Monitoring and early warning method and device for blockage of gas path pipeline of engine

Abstract

The invention provides an engine gas circuit pipeline blockage monitoring and early warning method and device, and the method comprises the steps: (1) arranging a vibration collection device on an engine gas circuit pipeline, obtaining vibration data of multiple working conditions, and dividing the vibration data into a training set, a verification set and a test set; (2) calculating a spectrum centroid, a short-time zero-crossing rate and a short-time energy feature, and fusing by using a cross attention mechanism to generate a multi-granularity feature; (3) introducing a channel attention module, a space attention module, a multi-view bidirectional LSTM module and a multi-scale self-attention module, and capturing subtle difference features of the vibration signals; (4) designing a conditional fault feature diffusion module and a progressive interlaced learning strategy, and improving the generalization ability and adaptability of the model on different models of engines through feature domain distribution bridging and alternate optimization; and (5) inputting test set data into the model, accurately judging the blockage degree according to classification scores, covering three states of unobstructed, semi-blocked and blocked, and realizing multi-level blockage monitoring and early warning of the gas path pipeline of the engine.

Description

Technical Field

[0001] The present invention relates to a method and device for monitoring and warning of engine gas pipeline blockage Background Art

[0002] At present, the blockage problem of engine gas pipelines is also a common and serious challenge. Due to the continuous changes in the use environment and the influence of various factors during the operation of the engine, the gas pipeline is easily affected by pollution, sediments, foreign matter, etc. and becomes blocked, which may cause poor air flow, affect combustion efficiency, and even cause engine performance degradation, failure or even damage, resulting in certain hidden dangers in use and the risk of loss of life and property.

[0003] In traditional methods, the detection of engine gas line blockage usually relies on vehicle performance degradation or the triggering of fault codes. This approach often means that diagnosis can only be performed when the problem has occurred and progressed to a certain stage. In addition, it is relatively difficult to detect small-scale or initial blockages, which leads to the need for real-time monitoring of the engine gas line condition and early warning.

[0004] In view of these problems, a method and device for monitoring and warning of engine gas line blockage can be used to monitor and warn of engine gas line blockage, which is beneficial to enhance the user's sense of security and improve the working efficiency of the engine. Summary of the invention

[0005] Purpose of the invention: In view of the problems existing in the prior art, the present invention provides an engine gas line blockage monitoring and early warning method. The present invention first calculates the spectral centroid, short-time zero-crossing rate and short-time energy characteristics of the real-time collected data, and obtains multi-granularity features by cross-attention fusion; then, the multi-granularity features are learned and trained through a multi-scale cascade attention network model, and the conditional fault feature diffusion module and the progressive staggered learning strategy are used to achieve adaptive recognition of different models of engines through distribution bridging and alternating optimization of the feature domain, so that the model can accurately learn the characteristic distribution state of blockage data of different models of engines, and save the optimal training network model parameters.

[0006] Technical solution: The engine gas line blockage identification method based on cascaded attention network of the present invention comprises:

[0007] (1) Multiple vibration sensors are arranged around the engine to intrusively collect vibration data in the target area, and the sampled data is quantized and converted into digital signals, and then preprocessed to extract vibration signal features to obtain the original feature domain;

[0008] (2) The extracted three features, namely, spectrum centroid, short-time zero-crossing rate and short-time energy, are fused through a cross-attention mechanism to generate multi-granularity vibration features;

[0009] (3) A multi-scale cascade attention network model is established, which consists of channel attention, spatial attention, multi-view bidirectional long short-term memory network and multi-scale self-attention network. For the input of generating multi-granularity vibration features, the channel attention module and the spatial attention module are used to find the fault-related part of each channel, and the latter is used to supplement the channel attention and find the fault-related part in the space. The self-attention operation result is F3; the multi-view bidirectional long short-term memory network includes two layers of LSTM networks, wherein the first layer of LSTM network is used to obtain the time dependency at the segment level, and then the time series at the segment level is moved by 3, 6 and 9 time steps respectively to obtain the time series of multiple different views, which are input into the next layer of LSTM network to obtain the time dependency at the sentence level. Finally, the feature sequences of these four different time steps are compressed using the maximum pooling layer and spliced ​​in the channel dimension to realize the multi-scale feature F B Extraction; The multi-scale self-attention network includes an input sequence F B The projected query (Q) and the keys (K) and values ​​(V) that are adaptive to different scales are output as multi-scale subtle difference fault features F after self-attention fusion. MSA ;

[0010] (4) The conditional fault feature diffusion module of the present invention uses a diffusion operator q to diffuse the input multi-scale subtle difference fault features, and then applies a single reverse operator p θ At the same time, a slightly concentrated and dispersed feature distribution or a diffuse and concentrated feature distribution can be achieved, and the diffuse (concentrated) feature distribution is encouraged to gradually transform into an unknown new feature distribution. Since it is only a slow diffusion (concentration) in small steps, there is no need to worry that the converted feature distribution will violate the properties of the source feature distribution. By using diffusion q and reverse operator p for each iterative step in the transition stage θ To bridge the distribution of two different feature domains and obtain the features under different diffusion steps Output; Finally, the progressive staggered learning strategy uses the transition process characteristics To guide the classification network (CN) to adapt to the complex feature distribution under the generated multiple original semantics, and use CN to supervise the learning of the conditional fault feature diffusion module, aiming to gradually improve the domain adaptation ability of the classification network while retaining the feature semantics. After multiple rounds of iterative training, the classification score is calculated by the softmax classifier. If the score is greater than the specified threshold, different congestion levels are output according to the score size, including unobstructed, semi-obstructed and blocked.

[0011] Furthermore, the original feature domain obtained by extracting the vibration signal features in step (1) is:

[0012] For the calculation of the spectral centroid of the time-frequency domain features of the input vibration signal, first, the dynamic difference is reduced by normalization; next, the power spectrum of each frame is calculated using the short-time Fourier transform (STFT) with a Hamming window length of 40ms and a window shift of 10ms; then, for the power spectrum P(f) of the vibration signal within a given short time window, the spectral centroid C can be calculated by the following formula:

[0013]

[0014] Where f is the frequency and P(f) is the power spectral density at the corresponding frequency.

[0015] For the calculation of the short-time zero-crossing rate feature in the time domain feature of the input vibration signal, first, a Hamming window with a length of 40ms is used to divide the vibration signal into short-time windows; then, the short-time energy is obtained for the zero-crossing rate of the signal in each window:

[0016]

[0017] Where E(i) is the short-time energy of the i-th frame, x(n) is the sampling value of the signal, and N is the window length.

[0018] Next, the zero-crossing point of the signal in each window is calculated by detecting the positive and negative changes between adjacent sampling points:

[0019]

[0020] Where ZCR(i) is the zero-crossing rate of the i-th frame, and sign() is the sign function.

[0021] Finally, the short-time zero-crossing rate is normalized to compare the changes in zero-crossing rate between different time windows. The normalization formula is:

[0022]

[0023] Among them, the obtained Normalized ZCR(i) feature is recorded as Where i represents the i-th sample.

[0024] For the calculation of the short-time energy feature in the energy feature of the input vibration signal, first, a Hamming window with a length of 40ms is used to divide the vibration signal into short-time windows; then, the signal in each window is squared to obtain the short-time energy. This step is usually used in conjunction with the calculation of the short-time zero-crossing rate to improve the robustness of the feature; finally, the obtained short-time energy is normalized to compare the energy changes between different time windows, wherein the moving average normalization method is optionally used to smooth the short-time energy to reduce sudden energy changes. The calculation formula is:

[0025]

[0026] Among them, the obtained Normalized Energy (i) feature is recorded as Where i represents the i-th sample.

[0027] Furthermore, the feature fusion described in step (2) through the cross-attention mechanism to generate multi-granularity vibration features is:

[0028]

[0029] In the formula, is cross attention, C, stpz, ste represent the three features of spectrum centroid, short-time zero-crossing rate and short-time energy respectively, and GA is the global average operation. are the spectral centroid feature and the weighted embedding feature of the two features, k is the sample number, and the output is the multi-granularity cross-attention fusion representation C k .

[0030] Furthermore, the channel attention weight F in step (3) c The calculation formula is:

[0031]

[0032] In the formula, σ represents the activation function, E represents the shared weight network, W1 and W0 represent the shared weights, avg() represents the average pooling method, and max() represents the maximum pooling method. By using the channel attention weight F c Multiply it with the original input feature map F0 to get the adjusted feature map F1.

[0033] The spatial attention weight F in step (3) s The calculation formula is as follows:

[0034]

[0035] In the formula, σ represents the activation function, conv represents the convolution operation with a convolution kernel size of 7*7, avg() represents the average pooling method, and max() represents the maximum pooling method. The obtained spatial attention weight F s Multiply it with F1 to calculate F2, and then use the skip link to sum the initial features F0 and F2 element by element as F3.

[0036] The multi-view bidirectional long short-term memory network described in step (3) is:

[0037] F v0 =LSTM2(LSTM1(F3))

[0038] F v3 =LSTM2(View(LSTM1(F3),3))

[0039] F v6 =LSTM2(View(LSTM1(F3),6))

[0040] F v9 =LSTM2(View(LSTM1(F3),9))

[0041] F B =cat(max(F v0 ),max(F v3 ),max(F v6 ),max(F v9 ))

[0042] Among them, LSTM represents a long short-term memory network with 128 hidden layers. View(·,n) is a shift operation, where n is the time step length of the shift. max(·) represents the maximum pooling method with a pooling kernel of 2, and cat(·) represents concatenating multiple input sequences in the channel dimension.

[0043] The multi-scale self-attention described in step (3) is

[0044]

[0045] Specifically, for different attention heads indexed by the i value, the key K and value V are downsampled to different scales:

[0046]

[0047] V i =V i +LE(V i )

[0048] In the formula, MTA(·,b i) indicates that the following sampling rate is b i Sampling is performed, of which is the linear mapping weight matrix in the i-th head. LE(·) is a local enhancement of the MTA result of the V value through a depthwise separable convolution to retain more fine-grained and low-level detail information to obtain high-quality fault features, and the output is a multi-scale subtle difference fault feature F X .

[0049] Furthermore, the conditional fault feature diffusion module in step (4) is:

[0050] The well-designed conditional fault feature diffusion module consists of two basic units: diffusion operator q and reverse operator p θ , which have the same definition as the traditional probabilistic diffusion model. Different from the standard diffusion model which separates the diffusion process and the reverse process, the present invention adopts the diffusion operator q and the reverse operator p in each iterative step in the transition phase of the diffusion model. θ To bridge the distribution of two different feature domains and obtain the features under different diffusion steps.

[0051] For a given source feature F output by the feature extraction network X , which is similar to the forward process feature map The relationship can be mathematically expressed as:

[0052]

[0053] Among them, For(,k) means using diffusion operator q to diffuse k consecutive steps, and t is the number of fragments contained in the feature. Correspondingly, the reverse process feature map The forward process feature map The denoising formula is as follows:

[0054]

[0055] Among them, Rev(,k) means using the reverse operator p θ Denoise k consecutive steps.

[0056] Furthermore, the progressive interleaving learning strategy in step (4) is:

[0057] The present invention proposes a progressive interleaved learning strategy (PILS) to maintain the semantics of transition feature mapping while gradually improving the domain adaptation ability of the classification network. The progressive interleaved learning strategy involves two learning directions (i.e., "CN-to-CFFD" and "CFFD-to-CN"). Among them, the "CN-to-CFFD" process uses the classification network trained under the diffusion feature distribution of the previous diffusion step to update the conditional fault feature diffusion module to constrain the feature semantics of the next diffusion feature distribution. Conversely, the "CFFD-to-CN" process uses the updated conditional fault feature diffusion module to promote the classification network to learn a new diffusion feature distribution, aiming to enhance its recognition performance. By iterating the above two processes in an interleaved manner, the classification network is driven to generalize on a small gap feature distribution, and finally a high recognition ability of different fault features is achieved.

[0058] The goal of network training is to minimize the total loss function of the network.

[0059] The present invention also discloses an engine gas line blockage identification device based on a cascade attention network, including a vibration data collector, a memory and a processor. The vibration data collector is composed of four vibration sensors for real-time data collection. The memory stores the data collected during the operation of the computer program and the model. The processor is used to implement the above method when executing the computer program.

[0060] Beneficial effects: Under normal living conditions, unobstructed engine gas pipelines will have certain regular vibrations, but when blockage or other faults occur, a vibration frequency that is significantly different from normal operation will be generated. At this time, the trained vibration model can be used to identify abnormal blockage signals, and the pipeline can be repaired. After the blockage is identified, an audible and visual alarm will be used to indicate the blockage, thereby increasing the service life of the engine. Since the amount of vibration data in the normal state of the pipeline is large and easy to collect, by collecting these vibrations in advance for training, the trained vibration model can quickly identify the blockage state and detect the blockage information in a timely manner. In addition, the blockage of the engine gas pipeline has special factors such as sporadic and concurrent factors. The use of the vibration model can improve the recognition rate of abnormal vibrations over a long period of time. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a logic flow chart of the present invention.

[0062] Figure 2 It is a heat map of the initial spectrum centroid feature map in the collected vibration data described in an embodiment of the present invention and three states of full blockage, half blockage and unobstructed after multi-scale cascade attention network training.

[0063] Figure 3 It is an experimental effect diagram of the present invention.

[0064] Figure 4 It is a structural diagram of the conditional fault feature diffusion module of the present invention.

[0065] Figure 5 It is the overall framework flow chart of the present invention.

[0066] Figure 6 It is the short-time energy curve of the congestion state.

[0067] Figure 7 It is a short-time energy curve diagram of the unobstructed state. DETAILED DESCRIPTION

[0068] This embodiment provides a method for monitoring and warning of engine gas pipeline blockage, such as Figure 1 , Figure 2 and Figure 3 As shown, including:

[0069] (1) Vibration sensors are arranged to collect vibration data of engine gas pipelines under different blockage states in a non-invasive manner, and a fault vibration database with corresponding category labels is established. The database is divided into training set, validation set, and test set at a ratio of 8:1:1. The vibration data is sampled and quantized into digital signals, and then preprocessed to extract three characteristics of the vibration signal: spectral centroid, short-time zero-crossing rate, and short-time energy.

[0070] Among them, for the calculation of the spectrum centroid in the time-frequency domain characteristics of the input vibration signal, first, the dynamic difference is reduced by normalization; next, the short-time Fourier transform (STFT) with a Hamming window length of 40ms and a window shift of 10ms is used to calculate the power spectrum of each frame; then, for the power spectrum P(f) of the vibration signal within a given short time window, the spectrum centroid C can be calculated by the following formula:

[0071]

[0072] Where f is the frequency and P(f) is the power spectral density at the corresponding frequency.

[0073] For the calculation of the short-time zero-crossing rate feature in the time domain feature of the input vibration signal, first, a Hamming window with a length of 40ms is used to divide the vibration signal into short-time windows; then, the short-time energy is obtained for the zero-crossing rate of the signal in each window:

[0074]

[0075] Where E(i) is the short-time energy of the i-th frame, x(n) is the sampling value of the signal, and N is the window length.

[0076] Next, the zero-crossing point of the signal in each window is calculated by detecting the positive and negative changes between adjacent sampling points:

[0077]

[0078] Where ZCR(i) is the zero-crossing rate of the i-th frame, and sign() is the sign function.

[0079] Finally, the short-time zero-crossing rate is normalized to compare the changes in zero-crossing rate between different time windows. The normalization formula is:

[0080]

[0081] Among them, the obtained Normalized ZCR(i) feature is recorded as Where i represents the i-th sample.

[0082] For the calculation of the short-time energy feature in the energy feature of the input vibration signal, first, a Hamming window with a length of 40ms is used to divide the vibration signal into short-time windows; then, the signal in each window is squared to obtain the short-time energy. This step is usually used in conjunction with the calculation of the short-time zero-crossing rate to improve the robustness of the feature; finally, the obtained short-time energy is normalized to compare the energy changes between different time windows, wherein the moving average normalization method is optionally used to smooth the short-time energy to reduce sudden energy changes. The calculation formula is:

[0083]

[0084] Among them, the obtained Normalized Energy (i) feature is recorded as Where i represents the i-th sample.

[0085] (2) For the three features of spectrum centroid, short-term zero-crossing rate and short-term energy, a method of fusing them using the cross-attention mechanism is used:

[0086] First, the three input features and Use three different fully connected layers to map them into different feature spaces and obtain three feature vectors f i , g j With h k :

[0087]

[0088] Where W f 、b f , W g 、b g , W h and b h are the learnable parameter matrices and bias vectors.

[0089] Next, we compute f by using dot product attention i , g j and h k The correlation score between ijk , the detailed calculation formula is as follows:

[0090]

[0091] Then, the scores are normalized using the Softmax function to obtain the attention weight a of each input element on the other two input elements. ij , b jk and c ki :

[0092]

[0093]

[0094] The calculated attention weights are used to fuse the three feature vectors by weighted summation to apply the attention weights to each feature vector:

[0095]

[0096] Finally, the three eigenvectors after fusion are Splice to obtain multi-granularity cross-attention fusion representation c k .

[0097] (3) For multi-dimensional feature c k For feature extraction, the present invention establishes a new multi-scale cascade attention network, which introduces channel attention network, spatial attention network and multi-head fusion self-attention mechanism to locate and learn target features.

[0098] First, for the input feature c with multiple channels k , using maximum pooling and average pooling for feature compression, the maximum pooling operation extracts the maximum eigenvalue in each channel to obtain finer features of each channel, while the average pooling operation calculates the average eigenvalue in each channel to aggregate the spatial information between channels, and obtains two feature descriptors respectively. and The feature descriptor is then applied to the entire shared weight network, and the channel weight is used to multiply the feature descriptor and then sum each element to obtain the channel attention weight F c Finally, using the channel attention weight F c Compared with the original input feature map c k Multiply them to get the adjusted feature map F1. The channel attention weight calculation formula is as follows:

[0099]

[0100] Among them, σ represents the activation function, E represents the shared weight network, W1 and W0 represent the shared weights, avg() represents the average pooling method, and max() represents the maximum pooling method.

[0101] In order to make fuller use of feature information, the present invention introduces spatial attention after channel attention to adaptively adjust the importance of different spatial positions in the input features to extract more interesting information from the existing features. In this way, the model can automatically focus on important positions and suppress unimportant positions when processing input, thereby improving the performance of the model.

[0102] First, the channel itself needs to be reduced in dimension to obtain the maximum pooling features. and average pooling features The feature maps are concatenated in the channel dimension, and then a convolutional layer is used to learn to obtain the spatial attention weight F s The calculation formula of spatial attention weight is as follows:

[0103]

[0104] Among them, σ represents the activation function, conv represents the convolution operation with a convolution kernel size of 7*7, avg() represents the average pooling method, and max() represents the maximum pooling method. s Multiply it with F1 to calculate F2, and then use the jump link to convert the initial feature C n The result of element-by-element summation with F2 is counted as F3 and input into the next module.

[0105] Then, a multi-view bidirectional long short-term memory network is used to extract multi-scale features. First, the output feature F3 of the previous module is divided into non-overlapping segments, and the temporal dependency at the segment level is obtained through the first layer of LSTM network. Then, the output time series is shifted by 3, 6, and 9 time steps respectively to obtain the time series features of different views. Then, these shifted time series are input into the next layer of LSTM network together with the original time series to obtain the temporal dependency at the sentence level. Finally, the maximum pooling layer is used to compress the features and splice them in the channel dimension to achieve the extraction of multi-scale features. The detailed calculation process is as follows:

[0106] F v0 =LSTM2(LSTM1(F3))

[0107] F v3 =LSTM2(View(LSTM1(F3),3))

[0108] F v6 =LSTM2(View(LSTM1(F3),6))

[0109] F v9 =LSTM2(View(LSTM1(F3),9))

[0110] F B =cat(max(F v0 ),max(F v3 ),max(F v6 ),max(F v9 ))

[0111] LSTM stands for a long short-term memory network with 128 hidden layers. View(·,n) is a shift operation, where n is the time step length of the shift. max(·) represents the maximum pooling method with a pooling kernel of 2, and cat(·) represents concatenating multiple input sequences in the channel dimension, and the output is a multi-scale feature F. B .

[0112] Finally, these multi-scale features are fused together using multi-scale self-attention, where the calculation formula for multi-scale self-attention is as follows:

[0113]

[0114] Where, d h Represents dimension.

[0115] Specifically, for the input feature F B Projected onto tensors of query (q), key (K), and value (V). Then, self-attention is calculated in parallel using H independent attention heads, and the length of K and V of different attention heads in the same attention layer is reduced and expanded by setting the downsampling rate b, so that on the same self-attention layer, the sizes of attention heads are different and the lengths vary in different heads of different scales of information. When the b value becomes larger, part of the feature maps in K and V are merged to shorten the length of K and V, so the computational cost is low, but the ability to capture large objects is still retained. On the contrary, when the b value becomes smaller, more detailed information is retained, but it brings more computational cost. Integrating different b values ​​into a self-attention layer enables it to have the ability to fuse multi-scale information. For different attention heads indexed by the i value, the key K and value V are downsampled to different scales:

[0116]

[0117] V i =V i +LE(Vi )

[0118] In the formula, MTA(·,b i ) indicates that the following sampling rate is b i Sampling is performed, of which is the linear mapping weight matrix in the i-th head. LE(·) is a local enhancement of the MTA result of the V value through a depthwise separable convolution to retain more fine-grained and low-level detail information to obtain high-quality fault features, and the output is a multi-scale subtle difference fault feature F MSA .

[0119] (4) Figure 4 As shown, unlike the standard diffusion model that separates the diffusion process and the reverse process, the conditional fault feature diffusion module designed in the present invention uses the diffusion operator q and the reverse operator p for each iterative step in the transition phase. θ To bridge the distributions of two different feature domains, we can obtain the features under different diffusion steps. It should be noted that a single step of the diffusion (or inversion) operator only slightly diffuses (centralizes) the distribution. Therefore, by applying a single inversion operator p θ A slightly concentrated and dispersed feature distribution or a diffuse and concentrated feature distribution can be achieved at one time, while encouraging the diffuse (concentrated) feature distribution to gradually transform into an unknown new feature distribution. Since it is only a slow diffusion (concentration) in small steps, there is no need to worry about the converted feature distribution violating the properties of the source feature distribution. It is worth noting that when k is small, the conditional fault feature diffusion module only makes small changes to the source feature distribution, obtaining a transition feature distribution that is less different from the source feature distribution. However, as k grows from 1 to K, the source feature distribution becomes more and more dispersed, and more and more reverse operations are performed, so that the difference between the output feature distribution and the source feature distribution gradually increases, thereby realizing the process of feature enhancement.

[0120] Given the source feature F output by the feature extraction network X , which is similar to the forward process feature map The relationship can be mathematically expressed as:

[0121]

[0122] Among them, For(,k) means using diffusion operator q to diffuse k consecutive steps, and t is the number of fragments contained in the feature. Correspondingly, the reverse process feature map The forward process feature map The denoising formula is as follows:

[0123]

[0124] Among them, Rev(,k) means using the reverse operator p θDenoise k consecutive steps.

[0125] Next, in order to enable the conditional fault feature diffusion module to have the ability to transform feature distribution without changing the semantics of the source feature distribution, the conditional fault feature diffusion module needs to be pre-trained. Specifically, the fault feature of the target domain is diffused for k diffusion steps, and then the diffusion features obtained by the forward process are used to pre-train the reverse operator using the reverse learning loss. The reverse learning loss can be calculated by the following mathematical formula:

[0126]

[0127] in, Represents the initial Gaussian distribution. After pre-trained reverse learning, given a random Gaussian noise fault feature distribution, the reverse operator p θ It can be gradually diffused (concentrated) into a complex feature distribution corresponding to the semantics in the corpus. This is because the goal of the reverse process learning is to obtain the negative log-likelihood -logp θ (f) is derived from the variational bounds of R Learn the generation of data f from various feature distributions. This feature of DDPM has been verified in previous studies and has been widely used in previous studies. It should be noted that the loss here is only used for the initialization training of the conditional fault feature diffusion module. The following invention introduces a progressive interleaving learning strategy to further train the proposed conditional fault feature diffusion module, aiming to ensure semantic preservation during feature distribution conversion.

[0128] (5) In the conditional fault feature diffusion module, the source features are transformed through a series of disturbances and confidence screening to obtain the characteristics of the transition process. The corresponding feature distribution is {X,D1,D2,…,D K}. The goal of this invention is to use this transition process to guide the classification network (CN) to adapt to the complex feature distribution under the generated multiple original semantics. Although the input of the conditional fault feature diffusion module only contains the source features F obtained from the feature extraction network X The corresponding source label Y X , but requires source label Y during training X The classification network can keep the original semantics unchanged during the transition feature mapping process and is applicable to all transition feature training.

[0129] Therefore, the present invention further proposes a progressive interleaved learning strategy (PILS) to maintain the semantics of the transition feature map while gradually improving the fault recognition ability of the classification network. The progressive interleaved learning strategy proposed in the present invention involves two learning directions (i.e., "CN-to-CFFD" and "CFFD-to-CN"). Among them, the "CN-to-CFFD" process uses the classification network trained under the diffusion feature distribution of the previous diffusion step to update the conditional fault feature diffusion module to constrain the feature semantics of the next diffusion feature distribution. Conversely, the "CFFD-to-CN" process uses the updated conditional fault feature diffusion module to promote the classification network to learn a new diffusion feature distribution, aiming to enhance its recognition performance. The present invention iterates the above two learning directions in an interleaved manner, driving the classification network to generalize on a small gap feature distribution, and ultimately achieving high recognition capabilities for features in different domains.

[0130] Specifically, the progressive interleaved learning strategy proposed in this invention mainly includes two parts: (1) the classification network guides the conditional fault feature diffusion module (the "CNto-CFFD" process); (2) the conditional fault feature diffusion module optimizes the classification network (the "CFFD to-CN" process). The above process can be simply described as follows: given a classification network C0, the model X The corresponding feature F X Then, the present invention freezes C0 and uses it to constrain the features from the next distribution D1 This is achieved by optimizing the loss function L CN-to-CFFD To achieve this, the calculation formula is as follows:

[0131] L CN-to-CFFD =L cls (C k -1(D k (F X )),Y X )+L R

[0132] Among them, L cls is the cross entropy loss function, C k-1 is the classification network frozen at the k-1th round and the model is already in the feature The training is completed, D k is the output feature distribution D k The conditional fault feature diffusion module module and D k (F X ) represents the corresponding feature under this feature distribution During the training of the conditional fault feature diffusion module through the above process, the classification loss forces the conditional fault feature diffusion module to retain The semantics of the feature, while LR The loss will distribute the original features D k Diffusion (concentration) to the new target feature distribution area.

[0133] After the first round of learning to keep the semantic labels, the source label Y X Can be used to update features The above process is the learning of “CN to-CFFD”. Then, the present invention converts the source feature F X Features after the first round of updates put together (i.e. ) to retrain the classification network and obtain a classification network C1 with good generalization performance on both distributions. This process is achieved through the loss function L CFFD-to-CN To constrain, it is defined as follows:

[0134]

[0135] Among them, C k is the new feature distribution D K Retrained classification network, when i = 0, there exists The first half of the above loss function is expressed as using the features of the previous training set To train the classification network, the second half is represented by using the newly generated features The classification network is trained to further enhance the generalization ability of the classification network.

[0136] Next, the "CN-to-CFFD" process is performed again using the newly trained classification network C1 to retain the features. semantics, and further expand the training set The classification network is retrained by executing the "CFFD-to-CN" process again. The present invention iterates the above two learning processes alternately, drives the classification network to generalize on the small gap feature distribution, gradually expands the high recognition ability of the classification network for different domain features, and finally obtains a classification network C with strong adaptability under the feature distribution of different models of engines. K , and used for evaluation on the test set.

[0137] In the test experiment, the output value of the model is set in the range of 0-1. The output value of 0-0.3 corresponds to the engine gas pipeline being unobstructed, the output value of 0.3-0.63 corresponds to the engine gas pipeline being semi-blocked, and the output value of 0.65-1 corresponds to the engine gas pipeline being blocked. Figure 3The figure shows the confusion matrix of the engine gas pipeline clear and blocked state corresponding to the true value and the engine gas pipeline clear and blocked state corresponding to the classification model output value. On the validation set, the probability that the engine gas pipeline clear and blocked state corresponding to the model prediction result and the true value are both clear state, semi-blocked state, and blocked state are 65%, 48%, and 83% respectively, among which the misidentification rate of clear state and blocked state is 0.

[0138] like Figure 6 , Figure 7 The short-time energy characteristic curve of vibration is shown in the figure. Figure 6 It is a curve diagram of the short-time energy of vibration under the blockage state. At position 1, the engine starts to take in air and the short-time energy increases. At position 2, due to the blockage, the short-time energy increases. At position 3, the engine stops working, the intake ends, and the short-time energy decreases rapidly. Figure 7 This is a short-time energy curve in an unobstructed state. At position 1, the engine starts to take in air and the short-time energy increases. At position 2, the engine stops working and the short-time energy slowly decreases until position 3. According to the decline of the short-time energy curve when the engine is about to end, the state of the engine gas pipeline can be preliminarily judged. The slope of the engine when it ends working in a blocked state decreases sharply, while the slope of the engine when it ends working in an unobstructed state decreases slowly.

[0139] The above disclosure is only a preferred embodiment of the present invention, which cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for monitoring and early warning of engine gas pipeline blockage, characterized in that: The following steps are involved: (1) Real-time collection of engine vibration data under different blockage states: Vibration collection devices are arranged on the engine gas pipelines in the source domain and the target domain to collect vibration data under the two environments in an intrusive manner. Specifically, vibration sensors are arranged on the engine in the source domain to collect vibration data under different blockage states, and a fault vibration database with corresponding category labels is established. These data will be used to train and optimize the monitoring model. Similarly, the same vibration sensors are arranged on the engine in the target domain to collect vibration data in real time. These data will be used for model testing and practical application. All collected vibration data must be sampled and quantized, converted into digital signals, and preprocessed to extract the spectral centroid, short-time zero-crossing rate, and short-time energy of the vibration signal, and construct the original feature domain. This ensures that the model can accurately monitor and identify the blockage state of the engine gas pipeline under different environments. (2) Generation of multi-granularity vibration features: The three features of spectrum centroid, short-time zero-crossing rate and short-time energy extracted from the source domain and target domain data are fused through the cross-attention mechanism to generate multi-granularity vibration features. Specifically, the spectrum centroid reflects the frequency range where the signal energy is concentrated, indicating the main distribution area of ​​the vibration energy; the short-time zero-crossing rate describes the waveform characteristics of the signal from the time domain perspective. Its level is related to the frequency component of the signal and can reflect the richness of the high-frequency components in the signal; the short-time energy measures the intensity or amplitude of the signal within a short time window and is used to detect sudden changes in the signal. Combining these three features can provide a comprehensive view of the vibration signal from multiple dimensions such as the relationship between frequency and energy, the complementarity of time domain and frequency domain, and the synergy of fault diagnosis, thereby improving the accuracy and reliability of the monitoring system in fault diagnosis; through the cross-attention mechanism fusion method, the blockage state of the engine gas pipeline can be more accurately identified and diagnosed, and the performance of the monitoring and early warning system can be enhanced; (3) Vibration feature extraction and anomaly recognition: A multi-scale cascade attention network model is established. The model consists of channel attention, spatial attention, multi-view bidirectional long short-term memory network and multi-scale self-attention. The input is the multi-granularity vibration features of the source domain, and the output is the multi-scale subtle difference fault features of the source domain. (4) Domain adaptive training and model alternating optimization: First, fault features are extracted on the target domain that needs to be generalized, and the conditional fault feature diffusion module is pre-trained. Then, the multi-scale subtle difference fault features of the source domain data extracted in S3 are input into the conditional fault feature diffusion module, and the diffusion operator q is used for K-step diffusion until the input fault feature distribution is converted to a random Gaussian noise distribution. Subsequently, the converted random Gaussian noise feature distribution is converted to a random Gaussian noise distribution using the inverse operator p. θ Gradually diffuse (concentrate) the complex feature distribution corresponding to the target domain and output the transition process characteristics Secondly, the model is trained alternately using the progressive interleaving learning strategy. Specifically, by designing the loss function L CN-to-CFFD Let the classification network supervise the learning of the conditional fault feature diffusion module ("CN to-CFFD" process): L CN-to-CFFD =L cls (C k -1(D k (F X )),Y X )+L R Where, L cls is the cross entropy loss function, C k-1 is the classification network frozen at the k-1th round, D k is the output feature distribution D k The conditional fault feature diffusion module and D k (F X ) represents the corresponding feature under this feature distribution By designing the loss function L CFFD-to-CN To optimize the learning of the classification network ("CFFD to-CN" process): In the formula, C k is the new feature distribution D K Retrained classification network, when i = 0, there exists The first half of the loss is expressed as using the features of the previous training set To train the classification network, the second half of the loss is expressed as using the newly generated features The classification network is trained to further enhance the generalization ability of the classification network. Finally, the classification score is calculated by the Softmax classifier during testing. If the score is greater than the specified threshold, different degrees of congestion are output according to the score, including unobstructed, semi-obstructed and blocked.

2. According to claim 1, a method for monitoring and early warning of engine gas pipeline blockage is characterized in that: The step (1) of collecting real-time vibration signals of the engine gas pipeline under different blockage states specifically includes the following steps: (1-1) The target area is the engine gas pipeline, and a group of small vibration collection devices are arranged on the outer wall of the main engine gas pipeline, wherein the vibration sensor array is composed of four vibration sensors at different positions; (1-2) The collected vibration data is sampled and quantized, converted into digital signals and preprocessed to obtain three vibration signal characteristics: spectrum centroid, short-time zero-crossing rate and short-time energy.

3. The engine gas line blockage monitoring and early warning method according to claim 2 is characterized by: In step (2), the three features obtained by extracting the vibration signal of the engine gas pipeline are fused through the cross-attention mechanism to generate multi-granularity vibration features, which specifically includes the following steps: (2-1) The three features of spectrum centroid, short-time zero-crossing rate and short-time energy extracted from the engine gas pipeline vibration data are cross-attentioned and global average pooling is performed to obtain weighted embedding features of the three types of features. (2-2) The three weighted embedding features are subjected to the cross-attention mechanism respectively to obtain the multi-granularity cross-attention fusion representation c k .

4. The engine gas line blockage monitoring and early warning method according to claim 3 is characterized by: The step (3) of establishing a cascaded attention network for vibration feature extraction and abnormality recognition specifically includes the following steps: (3-1) Use channel attention to locate the feature area in each channel and distinguish the importance of different channels to generate channel attention weight F c ; (3-2) Establish a multi-view bidirectional long short-term memory network, divide the time series F3 into non-overlapping segments, and input them into the first layer of LSTM network to obtain the temporal dependency at the segment level; then use the multi-view method to shift 3, 6, and 9 time steps respectively, and input the shifted time series together with the original time series into the next layer of LSTM network to obtain the temporal dependency at the sentence level; Finally, the feature sequences of these four different time steps are compressed using the maximum pooling method and concatenated in the channel dimension to obtain the multi-scale feature F B ; (3-3) Establish a multi-scale self-attention network and transform the input sequence F B Projected onto the tensors of query (Q), key (K), and value (V), multi-head self-attention is used to calculate self-attention in parallel using H independent attention heads, where the lengths of K and V of different attention heads in the same attention layer are reduced and expanded by setting the downsampling rate, so that the sizes of attention heads on the same self-attention layer are different, and the lengths vary in different heads used to capture information of different scales, so as to generate multi-scale subtle difference fault features F after self-attention fusion MSA ; (3-4) Use the Softmax function to classify the learned multi-scale subtle difference fault features and calculate the confusion matrix to obtain F1 Score The result is compared with the pre-trained threshold value to automatically determine the congestion status.

5. The engine gas line blockage monitoring and early warning method according to claim 4 is characterized by: In step (4), the domain adaptive training using the conditional fault feature diffusion module specifically includes the following steps: (4-1) Design a conditional fault feature diffusion module to diffuse the input multi-scale subtle difference fault features; use a diffusion operator to gradually diffuse the features to increase the distribution range and expression ability of the features; (4-2) In the process of feature diffusion, the reverse operator is introduced to dynamically adjust the feature distribution; when the feature distribution is too dispersed, the reverse operator can gradually centralize the features; When the feature distribution is too concentrated, the reverse operator can make the features gradually dispersed; This flexible adjustment of feature distribution ensures that features can be slowly transformed into unknown new feature distributions; since diffusion or concentration is carried out gradually, drastic changes in feature distribution can be avoided, ensuring that the converted features maintain the same properties as the original features; (4-3) In the feature distribution transition stage, by using the diffusion operator and the reverse operator simultaneously in each iteration step, the distribution of different feature domains is effectively bridged, and the difference between the source feature domain and the target feature domain is gradually narrowed; through multiple iterations, the feature output under different diffusion steps can be obtained, forming a smooth transition feature distribution from the source domain to the target domain; this method improves the adaptability and generalization ability of the model to various congestion states; (4-4) A progressive interleaved learning strategy is adopted to gradually guide the classification network to adapt to the generated complex feature distribution; the classification network not only classifies the feature distribution, but also supervises the learning process of the conditional fault feature diffusion module to ensure that the feature semantics remains unchanged; By gradually enhancing the domain adaptation capability of the classification network, it can still accurately identify and classify different types of blockage features; (4-5) In the later stage of model training, after multiple rounds of iterations, the classifier evaluates the gas pipeline characteristics and calculates the classification score; based on the size of the classification score, the blockage state of the engine gas pipeline is determined.

6. The engine gas line blockage monitoring and early warning device according to claim 5, characterized in that: It includes a vibration data collector, a memory and a processor. The vibration data collector is composed of multiple groups of vibration sensors with a sampling frequency of 1000HZ for real-time vibration data collection. The memory stores the data collected during the operation of the computer program and the model. The processor is used for executing the method of the computer program.

Images