An anti-theft detection method and device for ancient tombs based on an acceleration sensor

Abstract

The present invention discloses a method for detecting tomb theft based on an acceleration sensor, including: constructing a soil vibration event extraction model for screening out abnormal event soil vibration data, a soil vibration event classification model for identifying and classifying abnormal soil vibration data, and a soil vibration event early warning model for determining whether to give an alarm based on the event classification result, integrating and embedding the above three models into the acceleration sensor to obtain an anti-theft detection model that can independently complete soil event analysis and early warning. The present invention also provides a device for detecting tomb theft. This method integrates and embeds multiple lightweight models into the acceleration sensor, thereby improving the feedback speed of abnormal event recognition and the alarm accuracy rate of illegal tomb theft behaviors.

Description

Technical Field

[0001] The present invention relates to the technical field of anti-tomb-robbing detection of ancient tombs, and particularly relates to a method and device for anti-tomb-robbing detection of ancient tombs using an acceleration sensor. Background Art

[0002] Historical and cultural heritages are the witnesses of human history and the crystallization of the wisdom of the ancients. Modern archaeological and anthropological theories hold that cultural heritages mainly include cultural value, social value, personal value, and community value. The protection of historical and cultural heritages is a prerequisite for archaeological research, which is of great significance for us to inherit culture and understand history. Protecting historical and cultural heritages is the common responsibility of people all over the world.

[0003] An ancient tomb is a relic of the burial of the dead in a certain way in ancient times by humans. China has a long history, and a large number of precious cultural relics with high historical value are preserved in the ancient tombs of past dynasties. These cultural relics reflect the burial customs among different eras, regions, and social classes, and are important physical materials for exploring the social life conditions of the era to which they belong. Ancient tombs occupy an important position in historical and cultural heritages.

[0004] Because the cultural relics in ancient tombs have high economic value, tomb robbers take great risks and excavate them by illegal means, seriously damaging the cultural and economic values of ancient tombs. An important aspect of the protection of ancient tombs is to stop and punish tomb robbers in a timely manner. However, China has experienced the alternation of dozens of dynasties, and there are a large number of ancient tombs with a wide distribution range, usually in places where few people go and transportation is inconvenient. It is obviously difficult to detect tomb robbers in a timely manner through regular patrols by guards. The real-time detection of tomb robbing is mainly based on two types of information, one is image or video information, and the other is sensor information. The scheme based on image or video information can identify humans, as well as their postures and behaviors. It is applicable to the scenario where ancient tombs are densely distributed and power supply is convenient. The deployment cost of monitoring equipment is very high, and there is currently no good video detection algorithm for tomb robbing. The scheme based on sensor information, such as infrared monitoring, ultrasonic monitoring, radar monitoring, etc., can detect the entry of people, but cannot effectively identify tomb robbing behavior. And means such as optical fiber vibration can identify tomb robbing behavior, but the deployment is troublesome and the cost is high. Therefore, it is of great significance to study a high-precision anti-tomb-robbing detection method for ancient tombs with simple deployment and low cost.

[0005] Patent document CN112085914A discloses a highly sensitive anti-excavation protection device with network supervision and intelligent analysis, including an edge computing analysis and processing host, and buried digital sensors that provide analysis data for the edge computing analysis and processing host; after obtaining vibration information through the buried digital sensors, it is transmitted to the edge computing analysis and processing host for processing and identification. This method buries the sensors underground to avoid being discovered and circumvented. However, problems such as being easily intercepted or information loss may occur during the information transmission process between the sensors and the host, and the computing and analysis capabilities of the host are required to be very high, resulting in a large upfront investment.

[0006] Patent document CN106127135B discloses an algorithm for extracting and classifying invasion vibrations in a mausoleum area. It uses a multi-grating sensing detection system to collect vibration signals through a multi-grating sensor network, decomposes the signals using the EEMD algorithm to obtain IMF components, calculates the EEMD energy entropy, eliminates the interference of non-artificial signals, and performs feature extraction. Finally, it uses a support vector machine optimized by the particle swarm algorithm to classify and identify the invasion signals and give early warnings. This method uses an SVM optimized by the particle swarm algorithm to classify and identify the invasion signals. However, this method does not consider that the vibration forms and amplitudes are different in different terrains. For example, the vibration amplitudes are different at different positions in a sloping terrain, so problems of misjudgment will occur. Summary of the Invention

[0007] The present invention provides a method for detecting anti-excavation of ancient tombs based on an acceleration sensor. This method integrates multiple lightweight models into the acceleration sensor, thereby improving the feedback speed of abnormal event recognition and classification and the warning accuracy of illegal excavation behaviors.

[0008] A method for detecting anti-excavation of ancient tombs based on an acceleration sensor includes:

[0009] Step 1: Based on the working principle of the acceleration sensor, construct a soil vibration event extraction model for screening out soil vibration data of abnormal events. The soil vibration data includes the spatial acceleration direction and magnitude of soil movement;

[0010] Step 2: Through the soil vibration event extraction model in Step 1, collect soil vibration data of random abnormal events, and label the soil vibration data in two dimensions regarding excavation behavior events and other abnormal events. Combine the soil vibration data of the abnormal events with the labeled tags to form training samples;

[0011] Step 3: Construct a soil vibration event classification network, including a data preprocessing module, a feature extraction module, a data fusion module, and a classification module. The data preprocessing module is used to remove the event direction information in the soil vibration data provided by the soil vibration event extraction model, and input the processed soil vibration data into the feature extraction module. The feature extraction module is used to extract the multi-scale data features of the processed soil vibration data, and input the extracted multi-scale data features into the data fusion module. The data fusion module is used to fuse the multi-scale data features to obtain a fused feature, and input the fused feature into the classification module. The classification module is used to perform predictive calculations on the fused feature to output a classification result, and the classification result includes the classification result of the grave robbing behavior event and the classification result of other abnormal events;

[0012] Step 4: Use the training samples obtained in Step 2 to train the soil vibration event classification network. After the training is completed, obtain a soil vibration event classification model for identifying and classifying abnormal soil vibration data, and the abnormal soil vibration data is screened by the soil vibration event extraction model;

[0013] Step 5: Construct a soil vibration event warning model. The soil vibration event warning classification model is used to count the classification results output by the soil vibration event classification model within a certain period of time, and judge whether an alarm is needed according to the statistical results;

[0014] Step 6: Integrate and embed the soil vibration event extraction model in Step 1, the soil vibration event classification model in Step 4, and the soil vibration event warning model in Step 5 into the acceleration sensor to obtain an anti-grave robbing detection model that can independently complete soil event analysis and warning;

[0015] Step 7: Input a piece of soil vibration data to be identified into the anti-grave robbing detection model obtained in Step 6. After identification and analysis, output the result of whether an alarm is needed.

[0016] Preferably, for screening out the abnormal event soil vibration data in Step 1, the box plot method is used to judge whether an abnormal event occurs. Only when an abnormal event is determined will the corresponding soil vibration information be extracted. The feature in the determination process is the cosine similarity of the x-axis and y-axis data in the event signal:

[0017]

[0018] Among them, <a x ,a y > represents the inner product of the x-axis data and the y-axis data, and ||a x || and ||a y || represent the norms of the x-axis data and the y-axis data.

[0019] Specifically, the grave robbing behavior events in step 2 include excavation using a Luoyang shovel, and the other abnormal events include continuous jumping and walking by humans.

[0020] Preferably, the steps for removing the event direction information in the soil vibration data provided by the soil vibration event extraction model in step 3 are as follows:

[0021] Step 2.1: Perform a baseline subtraction operation on the soil vibration data provided by the soil vibration event extraction model to obtain baseline-subtracted data [a x , a y , a z T ;

[0022] Step 2.2: Calculate the modulus of the synthetic acceleration of the baseline-subtracted data:

[0023]

[0024] where ra is the modulus of the synthetic acceleration, a x is the x-axis data, a y is the y-axis data, a z is the z-axis data;

[0025] Step 2.3: Calculate the projection of the baseline-subtracted data in the direction of the gravitational acceleration:

[0026] v = [cosα, cosβ, cosγ][a x , a y , a z T

[0027] where v is the projection in the direction of the gravitational acceleration, α is the angle between the synthetic acceleration and the x-axis, β is the angle between the synthetic acceleration and the y-axis, and γ is the angle between the synthetic acceleration and the z-axis;

[0028] Step 2.4: Calculate the modulus of the projection vector of the baseline-subtracted data on the horizontal plane:

[0029]

[0030] where h is the modulus of the projection vector on the horizontal plane and v is the projection in the direction of the gravitational acceleration;

[0031] Step 2.5: Combine the data obtained in steps 2.2, 2.3, and 2.4 to form new soil vibration data [ra, v, h] T , and input it into the feature extraction module to avoid the event direction information affecting the final classification result.

[0032] ​​Specifically, the event direction information includes the buried angle of the wireless acceleration sensor and the event azimuth.

[0033] Preferably, the feature extraction module in step 3 includes two parts. The first part preliminarily extracts the data features of 4 scales based on the input soil vibration data. The second part performs convolution operations through two different types of channel attention time convolutional layers based on the preliminary extraction results to obtain the multi-scale data features after dimensionality reduction respectively. The data features output by the feature extraction module have the same dimension and size, thereby reducing the number of parameters of the soil vibration event classification model and improving the classification and recognition speed.

[0034] Preferably, the data features output by the feature extraction module are input to the data fusion module after splicing. The data fusion module uses group attention and channel shuffling to operate on the spliced data features to obtain the fusion features.

[0035] Specifically, for the training in step 4, for each classification dimension, the cross-entropy of the labeled label and the classification result is used as the loss function of a single classification dimension, and the parameters of the soil vibration event classification model are updated by comprehensively considering the loss functions of the two dimensions.

[0036] Specifically, for the training in step 4, the optimizer is selected as Adam, the initial learning rate is set to 0.001, and it decays by 0.2 times every 30 epochs.

[0037] Specifically, the soil vibration event warning model in step 5 is judged based on the naive Bayes algorithm, as follows:

[0038] When P(r1|t) > P(r2|t), an alarm instruction is issued;

[0039] When P(r1|t) ≤ P(r2|t), there is no need to issue an alarm instruction;

[0040] Wherein, t represents the classification result of the soil vibration event classification model within a certain period of time, r1 represents an alarm, r2 represents no alarm, and the judgment time interval is preset manually.

[0041] The present invention also provides an anti-theft excavation detection device for ancient tombs, which can quickly identify the behavior events corresponding to the soil vibration information, including:

[0042] A computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the anti-theft excavation detection model described above is adopted in the computer memory; when the computer processor executes the computer program, the following steps are implemented: input a section of soil vibration data to be identified into the anti-theft excavation detection model, and after identification and analysis, output the result of whether an alarm is required.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] (1) The present invention decomposes the traditional event analysis model into a soil vibration event extraction model and a soil vibration event classification model. First, the soil vibration event extraction model filters most non-abnormal events, thereby reducing the computing pressure on the soil vibration event classification model and further improving the feedback speed of recognition and classification.

[0045] (2) A set of data preprocessing methods is designed for the soil vibration event classification model to eliminate the event direction information that originally affects the recognition and classification results, improving the accuracy of recognition and classification.

[0046] (3) All models are integrated and embedded in the acceleration sensor, so that both information acquisition and analysis results are completed in the acceleration sensor, without being restricted by the communication network and devices, thereby making the accuracy of abnormal event alarms higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of the ancient tomb anti-theft detection method based on an acceleration sensor provided by the present invention;

[0048] Figure 2 is a schematic structural diagram of the soil vibration event classification network in this embodiment;

[0049] Figure 3 is a schematic structural diagram of the channel attention time convolution module of the soil vibration event classification network in this embodiment;

[0050] Figure 4 is a schematic structural diagram of the ordinary time convolution layer of the soil vibration event classification network in this embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] In order to solve the problems in the existing ancient tomb protection field that rely on manpower, have complex equipment deployment, and have high false alarm rates and missed detection rates, the embodiment provides an ancient tomb anti-theft detection method based on a wireless acceleration sensor.

[0052] As Figure 1 shown, an ancient tomb anti-theft detection method based on an acceleration sensor includes:

[0053] Step 1. Based on the working principle of the acceleration sensor, construct a soil vibration event extraction model for screening out abnormal event soil vibration data, where the soil vibration data includes the spatial acceleration direction and magnitude of soil movement:

[0054] Among them, the abnormal event soil vibration data is screened by using the box plot method to detect the abnormal cosine similarity value. Because when a vibration event occurs, the cosine similarity of the acceleration sensor data on the x-axis and y-axis will change: the lower quartile Q1 and the upper quartile Q3 can be obtained from all the cosine similarities in the queue, and IQR = Q3 - Q1, that is, the difference between the upper quartile and the lower quartile.

[0055] When the value of the cosine similarity is less than Q1 - 1.5 * IQR or greater than Q3 + 1.5 * IQR, it can be considered that there is a vibration segment with a length of win_size. Among them, the vibration segments within one win_size of adjacent ones can be directly spliced.

[0056] When the length of the spliced vibration segment reaches 3 * win_size, it can be considered that a vibration event has been extracted:

[0057] Filter out the vibration segments with very small event lengths, where the length of the vibration segment is less than 3 * win_size, because it is generally caused by short-term abnormal activities, such as small animals passing by, etc.;

[0058] Filter out the vibration events with very long event lengths, where the length of the vibration segment is greater than 5 * win_size, because it is generally caused by a car passing by or the continuous operation of agricultural machinery;

[0059] Extract the vibration events that meet the lengths of Luoyang shovel digging and general human activities, where the length of the vibration segment is greater than or equal to 3 * win_size and at the same time satisfies that the length of the vibration segment is less than or equal to 5 * win_size.

[0060] The corresponding soil vibration information will only be extracted when it is determined as an abnormal event. The feature in the determination process is the cosine similarity of the data on the x-axis and y-axis in the event signal:

[0061]

[0062] Among them, <a x ,a y > represents the inner product of the x-axis data and the y-axis data, and ||a x || and ||a y || represent the norms of the x-axis data and the y-axis data respectively.

[0063] Step 2: Through the soil vibration event extraction model in Step 1, collect the soil vibration data of random abnormal events, and label the soil vibration data with respect to two dimensions of the behavior events of using Luoyang shovels for digging and other abnormal events. Combine the soil vibration data of the abnormal events with the labeled tags to form a training sample, where other abnormal events include continuously performing jumping behaviors and / or walking behaviors artificially;

[0064] Step 3, as Figure 2 shown, construct a soil vibration event classification network based on time convolution and attention mechanism: including a data preprocessing module, a feature extraction module, a data fusion module and a classification module:

[0065] The data preprocessing module is used to remove the event direction information in the soil vibration data provided by the soil vibration event extraction model, and input the processed soil vibration data into the feature extraction module. Among them, the specific steps to remove the event direction information in the input soil vibration data are as follows:

[0066] Step 2.1, perform a baseline subtraction operation on the soil vibration data provided by the soil vibration event extraction model to obtain baseline-subtracted data [a x , a y , a z T ;

[0067] Step 2.2, calculate the modulus of the synthetic acceleration of the baseline-subtracted data:

[0068]

[0069] where ra is the modulus of the synthetic acceleration, a x is the x-axis data, a y is the y-axis data, a z is the z-axis data;

[0070] Step 2.3, calculate the projection of the baseline-subtracted data in the direction of the gravitational acceleration:

[0071] v = [cosα, cosβ, cosγ][a x , a y , a z T

[0072] where v is the projection in the direction of the gravitational acceleration, α is the angle between the synthetic acceleration and the x-axis, β is the angle between the synthetic acceleration and the y-axis, and γ is the angle between the synthetic acceleration and the z-axis;

[0073] Step 2.4, calculate the modulus of the projection vector of the baseline-subtracted data on the horizontal plane:

[0074]

[0075] where h is the modulus of the projection vector on the horizontal plane and v is the projection in the direction of the gravitational acceleration;

[0076] Step 2.5, form the new soil vibration data [ra, v, h] from the data obtained in Steps 2.2, 2.3 and 2.4 T ​​, input it into the feature extraction module to avoid the influence of event direction information on the final classification result.

[0077] The feature extraction module is used to extract multi-scale data features of the processed soil vibration information data. Among them, the structure of the channel attention time convolutional layer is as Figure 3 shown, including a normal time convolutional layer and an attention operation: The input first passes through the channel attention time convolutional module 1, the channel attention time convolutional module 2, the channel attention time convolutional module 3, and the channel attention time convolutional module 4 to extract the features of the first layer. Extracting enough features in the first layer is beneficial for subsequent classification. Then, the features are extracted in two paths. The purpose of extracting features in two paths is to reduce the number of model parameters. One path passes through the channel attention time convolutional module 5 and the channel attention depthwise separable time convolutional module 1, and the other path passes through the channel attention time convolutional module 6 and the channel attention time convolutional module 7.

[0078] Channel attention mainly includes a global average pooling layer, a fully connected layer, and an activation layer. The channel attention mechanism guides the computing resources to bias towards the parts with large amounts of information in the input signal. Considering that the information sizes on each channel of the sensor data are inconsistent, using the channel attention mechanism can improve the classification effect of the model.

[0079] The depthwise separable time convolutional module is to replace the convolutional block in the normal time convolutional module with a depthwise convolution and replace the feature concatenation with a 1×1 point convolution.

[0080] As Figure 4 shown, it is the structure of the normal time convolutional layer, including 1×3 convolution, 1×5 convolution, 1×11 convolution, 1×23 convolution, 1×3 dilated convolution, and the input data is concatenated after max pooling and average pooling, and then 1×5 convolution is performed. Convolution kernels of different sizes are beneficial for extracting information of different scales of the input data. Dilated convolution is equivalent to extracting information from the downsampled input, and max pooling and average pooling are equivalent to performing data augmentation on the input data, which is beneficial for improving the generalization performance of the model.

[0081] The data fusion module is used to fuse multi-scale data features to obtain fused features, including: using grouped attention and channel shuffle to increase the information interaction between channels, thereby improving the classification effect. The two shuffled paths of data continue to extract features respectively, passing through the channel attention depthwise separable time convolutional module 2 and the channel attention time convolutional module 8, and finally the features of the two paths are fused and input into the classification module.

[0082] The classification module is used to perform predictive calculations on the input features. The fused features pass through the global average pooling layer, the fully connected layer, and the sigmoid function layer to obtain the normalized probabilities for each category, so as to output the classification results. The classification results include the classification results of the act of robbing ancient tombs with a Luoyang shovel and the classification results of other abnormal events.

[0083] Step 4: Use the training samples obtained in Step 2 to train the soil vibration event classification model. After the training is completed, a soil vibration event classification model for identifying and classifying abnormal soil vibration data is obtained. During the training, for each classification dimension, the cross-entropy between the labeled label and the classification result is used as the loss function for a single classification dimension. Considering the loss functions of the two dimensions of the act of robbing ancient tombs with a Luoyang shovel and other abnormal events, the parameters of the soil vibration event classification model are updated. The optimizer is selected as Adam, the initial learning rate is set to 0.001, and it decays by 0.2 times every 30 epochs.

[0084] Step 5: Construct a soil vibration event warning model based on the Naive Bayes algorithm. The soil vibration event warning classification model is used to count the analysis results output by the soil vibration event classification model within a certain period of time, and determine whether to give an alarm according to the statistical results. The judgment process is as follows:

[0085] When P(r1|t)>P(r2|t), an alarm instruction is issued;

[0086] When P(r1|t)≤P(r2|t), no alarm instruction needs to be issued;

[0087] Among them, t represents the classification result of the soil vibration event classification model within a certain period of time, r1 represents an alarm, and r2 represents no alarm. The judgment time interval is preset manually.

[0088] Step 6: Integrate and embed the soil vibration event extraction model in Step 1, the soil vibration event classification model in Step 4, and the soil vibration event warning model in Step 5 into the acceleration sensor to obtain an anti-tomb-robbing detection model that can independently complete soil event analysis and warning;

[0089] Step 7: Input a piece of soil vibration data to be recognized into the anti-tomb-robbing detection model obtained in Step 6. After recognition and analysis, the result of whether to give an alarm is output.

[0090] The present invention also provides an anti-tomb-robbing detection device for ancient tombs, including:

[0091] A computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the anti-excavation detection model described above is adopted in the computer memory; when the computer processor executes the computer program, the following steps are implemented: input a segment of soil vibration data to be recognized into the anti-excavation detection model, and after recognition and analysis, output a result indicating whether an alarm is required.

[0092] Specific usage process: Install a wireless acceleration sensor in a certain piece of soil. The wireless acceleration sensor continuously collects acceleration data on the x, y, and z axes to reflect the vibration situation of the soil; during use, use the soil vibration event extraction model to screen out the soil vibration data of human behaviors such as continuous jumping, walking, and using a Luoyang shovel for excavation captured by the sensor; based on the screened soil vibration data, after recognition and analysis through the soil vibration event classification model, obtain the classification results of this data segment in the behaviors of using a Luoyang shovel for excavation, jumping, and walking; according to the classification results, after considering the excavation probability of multiple events in the recent time window through the soil vibration event warning model, output an instruction on whether to issue an alarm. When it is determined that an alarm needs to be issued, send the sensor location and alarm information to the terminal server to remind the operator to pay attention to abnormal events.

Claims

1. A method for detecting tomb theft prevention based on an acceleration sensor, characterized in that, Including: Step 1: Based on the working principle of the acceleration sensor, construct a soil vibration event extraction model for screening abnormal event soil vibration data, where the soil vibration data includes the spatial acceleration direction and magnitude of soil movement. Step 2: Through the soil vibration event extraction model in Step 1, collect the soil vibration data of random abnormal events, and label the soil vibration data in two dimensions regarding the excavation behavior event and other abnormal events. Combine the soil vibration data of the abnormal events with the labeled tags to form a training sample. Step 3: Construct a soil vibration event classification network, including a data preprocessing module, a feature extraction module, a data fusion module, and a classification module. The data preprocessing module is used to remove the event direction information in the soil vibration data provided by the soil vibration event extraction model and input the processed soil vibration data into the feature extraction module. The feature extraction module is used to extract the multi-scale data features of the processed soil vibration data and input the extracted multi-scale data features into the data fusion module. The data fusion module is used to fuse the multi-scale data features to obtain a fusion feature and input the fusion feature into the classification module. The classification module is used to perform predictive calculations on the fusion feature to output a classification result, and the classification result includes the classification result of the excavation behavior event and the classification result of other abnormal events. Step 4: Use the training sample obtained in Step 2 to train the soil vibration event classification network. After the training is completed, obtain a soil vibration event classification model for identifying and classifying abnormal soil vibration data, where the abnormal soil vibration data is screened by the soil vibration event extraction model. Step 5: Construct a soil vibration event warning model, which is used to count the classification results output by the soil vibration event classification model within a certain period of time and determine whether an alarm is needed according to the statistical results. Step 6: Integrate and embed the soil vibration event extraction model in Step 1, the soil vibration event classification model in Step 4, and the soil vibration event warning model in Step 5 into the acceleration sensor to obtain an anti-excavation detection model that can independently complete soil event analysis and warning. Step 7: Input a section of soil vibration data to be identified into the anti-excavation detection model obtained in Step 6. After identification and analysis, output the result of whether an alarm is needed.

2. The method for detecting tomb theft based on an acceleration sensor according to claim 1, wherein For screening the abnormal event soil vibration data in Step 1, the box plot method is used to determine whether an abnormal event occurs. Only when it is determined to be an abnormal event will the corresponding soil vibration information be extracted.

3. The method for detecting tomb theft based on an acceleration sensor according to claim 1, characterized in that The specific steps for removing the event direction information in the soil vibration data provided by the soil vibration event extraction model in Step 3 are as follows: Step 2.

1. Perform a baseline subtraction operation on the soil vibration data provided by the soil vibration event extraction model to obtain baseline-subtracted data [a x , a y , a z T ;​ Step 2.2: Calculate the modulus of the synthetic acceleration of the data after subtracting the baseline. Among them, ra is the modulus of synthetic acceleration, a x is the x-axis data, a y is the y-axis data, a z is the z-axis data; Step 2.3: Calculate the projection of the data after subtracting the baseline in the direction of the gravitational acceleration. v = [cosα, cosβ, cosγ][a x , a y , a z T ​ where v is the projection in the direction of the gravitational acceleration, α is the angle between the synthetic acceleration and the x-axis, β is the angle between the synthetic acceleration and the y-axis, and γ is the angle between the synthetic acceleration and the z-axis. Step 2.4: Calculate the modulus of the projection vector of the data after subtracting the baseline on the horizontal plane. Among them, h is the modulus of the projection vector on the horizontal plane, and v is the projection in the direction of the gravitational acceleration; Step 2.5: Combine the data obtained in Steps 2.2, 2.3, and 2.4 to form new soil vibration data [ra, v, h]. T Input it into the feature extraction module.

4. The ancient tomb anti-theft detection method based on an acceleration sensor according to claim 1, characterized in that, The feature extraction module in step 3 includes two parts. The first part preliminarily extracts the data features at 4 scales based on the input soil vibration data. The second part performs convolution operations through two different types of channel attention time convolutional layers based on the preliminary extraction results to obtain the multi-scale data features after dimensionality reduction respectively. The data features output by the feature extraction module have the same dimension and the same size.

5. The method for detecting tomb theft based on an acceleration sensor according to claim 4, wherein The data features output by the feature extraction module are input to the data fusion module after being concatenated. The data fusion module uses group attention and channel shuffling to operate on the concatenated data features to obtain the fusion features.

6. The ancient tomb anti-theft detection method based on an acceleration sensor according to claim 1, characterized in that For the training in step 4, for each classification dimension, the cross-entropy between the labeled label and the classification result is used as the loss function for a single classification dimension, and the parameters of the soil vibration event classification model are updated by comprehensively considering the loss functions of the two dimensions.

7. The ancient tomb anti-theft detection method based on an acceleration sensor according to claim 1, characterized in that For the training in step 4, the optimizer is selected as Adam, the initial learning rate is set to 0.001, and it decays by 0.2 times every 30 epochs.

8. The method for detecting tomb theft based on an acceleration sensor according to claim 1, characterized in that, The soil vibration event warning model in step 5 is judged based on the naive Bayes algorithm, specifically as follows: When P(r1|t) > P(r2|t), an alarm instruction is issued; When P(r1|t) ≤ P(r2|t), there is no need to issue an alarm instruction; Among them, t represents the classification result of the soil vibration event classification model within a certain period of time, r1 represents an alarm, and r2 represents no alarm. Among them, the judgment time interval is preset manually.

9. An anti-tomb-robbing detection device for ancient tombs, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that, The anti-excavation detection model described in claim 1 is adopted in the computer memory; when the computer processor executes the computer program, the following steps are implemented: input a segment of soil vibration data to be identified into the anti-excavation detection model, and after identification and analysis, output the result of whether an alarm is required.

Images