylinder head of a diesel engine;
[0040] S200, inputting the cylinder head vibration signal into MDUNN at every preset time interval;
[0041] S300 and MDUNN extract multi-depth features from the cylinder head vibration signal, adjust the reduced dimension at two weight layers, and obtain the fault type prediction results;
[0042] S400, output diesel engine fault type.
[0043] The acceleration sensor is a unidirectional acceleration sensor, and the unidirectional acceleration sensor is electrically connected to a multi-channel data acquisition system, and the multi-channel data acquisition system is electrically connected to acquire a cylinder head vibration signal through the unidirectional acceleration sensor.
[0044] The above technical solution provided by the embodiment of the present application has the following advantages compared with the prior art: the embodiment of the present application provides a diesel engine fault diagnosis method under variable load based on MDUNN, which enables the network to memorize and process multi-domain features through the embedded structure of multi-depth convolution and the consolidation learning method that imitates the subconscious mechanism, thereby avoiding network degradation caused by the reduction of common features of the variable load data set. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0047] Figure 1 A schematic diagram of the steps of a diesel engine fault diagnosis method under variable load based on MDUNN provided in an embodiment of the present application;
[0048] Figure 2 A schematic diagram of the structure of a diesel engine and an acceleration sensor in a diesel engine fault diagnosis method under variable load based on MDUNN provided in an embodiment of the present application;
[0049] Figure 3 A schematic diagram of a multi-depth convolutional network structure of a diesel engine fault diagnosis method under variable load based on MDUNN provided in an embodiment of the present application;
[0050] Figure 4 A schematic diagram of the MDUNN structure of a diesel engine fault diagnosis method under variable load based on MDUNN provided in an embodiment of the present application. DETAILED DESCRIPTION
[0051] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.
[0052] For ease of understanding, a diesel engine fault diagnosis method under variable load based on MDUNN provided in an embodiment of the present application is described in detail below. The full name of MDUNN is Multi-depth Unconscious Neural Network, which refers to a multi-depth subconscious neural network diagnostic model with multiple convolution kernels in the first layer, such as Figure 1 and Figure 4 As shown, the following steps are included:
[0053] S100: collecting vibration signals of different faults under different loads based on an acceleration sensor arranged on a cylinder head of a diesel engine.
[0054] like Figure 2As shown, the acceleration sensor is a unidirectional acceleration sensor, and the unidirectional acceleration sensor is electrically connected to a multi-channel data acquisition system, and the multi-channel data acquisition system is electrically connected to acquire a cylinder head vibration signal through the unidirectional acceleration sensor.
[0055] S200, segmenting vibration signals of different faults under different loads according to fixed signal lengths and establishing fault labels in one-to-one correspondence with fault types to obtain a fault data set, wherein the fault data set includes all fault labels.
[0056] S300, construct the MDUNN model, and use the fault data set to train the MDUNN in stages.
[0057] Specifically, Figure 3 As shown, step S300 includes the following steps:
[0058] S310. Build a multi-depth convolutional neural network.
[0059] Specifically, step S310 includes the following steps:
[0060] S311. Construct a convolutional neural network, increase the number of cores in the first layer of convolution to 128, and keep the number of cores in subsequent layers at 64. The first layer of convolution with 128 cores and the hidden layer with 64 cores constitute a feature extractor. Add a layer of convolution kernel of size 1 after the feature extractor, and adjust the number of channels to the preset number of categories. The number of categories is selected according to the fault type in the fault data set, so that the convolution output features can be directly used for decision-making.
[0061] S312. Reduce the convolution layer containing 64 convolution kernels in the above convolution structure, thereby reducing the degree of abstraction and forming convolution paths of different depths. The results of these convolution paths are directly subjected to global average pooling, which is referred to as GAP. The SoftMax classifier is used to perform gradient descent after taking the cross entropy with the fault label to complete the initial training, maximize the representation ability of each convolution depth, and avoid the interference of parameter-containing classifiers on non-intersection feature extraction. The learning effect of the feature depends only on the frequency and significance of the feature. After the training, the GAP and SoftMax classifiers are discarded.
[0062] S320, consolidating training of feature matrix weights.
[0063] Specifically, step S320 includes the following steps:
[0064] S321. The output features of convolution embeddings with different depths are dot-producted with matrices of the same shape and then averaged within the row, which is equivalent to adding a layer of selectively connected unbiased fully connected layer, so that the output still maintains category independence. The module output and the fault label are cross-entropy and gradient descent is performed through the SoftMax classifier. After training, the SoftMax classifier is discarded and the output is used as the feature input of the next repeated module. Under this structure, the learning of connection weights plays a role similar to feedback consolidation. Weights are generated under the input of full-load domain samples to adjust the expression intensity of convolution neurons, so that they can adapt to the multi-load average environment.
[0065] S322. After the stage training is completed, thresholds are assigned to different output features of the convolution. Features above the threshold will directly determine the output category advantage, while features below the threshold will serve to build thresholds for other features. The threshold of the final result of each feature value is calculated by the following formula:
[0066]
[0067] Among them, the number of rows of FWM is the number of categories, and the number of columns n is the number of features in a single channel;
[0068] FWM ab is the weight of the corresponding position in row a and column b of the FWM weight matrix;
[0069] FWM ij is the weight corresponding to the position in row i and column j of the FWM weight matrix except row a and column b;
[0070] O ij The features or score values outputted by row i and column j of the FWM weight matrix except row a and column b;
[0071] Thresold ab is the feature threshold corresponding to the position in row a and column b of the FWM weight matrix.
[0072] Specifically, the FWM weight matrix is constrained by the following formula:
[0073]
[0074] Among them, x is the additional registered parameter;
[0075] λ is the value range constraint coefficient;
[0076] σ(x) is the weight of the sigmoid constraint.
[0077] S330, consolidating the training of the depth weight matrix.
[0078] Specifically, step S430 includes the following steps:
[0079] A deep weight matrix of the same form as the feature weight layer is added after the feature weight layer, and the dot product with the output feature of the feature weight layer is performed and then summed within the row. Step S320 is repeated to obtain the deep weight layer. The output of the deep weight layer has the same shape as the fault label and can be used as the final output used in actual fault diagnosis to complete the training of MDUNN.
[0080] S400: Input the cylinder head vibration signal collected in real time into MDUNN, and output the diesel engine fault type.
[0081] Specifically, the S400 specifically includes:
[0082] S410: Acquire a cylinder head vibration signal based on an acceleration sensor disposed on the cylinder head of the diesel engine.
[0083] S420: Input the cylinder head vibration signal into MDUNN at every preset time interval.
[0084] S430 and MDUNN perform multi-depth feature extraction on the cylinder head vibration signal, adjust the reduced dimension in two weight layers, and obtain the fault type prediction result.
[0085] S440, output diesel engine fault type.
[0086] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.
[0087] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.
Claims
1. A diesel engine fault diagnosis method under variable load based on MDUNN, characterized in that: The following steps are involved: S100, collecting vibration signals of different faults under different loads based on an acceleration sensor provided on a cylinder head of a diesel engine; S200, segmenting vibration signals of different faults under different loads according to fixed signal lengths and establishing fault labels corresponding to fault types one by one, thereby obtaining a fault dataset, wherein the fault dataset includes all fault labels; S300: Build an MDUNN model and use the fault dataset to train MDUNN in stages. The specific steps include: S310, constructing a multi-depth convolutional neural network, including the following steps: S311. Construct a convolutional neural network, increase the number of kernels in the first convolution layer to 128, and maintain the number of kernels in subsequent layers at 64. The first convolution layer with 128 kernels and the hidden layer with 64 kernels constitute a feature extractor. Add a convolution kernel layer with a size of 1 after the feature extractor, and adjust the number of channels to a preset number of categories. The number of categories is selected based on the fault types in the fault dataset, so that the convolution output features can be directly used for decision making. S312: Reduce the convolution layer containing 64 convolution kernels in the above convolution structure, thereby reducing the level of abstraction and forming convolution paths of different depths. The results of these convolution paths are directly subjected to global average pooling (GAP). The global average pooling is referred to as GAP. The SoftMax classifier is used to perform gradient descent after taking cross entropy with the fault label to complete the initial training. After the training is completed, the GAP and SoftMax classifiers are discarded. S320, consolidating the feature matrix weights, including the following steps: S321: The output features of convolutional embeddings of different depths are dot-producted with matrices of the same shape and then averaged within the row. This is equivalent to adding a layer of selectively connected unbiased fully connected layer, so that the output still maintains category independence. The module output is cross-entropyed with the fault label through the SoftMax classifier and then gradient descent is performed. After training, the SoftMax classifier is discarded and the output is used as the feature input of the next repeated module. S322. After the stage training is completed, thresholds are assigned to different output features of the convolution. Features above the threshold will directly determine the output category advantage, while features below the threshold will serve to build thresholds for other features. The feature threshold is calculated using the following formula: Among them, the number of rows of FWM is the number of categories, and the number of columns n is the number of features in a single channel; FWM ab is the weight of the corresponding position in row a and column b of the FWM weight matrix; FWM ij is the weight of the corresponding position in row i and column j of the FWM weight matrix except row a and column b; O ij The feature or score value output for row i and column j except row a and column b of the FWM weight matrix; Thresold ab is the feature threshold corresponding to the position in row a and column b of the FWM weight matrix; The FWM weight matrix is constrained by the sigmoid function, and the formula is as follows: Among them, x is the additional registered parameter; λ is the value range constraint coefficient; σ(x) is the weight of the sigmoid constraint; S330, consolidating the training of the depth weight matrix, including the following steps: adding a depth weight matrix of the same form as the feature weight layer after the feature weight layer, and performing intra-row summation after performing dot product with the output feature of the feature weight layer, repeating step S320 to obtain a depth weight layer, the output of the depth weight layer has the same shape as the fault label, and can be used as the final output used in actual fault diagnosis, completing the training of MDUNN; S400: Input the cylinder head vibration signal collected in real time into MDUNN, and output the diesel engine fault type.
2. The diesel engine fault diagnosis method under variable load based on MDUNN according to claim 1 is characterized in that: The S400 specifically includes: S410, obtaining a cylinder head vibration signal based on an acceleration sensor provided on the cylinder head of the diesel engine; S420, inputting the cylinder head vibration signal into MDUNN at every preset time interval; S430 and MDUNN extract multi-depth features from the cylinder head vibration signal, adjust the reduced dimension at two weight layers, and obtain the fault type prediction result; S440, output diesel engine fault type.
3. The diesel engine fault diagnosis method under variable load based on MDUNN according to claim 1 is characterized in that: The acceleration sensor is a unidirectional acceleration sensor, which is electrically connected to a multi-channel data acquisition system. The multi-channel data acquisition system is electrically connected to obtain a cylinder head vibration signal through the unidirectional acceleration sensor.
