Aero-engine gas path fault diagnosis method based on multi-modal feature fusion

Abstract

The invention discloses a multi-modal feature fusion aero-engine gas path fault diagnosis method, and belongs to the technical field of aero-engine fault diagnosis and classification. According to the method, a multi-modal feature fusion aero-engine fault diagnosis model is designed, feature extraction is carried out in combination with an Informer network architecture, and a bidirectional long-short-term memory network (BiLSTM) is introduced before output, so that a bidirectional time sequence dependency relationship can be effectively captured, the capture capability of the model on a fault mode is further improved, and the fault diagnosis accuracy of the aero-engine is improved. Accurate recognition of aero-engine faults is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of aeroengine fault diagnosis and classification. Specifically, it relates to a diagnostic method for aeroengine component fault classification, and particularly to a diagnostic method for aeroengine gas path faults based on a deep learning multi-modal feature fusion model. Background Art

[0002] The present invention relates to the technical field of aeroengine fault diagnosis. Flight safety is the core guarantee factor for aircraft operation. Among them, as the key power device of the aircraft, the operating state of the aeroengine directly determines the safety performance and flight stability of the aircraft. As a complex pneumatic thermal system, the aeroengine is in an extreme working environment of high temperature, high pressure, high speed and high vibration for a long time, with a high failure rate and failure risk. These faults will not only reduce the overall performance indicators of the engine, but may also cause the aircraft to lose its normal flight ability in severe cases. Gas path component faults are the main inducement for major flight accidents such as in-flight shutdown.

[0003] During the whole life cycle of the aeroengine, performance degradation is an inevitable process. With the accumulation of the number of working cycles, under the action of complex gas path coupling effects, typical fault modes such as blade corrosion, fouling wear, tip clearance expansion and fuel injector blockage will gradually appear in the engine gas path system. These faults will cause the rapid decay of component performance parameters and may ultimately lead to serious system failures or even complete shutdown. Therefore, developing a high-timeliness and accurate aeroengine fault diagnosis method to achieve precise fault location has important engineering application value for improving the health management level of the engine and optimizing the maintenance strategy.

[0004] The aeroengine gas path fault diagnosis involved in the present invention mainly adopts the gas path analysis method. This method evaluates the performance changes of the main gas path components (such as efficiency, flow rate, effective area, etc.) by analyzing the data changes of sensors (monitoring parameters such as temperature, pressure, speed, etc.) arranged at different sections of the engine, so as to judge the fault mode of the engine. Currently, the aeroengine gas path fault diagnosis methods are mainly divided into two categories: physical model-based methods and data-driven methods.

[0005] The physical model-based method requires establishing an accurate digital model of the aeroengine, and requires in-depth understanding of the operating characteristics and structural functional characteristics of the engine. However, due to the complex nonlinearity and uncertainty of the engine, it is very difficult to establish an accurate mathematical model, and this method highly depends on the model accuracy. With the increasing complexity of the engine structure, the fault coupling phenomenon between components is more significant, and the physical model-based diagnostic method has great limitations in accurate modeling.

[0006] In contrast, with the continuous accumulation of engine monitoring data and the rapid development of computer technology, data-driven fault diagnosis methods have been widely applied in engineering and experiments. This method does not rely on the physical characteristics and thermodynamic cycles of the engine, and there is no need to establish a complex physical model. It can achieve effective fault diagnosis only by analyzing the engine monitoring data, with higher practicability and accuracy. Malhi et al. used a traditional RNN network in "Prognosis of Defect Propagation Based on Recurrent Neural Networks" to predict time series. In the face of the current situation where the amount of data is increasing and the model is becoming more complex, this research branch of deep learning has gradually developed. With the development of GPU computing power, deep learning-based methods have gradually become a reality and have more powerful performance than traditional shallow neural networks, thus becoming a research hotspot in the field of fault diagnosis. Zhao Hongli et al. proposed a method for fault diagnosis of aero engines based on the fusion of convolutional transformers in "Fault Diagnosis of Aero Engines Based on the Fusion of Convolutional Transformers". This method uses the self-attention mechanism to extract effective features and suppress redundant information. The research results show that this model can effectively improve the accuracy of fault recognition. However, in terms of fault diagnosis, the existing technical solutions are mainly limited to extracting fault features from one-dimensional time series data. This single-dimensional feature extraction method is difficult to comprehensively characterize the fault characteristics of aero engines. Considering the multi-dimensional and dynamic evolution characteristics of the degradation process of aero engines, it is necessary to construct a multi-dimensional feature extraction and analysis mechanism. However, the mainstream methods in the current field of aero engine fault diagnosis are still limited to single feature extraction, and this technical solution cannot achieve a comprehensive assessment and accurate diagnosis of the engine health status. Summary of the Invention

[0007] In order to fuse different features of aero engine fault data, a method for fault diagnosis of aero engines based on data-driven multi-modal feature fusion is proposed. This method designs a fault diagnosis model for aero engines with multi-modal feature fusion, as shown in the appendix Figure 1 shown, combines the Informer network architecture for feature extraction, and introduces a bidirectional long short-term memory network (BiLSTM) before the output, which can effectively capture bidirectional time series dependencies, further enhancing the model's ability to capture fault patterns and achieving accurate recognition of aero engine faults.

[0008] The technical solution of the present invention:

[0009] The method for diagnosing aero engine gas path faults with multi-modal feature fusion is as follows:

[0010] S1. Select the input variables of the fault diagnosis model and perform data preprocessing

[0011] First, according to different fault units of aero-engines, select the sensor parameters used for gas path fault diagnosis of turbofan engines as the input variables of the fault diagnosis model for fault diagnosis. The sensor parameter data is one-dimensional time-series fault data, including sampling time, features, and labels.

[0012] Before training the fault diagnosis model, it is necessary to perform normalization preprocessing on the features and labels of the one-dimensional time-series fault data.

[0013] Furthermore, the preprocessing uses the z-score method for normalization. The specific formula of z-score is:

[0014]

[0015] where, x i represents a value in the original data, n represents the number of original data, u is the mean of the original data, σ is the variance of the original data, and z i represents the value after normalization.

[0016] S2. Construct a fault diagnosis model, the specific steps are:

[0017] S2.1. Convert the preprocessed one-dimensional time-series fault data into two-dimensional images

[0018] Convert the preprocessed and normalized data into two types of two-dimensional images using Gramian Angular Field (GAF) and Recurrence Plot (RP) respectively.

[0019] Furthermore, the Gramian Angular Field (GAF) first encodes the values of the normalized canonical data as arccosine and encodes the sampling time as the radius to achieve the scaling of the time series. The specific expressions are:

[0020]

[0021] where, is the polar coordinate angle after normalizing the data, r is the radius, t i is the sampling timestamp corresponding to x i z i is the data after normalization preprocessing, Z i is the set of original data, and M is the span constant factor in the regularized polar coordinate system.

[0022] After converting the scaled time series into a polar coordinate system, from the perspective of angles, the time correlations within different time intervals are identified by considering the sum of the angles of the trigonometric functions and the angle sum and difference between each pair of points. The sum of the angles of the trigonometric functions and the angle difference between each pair of points are defined by the Gramian Angular Sum Field (GASF) and the Gramian Angular Difference Field (GADF) as follows:

[0023]

[0024] where j takes values from 1 to n;

[0025] The two-dimensional Gramian images are generated using the obtained Gramian Angular Sum Field (GASF) and Gramian Angular Difference Field (GADF).

[0026] Furthermore, the generation process of the two-dimensional recurrence plot of RP is as follows: First, using the RP method, the time domain space of the one-dimensional time series fault data obtained in step S1 is transformed into a phase space, and the distances between different data states are calculated to obtain the corresponding image features. For a given time series signal (x1 x2 x3...x n ), the appropriate embedding dimension m and delay time τ are determined, and then the time series is reconstructed. The trajectory extracted after reconstruction is:

[0027] v i =[x i x i+τ …x i+(m-1)τ (6)

[0028] Then the obtained recurrence plot can be expressed as:

[0029]

[0030] In the formula, ε is the threshold of the minimum distance; δ(·) is the Heaviside function.

[0031]

[0032] S2.2. Input the two two-dimensional images into the Informer network to extract features

[0033] The two two-dimensional images obtained through the Gramian Angular Field (GAF) and the recurrence plot (RP) are respectively input into the Informer network to extract features, obtaining GASF image features and RP image features; the normalized one-dimensional time series fault data is input into the GRU neural network to extract features, obtaining one-dimensional time series fault features.

[0034] S2.3. Fuse the one-dimensional time series fault features and the two-dimensional image features

[0035] Fuse the one-dimensional time-series fault features extracted by the GRU neural network in step 2.2, the GASF image features extracted by the Informer encoder, and the RP image features. Realize the fusion of time series, image space, and heterogeneous image features, and introduce BiLSTM before the output layer to enhance the fault recognition ability of the fault diagnosis model by capturing bidirectional time series dependencies.

[0036] S3. Design network evaluation metrics and apply the t-SNE algorithm for visualization

[0037] Train the fault diagnosis model constructed in step S2, and select 4 parameters to verify the performance of the fault diagnosis model, namely: Accuracy, Precision, Recall, and F1-Score.

[0038] Furthermore, the calculation formulas for the 4 parameters are as follows:

[0039]

[0040] Among them, TP i represents the number of cases where the true class and the model prediction result are the same; FP j represents the number of cases where the prediction result is this class and the true result is other classes; FN i represents the number of cases where the true result is this class and the prediction result is other classes. Accuracy represents the prediction accuracy of the model for all data; Precision represents the proportion of data that is truly this class among the results predicted by the model to be this class; Recall represents the proportion of data correctly predicted by the model among all results that are truly this class; F1-Score is a comprehensive evaluation metric that combines the results of Precision and Recall.

[0041] Use a visualization tool to visualize the classification results before and after the output, and more intuitively see the results before and after classification.

[0042] The present invention provides a multi-modal heterogeneous graph hybrid feature fusion method for aero-engine fault diagnosis, having the following significant advantages and technical effects:

[0043] First, the present invention innovatively converts aero-engine time-series fault data into GASF two-dimensional images and RP images respectively. Among them, GASF images are used to extract the global trend features of the data, and RP images are used to capture local non-linear features. The organic combination of the two not only enriches the feature space but also realizes the complementary fusion of global and local features, linear and non-linear features, thus significantly improving the comprehensiveness of feature expression.

[0044] Secondly, the present invention adopts a GRU network suitable for time series data analysis to effectively capture the dynamic change laws and trend characteristics in one-dimensional time series signals. By combining the GRU network and the Informer model in parallel, the collaborative analysis of time series features and spatial features is achieved, further enhancing the feature extraction ability of the model. In addition, a bidirectional long short-term memory network (BiLSTM) is introduced before the output layer, and by capturing bidirectional time series dependencies, the fault recognition accuracy and robustness of the model are significantly improved.

[0045] The experimental results show that compared with a single fault recognition network, the method proposed in the present invention significantly improves the recognition accuracy while effectively reducing the memory consumption and optimizing the calculation efficiency, thus comprehensively enhancing the fault diagnosis performance. This fully demonstrates the effectiveness and practicality of this method, providing a new technical reference and solution for the field of aeroengine fault diagnosis. Brief Description of the Drawings

[0046] Figure 1 It is a structural diagram of a multi-modal heterogeneous graph hybrid feature fusion network model

[0047] Figure 2 It is a structural diagram of an LSTM neural network.

[0048] Figure 3 It is a structural diagram of a GRU neural network

[0049] Figure 4 It is a structural diagram of an Informer neural network.

[0050] Figure 5 It is a structural diagram of an Informer EnconderStack.

[0051] Figure 6 It is a structural diagram of a BiLSTM neural network.

[0052] Figure 7 It is the loss function and accuracy curve of model training. Among them, (a) is the loss function curve of training, and (b) is the accuracy curve of training.

[0053] Figure 8 It is the t-SNE visualization before fault classification.

[0054] Figure 9 It is the t-SNE visualization after fault classification. Detailed Embodiments

[0055] The following further describes the detailed embodiments of the present invention in combination with the technical solutions and the drawings.

[0056] A fault recognition network model based on multi-modal heterogeneous graph hybrid feature fusion. The simulation verification is as follows:

[0057] In this patent, the dataset is generated by the continuous-time system model of an aero-engine. The faulty units of the aero-engine are set as fan fault, high-pressure compressor fault, combustor fault, high-pressure turbine fault, and low-pressure turbine fault.

[0058] Select the low-pressure rotor speed N1, high-pressure rotor speed N2, inlet total temperature T1, inlet total pressure P1, fan outlet total temperature T 21 , fan outlet total pressure P 21 , HPC inlet total temperature T 25 , HPC inlet total pressure P 25 , HPC outlet total temperature T3, HPC outlet total pressure P3, HPT outlet total temperature T 45 , HPT outlet total temperature total pressure P 45 , fuel flow rate W f as the input of the fault diagnosis model. And ten health indexes including efficiency index SE i and flow index SW i are used to characterize the gas path performance of the turbofan engine, where i = 0, 1, 2, 3, 4.

[0059] The sampling time of the sensor is 0.02 s, and a total of 8500 data pairs are collected. The division ratio of the training set, test set, and validation set is 7:2:1.

[0060] After normalizing the data, the data is transformed by the Gramian Angular Field (GAF), and the recurrence plot (RP) converts the time series into a two-dimensional image. Then, the gated recurrent unit (GRU) is used to extract the time series features, and Informer is used to extract the two-dimensional image features.

[0061] The recurrence plot (RP) is an important tool for visualizing the periodicity of trajectories through the phase space method. The main advantage of this method is that it can effectively retain the correlation information between the amplitude and phase in the original one-dimensional signal and enhance the distinguishability of fault features, thus significantly improving the expression ability of fault information.

[0062] As shown in the appendix Figure 3As shown in the figure, the gated recurrent unit (GRU), as an improved architecture of the recurrent neural network (RNN), can effectively solve the problems of gradient vanishing and gradient explosion in traditional RNNs. At the same time, it has excellent long-sequence correlation feature extraction capabilities. It has higher prediction accuracy compared to traditional RNNs and lower computational complexity compared to the long short-term memory network (LSTM). The core difference between GRU and LSTM lies in its more simplified internal structure, mainly by introducing the reset gate (Reset Gate) and the update gate (Update Gate) to control the information flow. Among them, the update gate combines the functions of the input gate and the forget gate in LSTM to adjust the update intensity of the previous hidden state at the current moment; the reset gate is used to control the update degree of the previous hidden state in the current candidate hidden state. The inputs of the above two gating structures are jointly determined by the input vector at the current moment and the hidden state at the previous moment, and their module outputs are calculated through the activation function of the fully connected layer.

[0063]

[0064] Among them, r t is the reset gate, and z t is the update gate. x t is the input vector at the current moment, and h t-1 is the state memory variable at the previous moment. Among them, the value range of r t is between [0, 1]. When it is close to 1, it means that past information is very important. h t ′ represents the updated value. And the value range of z t is also between [0, 1]. Being close to 1 means that more past information is retained. h t represents the candidate hidden state, and (I - z t ) × h t-1 means the selective discard of the original hidden state, and z t × h t ′ means the selective retention of the candidate hidden state of the current node. W r , W z , W h are their respective weight matrices, I represents the identity matrix, "×" represents matrix multiplication, and the activation functions used are the sigmoid and tanh functions. Their specific expressions are as follows:

[0065]

[0066] Informer is a multi-time series prediction model composed of multiple modules and is developed from Transformer. It includes an encoder (such as attached Figure 3shown), masked multi-head self-attention, convolutional layer, fully connected layer, probabilistic sparse self-attention mechanism, and adaptive global and local self-attention distillation mechanism. As shown in the appendix Figure 4 shown, the running efficiency of the Transformer model in long-term prediction tasks is not ideal. The Informer model replaces the full self-attention mechanism of the Transformer model by introducing a probabilistic sparse self-attention mechanism, avoiding the calculation of O(L) 2 time and space complexity, reducing the time and space complexity to O(LlogL). At the same time, the Informer model also proposes a self-attention distillation mechanism, avoiding the problem of excessive memory occupancy caused by multi-layer stacking.

[0067] Informer is a multi-time series prediction model composed of multiple modules, including an encoder (as shown in the appendix Figure 5 shown), masked multi-head self-attention, convolutional layer, fully connected layer, probabilistic sparse self-attention mechanism, and adaptive global and local self-attention distillation mechanism. The input form of the traditional self-attention mechanism is:

[0068]

[0069] where Q, K, and V are matrices composed of query vectors (queries), key vectors (keys), and value vectors (values) respectively, and d represents the feature dimension.

[0070] Next, the KL divergence can be used to measure the distance between these two distributions. The KL divergence (Kullback-Leibler divergence), also known as relative entropy, is a method to describe the difference between two probability distributions p(x) and q(x). The discrete KL divergence is defined as follows:

[0071]

[0072] where n is the number of sampling points, and x i are the sampling points of p(x) and q(x).

[0073] The extracted features are fused, and a BiLSTM is introduced before the output layer. By capturing bidirectional temporal dependencies, the fault recognition ability of the model is further improved.

[0074] The long short-term memory (LSTM) network adds memory cells that can store historical information in its structure, enabling selective utilization of this information, allowing the model to process long time series. The LSTM cell structure mainly consists of memory cells, forget gates, input gates, and output gates, effectively solving problems such as gradient vanishing and gradient explosion.

[0075] Among them, the hidden state h of the previous LSTM cell t-1 and the input value x of the sample at the current time t together constitute the input quantity of the current LSTM. The forget gate f t determines which information in the memory cell to forget, the input gate i t determines whether the current input information is written into the memory cell, and the output gate o t determines whether to use the output of the current memory cell as the output of the LSTM. The specific calculation formulas are as follows:

[0076]

[0077] Among them, W f , W i , W o are the weight matrices of the forget gate, input gate, and output gate respectively; b f , b i , b o are the bias vectors of the forget gate, input gate, and output gate respectively, σ is the sigmoid function, and tanh is the hyperbolic tangent function. C t is the state at time t, y t is the output at the current time t, is the state candidate value, W c is the weight of the state gates, and b c is the bias of the state gate.

[0078] However, LSTM can only learn the forward information of the time series and cannot utilize its backward information. BiLSTM improves LSTM. The structure diagram of the BiLSTM network is as shown in the appendix Figure 6 . The model learns the data information in two directions: forward in time and backward in time, and can obtain the output of the forward layer network and the output of the backward layer network respectively, which well solves this problem. BiLSTM combines the forward information and backward information of the data learned by the bidirectional network, fully utilizes the time order of the sequence, and obtains the final output result. BiLSTM can better deeply learn the long-term dependence relationship between features.

[0079] The experimental parameters are set as follows: the learning rate is 0.001. The training batch size selected each time is 32, the hidden state dimensions of GRU and Informer are 32, the number of GRU network layers is 2, and the number of multi-head attention heads of Informer is 8. At the same time, the dropout rate during each step of training of this neural network is set to 30%; in addition, we choose the Adam optimizer for the gradient descent algorithm of training the neural network. To further enhance the feature extraction ability of the input data. This network is built and trained on a computer with an Intel(R) Core(TM) i7-10700KF processor and an NVIDIA GeForce RTX 3090 GPU. The loss function and accuracy curve are as attached Figure 7 as shown. Use a visualization tool to visualize the classification results before and after output, and more intuitively see the results before and after classification. The T-SNE visualization diagrams before and after fault classification are as attached Figure 8 、attached Figure 9 as shown. t-Distributed Stochastic Neighbor Embedding (t-SNE) is a dimensionality reduction technique used to represent high-dimensional data sets in a low-dimensional space of two or three dimensions, so as to visualize them. To improve the interpretability of the model, apply the t-SNE interpretability algorithm to visualize the feature map; compare the sample distribution of the original samples and the samples after feature extraction by multi-modal fusion.

[0080] For verification, ablation experiments were conducted in this patent for comparison. Single-branch: time-series - GRU branch, GAF-Informer branch, RP-Informer branch for ablation experiments. And a Transformer neural network was selected for a comparative experiment on fault classification. Table 1 compares the estimation effects of different deep learning models on sensor faults. Among them, the accuracy (Accuracy), precision (Precision), recall (Recall), and F1-Score evaluation metrics were selected in Table 1, and the training time, accuracy, memory consumption, and total number of parameters were selected as evaluation metrics in Table 2:

[0081] Table 1 Comparison of evaluation metrics of the test data set in the ablation experiment

[0082]

[0083] Table 2 Comparison of performance metrics of the test data set of different deep learning networks

[0084]

[0085] It can be seen from the performance comparison in Table 1 that based on the same dataset, the accuracy rate of this feature fusion network is better than that of the single-branch network. It can be seen from the performance comparison in Table 2 that based on the same dataset, on the basis of ensuring the accuracy rate, the memory consumed by Informer is less than that of Transformer, and its comprehensive performance is the best.

Claims

1. A method for diagnosing a gas path fault of an aero-engine with multi-modal feature fusion, characterized in that, The steps are as follows: S1. Select the input variables of the fault diagnosis model and perform data preprocessing First, according to different fault units of the aero-engine, select the sensor parameters used for the gas path fault diagnosis of the turbofan engine as the input variables of the fault diagnosis model for fault diagnosis; The sensor parameter data is one-dimensional time-series fault data, including sampling time, features, and labels; Before training the fault diagnosis model, it is necessary to perform normalization preprocessing on the features and labels of the one-dimensional time-series fault data; S2. Construct the fault diagnosis model S2.

1. Convert the preprocessed one-dimensional time-series fault data into two-dimensional images Convert the preprocessed normalized data into two types of two-dimensional images using the Gramian Angular Field (GAF) and Recurrence Plot (RP) respectively; S2.

2. Input the two types of two-dimensional images into the Informer network to extract features Input the two types of two-dimensional images obtained through the Gramian Angular Field Recurrence Plot into the Informer network to extract features, obtaining GASF image features and RP image features; input the normalized one-dimensional time-series fault data into the GRU neural network to extract features, obtaining one-dimensional time-series fault features; S2.

3. Fuse the one-dimensional time-series fault features and two-dimensional image features Fuse the one-dimensional time-series fault features extracted by the GRU neural network in step 2.2, the GASF image features and RP image features extracted by the Informer encoder; Realize the fusion of time series, image space, and heterogeneous image features, and introduce BiLSTM before the output layer to enhance the fault recognition ability of the fault diagnosis model by capturing bidirectional time dependencies; S3. Design network evaluation metrics and apply the t-SNE algorithm for visualization.

2. The method for diagnosing a gas path fault of an aeroengine with multimodal feature fusion according to claim 1, wherein, The preprocessing is performed using the z-score method for normalization, and the specific formula of z-score is: where x i represents a value in the original data, n represents the number of the original data, u is the mean of the original data, σ is the variance of the original data, and z i represents the value after normalization.

3. The multi-modal feature fusion-based aero-engine gas path fault diagnosis method according to claim 1, wherein, The specific content of step S2.1 is as follows: The Gramian Angular Field first encodes the values of the normalized canonical data as arccosine and encodes the sampling time as the radius to achieve the scaling of the time series. The specific expressions are respectively: Among them, is the polar coordinate angle after normalizing the data, r is the radius, and t i is the i corresponding sampling timestamp, z i is the data after normalization preprocessing, and Z i is the set of original data, and M is the span constant factor in the regularized polar coordinate system; After converting the scaled time series into a polar coordinate system, use the angular perspective to identify the time correlations within different time intervals by considering the angular sum and angular sum and difference of the trigonometric functions between each point; the angular sum and angular sum and difference of the trigonometric functions between each point are defined by the Gramian Angular Sum Field and Gramian Angular Difference Field as follows: where j takes values from 1 to n; Generate a two-dimensional Gramian image using the obtained Gramian Angular Sum Field and Gramian Angular Difference Field; The generation process of the RP two-dimensional recurrence plot is as follows: First, using the RP method, the time domain space of the one-dimensional time series fault data obtained in step S1 is transformed into the phase space, and the distances between different data states are calculated to obtain the corresponding image features; for a given time series signal (x1x2x3...x n ), the appropriate embedding dimension m and delay time τ are determined, and then the time series is reconstructed. The trajectory extracted after reconstruction is: v i = [x i x i+τ …x i+(m-1)τ (6) The obtained recurrence plot can be expressed as: In the formula, ε is the threshold of the minimum distance; δ(·) is the Heaviside function; 4. The multi-modal feature fusion-based aero-engine gas path fault diagnosis method according to claim 1, wherein, The specific content of step S3 is as follows: Train the fault diagnosis model constructed in step S2, and select 4 parameters to verify the performance of the fault diagnosis model, namely: accuracy, precision, recall, and F1-Score; The calculation formulas of the 4 parameters are as follows: Among them, TP i represents the number of cases where the true category and the model prediction result are the same; FP j represents the number of cases where the prediction result is this category and the true result is other categories; FN i represents the number of cases where the true result is this category and the prediction result is other categories; Accuracy represents the prediction accuracy of the model for all data; Precision represents the proportion of data that is truly this category among the results predicted by the model to be this category; Recall represents the proportion of data correctly predicted by the model among all results that are truly this category; F1-Score is a comprehensive evaluation index that combines the results of Precision and Recall; Use a visualization tool to visualize the classification results before and after output, and more intuitively see the results before and after classification.

Images