A Fault Diagnosis Method for MOS Gas Sensor Arrays Oriented to Electronic Noses

Abstract

The present invention discloses a fault diagnosis method for a MOS gas sensor array for an electronic nose. The method comprises the following steps: First, construct an STW fault detection data set; Second, input the multi-dimensional response signal after STW segmentation processing into a GTN fault detection model to obtain a fault detection result. If the detection result is that a fault occurs, then execute step three; Third, construct a GADF-MTF fault identification data set; Fourth, construct a fault identification network based on IResnet; Fifth, input the GADF-MTF fault identification data set into the fault identification network based on IResnet, and use the fault identification network based on IResnet to identify the response signal type of each sensor according to the heat map feature, determine the position and fault type of the faulty sensor, and obtain a fault identification result. This method can effectively reduce the false alarm rate of fault detection, has good fault identification performance, and is suitable for long-term promotion and application.

Description

Technical Field

[0001] The present invention relates to a method for fault diagnosis of a gas sensor array, and particularly to a method for fault diagnosis of a MOS gas sensor array based on a Sliding Time Window algorithm (STW)-Gated Transformer Network (GTN) and GADF-MTF-IResnet (Gram Angle Difference Field-Markov Transition Field-Improved Deep Residual Network). Background Art

[0002] As a key information acquisition component of an electronic nose system, sudden failures of sensors will cause changes in the output signals of the sensor array, thus affecting the analysis accuracy of existing intelligent analysis models. With the increasingly complex application scenarios of the electronic nose system, during the long-term use process, the gas sensors will inevitably have sudden failures such as solder joint open circuit, heater degradation, power supply fluctuation, etc., resulting in unstable response output of the sensor array. Therefore, it is necessary to monitor the abnormal state of the gas sensor array in the electronic nose system. Timely fault detection, isolation and diagnosis of the sensors are of great significance for improving the reliability and maintainability of the electronic nose system.

[0003] Traditional fault detection, isolation and diagnosis methods based on PCA and similar methods use the normal signals of the MOS gas sensor array when no target gas is introduced for modeling. First, the SPE or T 2 statistic is used to achieve fault detection, then the contribution plot method is combined to isolate the faulty sensors, and finally a classification algorithm is used to solve the fault diagnosis problem. However, the response of the MOS sensor array to the target gas during operation will also cause sudden changes in the SPE or T 2 statistic, resulting in a certain degree of false alarms of faults. Most of the existing MOS gas sensor fault detection methods do not consider the fault detection under the normal response of the sensors to the gas. In addition, the fault isolation method based on the contribution plot has a high isolation accuracy for a single faulty sensor, but it cannot accurately isolate multiple faulty sensors in the multi-fault mode, and there are problems of false alarms or missed alarms in the fault location of the MOS gas sensor array. Finally, the existing research uses simulation software to superimpose simulation signals of different fault modes on the normal response signals to obtain experimental samples, which cannot fully reproduce the characteristics of the fault signals at the physical level. Summary of the Invention

[0004] Aiming at the above problems existing in the traditional MOS gas sensor array fault detection method and the contribution plot-based fault isolation method, the purpose of the present invention is to provide a method for fault diagnosis of a MOS gas sensor array for an electronic nose. This method can effectively reduce the false alarm rate of fault detection, has good fault recognition performance, and is suitable for long-term promotion and application.

[0005] The object of the present invention is achieved through the following technical solutions:

[0006] A fault diagnosis method for a MOS gas sensor array for an electronic nose, comprising the following steps:

[0007] Step 1, construct an STW fault detection data set, and convert the multi-dimensional response signal data set of the sensor array into a sliding time window data set;

[0008] Step 2, input the multi-dimensional response signal after STW segmentation processing into the GTN fault detection model, use the GTN fault detection model to implement window-by-window fault detection, accurately distinguish the multi-dimensional time series segments of normal signals and fault signals, and obtain the fault detection result. If the detection result is that a fault occurs, execute Step 3;

[0009] Step 3, construct a GADF-MTF fault identification data set, and convert the multi-dimensional response signal data set of the sensor array into a GADF-MTF heat map data set;

[0010] Step 4, construct a fault identification network based on IResnet;

[0011] Step 5, input the GADF-MTF fault identification data set into the fault identification network based on IResnet, and use the fault identification network based on IResnet to identify the response signal types of each sensor according to the heat map features, determine the positions and fault types of the fault sensors, and obtain the fault identification result.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] 1. The present invention uses a fault detection model based on STW and GTN to fully extract the time and space correlation features of the time window samples, thereby effectively distinguishing the normal response signals and fault signals of the sensor array, and solving the problem of high false alarm rate of the normal response of the sensor array to the target gas during the fault detection process.

[0014] 2. The GADF-MTF signal encoding method combines the time dependence feature of GADF for signals and the first-order Markov chain transition probability feature of MTF for signals, enhances the feature expression ability of the response signals of the sensor array, and effectively highlights the inter-class differences of the signals.

[0015] 3. The present invention uses PSA to improve the expression ability of the deep residual network for multi-scale information of the feature map, and at the same time enhances the effective attention of the network to important feature information through LIP, effectively improving the recognition accuracy and generalization ability of the fault diagnosis model for fault signal types. Description of the Drawings

[0016] Figure 1 This is the flowchart of the MOS gas sensor array fault diagnosis method for the electronic nose proposed by the present invention.

[0017] Figure 2 This is the overall network architecture diagram proposed by the present invention.

[0018] Figure 3 This is the structure diagram of the IResnet network proposed by the present invention.

[0019] Figure 4 This is the overall implementation flowchart of the present invention.

[0020] Figure 5 This is the structure diagram of the GTN fault detection model used in the present invention.

[0021] Figure 6 This is the implementation process of the LIP module.

[0022] Figure 7 This is the PSA module;

[0023] Figure 8 These are the fault detection results of six types of signals, a) heating power change, b) load resistance change, c) pin short circuit, d) power failure, e) voltage fluctuation, f) normal response. Detailed implementation manners

[0024] The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings, but are not limited thereto. Any modification or equivalent replacement of the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention shall be covered by the protection scope of the present invention.

[0025] The present invention provides a fault diagnosis method for a MOS gas sensor array for an electronic nose. First, STW is combined with GTN for fault detection. The multi-dimensional response signals of the sensor array are segmented into a set of time window samples with high time correlation features through STW, and then GTN is used to fully extract the time and space correlation features of the multi-variable time series, accurately distinguishing the multi-dimensional time series segments of normal signals and fault signals, and effectively solving the problem of high false alarm rate of normal gas response signals. Then, the Gram Angle Difference Field (GADF) is combined with the Markov Transition Field (MTF) for fault signal feature extraction, and the original multi-dimensional fault signals are converted into GADF-MTF heat images channel by channel. This image contains both the time-dependent features of the gas sensor response signals and the transition probability features of the first-order Markov chain, effectively increasing the inter-class difference of the fault signals. Finally, an improved deep residual neural network model is called channel by channel to realize the identification of the fault types of each sensor. This model uses ResNet50 as the backbone network, improves the network's expression ability for different scale features in the heat image by introducing a multi-scale channel attention mechanism (PSA), and then uses a local importance pooling module (LIP) to enhance the network's effective attention to important fault feature information, and can achieve effective fault identification. As Figure 1 shown, the specific steps of the method are as follows:

[0026] A1. Construct an STW fault detection data set, convert the multi-dimensional response signal data set of the sensor array into a sliding time window data set, and divide it into a training set and a test set according to a certain ratio. The specific steps are as follows:

[0027] B1. Obtain the multi-dimensional response fault signals and normal signals of the MOS gas sensor array.

[0028] In this step, the multi-dimensional response fault signals used are all obtained through an electronic nose experimental system equipped with a fault generation device.

[0029] B2. Use STW to segment the multi-dimensional response fault signals into continuous time window samples.

[0030] In this step, the original multi-dimensional response signal data set is segmented into a continuous sliding time window data set. The advantage of STW is that it can not only realize real-time fault detection window by window, but also retain the high spatio-temporal correlation features of the original multi-dimensional response signals.

[0031] B3. Assign labels to the segmented time window samples and randomly divide the multi-dimensional response fault signals into a training set and a test set.

[0032] In this step, the segmented time window samples are sequentially labeled as the fault class or the normal class, and a script is used to randomly divide the data set into a training set and a test set.

[0033] A2. Input the multi-dimensional response signal after STW segmentation into the GTN fault detection model to obtain the fault detection result. If the detection result indicates that a fault has occurred, execute step A3 and enter the fault identification and isolation phase.

[0034] Figure 5 It is the structure diagram of the GTN fault detection model. This model retains the encoder structure in the original Transformer, removes the decoder part, and expands the model into a two-tower structure by adding a set of channel encoders in parallel. At the same time, a Gate gating mechanism is introduced. By learning the output features of the two towers, a two-dimensional gating signal is generated to control the flow of feature information in the subsequent network to achieve the optimal feature fusion effect.

[0035] During the gating process, first concatenate the two-tower outputs Encoding1 and Encoding2 through Concat. Subsequently, obtain the two-dimensional gating signal through linear mapping by the Gate layer, and convert it into gating probabilities g1 and g2 using the Softmax function, which respectively correspond to the probability information of the first column and the second column of the two-dimensional gating signal. Finally, weight and combine Encoding1 and Encoding2 according to the gating probabilities g1 and g2 to obtain the fully fused two-tower feature output result Encoding. Process Encoding through the Linear linear layer to obtain the final output result of the GTN fault detection model, that is, the probability that the input sample is predicted as the corresponding category.

[0036] The entire gating process is expressed by formula (1) as follows:

[0037]

[0038] Among them, C and S respectively represent the two-tower output feature vectors of Encoding1 and Encoding2, W represents the weight configuration in the linear mapping of the Gate layer, b represents the bias vector in the linear mapping, h is the two-dimensional gating signal obtained through linear mapping, and g1 and g2 are the gating probabilities calculated by the Softmax function. Concatenate the two-tower output feature vectors according to the gating probabilities to obtain the final output vector that fully fuses the two-tower feature information y , and then perform linear processing on it to obtain the final classification result.

[0039] The two-tower structure of GTN enables the model to simultaneously capture the temporal and spatial correlation information of multi-dimensional time-series data, and the Gate gating mechanism can fully fuse the two-tower feature information to improve the final classification performance of the model.

[0040] A3. Construct a GADF-MTF fault recognition dataset, convert the multi-dimensional response signal dataset of the sensor array into a GADF-MTF heat image dataset, and divide it into a training set and a test set according to a certain ratio. The specific steps are as follows:

[0041] B4. Separate the multi-dimensional response fault signals into single-channel one-dimensional signals, and process the single-channel one-dimensional signals using the GADF and MTF signal encoding methods respectively.

[0042] In this step, GADF can retain the time-dependent features of the data, while having a large sparsity, and eliminating the redundant information between multi-modalities. MTF fully obtains the state change information of the time series at different times by calculating the transition probability between the current state and the other states.

[0043] B5. Perform weighted fusion on the GADF and MTF feature matrices to generate a GADF-MTF heat image.

[0044] In this step, the heat image generated by the weighted fusion of the feature matrices contains both the time-dependent features of the sensor response signal and the transition probability features of the first-order Markov chain, with richer feature information, effectively amplifying the inter-class differences of the fault signals.

[0045] B6. Assign fault category labels to the GADF-MTF heat image samples and randomly divide them into a training set and a test set.

[0046] In this step, fault category labels are assigned to the generated heat image samples, and the dataset is randomly divided into a training set and a test set using a script.

[0047] A4. Construct Figure 3 the fault recognition network based on IResnet as shown, using Resnet50 as the basic model, optimizing its convolutional layer network structure, and introducing the LIP and PSA modules at the same time. The specific steps are as follows:

[0048] B7. Change the network structure of the traditional Resnet50, replace the first convolutional layer of the traditional Resnet50 with two convolutional layers with a kernel size of 3×3 to enhance the expression ability in the initial stage of the network and reduce the model parameters.

[0049] B8. Introduce the LIP module to optimize the traditional pooling structure and residual structure.

[0050] As Figure 6 shown, the LIP module automatically enhances the recognition features by adaptively learning the importance weights of the input tensors during the downsampling process of the neural network. This module can determine which features are more important during the downsampling process, realizing the automatic enhancement and retention of key features.

[0051] The LIP module actively learns the feature discrimination criterion through network G, avoiding manual control of the importance function. G is named the logit module, which is implemented by a small fully convolutional network (FCN). The logit module can learn the features of each pixel point in the input image and generate an importance feature map. To make the importance weights non - negative and easy to optimize, an exp(·) operation is added after network G, obtaining:

[0052] W = F(I) = exp(G(I)) (2)

[0053] Figure 6 The implementation process of LIP is shown. First, the tensor I is input into the logit module to obtain the corresponding importance weights W. Then, average pooling operations are respectively performed on the product I×W of the importance weights and the input tensor and W. The result of the average pooling of I×W is divided by the result of the average pooling of W to obtain the final output of the LIP module:

[0054]

[0055] B9. Insert PSA modules before and after each residual module.

[0056] As Figure 7 shown, the PSA module can fully extract the multi - scale spatial information of the input feature map and enhance the expression ability of channel features. The implementation steps of this module are as follows:

[0057] (1) Use the SPC module to slice the input feature map X into S parts in the channel dimension, denoted as [X0, X1, …, X S-1 . The initial number of channels is C, and the number of channels of each part after slicing is Use multi - scale grouped convolutions with a convolution kernel size of K i (i = 0, 1, …, S - 1) to extract the different - scale spatial information of the channel feature maps of each part. The number of groups corresponding to each part is G i (i = 0, 1, …, S - 1), as shown in formula (4). Concatenate the output feature maps of each convolutional layer in the channel dimension to obtain a new feature map containing multi - scale channel features, as shown in formula (5).

[0058] F i = Conv(K i ×K i , G i )(X i ), i = 0, 1, 2, …, S - 1, F i ∈R C′×H×W (4)

[0059] F = Cat(F0, F1, …, F S-1 ) (5)

[0060] (2) Calculate the channel attention weights corresponding to the output feature maps of each convolutional layer using the SEWeight module, as shown in formula (6), and splice these weights in the channel dimension, as shown in formula (7).

[0061] Z i = SEWeight(F i ), i = 0, 1, 2, …, S - 1, Z i ∈R C′×1×1 (6)

[0062] Z = Z0 + Z1 + … + Z S-1 (7)

[0063] (3) Normalize the multi-scale channel attention vector using Softmax to obtain a new attention vector, as shown in formula (8).

[0064]

[0065] (4) Multiply the feature map F i with the attention vector att i obtained after normalization in the channel dimension to obtain the feature map after multi-scale channel attention weighting, as shown in formula (9). Splice the above feature maps in the channel dimension to output a feature map with richer multi-scale information, as shown in formula (10).

[0066] Y i = F i ⊙att i , i = 0, 1, 2, …, S - 1 (9)

[0067] Out = Cat(Y0, Y1, …, Y S-1 ) (10)

[0068] A5. Input the GADF-MTF fault recognition dataset into the fault recognition network based on IResnet to obtain the fault recognition result. As Figure 4 shown, it can be specifically divided into the following steps:

[0069] First, input the STW fault detection dataset after STW segmentation processing;

[0070] Then, use the GTN fault detection model to obtain the fault detection result;

[0071] Second, preprocess the response signal through the GADF-MTF method;

[0072] Then, input the GADF-MTF fault recognition dataset;

[0073] Finally, the fault type recognition result is obtained by using the fault recognition network based on IResnet.

[0074] Embodiment:

[0075] Dataset:

[0076] In this embodiment, an electronic nose experimental system is used in combination with a fault generation device to construct a fault detection dataset and a fault recognition dataset. Among them, the original fault signal dataset mainly includes five types of faults: power-off fault, voltage fluctuation fault, load resistance change fault, heating power change fault, and pin short-circuit fault.

[0077] In order to discuss the influence of the sliding time window parameters on the GTN fault detection model, in this embodiment, the sliding time window widths are respectively selected as 100, 200, and 300, and the window moving step sizes are 10 and 20 to perform sliding time window segmentation on the response signals of the MOS gas sensor array. Then, the GTN fault detection model is trained using the time window samples segmented from the training samples, and finally, the performance of the GTN fault detection model is tested using the test samples.

[0078] The GADF-MTF signal encoding method is used to convert the response signal of a single sensor into a heat map, and thus a fault recognition dataset is constructed. This dataset contains a total of 300 samples, including 210 training samples and 90 test samples.

[0079] Evaluation metrics:

[0080] (1) Fault detection evaluation metrics

[0081] Since the main goal of fault detection is to identify the samples with faults, the fault samples are usually regarded as Positive events, and the normal samples are regarded as Negative events. The present invention selects the fault detection rate Fault DetectionRate (FDR), the false positive rate False Positive Rate (FPR), and the false negative rate False Negative Rate (FNR) as the fault detection evaluation metrics, as shown in formulas (11), (12), and (13).

[0082]

[0083] Among them, the meanings of TP, TN, FP, and FN are expressed as follows:

[0084] (1) True Positive (TP): True positive example. It represents the situation where the actual sample is a fault sample and is correctly predicted as a fault sample by the model.

[0085] (2) True Negative (TN): True negative. It represents the situation where the actual sample is normal and is correctly predicted as normal by the model.

[0086] (3) False Positive (FP): False positive. It represents the situation where the actual sample is normal but is incorrectly predicted as a faulty sample by the model. This situation is usually referred to as a "false alarm" or "type I error".

[0087] (4) False Negative (FN): False negative. It represents the situation where the actual sample is faulty but is incorrectly predicted as normal by the model. This situation is usually called a "missed alarm" or "type II error".

[0088] (2) Fault identification evaluation metrics

[0089] Precision and Recall are used as model performance evaluation metrics for fault type identification.

[0090]

[0091] Among them, the meanings of TP, FP, and FN are expressed as follows:

[0092] (1) TP: It represents the situation where the samples in a specific fault category are correctly predicted as that category by the model.

[0093] (2) FP: It represents the situation where the samples in other fault categories are incorrectly predicted as a specific fault category by the model.

[0094] (3) FN: It represents the situation where the samples in a specific fault category are incorrectly predicted as other categories by the model.

[0095] Experimental results and analysis:

[0096] (1) Fault detection experiment

[0097] To verify the performance of the fault detection model, the samples are segmented using six different sizes of sliding time windows, and the average fault detection rate, average false alarm rate, and average missed alarm rate of various response signals are shown in Tables 1, 2, and 3.

[0098] Table 1 Average fault detection rate of the fault detection model under different sliding time window sizes

[0099]

[0100] As can be seen from Table 1, the GTN fault detection model trained with six different sliding window sizes can achieve a high fault detection rate for various response signals. As the values of the window width and the moving step size increase, the overall average fault detection rate increases. Among them, when using a sliding time window with a width of 100 to segment the signal, the fault detection rates of the two types of response signals, namely voltage fluctuation faults and load resistance change faults, are significantly lower than those of other types of signals. Both of the above two types of signal samples contain tiny fault samples. In the case of unstable power supply and environmental noise interference, it is difficult to distinguish the feature information contained in the time window samples with a width of 100 between tiny fault samples and normal samples, thus resulting in lower fault detection rates for the two types of response signals.

[0101] Table 2 Average false alarm rates of the fault detection model under different sliding time window sizes

[0102]

[0103] The magnitudes of the average false alarm rates shown in Table 2 reflect the recognition effects of the GTN fault detection model on normal response time window samples under different sliding time window sizes. As the width of the time window increases, the average false alarm rate generally shows a downward trend. Consistent with the average fault detection rate, the false alarm rates of the GTN fault detection model for voltage fluctuation faults and load resistance change faults are significantly higher than those of other types of signals.

[0104] Table 3 Average missed alarm rates of the fault detection model under different sliding time window sizes

[0105]

[0106] Table 3 reflects the missed alarm situations of the GTN fault detection model under six different sliding time window sizes. As the width of the window increases and the step size grows, the overall average missed alarm rate shows a downward trend.

[0107] When using a sliding time window with the size of (w = 300, s = 10), the fault detection effects of six types of samples are as Figure 8 shown. Among them, 0 represents the normal state and 1 represents the fault state. From Figure 8It can be seen that when using a well-trained GatedTransformer model to perform real-time fault detection on the signals collected from the sensor array, good detection effects can be achieved for the sensor signals of all six state types. Moreover, this model is sensitive to the fault occurrence time point and can relatively accurately judge the fault critical state and detect when the fault occurred. This model solves the problems that are difficult to solve by traditional fault detection models such as PCA and KPCA, that is, the problem of effectively distinguishing normal gas response signals and fault signals. The state of the normal gas response signal in each time window is shown as 0, that is, normal without faults, effectively solving the problem of high false positive rate of normal gas response faults.

[0108] (2) Fault identification experiment

[0109] To verify the superiority of the MOS gas sensor fault identification model proposed in the present invention, it is compared with a variety of other models.

[0110] Table 4 Comparison of fault identification results between IResnet and other models

[0111]

[0112] As can be seen from the results in Table 4, the precision and recall rates of the other nine fault identification models are lower than those of the IResnet fault identification model proposed in the present invention. When processing thermal images, the convolutional neural network model retains the translational invariance and the correlation features between local pixels, while the Transformer model lacks these features, resulting in poor recognition performance on datasets with relatively small specifications. Although a deeper network usually has stronger feature extraction ability, it is also more prone to overfitting, especially when the amount of training data is relatively small. Therefore, a model with a moderate network depth may be more suitable for solving recognition problems under limited training data. The fault identification model proposed in the present invention changes the initial layer convolutional structure on the basis of the traditional Resnet50 to enhance the network's ability to express image features in the initial stage. After introducing the attention mechanism, it effectively obtains and utilizes feature information of different scales to enrich the feature space, and uses local importance pooling to enhance the network's attention to important feature information, effectively improving the fault identification performance of the model.

[0113] The IResnet fault diagnosis model proposed in the present invention adds LIP and PSA modules on the basis of the original Resnet50 network, and replaces the 7×7 convolution of the first layer of the network with two 3×3 convolutions. Table 5 verifies the effectiveness of this improvement method through ablation experiments, where No means not performing this operation and Yes means performing this operation.

[0114] Table 5 Ablation experiment

[0115]

[0116] As can be seen from Table 5, when the three improvements in Model 2 to Model 4 are separately applied to the original model, certain improvement effects can be achieved. When the three operations are combined in pairs, the classification effect obtained is further improved compared to the single operation. When the three operations act on the original model together, the diagnostic precision rate and recall rate reach 100%. The above ablation experiments prove the effectiveness of the improved Resnet50 fault diagnosis model proposed by the present invention.

Claims

1. A MOS gas sensor array fault diagnosis method for electronic nose, characterized in that The method comprises the following steps: Step 1: Construct a STW fault detection dataset and convert the sensor array multidimensional response signal dataset into a sliding time window dataset; Step 2: Input the multidimensional response signal after STW segmentation processing into the GTN fault detection model, use the GTN fault detection model to implement window-by-window fault detection, accurately distinguish the multidimensional time series segments of normal signals and fault signals, and obtain the fault detection result. If the detection result is that the fault is detected, execute step 3; the GTN is a Gated Transformer Network, and the STW is a sliding time window; Step 3: Construct a GADF-MTF fault recognition dataset, and convert the sensor array multi-dimensional response signal dataset into a GADF-MTF thermal image dataset; Step 4: Build a fault identification network based on IResnet; Step 5: Input the GADF-MTF fault identification data set into the fault identification network based on IResnet. According to the heat map features, the fault identification network based on IResnet is used to identify the response signal type of each sensor, determine the location and fault type of the faulty sensor, and obtain the fault identification result.

2. The MOS gas sensor array fault diagnosis method for electronic nose according to claim 1 is characterized in that The specific steps of step one are as follows: Step 11, obtaining a multi-dimensional response fault signal and a normal signal of the MOS gas sensor array; Step 1 and 2: Use STW to segment the multidimensional response fault signal into continuous time window samples; Step 1: Assign labels to the segmented time window samples and randomly divide the multi-dimensional response fault signals into training sets and test sets.

3. The MOS gas sensor array fault diagnosis method for electronic nose according to claim 2 is characterized in that In the steps 1 and 3, the segmented time window samples are labeled with fault class or normal class in turn, and the data set is randomly divided into a training set and a test set using a script.

4. The MOS gas sensor array fault diagnosis method for electronic nose according to claim 1 is characterized in that The specific steps of step three are as follows: Step 31: Separate the multi-dimensional response fault signal into a single-channel one-dimensional signal, and process the single-channel one-dimensional signal using GADF and MTF signal coding methods respectively; Step 32: Perform weighted fusion of GADF and MTF feature matrices to generate a GADF-MTF thermal image; Step 3: Assign fault category labels to the GADF-MTF thermal image samples and randomly divide them into training sets and test sets.

5. The MOS gas sensor array fault diagnosis method for electronic nose according to claim 4 is characterized in that In step 33, the generated thermal image samples are labeled with fault categories, and a script is used to randomly divide the data set into a training set and a test set.

6. The MOS gas sensor array fault diagnosis method for electronic nose according to claim 1, characterized in that The specific steps of step 4 are as follows: Step 41: Change the network structure of the traditional Resnet50 and replace the first convolutional layer of the traditional Resnet50 with two convolutional layers with a convolution kernel size of 3×3 to enhance the expression ability of the network in the initial stage and reduce the model parameters; Step 42: Introduce the LIP module to optimize the traditional pooling structure and residual structure; Step 43: Insert PSA modules before and after each residual module.

7. The electronic nose-oriented MOS gas sensor array fault diagnosis method according to claim 6, characterized in that The LIP module actively learns the feature discrimination criterion through the network G, which is named the logit module. The exp(·) operation is added after the network G to obtain: W = F (I) = exp (G (I)) (2); Where I is the input tensor of the LIP module, and F(I) is the importance weight obtained by the LIP module function operation of I; First, the tensor I is input into the logit module to obtain the corresponding importance weight W. Then, the product of the importance weight and the input tensor I×W and W are average pooled, and the result of the I×W average pooling is divided by the result of the W average pooling to obtain the final output of the LIP module:

8. The method for fault diagnosis of MOS gas sensor array for electronic nose according to claim 6, characterized in that The implementation steps of the PSA module are as follows: (1) Use the SPC module to split the input feature map X into S parts in the channel dimension, represented as [X0,X1,…,X S-1 ], the initial number of channels is C, and the number of channels in each part after segmentation is Use a convolution kernel size of K i The multi-scale grouped convolution (i=0,1,…,S-1) extracts the spatial information of different scales of each part of the channel feature map, and the number of groups corresponding to each part is G i (i=0,1,…,S-1), as shown in formula (4), the output feature maps of each convolutional layer are concatenated in the channel dimension to obtain a new feature map containing multi-scale channel features, as shown in formula (5): F i =Conv(K i ×K i ,G i )(X i ),i=0,1,2,…,S-1,F i ∈R C′×H×W (4); F=Cat(F0,F1,…,F S-1 ) (5); Among them, F i It is the i+1th part of the input feature map X obtained by SPC segmentation and convolution kernel K i The feature information extracted by the convolution operation is F, which is the output feature map of each convolution layer. i A new feature map containing multi-scale channel features obtained by splicing in the channel dimension; (2) Use the SEWeight module to calculate the channel attention weights corresponding to the output feature maps of each convolutional layer, as shown in formula (6), and concatenate these weights in the channel dimension, as shown in formula (7): Z i =SEWeight(F i ),i=0,1,2,…,S-1,Z i ∈R C′×1×1 (6); Z=Z0+Z1+…+Z S-1 (7); Among them, Z i It is the output feature map F of the i+1th convolutional layer calculated using the SEWeight module i The corresponding channel attention weight, Z is the channel attention weight Z corresponding to the feature map obtained by S segmentation i The result of splicing in the channel dimension, Z0 is the channel attention weight corresponding to the output feature map F0 of the first convolution layer calculated by the SEWeight module, and Z1 is the channel attention weight corresponding to the output feature map F1 of the second convolution layer calculated by the SEWeight module. S-1 It is the Sth convolutional layer output feature map F calculated using the SEWeight module S-1 The corresponding channel attention weight; (3) Use Softmax to normalize the multi-scale channel attention vector to obtain a new attention vector, as shown in formula (8): (4) The feature map F i And the attention vector att obtained after normalization i Multiply them in the channel dimension to obtain the feature map after multi-scale channel attention weighting, as shown in formula (9). Concatenate the above feature maps in the channel dimension to output a feature map with richer multi-scale information, as shown in formula (10): Y i =F i ⊙att i ,i=0,1,2,…,S-1 (9); Out=Cat(Y0,Y1,…,Y S-1 ) (10); Among them, Y i The feature map F i And the attention vector att obtained after normalization i The feature map after multi-scale channel attention weighting obtained by multiplying in the channel dimension.

Images