Construction method, recognition method and system of a shooting recognition model for heavy weapons

Abstract

The present invention discloses a construction method, an identification method and a system for a shooting identification model of heavy weapons. The present invention adopts an identification method with an adaptive sliding window, that is, the collected acceleration sequence is divided into continuous windows for processing, and the identification model is dynamically updated during the processing of the acceleration sequence of each window, and the shooting acceleration physical characteristics of the equipment are used to dynamically identify the shooting events of the equipment. The method of the present invention can effectively improve the detection accuracy; can adapt to different types of equipment; can eliminate the influence of environmental factors, such as acceleration, deceleration and bumpy road surface and other factors.

Description

Technical Field

[0001] The present invention relates to the shooting detection technology of heavy weapons, and specifically to a method for constructing a shooting detection model of heavy weapons and a method for shooting detection using the constructed model. Background Art

[0002] As the most important firepower strike weapon of the army, the training of heavy weapons is much more difficult than that of light weapons, and the training cost is relatively high. Therefore, it is particularly important to detect the shooting of weapons in real time. Compared with the action detection scheme commonly used in the shooting of general light weapons, due to the very large recoil impact during shooting and the complex combat environment of heavy equipment, the shooting interference is very serious. Therefore, it is difficult to detect heavy weapons, and there are serious false alarms in the number of shootings. Summary of the Invention

[0003] Aiming at the defects or deficiencies of the prior art, on the one hand, the present invention provides a method for constructing a shooting recognition model of heavy weapons.

[0004] Therefore, the method for constructing a shooting recognition model of heavy weapons provided by the present invention includes:

[0005] Continuously collect the acceleration values of the turret; at the same time, divide the collected multiple acceleration values into time windows for processing in sequence. Each time window includes several sequentially collected acceleration values, and the acceleration sequence of any time window is , is any acceleration value in , ; The processing method of the acceleration sequence of each time window includes:

[0006] (1) Obtain the maximum acceleration value in the acceleration sequence of the current time window, and record the earliest occurrence position of the maximum acceleration value in the acceleration sequence of the current window; at the same time, obtain the minimum acceleration value in the acceleration sequence of the current time window;

[0007] (2) Calculate the volatility of the acceleration sequence of the current time window,

[0008]

[0009] where: ;

[0010] is the average value of the acceleration sequence of the current time window, ;

[0011] is the standard deviation of the acceleration sequence in the current time window, ;

[0012] (3) When holds, it is considered that the acceleration sequence in the current time window is an acceleration sequence with an acceleration event, and step (4) is executed; otherwise, the acceleration sequence of the next time window is processed;

[0013] (4) For the sequence with an acceleration event, iterate forward from the position of this sequence to obtain the position of the acceleration value at which the acceleration value starts to mutate in the acceleration mutation event sequence , and the acceleration at which the acceleration value starts to mutate satisfies the condition: , and then obtain the acceleration feature sequence of this sequence; is the first acceleration values in the next time window of the acceleration sequence with an acceleration mutation event;

[0014] (5) Incorporate the obtained acceleration feature sequence into the training data set, and use the training data set to update and train the neural network that was updated and trained last time to obtain an updated heavy weapon shooting recognition model; the training data set initially contains multiple groups of acceleration feature sequence samples, and the acceleration feature sequence samples are acceleration feature sequences obtained by taking steps (1)-(4) or artificially constructed acceleration feature sequences. Each acceleration feature sequence sample contains acceleration values, and the neural network is initially an untrained neural network.

[0015] Further, the method for constructing the heavy weapon shooting recognition model of the present invention further includes: (6) Verify the updated heavy weapon shooting recognition model, including: collecting the acceleration sequence of any window of the heavy weapon, and the number of acceleration values in this sequence is , use the collected acceleration sequence as the input of the updated heavy weapon shooting recognition model, and determine whether a shooting event occurs;

[0016] For the output result of judging that a shooting event occurs, if the displacement of the turret in the corresponding input time window is equal to the displacement of the turret in the previous time window of the corresponding input time window or there is a reasonable difference between the two, it is considered that the updated heavy weapon shooting recognition model is reliable; otherwise, the model needs to be updated and trained continuously.

[0017] Optionally, the neural network initially uses a feedforward neural network, and the feedforward neural network includes an input layer, a hidden layer, and an output layer connected in sequence.

[0018] The present invention also provides a method for identifying heavy weapon shooting. The identification method includes: collecting an acceleration sequence of any window of a heavy weapon, and the number of acceleration values in the sequence is , and inputting the collected sequence into the heavy weapon shooting detection model constructed by the above method to determine whether a shooting event occurs.

[0019] The present invention also provides a heavy weapon shooting identification system. The identification system includes: an acceleration acquisition module and a shooting identification module. The acceleration acquisition module is used to collect the acceleration values of the turret and divide the collected multiple acceleration values according to a time window; the shooting identification module uses the model constructed by the above method to identify the shooting of the heavy weapon, and simultaneously updates the model by using the above method.

[0020] Further, the shooting identification module also verifies the model by using the above method.

[0021] When a heavy equipment is moving, its acceleration can reach a relatively large value under certain conditions. At the same time, during the movement process, due to environmental factors such as road bumps, interference acceleration will be generated; during the signal sampling process, the sampling rate is limited by the highest sampling frequency of the acquisition device, resulting in the distortion of the amplitude of the acceleration signal during shooting; for different types of actual equipment, the calibers and launched ammunitions are different, resulting in different acceleration values and characteristics of the equipment. For the above reasons, it is difficult to select a reasonable threshold in actual use to distinguish shooting events of different equipment. Therefore, the threshold-based scheme will cause misjudgment and cannot provide accurate shooting detection. The method of the present invention can effectively improve the detection accuracy; can adapt to different types of equipment; can exclude environmental factors such as acceleration, deceleration, and bumpy roads. Description of the Drawings

[0022] Figure 1 is part of the original sampling data in the embodiment;

[0023] Figure 2 is part of the acceleration feature sequence processed by step (4) of the present invention in Embodiment 1;

[0024] Figure 3 is the loss curve during the model training process in Embodiment 1;

[0025] Figure 4 is any input sequence in Embodiment 1 (it is known that there is a shooting event in this sequence);

[0026] Figure 5 is another input sequence in Embodiment 1 (it is known that no shooting event has occurred in this sequence). Detailed Embodiments

[0027] Unless otherwise specified, the terms in this document are understood according to the knowledge of ordinary technicians in the relevant field.

[0028] It takes a certain amount of time for the shell of heavy weaponry to travel from launch to leaving the barrel. For example, it takes about 5 ms for the shell of a certain type of equipment to travel from launch to leaving the barrel. When the shell is moving inside the barrel, the turret acceleration signal, after passing through the barrel, hydraulic recoil system, and vehicle body, contains a large amount of noise components (mainly including irregular low-frequency interference signals during movement and high-frequency resonance signals during shooting, etc.), which bring random interference signals. Since the random interference signals have a wider frequency band than normal signals, the deviation between the sampling signal of individual measurement points and the baseline is very large. Directly using the collected acceleration values to identify shooting events will affect the accuracy of the analysis results. Based on this, the present invention adopts an identification method with an adaptive sliding window, that is, dividing the collected acceleration sequence into continuous windows for processing, and dynamically updating the identification model during the processing of the acceleration sequence of each window, and using the physical characteristics of the shooting acceleration of the equipment to dynamically identify the shooting events of the equipment. The present invention uses an acceleration sensor installed on the turret of the heavy weapon to continuously collect the acceleration values of the turret, and the negative direction of the Y-axis of the acceleration sensor is consistent with the muzzle direction.

[0029] The neural network of the present invention can adopt the network structure disclosed in the prior art, preferably a feedforward neural network, and the feedforward neural network includes an input layer, a hidden layer, and an output layer connected in sequence. The initial acceleration feature sequence samples included in the training set of the present invention are the acceleration feature sequences obtained by using the methods described in steps (1)-(4) or artificially constructed acceleration feature sequences. The acceleration values in the artificially constructed sample sequences are determined according to the value range of the acceleration feature sequences of specific heavy weapons. For example, the value range can be [-6, 8], and the unit is g .

[0030] Considering that there may be errors in the results when the detection model is first used on heavy weaponry, the model of the present invention can be verified using traditional neural network verification methods or the unique method of the present invention. The unique method is: based on the inventor's actual measurement, it is found that when the turret displacement is stable during the shooting process of most heavy weaponry, the shooting events output by the updated model of the present invention are valid. Therefore, the technical idea of whether the difference in turret displacement between adjacent time windows tends to be equal is used to verify the reliability of the model, that is: during the update process, the model after a certain update training is verified to determine whether the model is reliable. The specific update method is: collect the acceleration sequence of any window of the heavy weapon, and the number of acceleration values in this sequence is , input the collected sequence into the updated heavy weapon shooting detection model to determine whether a shooting event occurs; for the output result of the model judging that a shooting event occurs, if the displacement of the turret within the corresponding input time window is equal to the displacement of the turret within the previous time window of the corresponding input time window or there is a reasonable difference between the two (such as less than 30 cm or less than 50 cm, which can be determined according to the shooting stable displacement of the heavy weapon), it is considered that the updated heavy weapon shooting detection model is reliable, otherwise, the model needs to be updated and trained continuously. The displacement of the turret within a time window can be calculated from the acceleration sequence within this time window: Turret displacement , i = 1, 2, 3, …, L; f is the frequency of collecting acceleration.

[0031] Example:

[0032] In this example, during a certain live-fire shooting training, the method of the present invention and the method of manual recording are used to identify shootings:

[0033] Using some acceleration data collected by the acceleration sensor as Figure 1 shown (the horizontal axis is the number of acceleration values, and the vertical axis is the acceleration value), after being processed by step (4), multiple groups of acceleration feature sequences are extracted, such as Figure 2 shown (the horizontal axis is the number of acceleration values, and the vertical axis is the acceleration value), Figure 2 each curve in represents an acceleration feature sequence; ; f = 5 ms, from Figure 2 it can be seen that the acceleration feature sequences obtained through step (4) are relatively stable;

[0034] This example uses a feedforward neural network. The initial training set contains acceleration sequence samples randomly constructed according to the Figure 2 shown acceleration feature sequences. The network is updated and trained using the Figure 2 shown acceleration feature sequences. The training loss is as Figure 3 shown (the horizontal axis is the number of updates, and the vertical axis is the loss error). The model loss function uses the mean square error function (MeanSquare Error), and the learning rate is , after about 200 shootings, the model loss curve begins to converge; it can be seen that after training with acceleration feature sequences for a relatively large number of times, it can be clearly seen that as the number of shootings of the shooting recognition model increases, the shooting recognition error can be rapidly reduced;

[0035] Randomly select the acceleration sequences with shooting events, such as Figure 4As shown (the horizontal axis is the number of acceleration values, and the vertical axis is the acceleration value), when input into the model trained in this embodiment, the model determines that the shooting probability is 0.914; moreover, the displacement difference of the turret within the time window of the acceleration sequence where a shooting event occurs randomly selected and the displacement of the turret within the previous time window of the time window of this sequence is 25 cm, which is less than 30 cm;

[0036] Another acceleration sequence in the actual installation motion state (i.e., no shooting event occurs) is randomly selected, such as Figure 5 As shown (the horizontal axis is the number of acceleration values, and the vertical axis is the acceleration value), when input into the model trained in this embodiment, the model determines that the shooting probability is 0.401. It can be seen from this that the model of the present invention can effectively distinguish shooting characteristics.

Claims

1. A method for constructing a shooting recognition model of a heavy weapon, characterized in that the method Including: Continuously collect the acceleration values of the turret; at the same time, sequentially divide the multiple collected acceleration values into time windows for processing. Each time window includes several sequentially collected acceleration values, and the acceleration sequence of any time window is , is any one of the acceleration values in , ; The processing method for the acceleration sequence of each time window includes: (1) Obtain the maximum acceleration value in the acceleration sequence of the current time window , and record the earliest occurrence position of the maximum acceleration value in the acceleration sequence of the current window ; At the same time, obtain the minimum acceleration value in the acceleration sequence of the current time window ; (2) Calculate the volatility of the acceleration sequence in the current time window , Wherein: ; is the average value of the acceleration sequence for the current time window, ; is the standard deviation of the acceleration sequence in the current time window, ; (3) When is true, the acceleration sequence of the current time window is considered as the acceleration sequence with an acceleration event, and step (4) is executed; otherwise, the acceleration sequence of the next time window is processed; (4)For a sequence with an acceleration event, from the position of this sequence Iterate forward to obtain the position where the acceleration value starts to mutate in the sequence with an acceleration mutation event , the acceleration at which the acceleration value starts to mutate satisfies the condition: , and then obtain the acceleration feature sequence of this sequence ; is the previous acceleration values of the next time window of the acceleration sequence with an acceleration mutation event; (5) Incorporate the obtained acceleration feature sequence into the training data set, and use the training data set to update and train the neural network trained in the previous update to obtain an updated heavy weapon shooting recognition model; the training data set initially contains multiple groups of acceleration feature sequence samples, and the acceleration feature sequence samples are acceleration feature sequences obtained by taking steps (1)-(4) or artificially constructed acceleration feature sequences. Each acceleration feature sequence sample contains acceleration values, and the neural network is initially an untrained neural network.

2. The method for constructing a heavy weapon shooting recognition model according to claim 1, wherein, It also includes: (6) validating the updated heavy weapon firing recognition model, including: collecting an acceleration sequence of any window of the heavy weapon, the number of acceleration values in the sequence being , using the collected acceleration sequence as the input of the updated heavy weapon firing recognition model, and determining whether a firing event occurs; For the output result of judging a shooting event, if the displacement of the turret within the time window of the corresponding input is equal to the displacement of the turret within the previous time window of the corresponding input time window or there is a reasonable difference between the two, it is considered that the updated heavy weapon shooting recognition model is reliable; otherwise, the model needs to be continuously updated and trained.

3. The method for constructing a heavy weapon shooting recognition model according to claim 1, characterized in that The neural network initially adopts a feedforward neural network, and the feedforward neural network includes an input layer, a hidden layer, and an output layer connected in sequence.

4. A method for identifying heavy weapon shooting, characterized in that, The method includes: collecting an acceleration sequence of any window of a heavy weapon, where the number of acceleration values in the sequence is , and inputting the collected sequence into the heavy weapon shooting detection model constructed by the method described in claim 1 to determine whether a shooting event occurs.

5. A heavy weapon shooting recognition system, characterized in that, The system includes: an acceleration acquisition module and a shooting recognition module. The acceleration acquisition module is used to acquire the acceleration value of the turret and divide the acquired multiple acceleration values according to time windows; the shooting recognition module uses the model constructed by the method described in claim 1 to recognize the shooting of heavy weapons and simultaneously updates the model by the method described in claim 1.

6. The heavy weapon shooting recognition system according to claim 5, characterized in that The shooting recognition module also verifies the model by the method described in claim 2.

Images