Deep learning-based shock positioning method, device, equipment and medium

Abstract

The present application relates to the technical field of structural health monitoring, and particularly relates to an impact positioning method and device based on deep learning, equipment and a medium. The method comprises: acquiring impact signals monitored by a plurality of monitoring points in a target area; inputting the impact signals into a trained neural network model to obtain distance values from an impact source generating the impact signals to each monitoring point; determining existence trajectories formed by possible existence positions of the impact source relative to each monitoring point according to the distance values; eliminating interference points in all intersection points between the existence trajectories, and determining a center point of the remaining intersection points as a final center point; and outputting an impact positioning result by taking the final center point as an actual position of the impact source. The present application can be applied to various complex engineering structures, and ensures real-time performance and accuracy of engineering structure health monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structural health monitoring, in particular to an impact positioning method and device based on deep learning, equipment and a medium. BACKGROUND

[0002] Engineering structures may be subjected to sudden, low-speed, heavy-load and other impacts during service, which are likely to cause invisible matrix damage inside the structure. The damage will gradually accumulate, thereby bringing potential safety hazards to the structure. Therefore, it is necessary to perform structural health monitoring on the engineering structure: extracting parameter signals reflecting the characteristics of the structure from the engineering structure, and using a reasonable and effective information processing method to extract structural damage and aging information from the collected original data, to monitor the load borne by the structure and the operation status of the structure, and to provide a reference for the safe use and maintenance of the structure.

[0003] In a long-term research and practice of the prior art, the inventors of the present application found that the structural health monitoring means in the prior art either periodically, closely and offline uses ultrasonic, eddy current and the like to investigate the internal damage caused by the impact, or arranges a sensor array around the target area, collects signals before and after the impact occurs, and performs calculation and processing based on a certain physical model. The former reduces the use time of the structure and increases the maintenance cost, while the latter model needs to know the sound wave propagation parameters in each direction of the structure to be measured in advance, and the positioning effect has a large dependence on the experience of researchers, and it is difficult to be effectively applied in the field of structural health monitoring of engineering structures. SUMMARY

[0004] Based on the problems and shortcomings in the prior art, the present application provides an impact positioning method, device, equipment and medium based on deep learning, which can be applied to various complex engineering structures and ensure the real-time and accuracy of the structural health monitoring of the engineering structures.

[0005] One embodiment of the present application provides an impact positioning method based on deep learning, comprising:

[0006] obtaining impact signals monitored by a plurality of monitoring points in a target area;

[0007] inputting the impact signals into a trained neural network model to obtain distance values from an impact source generating the impact signals to each of the monitoring points;

[0008] determining existence trajectories formed by possible existing positions of the impact source relative to each of the monitoring points according to the distance values;

[0009] eliminating interference points among all intersection points of the existence trajectories, and determining a center point of the remaining intersection points as a final center point;

[0010] The final center point is taken as the actual position of the impact source, and an impact positioning result is output.

[0011] Optionally, interference points among all intersection points between the existing trajectories are removed, and a center point of the remaining intersection points is determined as a final center point, comprising:

[0012] A center point corresponding to an existing intersection point between the existing trajectories is determined.

[0013] An intersection point farthest from the center point is removed, and a new center point corresponding to the remaining intersection points is determined.

[0014] An intersection point farthest from the center point is regarded as an interference point and removed, and a new center point corresponding to the remaining intersection points is determined.

[0015] The above step is repeatedly executed until half of the interference points among all the intersection points are removed.

[0016] A center point corresponding to the remaining half of the intersection points is taken as a final center point.

[0017] Optionally, the existing trajectory formed by the possible existing positions of the impact source relative to each of the monitoring points according to the distance values comprises:

[0018] A circular trajectory is formed with each of the monitoring points as a center and with a corresponding distance value as a radius, and the circular trajectory is taken as the existing trajectory of the impact source relative to the monitoring point.

[0019] Optionally, the neural network model is a one-dimensional deep residual neural network, and the impact signal is a one-dimensional time domain signal.

[0020] Optionally, the one-dimensional deep residual neural network comprises, in sequence, a convolutional layer, a batch normalization layer, a first rectified linear unit, a maximum pooling layer, a plurality of average pooling layers, a first full connection layer, a second rectified linear unit, and a second full connection layer.

[0021] A connection branch is configured between the maximum pooling layer and each of the average pooling layers, and each connection branch contains a plurality of residual blocks connected in sequence.

[0022] Optionally, the residual block is a two-layer identity residual block or a three-layer identity residual block.

[0023] Optionally, the target area is arranged with a sensor array, and a position of an ultrasonic sensor of the sensor array is the monitoring point.

[0024] Based on the same inventive concept, the present application also provides a training method of a neural network model, comprising:

[0025] acquire impact signals monitored by a plurality of monitoring points in a target area;

[0026] input the impact signals into a neural network model to be trained to obtain distance values of an impact source generating the impact signals to each of the monitoring points as output results;

[0027] calculate a loss value of the output results and corresponding training labels by a loss function, the training labels being distance values of actual positions of the impact source to each of the monitoring points;

[0028] input the loss value as input, and optimize the neural network model by using an Adam optimizer and a back propagation algorithm;

[0029] repeat the above steps until the obtained loss value is less than a preset threshold.

[0030] Based on the same inventive concept, the present application also provides an impact positioning device based on deep learning, comprising:

[0031] a first acquisition module configured to acquire impact signals monitored by a plurality of monitoring points in a target area;

[0032] a first input module configured to input the impact signals into a trained neural network model to obtain distance values of an impact source generating the impact signals to each of the monitoring points;

[0033] a first calculation module configured to determine existence trajectories formed by possible existence positions of the impact source relative to each of the monitoring points according to the distance values;

[0034] a second calculation module configured to eliminate interference points among all intersection points of the existence trajectories, and determine a center point of the remaining intersection points as a final center point;

[0035] an output module configured to output an impact positioning result by taking the final center point as an actual position of the impact source.

[0036] Based on the same inventive concept, the present application also provides a training device of a neural network model, comprising:

[0037] a second acquisition module configured to acquire impact signals monitored by a plurality of monitoring points in a target area;

[0038] a second input module configured to input the impact signals into a neural network model to be trained to obtain distance values of an impact source generating the impact signals to each of the monitoring points as output results;

[0039] a third calculation module configured to calculate a loss value of the output result and a corresponding training label by using a loss function, the training label being a distance value from an actual position of the impact source to each of the monitoring points;

[0040] an optimization module configured to input the loss value into the neural network model and optimize the neural network model by using an Adam optimizer and a back propagation algorithm;

[0041] a judgment module configured to determine whether the loss value is less than a preset threshold, and stop the optimization of the neural network model when the loss value is less than the preset threshold.

[0042] Based on the same inventive concept, the present application further provides a computer device, comprising a processor, a memory and a computer program stored in the memory, the processor being coupled to the memory, and the processor being configured to execute the computer program to implement the above-mentioned impact positioning method based on deep learning, or the training method of the neural network model.

[0043] Based on the same inventive concept, the present application further provides a computer readable storage medium, the computer storage medium storing computer instructions, and the computer instructions being configured to make the computer execute the above-mentioned impact positioning method based on deep learning, or the training method of the neural network model when the computer instructions are executed by the computer.

[0044] One of the above-mentioned technical solutions has the following advantages and beneficial effects:

[0045] According to the embodiments of the present application, the multi-channel impact signals monitored by the plurality of monitoring points in the target region are processed by using the trained neural network model, the distance values from the impact source generating the impact signals to each of the monitoring points are output, the existence trajectories formed by the possible existence positions of the impact source relative to each of the monitoring points are determined according to the distance values, the interference points in all intersection points between the existence trajectories are eliminated, and the center point of the remaining intersection points is determined as the final center point, i.e. the actual position of the impact source. Based on this, the present application does not need to know the structural attributes and parameters in advance, can be applied to various complex engineering structures, and ensures the real-time performance and accuracy of the engineering structure health monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0046] The embodiments will be described in conjunction with the accompanying drawings. The drawings of the present application are only used for describing the embodiments and for the purpose of demonstration. Those skilled in the art can easily make other embodiments according to the steps described below without departing from the principles of the present application.

[0047] Figure 1 FIG. 1 is a flowchart of an impact positioning method based on deep learning according to an embodiment of the present application;

[0048] Figure 2 a structural schematic diagram of a neural network model in an embodiment of the present application;

[0049] Figure 3 a structural schematic diagram of a neural network model in an embodiment of the present application;

[0050] Figure 4 a flow schematic diagram of a shock positioning method based on deep learning in an embodiment of the present application;

[0051] Figure 5 a flow schematic diagram of a shock positioning method based on deep learning in an embodiment of the present application; Figure 1 or Figure 4 a schematic diagram of a central point clustering process corresponding to step S400 in the method;

[0052] Figure 6 a flow schematic diagram of a training method of a neural network model in an embodiment of the present application;

[0053] Figure 7 a block diagram of a shock positioning system capable of applying a shock positioning method based on deep learning and a training method of a neural network model in an embodiment of the present application;

[0054] Figure 8 a structural schematic diagram of a shock positioning device based on deep learning in an embodiment of the present application;

[0055] Figure 9 a structural schematic diagram of a training device of a neural network model in an embodiment of the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. It can be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application. In addition, it should be noted that, for the convenience of description, only parts related to the present application are shown in the drawings, but not all structures. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] The terms "first", "second", and the like in the present application are used to distinguish different objects, but not to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units not listed or optionally includes other steps or units inherent to the process, method, product or device.

[0058] Reference to“an embodiment” herein means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase“in one embodiment” in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily all referring to a common set of embodiments, although they can. As used herein, the term“or” as used herein, without further qualification, can be used to indicate exclusive- or inclusive-or. It will be apparent to one skilled in the art which

[0059] As shown in FIG. 1, one embodiment of the present application provides a deep learning based impact location method, including steps S100-S500. Figure 1

[0060] Step S100: Obtain impact signals monitored by a plurality of monitoring points in a target region.

[0061] In one embodiment, the target region is arranged with a sensor array, and the positions of the ultrasonic sensors of the sensor array are the monitoring points.

[0062] In order to effectively monitor the structural health of a target engineering structure, one or more detection regions need to be determined as target regions first. By arranging a sensor array around the target region, signals before impact and impact signals after impact can be collected.

[0063] The sensor array has different geometric designs, including linear, circular, planar, cylindrical, or spherical arrays. The advantage of using a sensor array instead of a single sensor is that the sensor array adds a new dimension to the observation, which helps to estimate more parameters and improve the estimation performance.

[0064] In one embodiment, the sensors in the sensor array are ultrasonic sensors.

[0065] Step S200: Input the impact signals into a trained neural network model to obtain distance values from an impact source generating the impact signals to each monitoring point.

[0066] In one embodiment, the neural network model is a one-dimensional deep residual neural network. The output obtained from the one-dimensional deep residual neural network is a 1xN feature vector (N is the total number of sensors in the sensor array), which corresponds to the distances from the impact source to the N sensor positions. Since the distances from the impact source to each sensor position can be different, the sound wave signals from different sensors can have different levels of impact on the impact location results. In order to alleviate the overfitting problem that may be caused, the convergence target is relaxed to N distance values from the impact point to each sensor.

[0067] ​In one embodiment, the neural network model is a one-dimensional deep residual neural network, and the impact signal is a one-dimensional time-domain signal. Based on this, without pre-knowledge of the sound wave propagation parameters in each direction of the target area, and without pre-knowledge of the properties and parameters of the engineering structure in the target area, the impact signal can be directly input into the trained neural network model for feature vector extraction without preprocessing, thereby reducing resource and time consumption, and improving data processing efficiency.

[0068] As shown in Figures 2-3 The one-dimensional deep residual neural network includes, in sequence, a convolutional layer, a batch normalization layer, a first rectified linear unit, a maximum pooling layer, a plurality of average pooling layers, a first full connection layer, a second rectified linear unit, and a second full connection layer.

[0069] The impact signal sequentially undergoes feature vector extraction of the convolutional layer, accelerated training of the batch normalization layer, nonlinear combination of the first rectified linear unit, and feature vector selection of the maximum pooling layer and the average pooling layer. Further, the selected feature vectors are subjected to regression processing via the first full connection layer and the second full connection layer which are separated by the second rectified linear unit (ReLU activation function), to obtain distance values of the impact source generating the impact signal to each of the monitoring points.

[0070] Each of the maximum pooling layer and the average pooling layer is configured with a connection branch, and each connection branch contains a plurality of residual blocks connected in sequence.

[0071] In one embodiment, three connection branches are configured between the maximum pooling layer and the average pooling layer, and each connection branch contains three residual blocks connected in sequence. The main difference among the three residual blocks is that the number of convolution kernels of the convolutional layer in each residual block increases and the number of feature maps decreases in sequence. The difference among the three connection branches is that the size of the convolution kernel increases from 3, 5 to 7, so that a series of features (rising edge, falling edge, sharp change, flat section, etc. in the signal) gradually increasing from low level to high level in the original input signal can be extracted, and these features will synergistically affect the prediction result of the impact positioning. At the end of the three connection branches, the average pooling layer is used to reduce the dimension of the extracted feature maps, and then all the extracted features are integrated into a large feature vector.

[0072] In one embodiment, the residual block is a two-layer identity residual block or a three-layer identity residual block. Each residual block is composed of two paths, one path for feature vector extraction and the other path for simple mapping of the input. The sum of the two paths is used as the output of the entire residual block, which can maintain good performance and avoid gradient instability when the number of network layers is large.

[0073] Step S300: Determine the existence trajectory formed by the possible existence positions of the impact source relative to each of the monitoring points based on the distance value.

[0074] In one embodiment, a circular trajectory is formed with each monitoring point as the center and the corresponding distance value as the radius, and the circular trajectory is used as the existence trajectory of the impact source relative to the monitoring point.

[0075] Step S400: Remove interference points from all intersections between the existing trajectories and determine the center point of the remaining intersections as the final center point.

[0076] like Figure 4 As shown, in one embodiment, step S400 includes steps S410-S440.

[0077] Step S410: Determine the center point corresponding to the existing intersection point between the existing trajectories.

[0078] Step S420: The intersection point farthest from the center point is regarded as an interference point and eliminated, and the new center point corresponding to the remaining intersection points is determined;

[0079] Repeat the previous step until half of the interference points out of all intersection points are eliminated.

[0080] The center point corresponding to the remaining half of the intersection points is taken as the final center point.

[0081] Step S430: When the number of remaining intersections is greater than half of the total number of intersections, repeat step S420. When the number of remaining intersections is equal to half of the total number of intersections, continue to execute step S440.

[0082] Step S440: Take the center point corresponding to the remaining half of the intersection points as the final center point.

[0083] For each sensor, a prediction circle can be drawn with the monitoring point where the sensor is located as the center and the distance value as the radius, representing the possible trajectory of the impact source. The prediction circles from multiple sensors will intersect, but due to prediction errors, all circles will not intersect at the same point; instead, each circle will have two intersection points, resulting in a total of [number missing]. There are N intersection points, where N is the number of sensors.

[0084] Understandable Of the intersection points, half are concentrated near the impact source, while the other half are distributed relatively far from the impact source. These latter half are considered interference points. Therefore, interference points are first eliminated by center point clustering, and then the geometric center of the remaining half of the intersection points is calculated as the final center point.

[0085] In one embodiment, for the remaining half of the intersection points, the horizontal coordinate average value and the vertical coordinate average value are taken as the coordinates of the geometric center of the remaining half of the intersection points, respectively.

[0086] As shown in Figure 5 , three sensors are taken as an example. First, the three prediction circles of the three sensors intersect to obtain six intersection points (for the case of not intersecting, the two points closest to the two circles are used instead); then the center points corresponding to the six intersection points are calculated; the intersection point farthest from the center point is removed, and the new center point corresponding to the remaining intersection points is calculated; after repeating the above step 3 times, all interference points (half of all intersection points) can be removed. The final center point is determined from the remaining three intersection points, which is the prediction result of the real impact point.

[0087] If N sensors are used in actual application, the center point clustering process needs to be performed to obtain the final center point.

[0088] In one embodiment, in order to remove the interference points from all intersection points between the existence trajectories, the interference points can also be determined and removed by other means.

[0089] For example: take the average value of the distances of all intersection points to the impact source, and remove the intersection points with a distance greater than the average distance.

[0090] S500: output the impact positioning result by taking the final center point as the actual position of the impact source.

[0091] In one embodiment, the impact positioning result is visualized.

[0092] In the above, the multi-channel impact signals monitored by the plurality of monitoring points in the target region are processed by the trained neural network model, the distance values of the impact source to each monitoring point are output, the existence trajectories of the impact source relative to each monitoring point are determined according to the distance values, the interference points from all intersection points between the existence trajectories are removed, and the center point of the remaining intersection points is determined as the final center point, i.e. the actual position of the impact source. Based on this, the embodiments of the present application can be applied to various complex engineering structures without knowing the structural properties and parameters in advance, and the real-time performance and accuracy of the engineering structure health monitoring can be ensured.

[0093] As shown in Figures 6-7 , based on the same inventive concept, the present application also provides a neural network model training method, which comprises steps S001-S004.

[0094] Step S001: obtain the impact signals monitored by the plurality of monitoring points in the target region.

[0095] In one embodiment, the target region is arranged with a sensor array, and the positions of the ultrasonic sensors of the sensor array are the monitoring points.

[0096] In order to effectively monitor the structural health of the target engineering structure, one or more detection regions need to be determined as target regions first. By arranging a sensor array around the target region, signals before impact and impact signals after impact can be collected.

[0097] The sensor array has different geometric designs, including linear, circular, planar, cylindrical, or spherical arrays. The advantage of using a sensor array instead of a single sensor is that the sensor array adds a new dimension to the observation, which helps to estimate more parameters and improve the estimation performance.

[0098] In one embodiment, the sensors in the sensor array are ultrasonic sensors.

[0099] Step S002: input the impact signal into the neural network model to be trained to obtain the distance value of the impact source generating the impact signal to each monitoring point as the output result.

[0100] In one embodiment, the neural network model is a one-dimensional deep residual neural network, and the output obtained from the one-dimensional deep residual neural network is a 1xN feature vector (N is the total number of sensors in the sensor array), which corresponds to the distance from the impact source to the N sensor positions. Since the distance from the impact source to each sensor position may be different, the sound wave signals from different sensors may have different levels of impact on the impact positioning result. In order to alleviate the possible overfitting problem, the convergence target is relaxed to the N distance values from the impact point to each sensor.

[0101] In one embodiment, the neural network model is a one-dimensional deep residual neural network, and the impact signal is a one-dimensional time domain signal. Based on this, without pre-knowledge of the sound wave propagation parameters in each direction of the target region, and without pre-knowledge of the properties and parameters of the engineering structure in the target region, the impact signal can be directly input into the trained neural network model for feature vector extraction without preprocessing, thereby reducing resource and time consumption and improving data processing efficiency.

[0102] As shown in Figures 2-3 The one-dimensional deep residual neural network includes sequentially connected convolutional layers, batch normalization layers, first rectified linear units, maximum pooling layers, multiple average pooling layers, first fully connected layers, second rectified linear units, and second fully connected layers.

[0103] The impact signal sequentially passes through feature vector extraction of a convolution layer, accelerated training of a batch normalization layer, nonlinear combination of a first rectified linear unit, feature vector selection of a max pooling layer and an average pooling layer. Further, the selected feature vectors are subjected to regression processing via a first fully connected layer and a second fully connected layer which are separated by a second rectified linear unit (ReLU activation function), to obtain distance values of the impact source generating the impact signal to each of the monitoring points.

[0104] Among the max pooling layer and each of the average pooling layers, a connection branch is configured respectively, and each connection branch contains a plurality of residual blocks which are sequentially connected.

[0105] In an embodiment, among the max pooling layer and the average pooling layers, three connection branches are configured, and each connection branch contains three residual blocks which are sequentially connected. The main difference among the three residual blocks is that the number of convolution kernels of the convolution layer in each residual block increases and the number of feature maps decreases. The difference among the three connection branches is that the size of the convolution kernel increases from 3, 5 to 7, so that a series of features (rising edge, falling edge, sharp change, flat section, etc. in the signal) gradually increasing from low level to high level in the original input signal can be extracted, and these features will synergistically affect the prediction result of the impact positioning. At the end of the three connection branches, the average pooling layer is used to reduce the dimension of the extracted feature maps, and then all the extracted features are integrated into a large feature vector.

[0106] In an embodiment, the residual block is a two-layer identity residual block or a three-layer identity residual block. Each residual block is composed of two paths, one path extracts feature vectors, and the other path simply maps the input. The sum of the two paths is the output of the entire residual block, which can maintain good performance and avoid gradient instability when the number of network layers is large.

[0107] Step S003: Calculate the loss value of the output result and the corresponding training label by a loss function, and the training label is the distance value from the actual position of the impact source to each of the monitoring points.

[0108] In an embodiment, before training the neural network model, a certain number of impact points need to be reasonably selected in the target area, and then a small energy non-destructive impact device such as a small ball is used to generate impact point by point. After the sensor array collects the impact signal, the impact signal is saved as the input during training, and the position of the impact point and the distance value from the impact point to each sensor are recorded as the training label. It can be understood that the position of the impact point is the actual position of the impact source.

[0109] In an embodiment, the loss function is a mean square error loss function.

[0110] Step S004: inputting the loss value into the neural network model to optimize the neural network model by using an Adam optimizer and a back propagation algorithm.

[0111] The Adam optimizer is used for gradient calculation and the back propagation algorithm to obtain a global optimal solution. In an embodiment, the learning rate is initialized as 0.002 and exponentially decays with the training step number to obtain a better convergence result, and the training step number is set as 200 steps.

[0112] The above steps S001-S004 are repeated until the obtained loss value is less than a preset threshold.

[0113] It can be understood that the use of the loss function is mainly in the training phase of the model. After each batch of training data is input into the model, the loss function calculates the difference value between the predicted value and the true value, that is, the loss value, by outputting the predicted value through forward propagation. After obtaining the loss value, the model updates each parameter through back propagation to reduce the loss between the true value and the predicted value, so that the predicted value generated by the model approaches the true value, thereby achieving the purpose of learning.

[0114] That is, the greater the difference between the true value and the predicted value, the greater the loss. The optimization goal is to reduce the loss.

[0115] The above can optimize the neural network model to not need to know the sound wave propagation parameters in each direction of the target region in advance, and not need to know the properties and parameters of the engineering structure in the target region in advance, can be applied to various complex engineering structures, and ensure the real-time and accuracy of engineering structure health monitoring, such as impact positioning.

[0116] As shown in Figure 8 Based on the same inventive concept, the application also provides an impact positioning device based on deep learning, which comprises a first acquisition module 10, a first input module 20, a first calculation module 30, a second calculation module 40 and an output module 50.

[0117] The first acquisition module 10 is used for acquiring impact signals monitored by a plurality of monitoring points in a target region.

[0118] In an embodiment, the target region is arranged with a sensor array, and the positions of ultrasonic sensors of the sensor array are the monitoring points.

[0119] In order to effectively monitor the structure health of the target engineering structure, one or more detection regions need to be determined as the target region first. A sensor array is arranged around the target region to collect signals before impact and impact signals after impact.

[0120] The sensor arrays can have different geometric designs, including linear, circular, planar, cylindrical, or spherical arrays. The advantage of using a sensor array instead of a single sensor is that the array adds a new dimension to the observation, facilitating the estimation of more parameters and improving estimation performance.

[0121] In one embodiment, the sensors in the sensor array are ultrasonic sensors.

[0122] The first input module 20 is used to input the impact signal into the trained neural network model to obtain the distance value from the impact source that generates the impact signal to each of the monitoring points.

[0123] In one embodiment, the neural network model is a one-dimensional deep residual neural network. The output from the one-dimensional deep residual neural network is a 1×N feature vector (where N is the total number of sensors in the sensor array), corresponding to the distance from the impact source to the N sensor locations. Since the distance from the impact source to each sensor location may be different, the acoustic signals from different sensors may have different levels of influence on the impact localization results. To mitigate the potential overfitting problem, the convergence target is relaxed to N distance values ​​from the impact point to each sensor.

[0124] In one embodiment, the neural network model is a one-dimensional deep residual neural network, and the impact signal is a one-dimensional time-domain signal. Based on this, it is not necessary to know the sound wave propagation parameters in each direction of the target area in advance, nor is it necessary to know the properties and parameters of the engineering structure in the target area in advance. The impact signal can be directly input into the trained neural network model for feature vector extraction without preprocessing, thereby reducing resource and time consumption and improving data processing efficiency.

[0125] like Figures 2-3 As shown, the one-dimensional deep residual neural network includes sequentially connected convolutional layers, batch normalization layers, first corrected linear units, max pooling layers, multiple average pooling layers, a first fully connected layer, a second corrected linear unit, and a second fully connected layer.

[0126] The impact signal sequentially undergoes feature vector extraction via convolutional layers, accelerated training via batch normalization layers, nonlinear combination of first modified linear units, and feature vector selection via max pooling and average pooling layers. Further, the selected feature vectors are regressed through a first fully connected layer and a second fully connected layer separated by a second modified linear unit (ReLU activation function) to obtain the distance value from the impact source generating the impact signal to each of the monitoring points.

[0127] Wherein, a connection branch is arranged between the maximum pooling layer and each of the average pooling layers, and each connection branch contains a plurality of residual blocks connected in sequence.

[0128] In one embodiment, three connection branches are arranged between the maximum pooling layer and the average pooling layer, and each connection branch contains three residual blocks connected in sequence. The main difference between the three residual blocks is that the number of convolution kernels of the convolution layer in each block increases and the number of feature maps decreases in turn. The difference between the three connection branches is that the size of the convolution kernel increases from 3 to 5 to 7, so that a series of features (such as rising edge, falling edge, sharp change, and gentle segment) gradually increasing from low level to high level in the original input signal can be extracted, and these features will work together to affect the prediction result of the impact location. At the end of the three connection branches, the average pooling layer is used to reduce the dimension of the extracted feature map, and then all the extracted features are integrated into a large feature vector.

[0129] In one embodiment, the residual block is a two-layer identity residual block or a three-layer identity residual block. Each residual block is composed of two paths, one for feature vector extraction and the other for simple mapping of the input. The sum of the two paths is the output of the entire residual block, which can maintain good performance and avoid gradient instability when the number of network layers is large.

[0130] The first calculation module 30 is configured to determine the existence trajectory formed by the possible existence position of the impact source relative to each monitoring point according to the distance value.

[0131] In one embodiment, the first calculation module 30 forms a circular trajectory with each monitoring point as the center and the corresponding distance value as the radius, and takes the circular trajectory as the existence trajectory of the impact source relative to the monitoring point.

[0132] The second calculation module 40 is configured to eliminate interference points among all intersection points between the existence trajectories, and determine the center point of the remaining intersection points as the final center point.

[0133] In one embodiment, the second calculation module 40 is specifically configured to:

[0134] determine the center point corresponding to the existing intersection point between the existence trajectories;

[0135] regard the intersection point farthest from the center point as an interference point for elimination, and determine the new center point corresponding to the remaining intersection points;

[0136] repeat the previous step until the number of eliminated intersection points is half of all intersection points;

[0137] take the center point corresponding to the remaining half of the intersection points as the final center point.

[0138] For each sensor, a prediction circle can be drawn as the possible existence trajectory of the impact source, with the monitoring point where the sensor is located as the center and the distance value as the radius. The prediction circles of multiple sensors intersect, but due to prediction errors, all circles do not intersect at the same point but have two intersection points each, and there are a total of intersection points.

[0139] where N is the number of sensors.

[0140] It can be understood that half of the intersection points are distributed near the impact source, and the other half of the intersection points are distributed relatively far away from the impact source. This half of the intersection points is regarded as interference points. Therefore, the interference points among all the trajectory intersection points are first removed by the center point clustering processing, and then the geometric center of the remaining half of the intersection points is calculated as the final center point.

[0141] In an embodiment, for the remaining half of the intersection points, the horizontal coordinate average value and the vertical coordinate average value are taken as the coordinates of the geometric center of the remaining half of the intersection points.

[0142] As Figure 5 shown, three sensors are taken as an example. First, the three prediction circles of the three sensors intersect to obtain six intersection points (for the case of not intersecting, the two points closest to the two circles are used instead); then the center points corresponding to the six intersection points are calculated; the intersection point farthest from the center point is removed, and the new center point corresponding to the remaining intersection points is calculated; after repeating the above step 3 times, all interference points (half of all intersection points) can be removed. The final center point is determined from the remaining three intersection points, which is the prediction result of the true impact point.

[0143] If N sensors are used in actual application, the center point clustering processing needs to be performed times to obtain the final center point.

[0144] In an embodiment, to achieve the purpose of removing the interference points among all the intersection points between the existence trajectories, the interference points can also be determined and removed by other ways.

[0145] For example: take the average value of the distances of all intersection points to the impact source, and remove the intersection points with a distance greater than the average distance.

[0146] The output module 50 is configured to output the impact positioning result by taking the final center point as the actual position of the impact source.

[0147] In an embodiment, the output module 50 is further configured to perform visual processing on the impact positioning result.

[0148] The above, through the trained neural network model, the multi-channel impact signal monitored by multiple monitoring points in the target area is processed, the distance value of the impact source to each monitoring point is output, the existence trajectory of the impact source relative to each monitoring point is determined according to the distance value, and the interference points in all intersection points between the existence trajectories are removed, and the center point of the remaining intersection points is determined as the final center point, that is, the actual position of the impact source. Based on this, the embodiment of the application does not need to know the structure attribute and parameter in advance, can be applied to various complex engineering structures, and ensures the real-time performance and accuracy of engineering structure health monitoring.

[0149] As Figure 9 shown, based on the same inventive concept, the application also provides a neural network model training device, comprising a second acquisition module 01, a second input module 02, a third calculation module 03, an optimization module 04 and a judgment module 05.

[0150] The second acquisition module 01 is used for acquiring the impact signal monitored by multiple monitoring points in a target area.

[0151] In one embodiment, the target area is arranged with a sensor array, and the positions of the ultrasonic sensors of the sensor array are the monitoring points.

[0152] In order to effectively monitor the structure health of the target engineering structure, one or more detection areas need to be determined as the target area first. By arranging a sensor array around the target area, signals before impact and impact signals after impact can be collected.

[0153] The sensor array has different geometric designs, including linear, circular, planar, cylindrical or spherical array. The advantage of using a sensor array instead of a single sensor is that the sensor array adds a new dimension to the observation, which helps to estimate more parameters and improve the estimation performance.

[0154] In one embodiment, the sensors in the sensor array are ultrasonic sensors.

[0155] The second input module 02 is used for inputting the impact signal into the neural network model to be trained, obtaining the distance value of the impact source generating the impact signal to each monitoring point as an output result.

[0156] In one embodiment, the neural network model is a one-dimensional deep residual neural network, and the output obtained from the one-dimensional deep residual neural network is a 1xN feature vector (N is the total number of sensors in the sensor array), which corresponds to the distances from the impact source to the N sensor positions. Since the distances from the impact source to the various sensor positions can be different, the sound wave signals from different sensors can have different levels of impact on the results of impact positioning. In order to alleviate the overfitting problem that can be caused, the convergence target is relaxed to the N distance values from the impact point to each sensor.

[0157] In one embodiment, the neural network model is a one-dimensional deep residual neural network, and the impact signal is a one-dimensional time domain signal. Based on this, without pre-knowledge of the sound wave propagation parameters in each direction of the target area, and without pre-knowledge of the properties and parameters of the engineering structure in the target area, the impact signal can be directly input into the trained neural network model for feature vector extraction without preprocessing, thereby reducing the consumption of resources and time, and improving the data processing efficiency.

[0158] As shown in Figures 2-3 The one-dimensional deep residual neural network includes, in sequence, a convolutional layer, a batch normalization layer, a first rectified linear unit, a maximum pooling layer, a plurality of average pooling layers, a first fully connected layer, a second rectified linear unit, and a second fully connected layer.

[0159] The impact signal sequentially undergoes feature vector extraction by the convolutional layer, accelerated training by the batch normalization layer, nonlinear combination by the first rectified linear unit, feature vector selection by the maximum pooling layer and the average pooling layers. Further, the selected feature vectors are subjected to regression processing via the first fully connected layer and the second fully connected layer, which are separated by the second rectified linear unit (ReLU activation function), to obtain the distance values from the impact source generating the impact signal to each of the monitoring points.

[0160] Each of the maximum pooling layer and the average pooling layers is configured with a connection branch, and each connection branch contains a plurality of residual blocks connected in sequence.

[0161] In one embodiment, three connection branches are configured between the maximum pooling layer and the average pooling layer, and each connection branch contains three residual blocks connected in sequence. The main difference between the three residual blocks is that the number of convolution kernels in the convolution layer increases and the number of feature maps decreases in sequence. The difference between the three connection branches is that the convolution kernel size increases from 3 to 5 to 7, so that a series of features (rising edge, falling edge, sharp change, flat section, etc.) gradually increasing from low to high in the original input signal can be extracted, and these features will work together to predict the impact location. At the end of the three connection branches, the extracted feature maps are reduced in dimension using the average pooling layer, and then all the extracted features are integrated into a large feature vector.

[0162] In one embodiment, the residual block is a two-layer identity residual block or a three-layer identity residual block. Each residual block is composed of two paths, one for feature vector extraction and the other for simple mapping of the input. The sum of the two paths serves as the output of the entire residual block, which can maintain good performance and avoid gradient instability when the number of network layers is large.

[0163] The third calculation module 03 is configured to calculate the loss value of the output result and the corresponding training label by a loss function, wherein the training label is the distance value from the actual position of the impact source to each monitoring point.

[0164] In one embodiment, before training the neural network model, a certain number of impact points need to be reasonably selected in the target area, and then a small energy non-destructive impact device such as a small ball is used to generate impact point by point. After the sensor array collects the impact signal, the impact signal is saved as the input during training, and the position of the impact point and the distance value from the impact point to each sensor are recorded as the training label. It can be understood that the position of the impact point is the actual position of the impact source.

[0165] In one embodiment, the loss function is a mean square error loss function.

[0166] The optimization module 04 is configured to input the loss value into the Adam optimizer and the back propagation algorithm to optimize the neural network model.

[0167] The Adam optimizer is used for gradient calculation and the back propagation algorithm to obtain the global optimal solution. In one embodiment, the learning rate is initialized to 0.002 and exponentially decays with the number of training steps to obtain better convergence results, and the number of training steps is set to 200 steps.

[0168] The judgment module 05 is configured to determine whether the loss value is less than a preset threshold, and stop the optimization action of the neural network model when the loss value is less than the preset threshold.

[0169] It can be understood that the use of the loss function is mainly in the training stage of the model. After each batch of training data is input into the model, the predicted value is output through forward propagation, and then the loss function calculates the difference value between the predicted value and the true value, that is, the loss value. After obtaining the loss value, the model updates the parameters through back propagation to reduce the loss between the true value and the predicted value, so that the predicted value generated by the model approaches the true value, thereby achieving the purpose of learning.

[0170] That is, the greater the difference between the true value and the predicted value, the greater the loss, and the optimization goal is to reduce the loss.

[0171] The above can optimize the neural network model to not need to know the sound wave propagation parameters in each direction of the target region in advance, and not need to know the properties and parameters of the engineering structure in the target region in advance, can be applied to various complex engineering structures, and ensure the real-time and accuracy of engineering structure health monitoring, such as impact positioning.

[0172] Based on the same inventive concept, the present application also provides a computer device, comprising: a processor, a memory and a computer program stored in the memory, the processor is coupled to the memory, and the processor executes the computer program to realize the above-mentioned deep learning-based impact positioning method when working.

[0173] Based on the same inventive concept, the present application also provides a computer device, comprising: a processor, a memory and a computer program stored in the memory, the processor is coupled to the memory, and the processor executes the computer program to realize the above-mentioned deep learning-based impact positioning method when working.

[0174] Based on the same inventive concept, the present application also provides a computer readable storage medium, the computer storage medium stores computer instructions, when the computer instructions are executed by the computer, the computer instructions make the computer execute the above-mentioned deep learning-based impact positioning method.

[0175] Based on the same inventive concept, the present application also provides a computer readable storage medium, the computer storage medium stores computer instructions, when the computer instructions are executed by the computer, the computer instructions make the computer execute the above-mentioned deep learning-based impact positioning method.

[0176] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network or other programmable apparatus. The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (for example: coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (for example: infrared, wireless, microwave, etc.) mode. The computer readable storage medium can be any available medium accessible by a computer, or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (for example: floppy disk, hard disk, magnetic tape), optical media (for example: digital versatile disc (DVD)) or semiconductor media (for example: solid state disk (SSD)) and the like.

[0177] The above is an embodiment provided by the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based shock location method, characterized in that, The method comprises the following steps: obtaining impact signals monitored by a plurality of monitoring points in a target area, wherein the target area is arranged with a sensor array, and positions of ultrasonic sensors of the sensor array are the monitoring points; inputting the impact signals into a trained neural network model to obtain distance values of an impact source generating the impact signals to each of the monitoring points, wherein the neural network model is a one-dimensional deep residual neural network, the impact signals are one-dimensional time domain signals, and the one-dimensional deep residual neural network comprises, in sequence, a convolution layer, a batch normalization layer, a first rectified linear unit, a maximum pooling layer, a plurality of average pooling layers, a first full connection layer, a second rectified linear unit and a second full connection layer; each of the maximum pooling layer and the average pooling layer is respectively configured with a connection branch, each connection branch contains a plurality of residual blocks connected in sequence, each residual block is a two-layer identity residual block or a three-layer identity residual block, and each residual block is composed of two paths, one path performs feature vector extraction, and the other path performs simple mapping on the input, and a sum of the two paths is taken as an output of the entire residual block; wherein the inputting the impact signals into the trained neural network model to obtain the distance values of the impact source generating the impact signals to each of the monitoring points comprises: sequentially performing feature vector extraction on the impact signals by the convolution layer, accelerating training by the batch normalization layer, nonlinear combination by the first rectified linear unit, and feature vector selection by the maximum pooling layer and the average pooling layer, and then performing regression processing on the selected feature vectors by the first full connection layer and the second full connection layer which are spaced apart by the second rectified linear unit to obtain the distance values of the impact source generating the impact signals to each of the monitoring points; determining existence trajectories formed by possible existing positions of the impact source relative to each of the monitoring points according to the distance values, comprising: forming a circular trajectory with each of the monitoring points as a center and with the corresponding distance value as a radius, and taking the circular trajectory as the existence trajectory of the impact source relative to the monitoring point; eliminating interference points in all intersection points between the existence trajectories, and determining a center point of the remaining intersection points as a final center point, comprising: determining a center point corresponding to an existing intersection point between the existence trajectories; regarding an intersection point farthest from the center point as an interference point for elimination; repeatedly performing the above steps until half of the intersection points are eliminated; and taking a center point corresponding to the remaining half of the intersection points as the final center point; taking the final center point as an actual position of the impact source, and outputting an impact positioning result.

2. A method for training a neural network model, comprising: The method comprises the following steps: obtaining impact signals monitored by a plurality of monitoring points in a target area, wherein the target area is arranged with a sensor array, and positions of ultrasonic sensors of the sensor array are the monitoring points; The impact signal is input into the neural network model to be trained to obtain distance values of an impact source generating the impact signal to each monitoring point as an output result, the neural network model is a one-dimensional deep residual neural network, the impact signal is a one-dimensional time domain signal, the one-dimensional deep residual neural network comprises sequentially connected convolution layers, batch normalization layers, first rectified linear units, a maximum pooling layer, a plurality of average pooling layers, a first full connection layer, second rectified linear units and a second full connection layer; each of the connection branches is configured between the maximum pooling layer and each average pooling layer, each connection branch contains a plurality of residual blocks connected in sequence, the residual block is a two-layer identity residual block or a three-layer identity residual block, each residual block is composed of two paths, one path performs feature vector extraction, and the other path performs simple mapping on the input, and the sum of the two paths is used as the output of the entire residual block; wherein the impact signal is input into the neural network model to be trained to obtain distance values of an impact source generating the impact signal to each monitoring point, comprising: the impact signal sequentially passes through feature vector extraction of the convolution layer, acceleration training of the batch normalization layer, nonlinear combination of the first rectified linear unit, feature vector selection of the maximum pooling layer and the average pooling layer, and then the selected feature vectors are subjected to regression processing through the first full connection layer and the second full connection layer separated by the second rectified linear unit to obtain distance values of an impact source generating the impact signal to each monitoring point; The loss value of the output result and the corresponding training label is calculated by a loss function, the training label is a distance value of an actual position of the impact source to each monitoring point; The loss value is input to optimize the neural network model by using an Adam optimizer and a back propagation algorithm; The above steps are repeated until the obtained loss value is less than a preset threshold.

3. A deep learning based shock location device, characterized by, Comprise: A first acquisition module is configured to acquire impact signals monitored by a plurality of monitoring points in a target region, the target region is equipped with a sensor array, and the positions of ultrasonic sensors of the sensor array are the monitoring points. The first input module is configured to input the impact signal into the trained neural network model to obtain distance values of an impact source generating the impact signal to each monitoring point, the neural network model is a one-dimensional deep residual neural network, the impact signal is a one-dimensional time domain signal, the one-dimensional deep residual neural network comprises sequentially connected convolution layers, batch normalization layers, first rectified linear units, a maximum pooling layer, a plurality of average pooling layers, a first full connection layer, second rectified linear units and a second full connection layer; each of the batch normalization layers is configured with a connection branch between the maximum pooling layer and each of the average pooling layers, each connection branch contains a plurality of residual blocks connected in sequence, the residual block is a two-layer identity residual block or a three-layer identity residual block, each residual block is composed of two paths, one path is used for feature vector extraction, and the other path is used for simple mapping of the input, and the sum of the two paths is used as the output of the entire residual block; wherein the impact signal is input into the trained neural network model to obtain the distance values of the impact source generating the impact signal to each monitoring point, which comprises that the impact signal sequentially passes through feature vector extraction of the convolution layer, acceleration training of the batch normalization layer, nonlinear combination of the first rectified linear unit, feature vector selection of the maximum pooling layer and the average pooling layer, and then the selected feature vectors are subjected to regression processing through the first full connection layer and the second full connection layer which are spaced apart by the second rectified linear unit, to obtain the distance values of the impact source generating the impact signal to each monitoring point; The first calculation module is configured to determine existence trajectories formed by possible existing positions of the impact source relative to each monitoring point according to the distance values, which comprises forming a circular trajectory with each monitoring point as the center and with the corresponding distance value as the radius, and taking the circular trajectory as the existence trajectory of the impact source relative to the monitoring point; The second calculation module is configured to eliminate interference points among all intersection points between the existence trajectories, and determine a center point of the remaining intersection points as a final center point, which comprises determining a center point corresponding to an existing intersection point between the existence trajectories; regarding the intersection point farthest from the center point as an interference point to be eliminated; repeatedly performing the above steps until half of the intersection points are eliminated; taking the center point corresponding to the remaining half of the intersection points as the final center point; The output module is configured to output a shock positioning result by taking the final center point as the actual position of the impact source.

4. A device for training a neural network model, characterized by, The second acquisition module is configured to acquire impact signals monitored by a plurality of monitoring points in a target region, the target region is arranged with a sensor array, and positions of ultrasonic sensors of the sensor array are the monitoring points. ​ The second input module is configured to input the impact signal into the neural network model to be trained to obtain distance values of an impact source generating the impact signal to each monitoring point as an output result. The third calculation module is configured to calculate a loss value of the output result and a corresponding training label by using a loss function, the training label being distance values of an actual position of the impact source to each monitoring point. The optimization module is configured to input the loss value as an input, and optimize the neural network model by using an Adam optimizer and a back propagation algorithm. The judgment module is configured to determine whether the loss value is less than a preset threshold, and stop the optimization of the neural network model when the loss value is less than the preset threshold.

5. A computer device, comprising: The computer program is stored in the memory and executed by the processor to implement the impact positioning method based on deep learning of claim 1 or the training method of the neural network model of claim 2. The computer storage medium stores computer instructions, when the computer instructions are executed by a computer, instructions for executing the impact positioning method based on deep learning of claim 1 or instructions for executing the training method of the neural network model of claim 2.

6. A computer-readable storage medium, characterized in that, ​

Images