Bolt group loosening positioning monitoring method based on deep learning and piezoelectric active sensing

Abstract

The present invention proposes a method for monitoring the loosening of bolt groups based on deep learning and piezoelectric active sensing. Based on the characteristics of piezoelectric sensors, it is proposed to use a sensor array to obtain the original piezoelectric signals of different bolt loosening of steel structure bolt groups, and construct a one-dimensional bolt group loosening index vector based on the energy value of three-layer wavelet decomposition as a representation of the loosening conditions of different bolts in the steel structure bolt group. Since convolutional neural networks have the characteristics of being able to retain and extract features in data and self-learning, the present invention establishes a mathematical model of loosening indicators and loose positions of different bolts in a bolt group based on the Lenet5‑CBAM deep learning model. The model can realize that the loosening indicators obtained according to the measured piezoelectric signals can directly determine the loose positions of steel structure bolt groups.

Description

Technical Field

[0001] The present invention relates to the field of health monitoring of bolted steel structures, and in particular to a method for monitoring the loosening of bolt groups based on deep learning and piezoelectric active sensing. Background Art

[0002] Bolt connection has the advantages of easy construction, removable and fatigue-resistant, and has become a major connection method for steel structures. However, due to fatigue, corrosion and other factors, the stress state of bolted connection parts of steel structures is very easy to change. Corresponding engineering practice has also proved that the damage of a single bolt in a bolt group will change the stress state of surrounding bolts. If the loose parts of the bolts are not found in time and replaced or repaired, it may have a significant impact on the structural performance.

[0003] This not only affects the normal service of the entire structure to a certain extent, but may even cause catastrophic accidents such as structural damage, resulting in significant economic losses and casualties. Therefore, real-time quantitative monitoring of the bolt connection status is of great engineering significance.

[0004] Traditional piezoelectric active sensing cannot locate the looseness of bolt groups in steel structures. Therefore, in order to solve the above problems, the present invention provides a bolt group looseness positioning monitoring method based on deep learning and piezoelectric active sensing. Deep learning and piezoelectric active sensing are applied to the monitoring of looseness of bolts in steel structure bolt groups, which can realize online monitoring of the loose position of bolts in steel structure bolt groups. Summary of the invention

[0005] In view of this, the present invention proposes a method for monitoring the loosening of bolt groups based on deep learning and piezoelectric active sensing. The piezoelectric active sensing method is used to obtain the changes in the structural nonlinear response signal and combined with the ability of deep learning to process big data, so as to perform real-time quantitative monitoring of the tightness of bolt connections in steel structures.

[0006] The technical solution of the present invention is implemented as follows: The present invention provides a method for monitoring the loosening of bolt groups based on deep learning and piezoelectric active sensing, comprising the following steps:

[0007] S1. Build a piezoelectric active sensing system, use piezoelectric active sensing to monitor the looseness of bolts in the bolt group, and obtain piezoelectric signals of different bolt looseness in the steel structure bolt group;

[0008] S2, calculating the wavelet decomposition energy value of the piezoelectric signal based on the wavelet packet decomposition method, and forming a one-dimensional looseness index vector with the wavelet energy values ​​calculated by the sensor array;

[0009] S3, building a one-dimensional convolutional neural network, and establishing the relationship between the bolt group tightness index vector and the loose position of the bolts in the bolt group through the one-dimensional convolutional neural network;

[0010] S4. Train and validate a one-dimensional convolutional neural network.

[0011] S5. Obtain the piezoelectric sensing signal of the current bolt and input it into the one-dimensional convolutional neural network to obtain the specific location of the loose bolt group.

[0012] On the basis of the above technical solution, preferably, S2 specifically includes the following steps:

[0013] S101, based on the collected piezoelectric sensor signal, calculating the three-layer wavelet packet decomposition into 8 sub-signals;

[0014] S102, calculating the wavelet packet decomposition signal energy component of the piezoelectric sensor signal and the wavelet packet decomposition sub-signal energy corresponding to each frequency band;

[0015] S103, calculating a one-dimensional looseness index vector of the piezoelectric sensor array signal according to the energy of the sub-signals decomposed by the wavelet packet;

[0016] S104, standardizing the one-dimensional looseness index vector data according to a standardization function; the standardization function is: Where E represents a one-dimensional looseness indicator vector; μ E represents the average energy of the sub-signal of the wavelet packet decomposition; σ E Represents the standard deviation of the energy of the sub-signals decomposed by the wavelet packet.

[0017] Based on the above technical solutions, preferably, the one-dimensional convolutional neural network in S3 includes: an input layer, a convolutional neural network Lenet5, a channel attention module, a spatial attention module and an output layer connected in sequence.

[0018] Based on the above technical solution, preferably, the convolutional neural network Lenet5 includes a convolutional layer 1, a ReLU layer, a maximum pooling layer, a convolutional layer 2, a ReLU layer and a pooling layer connected in sequence.

[0019] Based on the above technical solution, preferably, the calculation process of the channel attention module includes the following steps:

[0020] S201, channel attention mechanism extracts high-level features from the features extracted by the two convolutional layers and pooling layer of the convolutional neural network Lenet5 through maximum pooling and average pooling respectively;

[0021] S202, after passing through the shared multi-layer perceptron, the output elements are summed up, and then activated by the sigmoid function to finally generate the final channel attention weight;

[0022] S203: Perform a dot multiplication operation on the channel attention weight and the original input feature to generate the input feature required by the spatial attention mechanism.

[0023] Based on the above technical solution, preferably, the channel attention module is expressed as:

[0024] Among them, F represents the features extracted after the parallel convolution layer; Represents the feature value after global average pooling; Represents the feature value after global maximum pooling; W0 and W1 represent the parameters of the two layers in the multi-layer perceptron model respectively.

[0025] Based on the above technical solution, preferably, the calculation process of the spatial attention module includes the following steps:

[0026] S301, using the feature value output by the channel attention module as the feature map input by the spatial attention module, and performing global maximum pooling and average maximum pooling based on the channel;

[0027] S302, concat the two channels of the spatial attention module and the channel attention module, and then perform a convolution operation to reduce the dimension to a feature map with only one channel, and then generate the spatial attention weight through the sigmoid function, and finally perform a dot multiplication operation on the weight and the feature map output by the channel attention module to generate the final feature.

[0028] Based on the above technical solution, preferably, the spatial attention module is expressed as:

[0029]

[0030] Among them, M s represents the spatial attention weight, f 1*7 represents a 1*7 convolutional layer, Represents the feature value of the spatial attention module after global average pooling; Represents the feature value of the spatial attention module after global maximum pooling.

[0031] On the basis of the above technical solution, preferably, the neurons of the output layer are fully connected with the output layer of the spatial attention module, each neuron corresponds to the working condition of the loose bolts of the steel structure bolt group, and the softmax function is used as the activation function of the output layer;

[0032] The activation function is: In the formula, n and k represent the number of each neuron output, a n represents the output of the kth neuron in the output layer, K represents the number of output nodes, Softmax(an ) represents the output of the kth neuron after the activation function. The final output of the Lenet5-CBAM deep network is a 1*9 vector. Each value can be regarded as the confidence probability of the loosening of the bolts corresponding to the steel structure bolt group, and the tightness state corresponding to the maximum probability is taken as the result of fault diagnosis.

[0033] On the basis of the above technical solution, preferably, S4 specifically includes the following steps:

[0034] S401, repeatedly collect piezoelectric sensor signals under different tightness conditions of the bolt to obtain a large amount of data for training and verifying the model, and calculate the standardized multi-channel wavelet packet component energy according to step S2;

[0035] S402, dividing the standardized multi-channel wavelet packet component energy into training samples and test samples at a ratio of 8:2, using the training samples as the input of the one-dimensional convolutional neural network, and using the corresponding bolt loose position labels as the expected output of the dual-attention convolutional neural network model;

[0036] During the first training, the connection weights of each neuron in the one-dimensional convolutional neural network are set by random initialization, and the Adam gradient descent algorithm is used to update the connection weights of the one-dimensional convolutional neural network. Finally, the trained connection weights are saved to obtain a model for bolt loosening positioning of steel structure bolt groups.

[0037] S403, input the test sample into the one-dimensional convolutional neural network to test the generalization performance of the model. If the one-dimensional convolutional neural network meets the expected requirements, the one-dimensional convolutional neural network is saved; otherwise, the one-dimensional convolutional neural network is adjusted and the weight of the one-dimensional convolutional neural network is updated.

[0038] The bolt group loosening positioning monitoring method based on deep learning and piezoelectric active sensing of the present invention has the following beneficial effects compared with the prior art:

[0039] (1) The method based on piezoelectric active sensing is more sensitive to the loosening of steel structure bolts, and the energy based on wavelet packet decomposition can characterize the loosening state of bolts to a certain extent;

[0040] (2) Based on the characteristics of piezoelectric sensors, it is proposed to use a sensor array to obtain the original piezoelectric signals of different bolt loosening in a steel structure bolt group, and construct a one-dimensional bolt group loosening index vector based on the energy value of the three-layer wavelet decomposition as a representation of the different bolt loosening conditions of the steel structure bolt group. Since the convolutional neural network has the characteristics of being able to retain and extract features in the data and self-learning, the present invention establishes a mathematical model of loosening indicators and loose positions of different bolts in a bolt group based on the Lenet5-CBAM deep learning model. The model can realize that the loosening indicator obtained according to the measured piezoelectric signal can directly determine the loose position of the steel structure bolt group;

[0041] (3) Apply deep learning and piezoelectric active sensing to the quantitative monitoring and identification of the tightness of steel structure bolt connections. The principle is clear, the cost is low, the operation is simple and easy, and the identification is accurate. It can realize long-term online monitoring and identification of the tightness of steel structure bolts, thereby providing a scientific basis for the safety evaluation and life prediction of steel bolt connection nodes, and reducing the safety hazards and operational risks caused by loose bolt connections.

[0042] (4) The channel-space dual attention mechanism is introduced into the convolutional neural network to construct a lightweight and high-precision recognition network, which has a higher accuracy in identifying the tightness of bolts. By adding the spatial channel attention mechanism, the recognition effect of the convolutional neural network is improved, which improves the recognition ability of the model while improving the robustness of the model.

[0043] (5) A multi-channel monitoring scheme for bolt tightness is proposed. Based on the phenomenon of gravitational wave energy loss caused by bolt loosening, the wavelet packet component energy is used as the damage indicator. Combined with the powerful feature extraction and memory capabilities of the convolutional neural network, quantitative monitoring of bolt tightness is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 It is a flow chart of the bolt group loosening positioning monitoring method based on deep learning and piezoelectric active sensing of the present invention;

[0046] Figure 2 This is a flowchart of the Lenet5-CBAM model training in the present invention;

[0047] Figure 3 A one-dimensional convolutional neural network structure diagram in the present invention;

[0048] Figure 4 Schematic diagram of the channel space attention mechanism module in the present invention;

[0049] Figure 5 It is a structural diagram of the channel attention module and the spatial attention module in the present invention;

[0050] Figure 6 The accuracy curve and loss function diagram of the model training process and verification process in the present invention;

[0051] Figure 7The network model in the present invention is used to identify the effect diagram of the steel structure bolt torque and the confusion matrix diagram. DETAILED DESCRIPTION

[0052] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] In recent years, the main methods for monitoring bolt tightness include: detection methods based on structural vibration characteristics, detection methods based on piezoresistive impedance, and detection methods based on percussion signals. However, the existing methods can only detect whether the bolts are loose, but cannot determine the loose position. Therefore, in order to solve the above problems, Figure 1 As shown, this embodiment provides a bolt group loosening positioning monitoring method based on deep learning and piezoelectric active sensing, which includes the following steps:

[0054] S1. Build a piezoelectric active sensing system, use piezoelectric active sensing to monitor the looseness of bolts in the bolt group, and obtain piezoelectric signals of different bolt looseness in the steel structure bolt group;

[0055] The piezoelectric sensor used in this embodiment can be a piezoelectric ceramic, which has the characteristics of fast response, wide frequency range, easy cutting and low price, and is widely used in the field of engineering structure monitoring technology. Preferably, the piezoelectric sensor includes three piezoelectric ceramics, which are respectively installed on the left, top and right sides of the bolt connection part of the lower steel plate, wherein the installation position is not limited.

[0056] S2, calculating the wavelet decomposition energy value of the piezoelectric signal based on the wavelet packet decomposition method, and forming a one-dimensional looseness index vector with the wavelet energy values ​​calculated by the sensor array;

[0057] The specific steps include:

[0058] S101, based on the collected piezoelectric sensor signal, calculating the three-layer wavelet packet decomposition into 8 sub-signals;

[0059] The collected piezoelectric sensor signal is marked as X j , j represents the jth piezoelectric sensing signal;

[0060] The sub-signal components can be expressed as: X j =[X j,1 ,X j,2 …X j,m ](j=1,2,…,8), where m represents the number of signal points, X j,mRepresents the mth sub-signal component of the jth piezoelectric sensing signal.

[0061] S102, calculating the wavelet packet decomposition signal energy component of the piezoelectric sensor signal and the wavelet packet decomposition sub-signal energy corresponding to each frequency band;

[0062] According to the definition of signal energy, the energy of the sub-signal decomposed by wavelet packet can be expressed as: E j =X j,1 2 +X j,2 2 +…+X j,m 2 (j=1,2,…,8).

[0063] S103, calculating a one-dimensional looseness index vector of the piezoelectric sensor array signal according to the energy of the sub-signals decomposed by the wavelet packet;

[0064] In this embodiment, since three sensors are used, the wavelet base is db2, and the index used by the one-dimensional convolutional neural network is the combined value of the wavelet energy components of the three sensor signals, that is, it includes 24 wavelet energy values, so as to more comprehensively characterize the tightness of the bolt. Therefore, the final one-dimensional looseness index vector can be expressed as: E = [E1, ... E8, E9, ... E 16 ,E 17 ,…E 24 ].

[0065] S104, standardize the one-dimensional looseness index vector data according to a standardization function. The standardization function is: Where E represents a one-dimensional looseness indicator vector; μ E represents the average energy of the sub-signal of the wavelet packet decomposition; σ E Represents the standard deviation of the energy of the sub-signals decomposed by the wavelet packet.

[0066] S3, building a one-dimensional convolutional neural network, and establishing the relationship between the bolt group tightness index vector and the loose position of the bolts in the bolt group through the one-dimensional convolutional neural network;

[0067] In this embodiment, Figure 3 As shown in the figure, the one-dimensional convolutional neural network includes a sequentially connected input layer, a convolutional neural network Lenet5, a channel attention module, a spatial attention module and an output layer; each component is described in detail below:

[0068] In this embodiment, the convolutional neural network Lenet5 includes a convolutional layer 1, a ReLU layer, a maximum pooling layer, a convolutional layer 2, a ReLU layer, and a pooling layer, wherein the convolutional kernel size of the convolutional layer 1 is 1*3, the convolutional kernel size of the convolutional layer 2 is 1*5, the moving step size is 1, and the pooling layer adopts maximum pooling. The convolutional neural network Lenet5 has the advantages of simple structure, small number of parameters, fast training speed, etc., and has surpassed the traditional benchmark network model in many data sets.

[0069] The channel attention module and the spatial attention module obtain the characteristics of the loose bolt group.

[0070] The output layer has K=9 neurons, which are fully connected with the output layer of the spatial attention module. Each neuron corresponds to the loosening condition of the bolts in the steel structure bolt group, and the softmax function is used as the activation function of the output layer. The activation function is:

[0071] In the formula, n and k represent the number of each neuron output, a n represents the output of the kth neuron in the output layer, K represents the number of output nodes, Softmax(a n ) represents the output of the kth neuron after the activation function. The final output of the Lenet5-CBAM deep network is a 1*9 vector. Each value can be regarded as the confidence probability of the loosening of the bolts corresponding to the steel structure bolt group, and the tightness state corresponding to the maximum probability is taken as the result of fault diagnosis.

[0072] S4, train and verify the one-dimensional convolutional neural network; Figure 2 As shown, the specific steps include:

[0073] S401, repeatedly collect piezoelectric sensor signals under different tightness conditions of the bolt to obtain a large amount of data for training and verifying the model, and calculate the standardized multi-channel wavelet packet component energy according to step S2;

[0074] In this embodiment, 500 samples are sampled each time, and the wavelet decomposition energy vector of the original signal is obtained based on wavelet packet decomposition. A one-dimensional looseness index vector is constructed based on the piezoelectric sensor array. The index is converted into a numerical value with a mean of 0 and a variance of 1 using a standardized formula to speed up the iterative training of the neural network.

[0075] S402, dividing the standardized multi-channel wavelet packet component energy into training samples and test samples at a ratio of 8:2, using the training samples as the input of the one-dimensional convolutional neural network, and using the corresponding bolt loose position labels as the expected output of the dual-attention convolutional neural network model;

[0076] Among them, during the first training, the connection weights of each neuron in the one-dimensional convolutional neural network were set by random initialization, and the Adam gradient descent algorithm was used to update the connection weights of the one-dimensional convolutional neural network. The initial learning rate was set to 0.0001, and the number of training sample cycles was 400. The model was trained according to steps S1-S3 to obtain a model for locating loose bolts in steel structure bolt groups.

[0077] The one-dimensional looseness index vector obtained in step S401 is input into the deep learning model, and the model can output the corresponding 1*9 output vector, which corresponds to the predicted confidence probability of the 9 loose positions of the steel structure bolt group, where the bolt loose position corresponding to the maximum value is the actual bolt group loose position condition. In this way, the loose position of the bolts in the steel structure bolt group can be determined, providing a scientific basis for the replacement or repair of engineering management personnel.

[0078] S403, input the test sample into the one-dimensional convolutional neural network to test the generalization performance of the model. If the one-dimensional convolutional neural network meets the expected requirements, the one-dimensional convolutional neural network is saved; otherwise, the one-dimensional convolutional neural network is adjusted and the weight of the one-dimensional convolutional neural network is updated.

[0079] S5. Obtain the piezoelectric sensing signal of the current bolt and input it into the one-dimensional convolutional neural network to obtain the specific location of the loose bolt group.

[0080] The beneficial effects of this embodiment are as follows: the method based on piezoelectric active sensing is more sensitive to monitoring the loosening of bolts in steel structures, and the energy based on wavelet packet decomposition can characterize the loosening state of the bolts to a certain extent.

[0081] Based on the characteristics of piezoelectric sensors, the present invention proposes to use a sensor array to obtain the original piezoelectric signals of different bolt loosening of steel structure bolt groups, and construct a one-dimensional bolt group loosening index vector based on the energy value of three-layer wavelet decomposition as a representation of the different bolt loosening conditions of steel structure bolt groups. Since convolutional neural networks have the characteristics of being able to retain and extract features in data and self-learning, the present invention establishes a mathematical model of loosening indicators and loose positions of different bolts in a bolt group based on the Lenet5-CBAM deep learning model. The model can realize that the loosening indicators obtained according to the measured piezoelectric signals can directly determine the loose positions of steel structure bolt groups;

[0082] Applying deep learning and piezoelectric active sensing to the quantitative monitoring and identification of the tightness of steel structure bolt connections has a clear principle, low cost, simple operation, and accurate identification. It can achieve long-term online monitoring and identification of the tightness of steel structure bolts, thereby providing a scientific basis for the safety evaluation and life prediction of steel bolt connection nodes, and reducing safety hazards and operational risks caused by loose bolt connections.

[0083] The channel space dual attention mechanism is introduced into the convolutional neural network to construct a lightweight and high-precision recognition network, which has a higher accuracy in bolt tightness recognition.

[0084] A multi-channel monitoring scheme for monitoring the tightness of bolts is proposed. According to the phenomenon of gravitational wave energy loss caused by loose bolts, the wavelet packet component energy is used as the damage indicator. Combined with the powerful feature extraction and memory ability of convolutional neural networks, quantitative monitoring of bolt tightness is achieved.

[0085] Example 2

[0086] Based on Example 1, this example provides a specific calculation process of the channel attention module and the spatial attention module. Figure 5 .

[0087] Among them, the calculation process of the channel attention module includes the following steps:

[0088] S201, channel attention mechanism extracts high-level features from the features extracted by the two convolutional layers and pooling layer of Lenet5 through maximum pooling and average pooling respectively;

[0089] S202, after passing through the shared multi-layer perceptron, the output elements are summed up, and then activated by the sigmoid function to finally generate the final channel attention weight;

[0090] S203, perform a dot multiplication operation on the channel attention weight and the original input feature to generate the input feature required by the spatial attention mechanism. Finally, the channel attention mechanism can be expressed as:

[0091] Among them, M c represents the channel attention weight, and F represents the features extracted after the parallel convolution layer; Represents the feature value after global average pooling; Represents the feature value after global maximum pooling; W0 and W1 represent the parameters of the two layers in the multi-layer perceptron model respectively.

[0092] In addition, the calculation process of the spatial attention module includes the following steps:

[0093] S301, using the feature value output by the channel attention module as the feature map input by the spatial attention module, and performing global maximum pooling and average maximum pooling based on the channel;

[0094] S302, concat the two channels of the spatial attention module and the channel attention module, and then perform a convolution operation to reduce the dimension to a feature map with only one channel, and then generate the spatial attention weight through the sigmoid function, and finally perform a dot multiplication operation on the weight and the feature map output by the channel attention module to generate the final feature. Finally, the spatial attention mechanism can be expressed as:

[0095] Among them, f 1*7 represents a 1*7 convolutional layer, Represents the feature value of the spatial attention module after global average pooling; Represents the feature value of the spatial attention module after global maximum pooling.

[0096] The beneficial effect of this embodiment is that the recognition effect of the convolutional neural network is improved by adding a channel attention module and a spatial attention module. Because the attention mechanism can make the model more effectively focus on features that are helpful for the recognition effect, and the mechanism will automatically give a smaller weight to features that are not helpful for the model recognition effect, thereby improving the recognition ability of the model while improving the robustness of the model.

[0097] Example 3

[0098] On the basis of Example 2, in order to verify the effectiveness of Example 2, in this embodiment, each working condition is repeatedly collected 500 times, the loose index vector is calculated based on Example 2, and the training and verification sets are divided according to the principle of 8:2. Figure 2 The proposed training method trains a one-dimensional convolutional neural network, and determines the final model parameters according to the accuracy and loss value of the training set and the validation set. In order to finally test the performance of the model, 10 groups of piezoelectric signals are collected again under each working condition to obtain the loose index vector and input it into the constructed one-dimensional convolutional neural network, and the effectiveness of the present invention is tested according to the measured results. The training, validation, and test data sets are shown in Table 1:

[0099] Table 1

[0100]

[0101] The training set and validation set are input into the one-dimensional convolutional neural network respectively, such as Figure 6 As shown in FIG. 1 , after 400 iterations, the recognition accuracy of the one-dimensional convolutional neural network of Example 2 on the validation set can reach 100%, and the loss value is finally 0.025. The model recognition accuracy under this parameter is good. Figure 7As shown in the figure, the actual test results are expressed in the form of a confusion matrix. Except for the misdiagnosis rate of No. 4 and No. 6 bolts of 10%, the accuracy of bolt loosening judgment under other working conditions is 100%, and the overall loosening detection accuracy of the model is 98%.

[0102] In order to further verify the effectiveness of the introduced spatial channel attention module, we compared the recognition effects of the ordinary convolutional neural network and the one-dimensional convolutional neural network. As shown in Table 2, the accuracy of the model proposed in the present invention is higher than these models, which further proves the effectiveness of the model proposed in the present invention.

[0103] Table 2

[0104]

[0105] In this embodiment, the two convolutional neural network models of Lenet5 have the advantages of simple structure, small number of parameters, fast training speed, etc., and surpass the traditional benchmark network model in many data sets.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for monitoring the loosening of bolt groups based on deep learning and piezoelectric active sensing, characterized in that: The following steps are involved: S1. Build a piezoelectric active sensing system, use piezoelectric active sensing to monitor the looseness of bolts in the bolt group, and obtain piezoelectric signals of different bolt looseness in the steel structure bolt group; S2, calculating the wavelet decomposition energy value of the piezoelectric signal based on the wavelet packet decomposition method, and forming a one-dimensional looseness index vector with the wavelet energy values ​​calculated by the sensor array; S3, building a one-dimensional convolutional neural network, and establishing the relationship between the bolt group tightness index vector and the loose position of the bolts in the bolt group through the one-dimensional convolutional neural network; S4. Train and validate a one-dimensional convolutional neural network. S5, obtaining the piezoelectric sensing signal of the current bolt, and inputting it into a one-dimensional convolutional neural network to obtain the specific location where the bolt group is loose; The S2 specifically includes the following steps: S101, based on the collected piezoelectric sensor signal, calculating the three-layer wavelet packet decomposition into 8 sub-signals; S102, calculating the wavelet packet decomposition signal energy component of the piezoelectric sensor signal and the wavelet packet decomposition sub-signal energy corresponding to each frequency band; S103, calculating a one-dimensional looseness index vector of the piezoelectric sensor array signal according to the energy of the sub-signals decomposed by the wavelet packet; S104, standardizing the one-dimensional looseness index vector data according to a standardization function; the standardization function is: Where E represents a one-dimensional looseness indicator vector; μ E represents the average energy of the sub-signal of the wavelet packet decomposition; σ E Represents the standard deviation of the energy of the sub-signals decomposed by the wavelet packet; The one-dimensional convolutional neural network in S3 includes: an input layer, a convolutional neural network Lenet5, a channel attention module, a spatial attention module and an output layer connected in sequence; The calculation process of the channel attention module includes the following steps: S201, channel attention mechanism extracts high-level features from the features extracted by the two convolutional layers and pooling layer of the convolutional neural network Lenet5 through maximum pooling and average pooling respectively; S202, after passing through the shared multi-layer perceptron, the output elements are summed up, and then activated by the sigmoid function to finally generate the final channel attention weight; S203, performing a dot multiplication operation on the channel attention weight and the original input feature to generate the input feature required by the spatial attention mechanism; The S4 specifically includes the following steps: S401, repeatedly collect piezoelectric sensor signals under different tightness conditions of the bolt to obtain a large amount of data for training and verifying the model, and calculate the standardized multi-channel wavelet packet component energy according to step S2; S402, dividing the standardized multi-channel wavelet packet component energy into training samples and test samples at a ratio of 8:2, using the training samples as the input of the one-dimensional convolutional neural network, and using the corresponding bolt loose position labels as the expected output of the dual-attention convolutional neural network model; During the first training, the connection weights of each neuron in the one-dimensional convolutional neural network are set by random initialization, and the Adam gradient descent algorithm is used to update the connection weights of the one-dimensional convolutional neural network. Finally, the trained connection weights are saved to obtain a model for bolt loosening positioning of steel structure bolt groups. S403, input the test sample into the one-dimensional convolutional neural network to test the generalization performance of the model. If the one-dimensional convolutional neural network meets the expected requirements, the one-dimensional convolutional neural network is saved; otherwise, the one-dimensional convolutional neural network is adjusted and the weight of the one-dimensional convolutional neural network is updated.

2. The method for monitoring the loosening of bolt groups based on deep learning and piezoelectric active sensing as claimed in claim 1, characterized in that: The convolutional neural network Lenet5 includes a convolutional layer 1, a ReLU layer, a maximum pooling layer, a convolutional layer 2, a ReLU layer and a pooling layer which are connected in sequence.

3. The method for monitoring bolt group loosening positioning based on deep learning and piezoelectric active sensing as claimed in claim 1 is characterized in that: The channel attention module is expressed as: Among them, F represents the features extracted after the parallel convolution layer; Represents the feature value after global average pooling; Represents the feature value after global maximum pooling; W0 and W1 represent the parameters of the two layers in the multi-layer perceptron model respectively.

4. The method for monitoring bolt group loosening positioning based on deep learning and piezoelectric active sensing as claimed in claim 1 is characterized in that: The computation process of the spatial attention module includes the following steps: S301, using the feature value output by the channel attention module as the feature map input by the spatial attention module, and performing global maximum pooling and average maximum pooling based on the channel; S302, concat the two channels of the spatial attention module and the channel attention module, and then perform a convolution operation to reduce the dimension to a feature map with only one channel, and then generate the spatial attention weight through the sigmoid function, and finally perform a dot multiplication operation on the weight and the feature map output by the channel attention module to generate the final feature.

5. The method for monitoring bolt group loosening positioning based on deep learning and piezoelectric active sensing as claimed in claim 4 is characterized in that: The spatial attention module is expressed as: Among them, M s represents the spatial attention weight, f 1*7 represents a 1*7 convolutional layer, Represents the feature value of the spatial attention module after global average pooling; Represents the feature value of the spatial attention module after global maximum pooling.

6. The method for monitoring bolt group loosening positioning based on deep learning and piezoelectric active sensing as claimed in claim 2, characterized in that: The neurons of the output layer are fully connected with the output layer of the spatial attention module, each neuron corresponds to the working condition of the loose bolts of the steel structure bolt group, and the softmax function is used as the activation function of the output layer; The activation function is: In the formula, n and k represent the number of each neuron output, a n represents the output of the kth neuron in the output layer, K represents the number of output nodes, Softmax(a n ) represents the output of the kth neuron after the activation function. The final output of the Lenet5-CBAM deep network is a 1*9 vector. Each value can be regarded as the confidence probability of the loosening of the bolts corresponding to the steel structure bolt group, and the tightness state corresponding to the maximum probability is taken as the result of fault diagnosis.

Images