A tapping process tap breakage monitoring method and system based on an SSAE LSTM fusion network

Abstract

The application discloses a tapping process tap breakage monitoring method and system based on an SSAE-LSTM fusion network, and the method comprises the following steps: collecting a torque signal in a tapping process and classifying the torque signal into normal and breakage; extracting time domain, frequency domain and time-frequency domain features of the torque signal to form a feature vector set; fusing the feature vector by using an SSAE network, taking the fused feature vector as an input of an LSTM network to perform identification and calculate a loss function, reversely propagating the loss function to the LSTM network to optimize weight parameters, obtaining a monitoring input of the SSAE network according to the weight matrix optimized by the LSTM network, updating the loss function of the SSAE network and optimizing the weight of the loss function of the SSAE network, and forming an SSAE-LSTM fusion network model; training the model; and identifying a breakage state of the collected feature vector by using the SSAE-LSTM fusion network model to obtain a tap breakage monitoring result. The application fuses multiple feature domain features, effectively reduces data volume, enhances the correlation between features and categories, and improves the monitoring accuracy of the network.

Description

Technical Field

[0001] This invention belongs to the field of tap breakage monitoring technology in the tapping process, specifically relating to a tap breakage monitoring method based on SSAELSTM fusion network, and also relating to a tap breakage monitoring system based on SSAELSTM fusion network. Background Technology

[0002] Threaded connections are an important type of mechanical connection, widely used not only in daily life but also playing a crucial role in precision instruments, aerospace, and automotive industries. Tapping is a common method for machining internal threads, and it is usually the final step in part manufacturing; therefore, successful tapping is critical to the manufacturing process. However, in machining small-diameter internal threads (M5 and below), the tap breakage rate is very high. A broken tap requires machine shutdown and removal, and in severe cases, can even lead to the complete scrapping of the part. Therefore, research on tap condition monitoring during the tapping process has become an urgent problem to be solved.

[0003] Research on tap breakage monitoring is limited. This invention aims to achieve fault identification during the tapping process. Deep learning is widely used in related identification tasks, but existing deep learning methods suffer from limitations such as the need for manual selection in feature fusion, low intelligence, and the relative independence of feature fusion and fault identification. Fault identification cannot guide feature compression, leading to low accuracy in neural network recognition. Furthermore, feature extraction faces challenges, including difficulty in extracting features representing tap states and the reliance on manual selection for fault-related features. To address these issues, this invention proposes a novel deep learning-based fault identification method for the tapping process. It explores how to utilize the recognition results of a classification network to autonomously guide feature compression, eliminating manual selection and efficiently solving the problems of multi-feature autonomous compression and fault identification in the tapping process. This provides early warning information for the tapping process and offers a theoretical basis for intelligent tapping. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method and system for monitoring tap breakage during the tapping process based on SSAELSTM fusion network, which effectively solves the problems of difficulty in fusing features and low recognition accuracy during the tapping process of small-diameter internal threads.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: a tap breakage monitoring method based on SSAELSTM fusion network, the method comprising the following steps:

[0006] Step 1: Data Acquisition and Tap Status Classification: Acquire torque signals during the tapping process, and classify the tap status into two types based on the tapping mechanism and the acquired torque signals: normal and broken.

[0007] Step 2, Feature Vector Extraction: Extract the time-domain, frequency-domain, and time-frequency-domain features of the torque signal to form a feature vector set for the tapping process;

[0008] Step 3: Construction of the SSAELSTM Fusion Network Model: The feature vectors are fused using a stacked sparse autoencoder network (SSAE). The fused feature vectors are used as input to a long short-term memory network (LSTM) for identification and loss function calculation. The loss function is backpropagated to the LSTM network to optimize the weight parameters of the LSTM network. The backpropagation continues based on the optimized weight matrix of the LSTM network to obtain the monitoring input of the stacked sparse autoencoder network. The loss function of the stacked sparse autoencoder network is updated and the weights of the stacked sparse autoencoder network are optimized based on the loss function to form the SSAELSTM fusion network model.

[0009] Step 4: Train the model. The SSAELSTM fusion network model constructed in Step 3 is trained using the feature vector training set to obtain the optimal model.

[0010] Step 5: Monitoring and identification. The trained SSAELSTM fusion network model is used to identify the broken state of the collected feature vectors and obtain the tap breakage monitoring results.

[0011] Furthermore, the aforementioned Stacked Sparse Autoencoder (SSAE) network introduces sparsity into the hidden layer of the stacked encoder. After adding sparsity, the average activation of the hidden layer nodes is expressed as:

[0012] (5),

[0013] In the formula, , The total number of training samples. These represent the number of hidden layer nodes, h. j (x i ) is the input for the xth... i The activation level of the j-th hidden unit when there are n samples;

[0014] With the introduction of sparsity, the reconstruction error function of the stack encoder also needs to include a sparsity penalty term. Therefore, the reconstruction error is expressed as:

[0015] (6),

[0016] (7),

[0017] in, and These are the divergence parameter and the sparsity parameter, respectively. This represents the average activation level of the j-th hidden neuron; the sparsity parameter is characterized by the activation level. This represents the reconstruction error without the addition of a sparse penalty term. Indicates KL divergence;

[0018] Furthermore, the method for constructing the SSAELSTM fusion network model in step three above is as follows:

[0019] The loss function of the SSAELSTM fusion network model is the classification error E1, which is defined as:

[0020] (14)

[0021] In the formula, y al The marked tap condition; if the tap condition is normal, y al =0, if the tap breaks y al =1, y pl The category to be monitored; during training, the classification error E1 is continuously updated with the increase of the number of iterations and used to optimize the weight parameters of the SSAELSTM fusion network. In order to minimize the loss function of the constructed LSTM neural network, the weight matrix w of the Long Short Time Memory network is... l The update equation is:

[0022] (15)

[0023] in, Let q be the classification error of the short-term memory network during the qth iteration. Let be the learning rate at the q-th iteration, with an initial value of 0.0005. The weights at the q-th iteration are used. The classification error E1 is backpropagated to the Long Short-Term Memory (LSTM) network, and the weight parameters of the LSM network are optimized according to formula (15). And continue backpropagation to obtain the monitoring input of the stacked sparse autoencoder network. At this point, the loss function E2 is the loss function of the stacked sparse autoencoder network. The loss function E2 includes the error gradient propagated from the long short-term memory network and the reconstruction error of the encoding and decoding of the stacked sparse autoencoder network. Combining formulas (6)-(7), the loss function E2 is expressed as:

[0024] (16)

[0025] in, The monitored values ​​are obtained after the error is propagated to the stacked sparse autoencoder network. , This is the theoretical value. Indicates the first After the next iteration, the loss function E2 is updated to optimize the weight matrix w of the stacked sparse autoencoder network.s The stacked sparse autoencoder network has 2 hidden layers, meaning it consists of two encoder stacks. Therefore, the weight matrix w s The first encoder weight matrix w of SSAE a and the second encoder weight matrix w k Together they form a whole, and their update equation is:

[0026] ,

[0027] in, and represents the weight parameters and reconstruction error of the stacked sparse autoencoder network at the q-th iteration.

[0028] Furthermore, in step three above, the training model updates the parameters of the SSAE network through formula (17), and the error continues to propagate on the SSAELSTM fusion network. Training stops after reaching the set number of iterations or the error, and the optimal model of the SSAELSTM fusion network model under the current structure is obtained.

[0029] A tap breakage monitoring system based on SSAELSTM fusion network, comprising,

[0030] The data acquisition and tap state classification module is used to acquire torque signals during the tapping process and classify the tap state into two types: normal and broken, based on the tapping mechanism and the acquired torque signals.

[0031] The feature vector extraction module is used to extract the time domain, frequency domain, and time-frequency domain features of the torque signal to form a feature vector set for the tapping process;

[0032] The SSAELSTM fusion network model building module is used to fuse feature vectors using a stacked sparse autoencoder network (SSAE). The fused feature vectors are used as input to a long short-term memory network (LSTM) for identification and loss function calculation. The loss function is backpropagated to the LSTM network to optimize the weight parameters of the LSTM network. The backpropagation continues based on the optimized weight matrix of the LSTM network to obtain the monitoring input of the stacked sparse autoencoder network. The loss function of the stacked sparse autoencoder network is updated and the weights of the stacked sparse autoencoder network are optimized based on the loss function to form the SSAELSTM fusion network model.

[0033] The training model module is used to train the constructed SSAELSTM fusion network model using a feature vector training set to obtain the optimal model;

[0034] The identification and monitoring module is used to identify the breakage status of the collected feature vectors using a trained SSAELSTM fusion network model, and obtain the tap breakage monitoring results.

[0035] The beneficial effects of this invention are as follows: Compared with the prior art, the model constructed in this invention effectively reduces the amount of data, improves the correlation between categories, and makes feature fusion easier. The SSAELSTM fusion network improves accuracy and robustness, effectively solving the problems of high tap breakage rate, difficulty in feature extraction and fusion to characterize tap fracture, and low recognition accuracy during small-diameter internal thread tapping. The specific effects are as follows:

[0036] (1) Measure the torque during the tapping process, define the tapping state based on the measurement signal and tapping mechanism, extract the local mixed domain features of the torque signal, construct a set of categorized feature vectors, and provide a sample set for subsequent training and testing of the neural network;

[0037] (2) A new deep learning-based tap breakage detection method is proposed. The SSAE network in the fusion model can fuse features from multiple feature domains, effectively reducing the amount of data, enhancing the correlation between features and categories, and improving the network's detection accuracy.

[0038] (3) The proposed method uses the recognition results of the classification network to autonomously guide feature fusion, eliminating the need for manual selection, and efficiently solving the problems of autonomous fusion of multiple features and tap breakage monitoring, providing early warning to avoid tap breakage. Attached Figure Description

[0039] Figure 1 This is an overall flowchart of a tap breakage monitoring method based on SSAELSTM fusion network;

[0040] Figure 2 This is a diagram of the sparse stack autoencoder structure.

[0041] Figure 3 This is a diagram of the internal structure of an LSTM cell.

[0042] Figure 4 This is a diagram of the tap breakage monitoring network structure based on the SSAELSTM fusion network.

[0043] Figure 5 This serves as an experimental platform for monitoring tap breakage during the tapping process.

[0044] Figure 6 A comparison chart of experimental results showing tap breakage;

[0045] Figure 7 Force diagram illustrating the working principle of the tapping process;

[0046] Figure 8 This is a graph showing the torque variation during the tapping process;

[0047] Figure 9The training accuracy variation of SSAELSTM fusion networks with different numbers of SSAE network layers;

[0048] Figure 10 The iterative convergence curves of fusion networks with different numbers of SSAE network layers;

[0049] Figure 11 A graph showing the variation in training accuracy of the SSAELSTM fusion network with different numbers of output layer nodes in the SSAE.

[0050] Figure 12 Iterative convergence curves of SSAELSTM fusion networks with different numbers of SSAE output layer nodes;

[0051] Figure 13 The curve showing the change in training accuracy as a function of the number of neurons;

[0052] Figure 14 The iterative convergence curves of the SSAELSTM fusion network with different numbers of LSTM neurons;

[0053] Figure 15 The figures show the test results of the constructed model for tap breakage monitoring; in the figure, (a) the original torque signal of test experiment 1, (b) the original torque signal of test experiment 2, (d) the test results of test experiment 1, and (d) the test results of test experiment 2.

[0054] Figure 16 The visualization results are shown before and after feature fusion; in the figure, (a) are the original features, and (b) are the fused and compressed features.

[0055] Figure 17 A comparison chart of correlations after feature fusion from different models;

[0056] Figure 18 Figures show the tap breakage monitoring results for different models; in the figures, (a) is the torque signal of the test experiment, (b) is the prediction result based on the LSTM network, (c) is the prediction result based on the KAPCLSTM fusion network, (d) is the prediction structure based on the SSAECNN fusion network, and (e) is the prediction result based on the SSAELSTM fusion network.

[0057] Figure 19 A comparison chart of precision and recall;

[0058] Figure 20 A comparison chart of F1 scores;

[0059] Figure 21 The algorithm flow is based on the SSAELSTM fusion network. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0061] Before introducing the embodiments of this application, let me first introduce the inventive concept of this invention: Addressing the technical problem in the prior art of "difficulty in fusing tap breakage features and low recognition rate during tapping," this invention proposes a process of signal acquisition and tap state division, feature extraction, construction of a fusion network model, and tap breakage recognition. The fusion network model uses a stacked sparse autoencoder network (SSAE) to fuse feature vectors. The fused feature vectors are used as input to a long short-term memory network (LSTM) for recognition and loss function calculation. Backpropagation is performed using the loss function to first optimize the weight parameters of the long short-term memory network; then the LSTM... The layer gradient continues to propagate backward to the stacked sparse autoencoder network, updating the SSAE network loss function and completing its weight parameter iterative optimization to form the SSAELSTM fusion network model. The SSAELSTM fusion network model autonomously guides feature fusion using the recognition results of the classification network, eliminating the need for manual selection and efficiently solving the problems of autonomous fusion of multiple features and tap breakage detection, providing early warning to avoid tap breakage. In the SSAELSTM fusion network model, the SSAE network fusion model integrates features from multiple feature domains, effectively reducing the amount of data, enhancing the correlation between features and categories, and improving the network's monitoring accuracy. For a detailed description of the specific scheme, please refer to Examples 1-2.

[0062] Example 1: As Figure 1 As shown, a method for monitoring tap breakage during the tapping process based on a SSAELSTM fusion network is described. The method includes the following steps:

[0063] Step 1: Data Acquisition and Tap Status Classification: Acquire torque signals during the tapping process, and classify the tap status into two types based on the tapping mechanism and the torque signals obtained from the experiment: normal and broken.

[0064] Step 2, Feature Vector Extraction: Extract the time domain, frequency domain, and time-frequency domain features of the torque signal collected in Step 1 to form a feature vector set for the tapping process; based on the number of layers in the SSAE network and the number of neurons in the LSTM network, 31 different network models were constructed, and the constructed SSAELSTM fusion network was trained. The final tap breakage monitoring model was determined based on the training results.

[0065] Step 3: Constructing the SSAELSTM Fusion Network Model: The feature vectors are fused using a stacked sparse autoencoder network (SSAE). The fused feature vectors are then used as input to a long short-term memory (LSTM) network for identification and loss function calculation. The loss function is backpropagated to the LSTM network to optimize its weight parameters. The resulting gradient from the LSTM network is then backpropagated to the stacked sparse autoencoder network to obtain its monitoring input. The loss function of the stacked sparse autoencoder network is updated, and its weights are optimized, thus forming the SSAELSTM fusion network model.

[0066] Feature fusion principle based on SSAE: An autoencoder is a data dimensionality reduction algorithm that can compress high-dimensional input and convert it into a reconstructed vector. The process of high-dimensional input compression is called encoding, and the process of converting the compressed input into a reconstructed vector is called decoding. This encoding and decoding process can be used in feature fusion techniques to find more valuable latent features between mixed domains; encoding and decoding Z... The calculation process is as follows:

[0067] (1),

[0068] (2),

[0069] Where Z is the encoded feature, i.e. the fused feature, and X is the input feature (the time-domain and time-frequency domain features extracted from the torque signal, as shown in Table 1 and the wavelet part). These are the reconstructed features. and These are the weights and biases of the encoding part, respectively. and These are the weights and biases of the decoding part, respectively.

[0070] A single autoencoder can fuse lower-level features, while stacking multiple autoencoders to form a stacked autoencoder can fuse deeper features. For example... Figure 2 This is a schematic diagram of a neural network structure formed by stacking three autoencoders. Its goal is to continuously optimize the weights through training to minimize the reconstruction error.

[0071] The reconstruction error function of the SSAE network is:

[0072] (3),

[0073] (4),

[0074] Where m is the total number of training samples.

[0075] This type of stacked network suffers from excessively large hidden layer nodes. To address this issue, sparsity is introduced into the hidden layers, resulting in a network known as a stacked sparse autoencoder. The sparsity is achieved by using regularization to activate some neurons and deactivate others, effectively preventing the encoder from getting trapped in local minima and overfitting.

[0076] Stacked Sparse Autoencoder (SSAE) introduces sparsity into the hidden layers of a stacked encoder. The average activation of the hidden layer nodes after adding sparsity is expressed as:

[0077] (5),

[0078] In the formula, , The total number of training samples. These represent the number of hidden layer nodes, h. j (x i ) is the input for the xth... i The activation level of the j-th hidden unit when there are n samples;

[0079] With the introduction of sparsity, the reconstruction error function of the stack encoder also needs to include a sparsity penalty term. Therefore, the reconstruction error is expressed as:

[0080] (6),

[0081] (7),

[0082] in, and These are the divergence parameter and the sparsity parameter, respectively. This represents the average activation level of the j-th hidden neuron. This represents the reconstruction error without the addition of a sparse penalty term. The KL divergence is represented by this unique structure, which can effectively reduce the feature dimension and has good fusion performance. It has been widely used in feature fusion in mixed domains. Therefore, this application selects the SSAE network as the network for the feature fusion part, and uses the features extracted from the original signal as the input of the SSAE network to obtain the dimensionality-reduced features. This process eliminates redundant features and the time required for manual selection.

[0083] To achieve sparse representation in the stack encoder, a penalty factor is added to impose sparse constraints on the hidden layer neurons. KL divergence is chosen as the penalty factor, meaning a sparse penalty term needs to be added to the reconstruction error function. KL is the KL divergence.

[0084] After extracting multi-dimensional features from the torque signal during the tapping process, a feature vector set is constructed. This feature vector set is used as input to a fusion neural network and compressed through an SSAE network. The compressed features are then used as input to an LSTM network. The LSTM learns the relationship between the feature vector set and the categories. Based on the compressed temporal feature vector set, it determines the category-related information that needs to be retained. This information is then input to a fully connected layer, and a Softmax layer calculates the probability of belonging to different categories. The probability calculation method can be expressed as follows:

[0085]

[0086] In the formula:

[0087] The predicted probability for the category;

[0088] z i This is the output of the fully connected layer;

[0089] j=0,1 represents two categories: normal and broken.

[0090] The classification and recognition principle based on LSTM: Long Short-Term Memory (LSTM) network is a special type of recurrent neural network. The structure of an LSTM network consists of multiple cells, and the internal structure of each cell is as follows: Figure 3 As shown, a unique gating mechanism is used to control the degree of information retention in the input unit. The gates are divided into three types: "forget gate", "input gate" and "output gate".

[0091] (1) Forget gate: This part uses the sigmoid function to process the input x at time t. t The calculation is performed using the input at time t-1 to determine the information that needs to be filtered and the information that needs to be retained. The specific calculation process is as follows:

[0092] (8),

[0093] Among them, w f b is the weight matrix of the forget gate; f Let be the bias matrix of the forget gate. This represents the sigmoid function. The output is represented by t-1;

[0094] (2) Input gate: The "input gate" processes the input information through a sigmoid layer and a tanh layer to obtain i. t and ∆c t This determines the input information that needs to be retained.

[0095] (9),

[0096] (10)

[0097] Among them, w i and w c b is the weight matrix of the input gate; i and b c This is the bias matrix of the input gate;

[0098] (3) Update of cell state: The update of cell state is based on the cell information c from the previous time step. t-1 and the input information at the current moment Decision. A new unit state c is obtained after two parts of linear computation. t :

[0099] (11),

[0100] (4) Output gate: The final output of this unit is determined by the sigmoid function and the tanh function:

[0101] (12)

[0102] (13)

[0103] Among them, w o b is the weight matrix of the output gate; o This is the bias matrix for the output gate. The unique gating mechanism of LSTM networks can capture long-term dependencies. Breakage detection during the tapping process cannot predict the next stage's category solely based on the previous stage's category; instead, it requires capturing the dependencies within the previous time-series signals to predict the next stage's category. This aligns perfectly with the characteristics of LSTM neural networks, hence this application chose LSTM as the network for breakage detection. Multiple LSTM units, fully connected layers, and Softmax layers together constitute a classification and recognition model. This model identifies the features fused from the SSAE network, monitors whether the tap will break, and provides early warning for the tapping process.

[0104] A Tap Breakage Detection Model Based on SSAELSTM Fusion Network Model is Established: Traditional feature selection methods involve first extracting a large number of features from the signal, then evaluating the correlation between features and categories, and finally selecting the feature with the strongest correlation. This method is time-consuming, has uncertain correlation, and low recognition accuracy. This application combines the advantages of SSAE and LSTM neural networks, proposing an innovative tap breakage recognition method based on an SSAELSTM fusion network. This method can autonomously fuse features and identify the tap breakage state. The SSAE network fuses the input multi-feature domain features, and the fusion result is input into the LSTM network to identify the tap state. The classification error E1 is calculated based on the recognition result of the LSTM neural network, and the error E2 of the SSAE network is obtained through backpropagation. The feature fusion of the SSAE network is optimized, and the optimal network model is obtained through iteration, making the fused features have more obvious fault characteristics and higher recognition accuracy. The proposed network structure and algorithm for tap breakage detection are as follows: Figure 4 and 21 As shown.

[0105] The error between the output class probability and the true class is calculated. The loss function of the SSAELSTM fusion network model is the classification error E1, which is defined as follows:

[0106] (14)

[0107] In the formula, y al The marked tap condition; if the tap condition is normal, y al =0, if the tap breaks y al =1, y pl The category to be monitored; during training, the classification error E1 is continuously updated with the increase of the number of iterations and used to optimize the weight parameters of the SSAELSTM fusion network. The LSTM recognition network consists of LSTM layers, fully connected layers, and softmax layers. In order to minimize the loss function of the constructed LSTM neural network, the weight matrix w l The weight matrix w of LSTM L and the fully connected layer matrix w q Together they form the weight matrix w l The update equation is:

[0108] (15)

[0109] in, Let q be the classification error of the short-term memory network during the qth iteration. Let be the learning rate at the q-th iteration, with an initial value of 0.0005. The weights at the q-th iteration are used. The classification error E1 is backpropagated to the Long Short-Term Memory (LSTM) network, and the weight parameters of the LSM network are optimized according to formula (15). After further backpropagation, the monitoring input of the stacked sparse autoencoder network is obtained. At this point, the updated loss function E2 is calculated. Loss function E2 is the loss function of the stacked sparse autoencoder network. Loss function E2 includes the error gradient propagated from the long short-term memory network and the reconstruction error of the encoding and decoding of the stacked sparse autoencoder network. Combining formulas (6)-(7), loss function E2 is expressed as:

[0110] (16)

[0111] in, The monitored values ​​are obtained after the error is propagated to the stacked sparse autoencoder network. , This is the theoretical value. Indicates the first After the next iteration, E2 is updated to optimize the weight matrix w of the stacked sparse autoencoder network. s Weight matrix w s Composed of the weight matrices of each encoder in SSAE, Figure 4 In this context, the stacked sparse autoencoder network has 2 hidden layers, meaning it consists of two encoder stacks. Therefore, the weight matrix w... s The first encoder weight matrix w of SSAE a and the second encoder weight matrix w k Together they form a whole, and their update equation is:

[0112] ,

[0113] in, and For the stacked sparse autoencoder network weight parameters and reconstruction error at the qth iteration;

[0114] The parameters of the SSAE network are updated by formula (17). The error continues to propagate on the SSAELSTM fusion network. Training stops when the set number of iterations or error is reached, and the optimal model under the current structure is obtained.

[0115] Step 4: Train the model. The SSAELSTM fusion network model constructed in Step 3 is trained using the feature vector training set to obtain the optimal model.

[0116] The training model updates the parameters of the SSAE network using formula (17). The error propagates continuously on the SSAELSTM fusion network. Training stops after reaching the set number of iterations or the error, thus obtaining the optimal model of the SSAELSTM fusion network under the current structure. This application obtains multi-domain features and categories by analyzing the torque signal measured during the tapping process, constructs a set of categorized feature vectors, and uses them as input to the stacked sparse autoencoder-based long short-term memory (SSAELSTM) fusion network. The performance of fusion models with different structures is compared and analyzed to obtain the optimal tap breakage recognition model. Figure 21 As shown, the steps of the tapping process fault identification algorithm based on the SSAELSTM fusion network are as follows:

[0117] (1) Preprocess the torque signal of the tapping process, extract the multidimensional features of the torque signal of the tapping process, and construct a training set and a test set with categories;

[0118] (2) Initialize the parameters of the fusion neural network, and define the number of network layers, output layers and neurons of the SSAE network;

[0119] (3) The training set data is input into the SSAELSTM fusion network for classification prediction, and the classification error E1 is calculated according to equation (14);

[0120] (4) Update the weight parameter matrix wl of the LSTM network according to equation (15) using the backpropagation algorithm;

[0121] (5) Backpropagate to the SSAE network and calculate the error E2 according to equation (16);

[0122] (6) Update the weight parameter matrix ws of the SSAE network according to equation (17);

[0123] (7) Determine whether the model has converged. If the model has converged, training ends and the model is saved. If the model has not converged, repeat steps (3)-(6) to continue iterating.

[0124] (8) In order to verify the recognition performance of the trained neural network model, the test set is input into the trained SSAELSTM fusion network model to predict the category of each input data in the test set. If the accuracy does not meet the requirements, repeat steps (3)-(6) to adjust the neural network. If the requirements are met, the training is completed.

[0125] Step 5: Monitoring and identification. The trained SSAELSTM fusion network model is used to identify the broken state of the collected feature vectors and obtain the tap breakage monitoring results.

[0126] Example 2: A tap breakage monitoring system based on SSAELSTM fusion network, comprising,

[0127] The data acquisition and tap state classification module is used to acquire torque signals during the tapping process and classify the tap state into two types: normal and broken, based on the tapping mechanism and the acquired torque signals.

[0128] The feature vector extraction module is used to extract the time domain, frequency domain, and time-frequency domain features of the torque signal to form a feature vector set for the tapping process;

[0129] The SSAELSTM fusion network model building module is used to fuse feature vectors using a stacked sparse autoencoder network (SSAE). The fused feature vectors are used as input to a long short-term memory network (LSTM) for identification and loss function calculation. The loss function is backpropagated to the LSTM network to optimize the weight parameters of the LSTM network. The backpropagation continues based on the optimized weight matrix of the LSTM network to obtain the monitoring input of the stacked sparse autoencoder network. The loss function of the stacked sparse autoencoder network is updated and the weights of the stacked sparse autoencoder network are optimized based on the loss function to form the SSAELSTM fusion network model.

[0130] The training model module is used to train the constructed SSAELSTM fusion network model using a feature vector training set to obtain the optimal model;

[0131] The identification and monitoring module is used to identify the breakage status of the collected feature vectors using a trained SSAELSTM fusion network model, and obtain the tap breakage monitoring results.

[0132] To illustrate the effectiveness of the method of the present invention, the following embodiments were performed for verification:

[0133] 1. Acquisition of torque signals during tapping process and construction of sample set for fracture monitoring

[0134] 1.1 Experimental equipment for signal acquisition

[0135] This application classifies the tap state during the tapping process by combining experiments and tapping mechanism analysis: First, a tapping experiment is conducted to obtain torque signals, and a system is built as follows... Figure 5 The tap breakage monitoring experimental platform shown depicts a tapping process where tapping is performed using a tapping tool holder connected to a CNC machine tool. A torque sensor (Kistler 9271A) is fixed to the machine tool's worktable using a fixture, and the workpiece is fixed to the torque sensor. The torque signal is collected and transmitted to the signal acquisition software (Vib'SYS) on the computer via the torque sensor and a signal acquisition instrument (WS-5921). The machining tool is an M3 high-speed steel straight flute tap, the tapped hole is a 2.6mm through hole, the workpiece dimensions are 50 mm × 50 mm × 10 mm, the material is 304 stainless steel, and the sampling frequency is 1000 Hz.

[0136] 1.2 Classification of Tap Breakage Conditions

[0137] Torque signals during tap breakage and normal operation were obtained through tapping experiments. The workpiece after tap breakage was observed using a Keyence microscope. The results are as follows: Figure 6 As shown, from Figure 6 The broken tap photograph shows severe chip blockage when the tap broke. The second set of chips completely blocked the hole, while the first and third sets were not completely blocked, with some chip grooves still allowing light to pass through, but there were still a lot of chips. The corresponding torque signal during the tapping process is as follows: Figure 6 As shown, observation of the torque signal obtained from the experiment revealed that although tap breakage is an instantaneous event, the breakage of the tap follows a pattern, and its breakage is closely related to the torsion of the tap during the tapping process.

[0138] The torque signal data shows that significant and drastic fluctuations in torque occurred before the tap broke. Figure 6 The image of the broken tap shows a large amount of chips in the chip flute when the tap breaks. A torsional analysis of the tap, based on the tapping mechanism, reveals that the tap consists of cutting teeth and straightening teeth. The tap rotates in a self-guided manner to tap, with each cutting tooth bearing a portion of the cutting amount. After all the cutting teeth have completed their cutting, the straightening teeth refine the tooth profile. Due to the large number of cutting teeth, the entire process is complex and variable; the forces acting on the tap during tapping are as follows... Figure 7 As shown, the torque of each cutting tooth is determined by the cutting torque M. c and friction torque M f The composition involves the gradual increase in the number of cutting teeth during the cutting process, and the cutting torque M... c and friction torque M f As the cutting edge penetrates fully, the cutting torque remains unchanged; only the frictional torque changes. The frictional torque can be divided into the contact friction between the tool and the machined surface and the friction between the cutting teeth and the chips. The frictional torque M... f The torque should increase slowly as the cutting progresses; however, in actual machining, since the entire cutting process is a relatively closed process, the formation and removal of chips are random, resulting in uncertain friction torque and thus unstable torque signal.

[0139] like Figure 8 The figure shows the torque signal changes of a tap from its entry to its breakage. The area highlighted by the dashed line is the region where the torque fluctuates. The signal within the green dashed line represents a relatively short fluctuation. From a mechanistic perspective, this phenomenon is caused by chips accumulating in the chip flutes, creating significant resistance and making tapping difficult. However, the machining process does not stop; the tap continues to rotate. Under the combined action of cutting force and resistance, the tap exhibits torsional vibration, releasing it from its locked state and allowing some chips to be discharged smoothly. The frictional torque M...f The torque signal decreases, and the tap continues to rotate and tap normally, with a normal signal and no further fluctuations. The red dotted box represents the torque signal fluctuation area before the tap broke. This fluctuation occurs because the tap is constantly in a torsional vibration phase to counteract chip resistance. After prolonged torsional vibration, the excessive resistance cannot be resolved, ultimately leading to tap breakage. After the tap breaks, the torque signal no longer changes.

[0140] To ensure real-time monitoring, a 0.5s signal interval is defined. Fluctuations shorter than 0.5s, as shown in the green dashed box in the figure, pose a risk of tap breakage, but the risk is low. To reduce unnecessary losses, this low-risk or no-risk phase is defined as the normal phase. Fluctuations longer than 0.5s, as shown in the red dashed box in the figure, pose a high risk of tap breakage. Defining only the 0.5s torque signal before tap breakage as the breakage phase would reduce reaction time and make timely processing difficult. Therefore, this application defines both the high-risk torque signal fluctuation phase and the 0.5s torque signal before tap breakage as the breakage phase. As shown in Table 1, the duration of the breakage phase varies for each signal group. Defining the breakage phase for each group based on its specific time reduces misjudgments in subsequent identification and improves identification accuracy.

[0141] 1.3 Torque signal feature extraction and sample set construction

[0142] Due to the imbalance of class samples, a sliding interval is used for value selection during the anomaly phase. To maintain consistency with the signal during the normal phase, the interval length is set to 500, and the window type is a rectangular window, whose expression is:

[0143] (18)

[0144] Where u(n) is the Heaviside function, and n slides through values ​​in increments of 100, meaning there is a T-value between adjacent samples. j The signal obtained from the tapping experiment was divided into 987 samples after an overlap of 0.4 s, and features were extracted from them. This application extracted a total of 36 features, including time-domain, frequency-domain, and time-frequency-domain features, reflecting the characteristics of the signal from different perspectives. Sixteen features were extracted from the time domain, including peak value, variance, and RMS value; twelve features were extracted from the frequency domain, including root-mean-square frequency, variance frequency, and root-mean-square frequency. The calculation methods for the time-domain and frequency-domain features are shown in Table 1.

[0145] Table 1. Time-domain and frequency-domain characteristics

[0146]

[0147] In the time-frequency domain, wavelet packet decomposition is selected to process the signal. The decomposition involves three layers of wavelet packets, with the wavelet function being DB1. This yields two... 3 The energy values ​​of each wavelet packet. The energy of each frequency band of the wavelet packet can be expressed as:

[0148] (19)

[0149] Among them, S j For the reconstructed wavelets of each frequency band, the 36-dimensional feature vectors after feature extraction, together with the normal (0) and broken (1) states divided in the previous part, constitute a set of categorized feature vectors for subsequent training and learning of the neural network.

[0150] Twenty tapping signals were obtained from the tapping experiment, including 10 signals of tap breakage and 10 signals of tap success. After feature extraction, the total number of samples was 987. The training set consisted of 7 breakage signals and 7 success signals, with a total of 635 samples. The test set consisted of 3 breakage signals and 3 success signals, with a total of 352 samples.

[0151] 2. Determining the optimal structure of the SSAELSTM fusion network model

[0152] The network model constructed in this application is a fusion of the SSAE network and the LSTM network. Therefore, the construction of the network structure is divided into two parts: SSAE network structure design and LSTM network structure design. The number of hidden layers and nodes in the SSAE network and the number of neurons in the LSTM network have a significant impact on the recognition accuracy. Therefore, this application constructs different network model structures from these three aspects, and determines the final structure based on performance indicators such as training accuracy, testing accuracy, and training time. The method for determining the optimal network structure is as follows: the LSTM network structure remains unchanged while the SSAE network structure is adjusted to determine the initial structure based on performance indicators. Then, the LSTM network structure is adjusted to find the final optimal network structure.

[0153] 2.1 Determining the Number of Layers in the SSAE Network

[0154] Based on the characteristics of the greedy layer-by-layer mechanism of the SSAE network, the number of nodes must be designed in a layer-by-layer decreasing manner. Changes in the number of layers and the number of output nodes will cause changes in the number of nodes. Therefore, the initial setting is to set the number of neurons in the LSTM network to 70, the number of output nodes in the SSAE network to 6, and the number of layers in the SSAE network to 2-6. Based on the above, five network models are constructed as shown in Table 2.

[0155] Table 2. Fusion model structure with different numbers of SSAE network layers

[0156]

[0157] The parameters are set as follows: iteration count is 1000, expected error is 0.005, sparsity is 6, sparsity ratio is 0.05, learning rate is 0.0005, minimum training batch size is 4, and the optimizer is Adam. The constructed network is then trained using the training set. Figure 9 It represents the variation in training accuracy of fusion network models with different numbers of SSAE network layers, from Figure 9 As can be seen, when the number of hidden layers is 6, the model accuracy is poor, and the overall performance is... Figure 10 The accuracy variation curves for 2-5 hidden layers show that the model is unstable when the number of hidden layers is 4 and 5. Comparing the performance of the fusion model with different numbers of SSAE network layers in Table 3, and considering the training accuracy, training time, and test accuracy results, 3 hidden layers is the optimal choice. Therefore, the number of network layers for SSAE is determined to be 3.

[0158] Table 3 Performance comparison results of fusion network models with different numbers of SSAE network layers

[0159]

[0160] 2.2 Determining the Number of Nodes in the SSAE Output Layer

[0161] Based on the characteristics of the SSAE network, the number of nodes in each layer will vary with the number of nodes in the output layer and the number of layers in the SSAE network. The previous part studied the influence of the number of network layers. This part mainly studies the number of nodes in the output layer. When the number of LSTM neurons is 70 and the number of hidden layers is 3, the range of the number of nodes in the output layer is set to 1-10, and 10 different fusion models are constructed as shown in Table 4.

[0162] Table 4. Different structures of SSAELSTM fusion models with different numbers of SSAE output layer nodes.

[0163]

[0164] Figure 11 This is the curve showing the change in accuracy of the fusion network model as the number of output layer nodes changes. Figure 11 As can be seen, when the number of output nodes is 1, the network model does not converge. When the number of output layer nodes is 3, 7, and 8, the training accuracy of the model is relatively high, all above 95%. Figure 12 The accuracy iteration curves of these three models are shown. The model converges fastest when the number of nodes is 3. The more output nodes there are, the longer the training time. Considering the test accuracy and training time in Table 5, the output layer of the SSAE network has 3 nodes.

[0165] Table 5. Performance Comparison Results of SSAELSTM Fusion Network Models with Different Number of SSAE Network Layers

[0166]

[0167] 2.3 Determining the Number of LSTM Neurons

[0168] After determining the SSAE network structure, the structure of the LSTM network must be determined; the number of LSTM neurons was varied from 10 to 180, and 18 different network models were built and trained respectively. Figure 13 The curve showing the change in accuracy as the number of neurons changes is presented, from... Figure 13 As can be seen, the training accuracy fluctuates with the number of neurons, reaching its maximum when the number of neurons is 90, and remaining almost unchanged after 150.

[0169] The training accuracy iteration curve of a network model with a training accuracy of over 97% is as follows: Figure 14 As shown, from Figure 14 As can be seen, when the number of neurons is 90, the training accuracy converges the fastest and is significantly higher than other models, reaching 98.02%. The performance comparison results are shown in Table 6. The comparison shows that although the training accuracy of the five structures is not much different, the test accuracy is significantly different. Models H2N9O3 and H2N10O3 have higher test accuracy. Considering the overall training time, the network structure of model 11 is selected as the final network model.

[0170] Table 6. Performance comparison results of SSAELSTM fusion networks with different numbers of LSTM neurons.

[0171]

[0172] By constructing network model structures with different numbers of SSAE network layers, nodes, and LSTM network neurons, training and testing were conducted on different structures. The training accuracy, testing accuracy, and training time were compared and analyzed. The optimal network structure for the tap breakage monitoring model based on the SSAELSTM fusion network was determined to be 36-25-14-3-90-2.

[0173] 3. Results Analysis and Discussion Based on SSAELSTM Fusion Network

[0174] 3.1 Analysis of Tap Breakage Monitoring Results Based on SSAELSTM Fusion Network

[0175] The constructed test feature vector set was input into the optimal network structure for fracture detection, achieving a test accuracy of 91.92%. The monitoring results of two sets of fracture signals selected from the test set are as follows: Figure 15As shown, the torque signal in Test Experiment 1 exhibited multiple short-duration torsional vibrations, which could easily lead to misjudgments. However, the predicted categories effectively avoided aliasing, and all categories were successfully identified. For the torque signal in Test Experiment 2, the misjudgment occurred in the later stages and only happened once. In real production, if continuous warnings were issued, measures would be taken to avoid them, so this misjudgment would have minimal impact. Therefore, the method proposed in this application can essentially monitor the high-risk breakage intervals during the tapping process and provide early warning information.

[0176] 3.2 Feature Fusion Performance Analysis Based on SSAELSTM Fusion Network

[0177] To evaluate the performance of feature fusion, Spearman's rank correlation coefficient was introduced to assess the correlation between features and categories. A correlation coefficient closer to 1 indicates a stronger correlation; a positive coefficient represents a positive correlation, and a negative coefficient represents a negative correlation. The correlation coefficients between features and categories before and after feature fusion were calculated, and the comparative analysis results are shown in Table 7. The maximum correlation coefficient after fusion increased from 0.81 to 0.89, and the minimum correlation coefficient significantly increased from 0.02 to 0.85. The data volume decreased from 33558 to 2952. The maximum correlation coefficient between features and categories before fusion was 0.81, and the minimum correlation coefficient after fusion was 0.85. The correlation between features and categories after fusion was significantly improved, and the data volume was effectively reduced, resulting in higher computational efficiency.

[0178] Table 7 Comparison Analysis Before and After Feature Fusion

[0179]

[0180] To visualize the performance of feature fusion, the nonlinear algorithm t-Distributed Stochastic Neighbor Embedding (t-SNE) was used to perform cluster analysis on the features before and after fusion. The same parameters were used to visualize the 36-dimensional features before fusion and the 3-dimensional features after fusion in a 2D space, as shown below. Figure 16 As shown, the original 36-dimensional feature clustering results are mostly mixed and difficult to distinguish. The clustering results after feature fusion clearly show two categories. Although there is some overlap, these categories can be distinguished in the subsequent neural network. This result indicates that the fused features have more obvious category information and are easier to identify.

[0181] To determine whether the constructed network model can effectively guide feature fusion and improve the correlation between fused features and categories, a network model that does not optimize the feature fusion process through classification results, namely the SSAE-LSTM network, was constructed. The SSAE-LSTM network structure is the same as the SSAELSTM fusion network structure constructed above. The feature fusion results of the two network models are output, and the Spearman's rank correlation coefficient between the fused features and categories is calculated. The correlation between the two models is compared as follows: Figure 17 As shown, from Figure 17 As can be seen, the features fused by the SSAE-LSTM fusion network model are positively correlated with the category, with a correlation coefficient close to 1, which is much larger than that of the SSAE-LSTM network model. This indicates that the correlation between the fused features obtained through the autonomous guidance of the classification network and the category is significantly enhanced, and the fusion network can effectively guide the feature fusion network.

[0182] 3.3 Comparative Analysis of Monitoring Performance Based on SSAELSTM Fusion Network

[0183] To determine whether the constructed network model can effectively improve the accuracy of tap breakage recognition, the tap breakage recognition model based on the SSAELSTM fusion network was compared with other methods. Other methods included LSTM, Kernel Principal Component Analysis Long Short-Term Memory Neural Network (KPCALSTM), and Stacked Sparse Autoencoder Convolutional Network (SSAECNN). The results are shown in Table 8. The comparison shows that the proposed SSAELSTM fusion model outperforms other network structures in both training and testing accuracy.

[0184] Table 8 shows the comparison results between the constructed model and other models.

[0185]

[0186] These four tap breakage monitoring and identification models were used to... Figure 18 Monitoring the torque signal from Figure 18As can be seen, the LSTM network identifies many normal breaks as breakage stages, resulting in severe aliasing. The KPCALSTM network model has a high misclassification rate in the defined tap breakage stage. Comparative analysis of Figures (c) and (e) shows that the SSAELSTM fusion network's recognition performance is significantly better than the KPCALSTM network, indicating that the SSAE network's fusion effect is superior to that of KPCA. Figures (d) and (e) show the monitoring results of the SSAECNN fusion network and the SSAELSTM fusion network, respectively. The SSAECNN fusion network's misclassification occurs in the defined tap breakage stage, indicating a high risk, while the SSAELSTM fusion network's monitoring results are accurate, with no misclassifications. This shows that the LSTM network's recognition performance is significantly better than that of the CNN network. Comparative experiments demonstrate that the SSAELSTM fusion model has a significant advantage in tap breakage monitoring.

[0187] Figure 19 For precision and recall results, from Figure 19 As can be seen, the LSTM fusion network model has the highest precision, followed by the SSAELSTM fusion network model, and the SSAECNN model has the lowest precision. Conversely, the SSAELSTM fusion network model has the highest recall, followed by the SAECNN fusion network model, and the LSTM model has the lowest recall. Considering both precision and recall, the SSAELSTM fusion network model has a clear advantage. Figure 20 By comparing the F1 scores of different models, it can be seen that the F1 score of the SSAELSTM fusion network model is superior to that of other network models. Therefore, the network model proposed in this application, which optimizes the feature selection process based on fault identification results, can improve the accuracy of tap breakage identification and its performance is significantly better than other models. This method does not rely on manual selection, eliminates human intervention, and to a certain extent solves the problem of difficult feature extraction and fusion for characterizing tap breakage. It effectively monitors the tap breakage state and provides a possible solution for intelligent tapping processes.

[0188] In summary, the advantages of the method in this application are as follows:

[0189] (1) The tap breakage monitoring method based on SSAELSTM fusion network built in this application realizes autonomous feature fusion. The feature fusion process of SSAE neural network is optimized by the recognition result of LSTM neural network, which provides guidance for the feature fusion process, eliminates human interference, and efficiently solves the problems of multi-feature domain fusion, autonomous feature fusion and state recognition.

[0190] (2) Based on the torque variation law and tapping mechanism in the tapping experiment, the analysis of the torque signal and broken wire images in the tapping process revealed that before the tap breaks, a large amount of chips accumulate, the tap exhibits torsional vibration, the torque signal fluctuates violently, and the tap is prone to breakage. The tap state is divided by the fluctuation time, providing a classification label for tap breakage monitoring. 36 feature vectors of the multiple feature domains of the torque signal are extracted to form a labeled feature vector set.

[0191] (3) By constructing 31 different network models, the optimal network model structure was obtained through training and testing: 34-25-14-3-90-2. Compared with the features before fusion, the features after fusion of the constructed model effectively improved the correlation between the features and the categories. Compared with the existing monitoring methods, the proposed model effectively solved the problems of difficult feature fusion and low recognition rate. It predicted the interval with high risk of breakage during the tapping process, provided breakage warning information for the tapping process, and laid the foundation for realizing intelligent tapping process.

[0192] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the claims.

Claims

1. A method for monitoring tap breakage during tapping based on a SSAELSTM fusion network, characterized in that, The method includes the following steps: Step 1: Data Acquisition and Tap Status Classification: Acquire torque signals during the tapping process, and classify the tap status into two types based on the tapping mechanism and the acquired torque signals: normal and broken. Step 2, Feature Vector Extraction: Extract the time-domain, frequency-domain, and time-frequency-domain features of the torque signal to form a feature vector set for the tapping process; Step 3: Constructing the SSAELSTM Fusion Network Model: The feature vectors are fused using a stacked sparse autoencoder network. The fused feature vectors are used as input to the long short-term memory network for identification and loss function calculation. The loss function is backpropagated to the long short-term memory network to optimize its weight parameters. The optimized weight matrix of the long short-term memory network is then backpropagated to obtain the monitoring input of the stacked sparse autoencoder network. The loss function of the stacked sparse autoencoder network is updated, and its weights are optimized based on this loss function to form the SSAELSTM fusion network model. Step 4: Train the model. The SSAELSTM fusion network model constructed in Step 3 is trained using the feature vector training set to obtain the optimal model. Step 5: Identification and monitoring. The trained SSAELSTM fusion network model is used to identify the broken state of the collected feature vectors and obtain the tap breakage monitoring results.

2. The tap breakage monitoring method based on SSAELSTM fusion network according to claim 1, characterized in that, The stack sparse autoencoder network introduces sparsity into the hidden layers of the stack encoder. After adding sparsity, the average activation of the hidden layer nodes is expressed as: （5）， In the formula, , The total number of training samples. These represent the number of hidden layer nodes, h. j (x i ) is the input for the xth... i The activation level of the j-th hidden unit for each sample; With the introduction of sparsity, the reconstruction error function of the stack encoder also needs to include a sparsity penalty term. Therefore, the reconstruction error is expressed as: （6）， （7）， in, and These are the divergence parameter and the sparsity parameter, respectively. This represents the average activation level of the j-th hidden layer node. This represents the reconstruction error without the addition of a sparse penalty term. This represents the KL divergence.

3. The tap breakage monitoring method based on SSAELSTM fusion network according to claim 2, characterized in that, The method for constructing the SSAELSTM fusion network model in step three is as follows: The loss function of the SSAELSTM fusion network model is the classification error E1, which is defined as follows: （14）， In the formula, y al The marked tap condition; if the tap condition is normal, y al =0, if the tap breaks y al =1, y pl The category to be monitored; during training, the classification error E1 is continuously updated with the increase of the number of iterations and used to optimize the weight parameters of the SSAELSTM fusion network and the weight matrix w of the Long Short Time Memory network. l The update equation is: （15）， in, Let q be the classification error of the short-term memory network during the qth iteration. Let be the learning rate at the q-th iteration. The weights at the q-th iteration are used. The classification error E1 is backpropagated to the Long Short-Term Memory (LSTM) network, and the weight parameters of the LSM network are optimized according to formula (15). And continue backpropagation to obtain the monitoring input of the stacked sparse autoencoder network. At this point, the loss function E2 is the loss function of the stacked sparse autoencoder network. The loss function E2 includes the error gradient propagated from the long short-term memory network and the reconstruction error of the encoding and decoding of the stacked sparse autoencoder network. Combining formulas (6)-(7), the loss function E2 is expressed as: （16）， in, The monitored values ​​are obtained after the error is propagated to the stacked sparse autoencoder network. , This is the theoretical value. Indicates the first After the next iteration, the loss function E2 is updated to optimize the weight matrix w of the stacked sparse autoencoder network. s The stacked sparse autoencoder network has 2 hidden layers, meaning it consists of two encoder stacks. Therefore, the weight matrix w s The update equation is composed of the weight matrices of the first and second encoders of the stacked sparse autoencoder network: ， in, and represents the weight parameters and reconstruction error of the stacked sparse autoencoder network at the q-th iteration.

4. The tap breakage monitoring method based on SSAELSTM fusion network according to claim 3, characterized in that, In step three, the training model updates the parameters of the SSAE network using formula (17). The error continues to propagate on the SSAELSTM fusion network. Training stops when the set number of iterations or error is reached, and the optimal model of the SSAELSTM fusion network under the current structure is obtained.

5. A tap breakage monitoring system based on SSAELSTM fusion network during tapping process, characterized in that, include, The data acquisition and tap state classification module is used to acquire torque signals during the tapping process and classify the tap state into two types: normal and broken, based on the tapping mechanism and the acquired torque signals. The feature vector extraction module is used to extract the time domain, frequency domain, and time-frequency domain features of the torque signal to form a feature vector set for the tapping process; The SSAELSTM fusion network model building module is used to fuse feature vectors using a stacked sparse autoencoder network. The fused feature vectors are used as input to a long short-term memory network for identification and loss function calculation. The loss function is backpropagated to the long short-term memory network to optimize its weight parameters. The optimized weight matrix of the long short-term memory network is then backpropagated to obtain the monitoring input of the stacked sparse autoencoder network. The loss function of the stacked sparse autoencoder network is updated, and its weights are optimized based on this loss function to form the SSAELSTM fusion network model. The training model module is used to train the constructed SSAELSTM fusion network model using a feature vector training set to obtain the optimal model; The identification and monitoring module is used to identify the breakage status of the collected feature vectors using a trained SSAELSTM fusion network model, and obtain the tap breakage monitoring results.

Images