Decoy jamming bomb discrimination model generation method and device

Abstract

The application discloses a generation method and equipment of a decoy jamming projectile discrimination model, and belongs to the technical field of infrared homing terminal guidance. The method comprises the following steps: obtaining time-series radiation sequences and attitude sequences based on original images output by an infrared focal plane array in a seeker and attitude data of a projectile output by an inertial measurement unit; coupling the radiation sequences and the attitude sequences by using a first encoder in a preset network to obtain space-time embedding vectors; adjusting a single-layer full connection classification head in the preset network according to the space-time embedding vectors; constructing an initial discrimination model according to the first encoder and the adjusted single-layer full connection classification head; obtaining a decoy sequence according to limit values of physical parameters of a decoy jamming projectile; obtaining an intermediate discrimination model according to the initial discrimination model and the decoy sequence; and training the intermediate discrimination model by using the decoy sequence, the time-series radiation sequences and the attitude sequences to obtain a target discrimination model, which can reduce the false alarm rate of the decoy jamming projectile.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of infrared homing terminal guidance, and particularly relates to a method and device for generating a decoy jamming bomb discrimination model. BACKGROUND

[0002] In the infrared homing terminal guidance phase, the missile relies on the infrared focal plane array in the seeker to detect, track and lock the target. In a real tactical scenario, the enemy often releases decoy jamming bombs (such as high-temperature magnesium jamming bombs, metal foil jamming bombs, etc.) to interfere with the guidance system of the missile, causing it to deviate from the real target and seriously reducing the guidance accuracy and strike effectiveness.

[0003] In related technologies, the idea of combining spectral difference and motion consistency is usually used to deal with the true target-decoy jamming bomb discrimination problem in the terminal guidance phase. The typical process is: first, all high-light candidate points are extracted from the monocular infrared sequence through multi-threshold segmentation and morphological filtering, then an energy spectrum curve and a short-time velocity vector are constructed for each candidate point, these two types of features are mapped to a two-dimensional plane, and a Bayesian classification or a support vector machine is used to complete the preliminary screening; a parallel branch uses Kalman to fit the trajectory of the candidate points in consecutive frames, and outputs the targets whose trajectory direction and visual axis are consistent for a long time as a high-confidence set, and gives a true or false label after decision by a rule tree.

[0004] Although the above scheme introduces spectral difference and motion consistency at the same time, the two are usually run in independent channels, and the radiant energy and attitude dynamics are not fully coupled in the same feature representation. When the target enters a rapid rolling or lateral maneuver, the correlation between the thermal field morphology and the velocity field is weakened, and the classifier can only rely on threshold and rule tree for back-end correction. The discrimination degree of decoy jamming bombs and real targets in the single-frame scale decreases, and the false positive rate under complex cloud conditions or parallax conditions still increases significantly. SUMMARY

[0005] Embodiments of the present application provide a method and device for generating a decoy jamming bomb discrimination model, which can accurately discriminate decoy jamming bombs to some extent and reduce the false positive rate of decoy jamming bombs.

[0006] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the present application.

[0007] According to a first aspect of the embodiments of the present application, a method for generating a decoy jamming bomb discrimination model is provided, comprising:

[0008] Based on the original image output by the infrared focal plane array in the seeker and the attitude data of the missile body output by the inertial measurement unit, a time-series consistent radiation sequence and an attitude sequence are obtained;

[0009] coupling the radiation sequence and the attitude sequence by using a first encoder in a preset network to obtain a spatio-temporal embedding vector;

[0010] adjusting a single-layer fully connected classification head in the preset network according to the spatio-temporal embedding vector;

[0011] constructing an initial discrimination model according to the first encoder and the adjusted single-layer fully connected classification head;

[0012] obtaining a decoy sequence according to limit values of physical parameters of the decoy jamming projectile, wherein the decoy sequence is used to represent the radiation sequence and the attitude sequence of the decoy jamming projectile;

[0013] obtaining an intermediate discrimination model according to the initial discrimination model and the decoy sequence;

[0014] training the intermediate discrimination model by using the decoy sequence, the radiation sequence and the attitude sequence in time sequence to obtain a target discrimination model.

[0015] In some embodiments, the infrared focal plane array in the seeker includes an infrared detector, the inertial measurement unit includes a three-axis gyroscope, the interrupt trigger time of the infrared detector when the integration reading is completed is the same as the interrupt trigger time of the three-axis gyroscope when the sampling is completed, and the radiation sequence and the attitude sequence in time sequence are obtained based on the raw image output by the infrared focal plane array in the seeker and the attitude data of the projectile output by the inertial measurement unit, and include:

[0016] correcting the gain matrix and the bias matrix of the infrared focal plane array in the seeker based on the focal plane temperature of the infrared focal plane array in the seeker;

[0017] determining a linear radiation matrix according to the raw image, the corrected gain matrix and the corrected bias matrix;

[0018] rotating and projecting the current attitude data to obtain a displacement field corresponding to the current attitude data;

[0019] determining a target radiation matrix corresponding to the current attitude data by using the displacement field, the linear radiation matrix and a quadratic B-spline interpolation algorithm;

[0020] obtaining the radiation sequence and the attitude sequence in time sequence according to the current attitude data and the target radiation matrix corresponding to the current attitude data.

[0021] In some embodiments, the radiation sequence and the attitude sequence are coupled by using a first encoder in a preset network to obtain a spatio-temporal embedding vector, including:

[0022] determining a plurality of minimum prediction units according to the radiation sequence and the attitude sequence;

[0023] The pose data in each minimum prediction unit is mapped to a planar vector field, and is spliced with a corresponding target radiation matrix in the channel dimension to obtain a plurality of multi-channel tensors;

[0024] The initial network is trained by using the multi-channel tensors to obtain a preset network, wherein the initial network includes an initial encoder, a radiation decoding head and an optical flow decoding head, and the preset network includes a first encoder, the radiation decoding head and the optical flow decoding head;

[0025] The multi-channel tensors are input into the preset network to obtain a first feature vector output by the first encoder, and the first feature vector is determined as a spatio-temporal embedding vector.

[0026] In some embodiments, the initial network is trained by using the multi-channel tensors to obtain the preset network, including:

[0027] The multi-channel tensors are input into the initial network to obtain a second feature vector output by the initial encoder, a radiation matrix prediction value output by the radiation decoding head and an optical flow prediction value output by the optical flow decoding head;

[0028] A same-group positive clustering center is determined according to the second feature vector corresponding to the multi-channel tensors of the current minimum prediction unit;

[0029] A cross-group negative clustering center is determined according to the second feature vector corresponding to the multi-channel tensors of any minimum prediction unit except the current minimum prediction unit in the plurality of minimum prediction units;

[0030] The radiation matrix prediction value and the optical flow prediction value corresponding to the multi-channel tensors of the current minimum prediction unit, the same-group positive clustering center and the cross-group negative clustering center are substituted into a loss function of the initial network, and in a case where a value of the loss function is lower than a preset threshold, it is determined that the loss function converges, and the preset network is obtained.

[0031] In some embodiments, the single-layer fully connected classification head in the preset network is adjusted according to the spatio-temporal embedding vector, including:

[0032] The spatio-temporal embedding vectors are clustered and screened to obtain a first spatio-temporal embedding vector, and a label corresponding to the first spatio-temporal embedding vector is obtained, wherein the label includes a first label for representing a real target and a second label for representing a decoy jamming bomb;

[0033] In a case where parameters of the first encoder are unchanged, the single-layer fully connected classification head in the preset network is adjusted according to the first spatio-temporal embedding vector and the label corresponding to the first spatio-temporal embedding vector;

[0034] Vectors in the spatio-temporal embedding vectors except the first spatio-temporal embedding vector are taken as second spatio-temporal embedding vectors, and a weak augmentation process is performed on original images corresponding to the second spatio-temporal embedding vectors to obtain a pseudo-label candidate set;

[0035] screening weak augmented images and pseudo labels corresponding to the weak augmented images from the pseudo label candidate set according to a preset decreasing threshold sequence;

[0036] performing strong augmentation processing on the weak augmented images to obtain strong augmented images;

[0037] adjusting a single-layer fully connected classification head in the preset network according to the pseudo labels corresponding to the weak augmented images and the strong augmented images, under the condition that parameters of the first encoder remain unchanged.

[0038] In some embodiments, the decoy sequence is obtained according to limit values of physical parameters of the decoy jamming bomb, and includes:

[0039] constructing a parameter generator according to the limit values of the physical parameters of the decoy jamming bomb;

[0040] obtaining the decoy sequence according to parameters output by the parameter generator.

[0041] In some embodiments, the intermediate discriminant model is obtained according to the initial discriminant model and the decoy sequence, and includes:

[0042] unfreezing the second-to-last layer of the first encoder in the initial discriminant model and the single-layer fully connected classification head in the initial discriminant model to construct the adversarial discriminator;

[0043] determining a confusion probability of the decoy sequence according to the adversarial discriminator;

[0044] training the adversarial discriminator according to the confusion probability, continuing to train the adversarial discriminator using the decoy sequence, the time-sequentially consistent radiation sequence and the attitude sequence after a first preset condition is reached, and constructing the intermediate discriminant model according to the first encoder and the trained adversarial discriminator after a second preset condition is reached.

[0045] In some embodiments, the target discriminant model is obtained by training the intermediate discriminant model using the decoy sequence, the time-sequentially consistent radiation sequence and the attitude sequence, and includes:

[0046] inserting the decoy sequence into the time-sequentially consistent radiation sequence and the attitude sequence according to a time axis to obtain a synthetic sequence;

[0047] adjusting the last two layers of the first encoder in the intermediate discriminant model according to the synthetic sequence to obtain a second encoder;

[0048] adjusting the single-layer fully connected classification head of the adversarial discriminator in the intermediate discriminant model according to the original image corresponding to the first spatio-temporal embedding vector;

[0049] constructing the target discriminant model according to the adjusted second encoder and the adjusted adversarial discriminator.

[0050] In some embodiments, the method for generating the decoy jamming projectile discrimination model further comprises:

[0051] decomposing the three-dimensional convolution kernel of the target discrimination model into an intra-frame two-dimensional convolution kernel and an inter-frame one-dimensional convolution kernel;

[0052] According to the contribution degree of the output channel of the target discrimination model to the back propagation gradient corresponding to the difficult example sequence, the output channel is pruned, wherein the difficult example sequence is the decoy sequence with a medium confidence degree;

[0053] The target discrimination model is subjected to inference delay time and anti-decoy precision detection;

[0054] In the case that the inference delay time is less than the preset time and the anti-decoy precision is greater than the preset precision, the target discrimination model is deployed.

[0055] According to a second aspect of the embodiments of the present application, a device for generating a decoy jamming projectile discrimination model is provided, comprising a processor and a memory, the memory storing computer program instructions capable of being executed by the processor, and the processor implements the steps of the method of any one of the above first aspect when executing the computer program instructions.

[0056] In the present application, based on the original image output by the infrared focal plane array in the seeker and the attitude data of the projectile output by the inertial measurement unit, time-sequential radiation sequences and attitude sequences are obtained; the radiation sequences and the attitude sequences are coupled by using a first encoder in a preset network to obtain space-time embedding vectors; a single-layer fully connected classification head in the preset network is adjusted according to the space-time embedding vectors; an initial discrimination model is constructed according to the first encoder and the adjusted single-layer fully connected classification head; a decoy sequence is obtained according to the limit value of the physical parameters of the decoy jamming projectile; an intermediate discrimination model is obtained according to the initial discrimination model and the decoy sequence; the intermediate discrimination model is trained by using the decoy sequence, the time-sequential radiation sequences and the attitude sequences to obtain a target discrimination model, which can reduce the false positive rate of the decoy jamming projectile.

[0057] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0058] The drawings herein are incorporated into the specification and form a part of the specification, show embodiments consistent with the present application, and together with the specification serve to explain the principles of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. In the drawings:

[0059] Figure 1A flowchart of a method for generating a decoy jamming bomb discrimination model is shown according to some embodiments of the present application.

[0060] Figure 2 A block diagram of a device for generating a decoy jamming bomb discrimination model is shown according to some embodiments of the present application.

[0061] Figure 3 A structural diagram of a device for generating a decoy jamming bomb discrimination model is shown according to some embodiments of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0063] In addition, the described features, structures or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, many specific details are provided to give a sufficient understanding of the embodiments of the present application. However, those skilled in the art will realize that the technical solutions of the present application can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. can be used. In other cases, well-known methods, devices, implementations or operations are not shown or described in detail to avoid obscuring the aspects of the present application.

[0064] The block diagrams shown in the drawings are only functional entities, which do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0065] The flowcharts shown in the drawings are only exemplary illustrations, which do not necessarily include all contents and operations / steps, and are not necessarily executed in the order described. For example, some operations / steps can be further decomposed, and some operations / steps can be combined or partially combined, so that the actual execution order can be changed according to the actual situation.

[0066] Figure 1 A flowchart of a method for generating a decoy jamming bomb discrimination model is shown according to some embodiments of the present application. As shown, Figure 1 a method for generating a decoy jamming bomb discrimination model is provided, which can include the following steps 101 to 107.

[0067] In step 101, based on the original image output by the infrared focal plane array in the seeker and the attitude data of the projectile output by the inertial measurement unit, a time-sequential radiation sequence and an attitude sequence are obtained.

[0068] In the formula, the infrared focal plane array in the seeker can include a plurality of infrared detectors for outputting the original image; the inertial measurement unit can include a three-axis gyroscope and a digital signal processor (DSP); and the attitude data can be attitude quaternions. After the three-axis angular velocity is sampled by the three-axis gyroscope, the three-axis angular velocity is trapezoidal integrated by the DSP, and the attitude quaternions can be obtained.

[0069] In the implementation process, a single crystal oscillator clock can be configured in the seeker, and all peripheral interrupts are driven by the clock. An interrupt is triggered when the infrared detector completes integration and reading and the time T IR (t) is recorded, and an interrupt is triggered when the three-axis gyroscope completes a sampling and the time T IMU (k) is recorded. The two types of interrupts are written into the same time stamp register, so that T IR (t) = T IMU (k) = τ c . In this way, the original image and the attitude data share the global clock.

[0070] In some embodiments, monotonicity detection can be performed on the τ c sequence on the data bus side. If τ c (t+1) - τ c (t) ≠ 1 / F IR (F IR is the imaging frame rate) or any out-of-order event is detected, a placeholder frame D gap can be immediately inserted at the gap position, and a mask M gap = 1 is generated.

[0071] In some embodiments, the gain matrix and the bias matrix of the infrared focal plane array in the seeker can be corrected based on the focal plane temperature of the infrared focal plane array in the seeker; a linear radiation matrix can be determined according to the original image, the corrected gain matrix and the corrected bias matrix; a current attitude data is rotated and projected to obtain a displacement field corresponding to the current attitude data; a target radiation matrix corresponding to the current attitude data is determined by using the displacement field, the linear radiation matrix and a quadratic B-spline interpolation algorithm; and a time-sequential radiation sequence and an attitude sequence are obtained according to the current attitude data and the target radiation matrix corresponding to the current attitude data.

[0072] It is understandable that infrared focal plane arrays undergo multi-temperature calibration using a blackbody source at the factory to obtain the gain matrix G0 and bias matrix O0. During online operation, due to fluctuations in operating temperature, the pixel response exhibits linear drift. The focal plane temperature Θ can be acquired simultaneously with each frame of data readout. t And utilize gain drift ΔG(Θ) t ) and bias drift ΔO(Θ t Incremental corrections are applied to the gain matrix G0 and the bias matrix O0 respectively. Then, the linear radiation matrix is ​​calculated based on the original digital counting matrix of the original image, the corrected gain matrix, the corrected bias matrix, and the following formula:

[0073] R t =[G0+ΔG(Θ)] t )]°D t +[O0+ΔO(Θ t )];

[0074] in, It is a linear radiation matrix. This is the original numerical counting matrix, where ° represents element-wise multiplication, and H×W is the focal plane size. Through linear mapping, any radiation source within the scene can be represented within the thermal drift range (Θ). min ,Θ max Maintaining a constant grayscale scale within the range provides a radiation conservation premise for comparison pre-training.

[0075] Focal plane temperature Θ t The closed-loop temperature control module can update at 100Hz. The device generating the decoy / flare discrimination model can periodically read the focal plane temperature Θ. t And calculate the gain drift ΔG(Θ) t ) = A G Θ t +B G Bias drift ΔO(Θ) t ) = A O Θ t +B O Matrix A G B G A O B O All data are derived from least squares regression results of experimental data from the full-temperature range of the hot plate test. The EEPROM calibration coefficients are checked daily to prevent aging.

[0076] The inertial measurement unit outputs the current attitude data Q. t Then, it can be compared with the pose data Q of the original image from the previous frame. t-1 Constructing relative rotation Let the horizontal field of view of the image be f ov The focal plane size is H×W. Using the projection operator Π(·), Q is...Δt mapping to the imaging plane, obtaining a sub-pixel level displacement field (u t ,v t ), wherein Π(Q Δt ,f ov ,H,W) is directly calculated by a three-dimensional rotation-perspective projection equation.

[0077] The above displacement field is applied to a linear radiation matrix R t , and a quadratic B-spline interpolation kernel is used to realize sub-pixel level resampling to obtain a target radiation matrix wherein is a sub-pixel level rearrangement operator. In hardware implementation, a field-programmable gate array (FPGA) can perform interpolation on each pixel scan line in a streaming pipeline manner, realizing sub-pixel level resampling of the thermal field in the FPGA pipeline and eliminating optical axis jitter.

[0078] In the implementation process, the radiation sequence and the attitude sequence can be output in a sliding window manner. In some examples, the length N s of the sliding window can be defined as 16 frames, and a cache queue The target radiation matrix S t after jitter elimination and the attitude data Q t are written in time sequence using a double-ended ring structure. When , a is generated and pushed to a thermal-motion coupling self-supervised pre-training module. At the same time, the radiation sequence and the attitude sequence are provided as two information streams, so that the encoder can perform cross-modal contrastive learning on the same alignment reference in the subsequent training process, realizing high-order coupling of thermal motion.

[0079] Through the above processing, it can be ensured that the infrared radiation matrix and the motion vector field are strictly aligned at each time index, and the jitter error is compressed to a sub-pixel level scale, providing stable input for subsequent cross-modal feature coupling, which is different from the loose coupling mode of the prior art in which the two data are processed and then synchronized.

[0080] In step 102, the radiation sequence and the attitude sequence are coupled using a first encoder in a preset network to obtain a spatio-temporal embedding vector.

[0081] The preset network can include a first encoder E φ , a radiation decoding head , and an optical flow decoding head The initial network is trained (including an initial encoder ε φ , a radiation decoding head , and an optical flow decoding head It is obtained that the encoders in both networks can output feature vectors, the radiation decoder can output radiation matrix predictions, and the optical flow decoder can output optical flow predictions. The radiation decoder and the optical flow decoder can share all backbone parameters. Step 102 can be completed in the thermal-motion coupled self-supervised pre-training module.

[0082] In some embodiments, step 102 may include the following sub-steps:

[0083] Step 1021: Determine multiple minimum prediction units based on the radiation sequence and attitude sequence;

[0084] Step 1022: After mapping the attitude data in each minimum prediction unit to the planar vector field, concatenate it with the corresponding target radiation matrix according to the channel dimension to obtain multiple multi-channel tensors;

[0085] Step 1023: Train the initial network using multi-channel tensors to obtain the preset network. The initial network includes an initial encoder, a radiative decoding head, and an optical flow decoding head. The preset network includes a first encoder, a radiative decoding head, and an optical flow decoding head.

[0086] Step 1024: Input the multi-channel tensor into the preset network to obtain the first feature vector output by the first encoder, and determine the first feature vector as the spatiotemporal embedding vector.

[0087] In step 1021, it is possible to retrieve from the queue The radiation-attitude sequence corresponding to k consecutive original images is sampled using frame order indexing {(S t-k+1 Q t-k+1 ),…,(S t Q t )}, and reserve an empty slot for the next frame. Combined into the smallest prediction unit

[0088] In step 1022, the pose data Q corresponding to each frame of the original image can be processed. τ Calculate the instantaneous angular velocity Ω τ It is then transformed into a planar vector field through perspective mapping Π(·). Then, the corresponding target radiation matrix S is assembled according to the channel dimension. τ With planar vector field Obtain a two-dimensional multichannel tensor It achieves pixel-level coupling of thermal and motion information, so that the multi-channel tensor contains both the spatial distribution characteristics of the thermal field (i.e., heat) and the motion vector information that evolves over time (i.e., motion).

[0089] In step 1023, the multi-channel tensor can be input into the initial network to obtain the second feature vector of the initial encoder output, the radiation matrix prediction value of the radiation decoding head output, and the optical flow prediction value of the optical flow decoding head output; the same group positive clustering center is determined according to the second feature vector corresponding to the multi-channel tensor of the current minimum prediction unit; the cross-group negative clustering center is determined according to the second feature vector corresponding to the multi-channel tensor of any minimum prediction unit except the current minimum prediction unit in the plurality of minimum prediction units; the radiation matrix prediction value and the optical flow prediction value corresponding to the multi-channel tensor of the current minimum prediction unit, the same group positive clustering center, and the cross-group negative clustering center are substituted into the loss function of the initial network, and in the case that the value of the loss function is lower than a preset threshold, it is determined that the loss function converges, and a preset network is obtained.

[0090] Wherein, the initial encoder is only different in parameters compared with the first encoder, and the others are the same.

[0091] It can be understood that the input multi-channel tensor X t-k+1:t After that, the radiation matrix prediction value The initial encoder learns the time migration characteristics of the heat distribution along the inertial path guided by the L1 distance, and the goal is to make the initial network not only remember the static texture, but also perceive the evolution direction and rate of the radiation intensity under the dynamics driving.

[0092] Based on the optical flow decoding head, the current optical flow prediction value The optical flow true value F t The self-supervised estimator based on three-frame difference can be used for online generation, without external labeling.

[0093] The parameters of the dual decoding head shared encoder force the initial encoder ε φ The heat and motion are represented in the same latent space at the same time, avoiding the defect that the traditional self-supervised method only captures static texture and ignores dynamic information for small size targets.

[0094] The second feature vectors output by the initial encoder for the internal frames of the same minimum prediction unit The intra-group average is performed to obtain the same group positive clustering center The second feature vectors output by the initial encoder for the internal frames of the minimum prediction unit of another time period are randomly selected again, and the intra-group average is performed to obtain the cross-group negative clustering center The contrast loss reduces z τ The Euclidean distance between z and z τ is compared, and the distance between z and z τ is enlarged, further suppressing the background static redundancy and highlighting the maneuver-radiation coupling features.

[0095] The loss function of the initial network can be set as:

[0096]

[0097] wherein, is the radiometric matrix prediction value; S τ+1 is the radiometric matrix ground truth value; is the optical flow prediction value; F τ is the optical flow ground truth value; z τ is the second feature vector; is the same-group positive clustering center; is the cross-group negative clustering center; a, b, g are weight coefficients; m is the contrast loss safety interval; [·] + denotes the ReLU operation.

[0098] By minimizing the encoder can obtain a heat-motion coupling representation capability under the condition of no artificial labeling, and provide a high-fidelity feature basis for subsequent few-labeled semi-supervised discriminant training.

[0099] In each iteration cycle, the errors of the radiometric matrix prediction value and the optical flow prediction value can also be counted, and the weight coefficients of the top p high-loss samples in the error ranking are refreshed and backfilled into the training list, so as to accelerate the adaptation of the model to high-acceleration maneuver or sharp temperature rise decoy scenes through a hard excavator mechanism, and improve the memory persistence of the encoder to extreme terminal guidance working conditions.

[0100] In step 1024, when the value of the loss function is lower than the set threshold value ∈ stop , it indicates that the initial network has been trained, at this time the parameters of the encoder φ are frozen, and the multi-channel tensor is input into the preset network, the first encoder E φ will output the spatio-temporal embedding vector z τ =E φ (X τ ) for subsequent few-labeled semi-supervised discriminant training. The spatio-temporal embedding vector simultaneously encodes the radiation field pattern, the instantaneous motion vector and the local contrast semantics in the same vector space, forming a stable foundation for few-labeled learning.

[0101] In step 103, the single-layer fully connected classification head in the preset network is adjusted according to the spatio-temporal embedding vector.

[0102] In step 104, an initial discriminant model is constructed according to the first encoder and the adjusted single-layer fully connected classification head.

[0103] In the implementation process, representative vectors can be extracted from the space-time embedding vectors, and binary artificial labels are injected into the original images corresponding to the vectors, including a first label for representing a real target and a second label for representing a decoy jamming bomb. In the case of freezing the parameters of the encoder, a single-layer fully connected classification head is added in the preset network, and the single-layer fully connected classification head is adjusted by using the labels. Steps 103 and 104 can be completed in the few-labeled semi-supervised discrimination training module, which generates an initial discrimination model in the case of few labels.

[0104] In some embodiments, step 103 can include the following sub-steps:

[0105] Step 1031, clustering and screening the space-time embedding vectors to obtain first space-time embedding vectors, and obtaining labels corresponding to the first space-time embedding vectors, wherein the labels include a first label for representing a real target and a second label for representing a decoy jamming bomb;

[0106] Step 1032, in the case of unchanged parameters of the first encoder, adjusting a single-layer fully connected classification head in the preset network according to the first space-time embedding vectors and the labels corresponding to the first space-time embedding vectors;

[0107] Step 1033, taking vectors other than the first space-time embedding vectors in the space-time embedding vectors as second space-time embedding vectors, and performing weak augmentation processing on original images corresponding to the second space-time embedding vectors to obtain a pseudo-label candidate set;

[0108] Step 1034, according to a preset decreasing threshold sequence, screening a weak augmented image and a pseudo-label corresponding to the weak augmented image from the pseudo-label candidate set; performing strong augmentation processing on the weak augmented image to obtain a strong augmented image;

[0109] Step 1035, in the case of unchanged parameters of the first encoder, adjusting the single-layer fully connected classification head in the preset network according to the pseudo-label corresponding to the weak augmented image and the strong augmented image.

[0110] It should be noted that, in order to make up for the problem of scarce real tactical scene labels, the industry generally adopts the strategy of large-scale synthetic sequence pre-training plus a small amount of field data migration. However, the synthetic sequence has an indiscernible domain difference in spectral distribution, tail shape and ballistic noise compared with the live scene, which makes it difficult for the model to maintain the discrimination boundary during the training period within the real combat weather window. The limitation of computing power and safety gap leads to deep pruning and solidification of the backbone network, allowing only the end classification layer to be fine-tuned online. This local update lacks adaptive ability to sudden changes in the coupling of jamming bomb radiation surface and motion.

[0111] In step 1031, the entire space-time embedding vector sequence can be clustered and screened Compute the pairwise Euclidean distance matrix of all spatio-temporal embedding vectors, then perform K-Medoids clustering with the number of cluster centers Thus, global coverage is achieved in three dimensions: scene, weather type, and maneuvering posture.

[0112] Lock the original image frame index corresponding to each cluster center The field test operator only reviews these original image frames and assigns a binary label y τ ∈{1,0}, where 1 represents a real target and 0 represents a decoy jamming bomb.

[0113] In step 1032, the first encoder is frozen and a single-layer fully connected classification head is trained. In this sub-step, the parameters φ of the first encoder are all frozen, and a single-layer fully connected classification head is added to the pre-set network The input dimension is consistent with the embedding dimension, and the output dimension is 2.

[0114] The spatio-temporal embedding vectors e i of all labeled original image frames are obtained. i The weighted cross-entropy objective function is constructed as:

[0115]

[0116] where J init is the initial loss function; N L is the number of manually labeled samples; is the adaptive weight based on the frequency of class occurrence; and represent the prediction probabilities of the classification head for the real target class and the decoy jamming bomb class, respectively.

[0117] By minimizing J init , the first version of the decision hyperplane is given under the condition of sparse labels, laying the foundation for subsequent confidence assessment.

[0118] In step 1033, pseudo labels are generated by weak augmentation. In this sub-step, the weights of the single-layer fully connected classification head are kept fixed, and two kinds of weak augmentation processing, brightness jitter and fine-grained rotation, are applied to the unlabeled original image frames to obtain weakly augmented images A w (·). Then, the output probability is calculated. The highest output probability is recorded as the confidence, and the triple is included in the pseudo label candidate set U1, where

[0119] Since weak augmentation does not destroy the geometric characteristics of the scene, the confidence can truly reflect the decision reliability of the classification head on the original features.

[0120] In step 1034, a decreasing threshold sequence {θ (r)} is set for the pseudo label candidate set U1, and in each iteration process, the threshold θ (r) is adjusted according to the current high confidence ratio σ (r+1) = max(θ floor , θ (r) - λσ (r) ), where θ floor is a safe lower bound that cannot be lowered; λ is the decay factor. Only when c τ ≥ θ (r+1) and the proportion of two-class samples is kept in the interval [ρ, 1-ρ], the weak augmented image x τ and the pseudo label are supplemented into the training buffer pool In this way, the mislabeling flow can be controlled, and the class balance can be stabilized to avoid the encoder forgetting the rare classes.

[0121] In step 1035, each frame of image in the training buffer pool is additionally applied with occlusion, random cropping and time disorder combination to form a strong augmented image A s (·), and the prediction distribution of the same frame image in weak and strong views remains consistent. The Kullback-Leibler divergence is used as a regularization term to join the loss, which improves the robustness of the model to occlusion, flicker and sudden changes in view, and prevents the weight from deviating by the consistency constraint.

[0122] In step 1036, after training a round using the strong augmented image and the weak augmented image corresponding to the pseudo label, the false positive rate ε (r) is calculated on the validation set. If ε (r) > ε (r-1) , the classifier weight is immediately rolled back to the last snapshot, and the threshold is raised to θ (r+1) = θ (r) + 2λ.

[0123] In the implementation process, the last layer of the first encoder can also be unfrozen and fine-tuned jointly, and when the absolute value of the difference between two consecutive rounds |ε (r) - ε (r-1) | < δ stop (where δ stop is a preset stop threshold), the parameters φ tail of the last layer of the first encoder and the parameters of the single-layer fully connected classification head are unlocked. The parameters are fine-tuned with a low learning rate and WeightDecay until the validation loss no longer decreases, and finally the parameters are frozen to obtain the initial discrimination model where φ ＊ , represents the global optimal parameter. The initial discrimination model has both thermal-movement prior and sparse supervision adaptation capabilities, and can provide stable and high-confidence real target / decoy jamming bomb discrimination base in the subsequent adversarial reasoning stage.

[0124] By performing a small amount of binary labeling on representative original images, freezing the encoder and training a single-layer fully connected classification head, and then expanding the labeling domain with weak augmented pseudo labels and strong augmented consistency regularization, the constraint of difficult labeling in real tactical scenarios is broken.

[0125] In step 105, according to the limit value of the physical parameter of the decoy jamming bomb, the decoy sequence is obtained, wherein the decoy sequence is used to represent the radiation sequence and the attitude sequence of the decoy jamming bomb.

[0126] In step 106, according to the initial discrimination model and the decoy sequence, the intermediate discrimination model is obtained.

[0127] Steps 105 and 106 can be completed in the adversarial difficult example incremental training module, which generates an adversarially enhanced intermediate discrimination model using the decoy sequence.

[0128] Wherein, the physical parameters of the decoy jamming bomb can include initial temperature T0, temperature decay constant k T , effective area of burning surface A f , initial speed of ejection v e , charge mass m f , tail evaporation rate η b , particle dispersion scale σ s , spectral band specific radiation coefficient ε, etc.

[0129] In the implementation process, the infrared radiation curve, the flame spectrum and the trajectory of the target range measured magnesium jamming bomb can be obtained, and the related data of the above eight physical parameters can be counted, and then the upper and lower bound vectors

[0130] L≤p=[T0,k T ,A f ,v e ,m f ,η b ,σ s ,ε] T ≤U

[0131] Wherein, p is the physical parameter vector to be generated. The inequality ensures that the subsequent samples remain tactically realistic and do not violate the law of thermal conservation.

[0132] In some embodiments, step 105 can include the following sub-steps:

[0133] Step 1051, constructing a parameter generator according to the limit value of the physical parameter of the decoy jamming bomb;

[0134] Step 1052, obtaining the decoy sequence according to the parameter output by the parameter generator.

[0135] Defining the parameter generator where d z represents the Gaussian latent code dimension, 6-dimensional prior vector consists of the distance d t between the current true target and the seeker, radial velocity and attitude Euler angles (α t , β t , γ t ), and the remaining flight time ξ t .

[0136] Each round is sampled from a normal distribution After Output and combined with the upper and lower bound vectors to force p to be within the physical boundary.

[0137] In step 1052, the parameter p output by the parameter generator and the attitude prior can be sent to a fully differentiable infrared renderer to render the decoy sequence The resolution and timestamp of the decoy sequence are completely aligned with the real data stream.

[0138] The infrared renderer can include the following sub-modules inside: an energy-conservative radiation transmission sub-module to convert T0,k T , A f , m f , ε into a time-gray surface; a ballistic sub-module based on Newtonian aerodynamics to calculate a three-dimensional trajectory using v e , σ s ; and a wake sub-module based on a volatile decay model to modulate the decay radius with η b .

[0139] In some embodiments, step 106 can include the following sub-steps:

[0140] Step 1061, thawing the second-to-last layer of the first encoder in the initial discriminant model and the single-layer fully connected classification head in the initial discriminant model to construct an adversarial discriminator;

[0141] Step 1062, determining the confusion probability of the decoy sequence according to the adversarial discriminator;

[0142] At step 1063, the adversarial discriminator is trained according to the confusion probability. After a first preset condition is reached, the adversarial discriminator is continuously trained using the decoy sequence, the time-consistent radiation sequence, and the attitude sequence. After a second preset condition is reached, the intermediate discrimination model is constructed according to the first encoder and the trained adversarial discriminator.

[0143] In step 1061, the decoy sequence can be input into the first encoder of the initial discrimination model to obtain a spatiotemporal embedding vector The parameters φ of the second-to-last layer of the first encoder of the initial discrimination model are unfrozen tail The parameters of the single-layer fully connected classification head of the initial discrimination model are The adversarial discriminator is composed of

[0144] In step 1062, the label y = 1 of the real target is taken for each spatiotemporal embedding vector in step 1061, and the confusion probability is calculated The closer the confusion probability is to 1, the more the decoy jamming projectile resembles the real target.

[0145] In step 1063, the parameter generator and the adversarial discriminator are game-optimized. The adversarial target joint optimization is adopted to maximize the confusion of the parameter generator and minimize the confusion of the adversarial discriminator while maintaining the physical rationality. The adversarial target joint training is as follows:

[0146]

[0147] Wherein, λ phys is a physical rationality penalty coefficient; the first term measures the degree of model being misled by the decoy; and the second term constrains p back to the legal interval in the gradient backpropagation process to prevent extreme parameters that violate the laws of thermodynamics. The game iteration is performed until The distance to the boundary of the real target cluster is reduced to a preset threshold δ adv (i.e., the first preset condition).

[0148] It can be understood that the time-consistent radiation sequence and the attitude sequence are real sequences. The decoy sequence and the real sequence can be mixed at a ratio of 1:1 to continuously train the unfrozen parameter group until the false positive rate of the verification set decreases and stabilizes (i.e., the second preset condition). This process expands the decision boundary to the safe side, weakens the false positive risk of high-similarity decoy jamming projectiles, and finally outputs the intermediate discrimination model that is strengthened by the confrontation

[0149] In step 107, the intermediate discrimination model is trained using the decoy sequence, the time-consistent radiation sequence, and the attitude sequence to obtain the target discrimination model.

[0150] Understandably, the decoy sequence can be inserted into the real sequence to obtain a synthetic sequence, which can then be used to train the intermediate discrimination model again, ultimately yielding the target discrimination model. After obtaining the target discrimination model, it can be used to distinguish decoy flares during the terminal guidance phase of infrared homing. Step 107 can be completed in the difficult example insertion and re-pre-training module.

[0151] In some embodiments, step 107 may include the following sub-steps:

[0152] Step 1071: Insert the decoy sequence into the radiation sequence and attitude sequence that are in the same time sequence according to the time axis to obtain the synthetic sequence;

[0153] Step 1072: Adjust the last two layers of the first encoder in the intermediate discriminant model according to the synthesized sequence to obtain the second encoder;

[0154] Step 1073: Adjust the single-layer fully connected classification head of the adversarial discriminator in the second encoder and intermediate discriminant model according to the original image corresponding to the first spatiotemporal embedding vector;

[0155] Step 1074: Construct a target discrimination model based on the adjusted second encoder and the adjusted adversarial discriminator.

[0156] In step 1071, a time index set T of the real sequence can be constructed. real ={t1,t2,…,t M}, where t i = i·Δt, where Δt is a fixed frame period; Read the decoy sequence Search for the closest one in sequence Original empty frame index If the position is already occupied, search forward along the timeline until an unfilled slot is found; insert the difficult-to-insert frame at the desired location. Then, the synthetic sequence is regenerated. Where y t Label the target with 0-1 tags (1 represents the real target, 0 represents the decoy / flare); record the interpolation label matrix M∈{0,1} M , used for explicit grouping of real target-decoy frames.

[0157] In step 1072, a partial thawing thermal-motor contrast retraining is performed. In this sub-step, the first L-2 layers of the first encoder are frozen, and only the parameter sets of the last two layers are unlocked. exist The thermal-motion dual-contrast pre-task is executed in parallel, specifically: the temperature surface prediction sub-task uses the target radiation matrix S. t The supervised network captures the temporal patterns of radiation; the ballistic vector prediction subtask uses attitude data Q.t Motivate network to reserve motion dynamics prior; add two positive and two negative difficult case separation regularization: construct positive-negative pair for real target-decoy frames in the same moment, let network actively pull apart the distance between them in local embedding space. Overall multi-target loss is:

[0158]

[0159] wherein: is the re-pretrained prediction output; are the embedding vectors of the same frame before and after re-pretraining, respectively; denotes the real target frame index set, denotes the decoy frame index set; is the real target embedding centroid; α, β, γ, κ are hyperparameters.

[0160] The first two terms maintain thermal-motion physical consistency, the third term protects old memories, and the fourth term deliberately punishes difficult cases from approaching real targets.

[0161] The decoy average embedding distance increment ΔD can be calculated in real time by the following formula, and the exponential moving average is used to eliminate random oscillations:

[0162]

[0163] When ΔD≥δ gap (where δ gap is the preset separation threshold) and When a monotonic decreasing inflection point appears in the last K mini-batches, all trainable weights are immediately frozen to obtain the second encoder

[0164] The separation effect of decoy interference bombs and real targets is quantified in real time, and training is stopped as soon as the separation reaches the threshold to avoid overfitting and unnecessary iterations.

[0165] In step 1073, the original image (which can also be understood as a representative frame) set corresponding to the first spatiotemporal embedding vector pair is obtained The set is used for weak-strong augmentation consistency training. Specifically, weak augmentation (brightness jitter, fine angle rotation) is used to ensure stable basic distribution; strong augmentation (occlusion, random cropping, time sequence disorder) is used to force the single-layer fully connected classification head of the intermediate discriminant model's adversarial discriminator to maintain consistent prediction under large deformation scenarios; only the two-layer classification head is optimized Weighted cross-entropy plus KL consistency regularization is used without modifying the encoder main body to force the prediction distribution under strong and weak views to remain consistent; the false positive rate ε val is monitored on the validation set, and if there is no significant decrease in 20 consecutive mini-batches, the fine-tuning is ended.

[0166] After the fine-tuning is completed, the remaining difficult example sequences can also be scanned and progressively superimposed using the new model Full-volume scan data pool: record the confidence of all decoy sequences The samples with c t ∈(0.45,0.55) are merged into the difficult example set as difficult example sequences For Backward search generates parameters And stores them in the parameter pool The Join the training buffer to provide a more acute sample base for the next round of circulation. Circulation steps 1072 and 1073 are executed, and the following double-standard judgment is used to determine the end of training: the proportion of residual difficult example sequences decreases: (Where M is the total number of samples in a full-volume scan, ε stop is the preset false alarm rate) for two consecutive rounds; and the verification set false alarm rate is stable: (Where η flat is the flat threshold of the verification set false alarm rate) meets the two conditions to solidify the encoder version And the final classification head

[0167] In step 1074, the target discrimination model is constructed according to the adjusted encoder And the adjusted adversarial discriminator

[0168] The target discrimination model can be closed-loop verified in the terminal guidance simulation chain. Specifically, the target discrimination model can be embedded in a six-degree-of-freedom terminal guidance simulation chain, keeping the embedded-decision module and the guidance loop synchronized at a frequency of 200Hz; a strong interference scene matrix Ω sim is constructed, including multiple ballistic incidence angles, rapid attitude roll, and thermal-cloud quantity double masking cap Run the simulation for each scene and real-time statistics of capture delay τ sim , decoy false capture rate ε sim , and miss distance Δr; compare with the control model for evaluation, if ε baseline ≤0.4ε cap and the increase of τ J is less than 3%, it is determined that the technical gain meets the engineering indicators; output the test report and software version label, and complete the whole chain delivery.

[0169] The embodiment of the application adopts a multi-channel tensor entering a double-task self-supervised framework based on thermal field-motion field channel splicing, the same encoder simultaneously undertakes future radiation prediction and current optical flow estimation, and two decoding heads share all backbone parameters. During the training process, the frame-level feature vector output by the encoder performs intra-group averaging within a sliding window to form a same-group forward clustering center, and then is compared with a randomly selected cross-group negative clustering center, so that radiation and dynamics are jointly represented in a unified latent space through a joint loss function. After self-supervision, only a small amount of binary labeling is performed on the representative frames, the encoder is frozen and a single-layer fully connected classification head is trained, and then weakly augmented pseudo labels and strongly augmented consistency regularization are used to progressively expand the labeling domain. Further, decoy sequences are generated through a fully differentiable physical rendering pipeline, and participate in the adversarial game with the real sequences; after convergence, the decoy sequences are inserted back into the timeline of the real sequences, the tail layer parameters are partially unfrozen, and two positive and two negative difficult example separation regularizations are added to continuously increase the distance between the real target and the decoy in the embedding space. The integrated training chain realizes high-order coupling of radiation and motion at the encoder level, adversarial difficult example reinforcement, and sparse supervision adaptation, which is significantly different in structure from the existing methods of spectral threshold, motion filtering, and experience rule partitioning, and does not help each other. The decoy jamming bomb determination accuracy is effectively improved, and the false positive rate is reduced.

[0170] After obtaining the target model, the method for generating the decoy jamming bomb discrimination model can further include step 108 of compressing and deploying the target discrimination model. In some embodiments, step 108 can include the following sub-steps:

[0171] Step 1081, decomposing the three-dimensional convolution kernel of the target discrimination model into an intra-frame two-dimensional convolution kernel and an inter-frame one-dimensional convolution kernel;

[0172] Step 1082, pruning the output channels according to the contribution degree of the back propagation gradient corresponding to the difficult example sequence to the output channels, wherein the difficult example sequence is a decoy sequence with a confidence in the middle;

[0173] Step 1083, performing inference delay time and anti-decoy accuracy detection on the target discrimination model;

[0174] Step 1084, deploying the target discrimination model in the case that the inference delay time is less than the preset time and the anti-decoy accuracy is greater than the preset accuracy.

[0175] In step 1081, the target discrimination model is subjected to hierarchical time consumption profiling, and in this sub-step, the hardware counters of the FPGA-DSP collaborative environment can be collected to record the measured time consumption vector l = [l1, l2, …, ln] of each layer of the target discrimination model. J ] Twhere J is the number of network layers. The temporal grouping of the three-dimensional convolution kernel of the target discrimination model is reconstructed, and the original three-dimensional convolution kernel is decomposed into an intra-frame depth two-dimensional convolution kernel where C in is the number of input feature channels, C out is the number of output channels, which is also the number of convolution kernels of the current layer, and k is the length of the time dimension convolution kernel. and the heat field channel and the optical flow channel The output feature Y t at the same time index t after conversion is written as:

[0176]

[0177] wherein, is the input of the current frame of the cth channel, is the buffered feature after channel splicing, * represents two-dimensional convolution, and represents element-wise multiplication and accumulation.

[0178] By decomposition, the original three-dimensional convolution kernel memory access is divided into spatial locality + time sparsity twice, so that the single-frame off-chip bandwidth demand is reduced, and the FPGA and DSP parallel operator library are reused.

[0179] In step 1082, the gradient-sensitive channel contribution is evaluated, and in this sub-step, the back propagation gradient The contribution degree of each output channel c is counted in the whole training set:

[0180]

[0181] wherein, is the sample set, Ω is the spatial position index of the feature map, is the activation value of the dth sample at the output channel c and the position p.

[0182] According to Γ c from small to large, the output channels with low contribution degree enter the pruning candidate set to ensure that the output channels with high contribution degree are completely retained, which is crucial to avoid false deletion of decoy interference bombs to lure the discrimination of the decoy optical mode.

[0183] The candidate set can be pruned layer by layer with a 10% step size: each round removes a subset of candidate channels The FPGA logic resource table is statically recompiled to immediately shrink the on-chip SRAM and convolution array; only three small batch fine-tuning is performed on the last two layers of parameters until the anti-decoy precision is verified by the verification set ​Hold

[0184] Repeat sub-steps 1081-1082 to measure the end-to-end inference delay time τ in real time e2e : If τ e2e ≤ 4ms and (η min is fixed at 97%), stop iteration and solidify the lightweight model M lite ; if the real-time performance is affected due to continued pruning Roll back to the previous version of parameters and structures to ensure real-time performance and robustness.

[0185] In step 1083, the convolution kernel, bias and normalization scale of the lightweight model M lite may be mapped to 8-bit integers and mapped to the multiplier array of the FPGA-DSP collaborative platform in fixed-point format (Q2.6); a three-stage pipeline buffer is constructed synchronously: input cache→ deep convolution→ grouped convolution.

[0186] It should be noted that in terms of model deployment, the conventional practice in the industry is to split the three-dimensional convolution kernel of the model into spatial two-dimensional convolution and temporal recursive update, supplemented by 8-bit fixed-point quantization, on-chip double-ended cache and Ping-Pong scheduling to reduce external bandwidth overhead. This approach will weaken the expression of spatial-temporal correlation across frames, and when a high-speed cloud curtain sweeps or the background radiation fluctuates rapidly, the recursive state will lag, further amplifying the risk of terminal lock. Overall, existing technologies are still difficult to stably reduce the decoy false alarm rate in the terminal guidance phase to a more stringent numerical target under the hard constraint of millisecond-level inference delay.

[0187] By performing hierarchical time consumption analysis on the target discrimination model, the cross-frame three-dimensional convolution kernel is decomposed into intra-frame two-dimensional convolution kernel and one inter-frame grouped one-dimensional convolution kernel, the contribution degree of the difficult example sequence corresponding to the back propagation gradient is counted in the entire training set, and the low contribution channels are pruned in increasing proportion, then the convolution kernel and bias are quantized to 8-bit fixed-point parameters and mapped to the three-stage pipeline of the FPGA-DSP collaborative platform, realizing end-to-end millisecond-level inference.

[0188] After the target model deployment is completed, the method for generating the decoy jamming bomb discrimination model can further include step 109, on-orbit micro-updating and maintaining the target discrimination model. In some embodiments, step 109 can include the following sub-steps:

[0189] Step 1091, extracting the latest ten frames of original images and corresponding optical flow fields in time sequence during the flight of the missile;

[0190] Step 1092, calculating the drift intensity of the current window according to the extracted original images and corresponding optical flow fields;

[0191] Step 1093, construct a weighted consistency loss function according to the drift intensity, and apply the weighted consistency loss function only to the statistical layer;

[0192] Step 1094, only perform a limited number of micro-updates on the mean and variance of each layer BatchNorm;

[0193] Step 1095, adjust the classification threshold according to the window F1 score curve sliding;

[0194] Step 1096, if the continuous verification hits the real target and there is no new false alarm, solidify the update, otherwise roll back to the last version, and return the update record to the ground station through the telemetry channel.

[0195] In step 1091, the latest ten frames of raw images I t and the corresponding optical flow field Ensure that the image and the optical flow are aligned at the same sampling timestamp, and prepare complete thermal-movement co-window observations for this round of micro-update.

[0196] In step 1092, ten frames of raw images-optical flow field pairs are injected into the lightweight model M lite After that, the thermal branch embedding and the motion branch embedding (b is the pixel index) are extracted respectively. The residual error is calculated pixel by pixel and accumulated in the time dimension to obtain the drift intensity of the current window:

[0197]

[0198] Where Ω is the set of single-frame pixel indices. The larger the value of D, the more significant the drift of the center waveband of the atmospheric transmission window, resulting in thermal-movement coupling mismatch.

[0199] In step 1093, calculate the weight ω = min(κD, 1) (κ is a proportionality constant) according to the drift intensity D, and construct the weighted consistency loss function as follows:

[0200]

[0201] Only is applied to the statistical layer to avoid damaging the compressed convolution kernel weights.

[0202] In step 1094, freeze all convolution kernels and biases, and only perform three rounds of micro-updates on the mean vector μ l and the variance vector of each layer BatchNorm of the target discrimination model: use learning rate η bn (<10 -3 ) gradient descent correction The activation is renormalized immediately after each update, ensuring that the new distribution does not have a spike; constraints Avoid excessive drift.

[0203] In step 1095, the positive and negative class output score sets of the ten frames after the micro-update are collected The F1 score curve F1(θ) is calculated according to the separation threshold θ. The full connection layer discriminant threshold is slid to the θ corresponding to the F1 peak ★ , offsetting the global bias introduced by cloud thickness and humidity changes.

[0204] In step 1096, the positive and negative class output score sets of the ten frames after the micro-update are collected t′ Run the new threshold θ ★ If the real target is hit continuously and no new false positives are added, solidify Otherwise, immediately roll back to the previous version of the parameter to avoid the impact of false updates on the safety of the guidance closed loop.

[0205] After completing three successful solidifications, the corresponding time stamp, drift index {t, D} sequence is written Write to the encrypted telemetry channel back to the ground station for subsequent offline batch simulation reproduction, and update the model repository synchronously, so that the subsequent firmware version inherits the latest statistical prior.

[0206] The above scheme realizes, in the infrared homing terminal guidance phase, under the condition of very few annotations, within the low-power millisecond-level inference framework defined by the flight platform, to construct a real-time discriminant model that simultaneously describes the coupling relationship between thermal radiation distribution and maneuver dynamics, so that it can still accurately distinguish real targets from decoy interference ammunition in the presence of multi-spectral, complex atmospheric windows, and high-intensity magnesium interference scenes, and control the false alarm rate of the whole process within the predetermined threshold.

[0207] The device embodiment of the present application is introduced below, which can be used to execute the decoy interference ammunition discriminant model generation method in the above-mentioned embodiments of the present application. For details not disclosed in the device embodiment of the present application, please refer to the above-mentioned embodiments of the decoy interference ammunition discriminant model generation method.

[0208] Referring to Figure 2 , a block diagram of a decoy interference ammunition discriminant model generation device in an embodiment of the present application is shown.

[0209] As Figure 2As shown, the bait decoy missile discrimination model generation device of the embodiment of the present application includes: a sequence acquisition module 201, a sequence coupling module 202, an initial model construction module 203, an intermediate model construction module 204, and a target model construction module 205, wherein the sequence acquisition module 201 is configured to obtain time sequence consistent radiation sequences and attitude sequences based on original images output by an infrared focal plane array in a seeker and attitude data of a missile body output by an inertial measurement unit; the sequence coupling module 202 is configured to couple the radiation sequences and the attitude sequences by using a first encoder in a preset network to obtain a space-time embedding vector; the initial model construction module 203 is configured to adjust a single-layer full connection classification head in the preset network according to the space-time embedding vector; the initial model construction module 203 is further configured to construct an initial discrimination model according to the first encoder and the adjusted single-layer full connection classification head; the intermediate model construction module 204 is configured to obtain a decoy sequence according to limit values of physical parameters of the bait decoy missile, wherein the decoy sequence is configured to represent the radiation sequences and the attitude sequences of the bait decoy missile; the intermediate model construction module 204 is further configured to obtain an intermediate discrimination model according to the initial discrimination model and the decoy sequence; and the target model construction module 205 is configured to train the intermediate discrimination model by using the decoy sequence, the time sequence consistent radiation sequences, and the attitude sequences to obtain a target discrimination model.

[0210] Based on the same inventive concept, the embodiment of the present application also provides a bait decoy missile discrimination model generation device, which refers to Figure 3 The structure of the bait decoy missile discrimination model generation device in the embodiment of the present application is shown in a structural schematic diagram, which includes one or more memories 304, one or more processors 302, and at least one computer program (computer program instructions) stored in the memory 304 and executable on the processor 302, and the processor 1202 executes the computer program to implement the method as described above.

[0211] In the above method, Figure 3 The bus architecture (represented by the bus 300) can include any number of interconnected buses and bridges, which link various circuits including one or more processors represented by the processor 302 and the memory represented by the memory 304. The bus 300 can also link various other circuits such as peripheral devices, voltage stabilizers, and power management circuits, which are well known in the art, and therefore, will not be further described herein. The bus interface 305 provides an interface between the bus 300 and the receiver 301 and the transmitter 303. The receiver 301 and the transmitter 303 can be the same element, i.e., a transceiver, which provides a unit for communicating with various other devices on a transmission medium. The processor 302 is responsible for managing the bus 300 and general processing, while the memory 304 can be used to store data used by the processor 302 in performing operations.

[0212] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, cause the processor to perform the steps of the method described above.

[0213] Based on the same inventive concept, embodiments of this application provide a computer program product, including a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0214] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored as one or more instructions or codes on or transmitted via a computer-readable medium. Other examples and embodiments are within the scope and spirit of this application and the appended claims. For example, due to the nature of software, the functions described above may be implemented using software executed by a processor, hardware, firmware, hardwired, or any combination thereof. Furthermore, the functional units may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit.

[0215] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0216] The units described as separate components may or may not be physically separate. Similarly, the components of the control device may or may not be physical units; they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0217] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk and various computer program instruction storage media.

[0218] The above only describes the embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A method for generating a decoy jamming bomb discrimination model, characterized in that, The method comprises: obtaining a time-series consistent radiation sequence and a posture sequence based on raw images output by an infrared focal plane array in a seeker and posture data of a projectile output by an inertial measurement unit; determining a plurality of minimum prediction units according to the radiation sequence and the posture sequence; performing channel dimension splicing on posture data in each of the minimum prediction units after converting the posture data into a plane moment field through a transmission mapping, and a corresponding target radiation matrix, to obtain a plurality of multi-channel tensors; training an initial network using the multi-channel tensors to obtain a preset network, wherein the initial network comprises an initial encoder, a radiation decoding head and an optical flow decoding head, and the preset network comprises a first encoder, the radiation decoding head and the optical flow decoding head; inputting the multi-channel tensors into the preset network to obtain a first feature vector output by the first encoder, and determining the first feature vector as a space-time embedding vector; adjusting a single-layer fully connected classification head in the preset network according to the space-time embedding vector; constructing an initial discrimination model according to the first encoder and the adjusted single-layer fully connected classification head; obtaining a decoy sequence according to limit values of physical parameters of a decoy jamming projectile, wherein the decoy sequence is used to represent a radiation sequence and a posture sequence of the decoy jamming projectile; obtaining an intermediate discrimination model according to the initial discrimination model and the decoy sequence; training the intermediate discrimination model using the decoy sequence, the time-series consistent radiation sequence and the posture sequence to obtain a target discrimination model.

2. The method of claim 1, wherein the decoy jamming bomb discrimination model is generated based on a plurality of training data, and the plurality of training data comprises a plurality of training decoy jamming bomb data and a plurality of training target data. The infrared focal plane array in the seeker comprises an infrared detector, and the inertial measurement unit comprises a three-axis gyroscope, an interrupt trigger time of the infrared detector when integral reading is completed is the same as an interrupt trigger time of the three-axis gyroscope when sampling is completed, and the method of obtaining a time-series consistent radiation sequence and a posture sequence based on raw images output by an infrared focal plane array in a seeker and posture data of a projectile output by an inertial measurement unit comprises: correcting a gain matrix and a bias matrix of the infrared focal plane array in the seeker based on a focal plane temperature of the infrared focal plane array in the seeker; determining a linear radiation matrix according to the raw images, the corrected gain matrix and the corrected bias matrix; performing rotation and projection processing on current posture data to obtain a displacement field corresponding to the current posture data; determining a target radiation matrix corresponding to the current posture data using the displacement field, the linear radiation matrix and a quadratic B-spline interpolation algorithm; obtaining a time-series consistent radiation sequence and a posture sequence according to the current posture data and the target radiation matrix corresponding to the current posture data.

3. The method of claim 1, wherein the decoy jamming bomb discrimination model is generated by: The method of training an initial network using the multi-channel tensors to obtain a preset network comprises: inputting the multi-channel tensors into the initial network to obtain a second feature vector output by the initial encoder, a radiation matrix prediction value output by the radiation decoding head and an optical flow prediction value output by the optical flow decoding head; determining a same-group forward clustering center according to a second feature vector corresponding to a multi-channel tensor of a current minimum prediction unit; determine a cross-group negative cluster center according to a second feature vector corresponding to a multi-channel tensor of any minimum prediction unit except the current minimum prediction unit in the plurality of minimum prediction units; put the radiation matrix prediction value and the optical flow prediction value corresponding to the multi-channel tensor of the current minimum prediction unit, the same-group positive cluster center, and the cross-group negative cluster center into a loss function of the initial network, and determine that the loss function converges when a value of the loss function is lower than a preset threshold, to obtain the preset network.

4. The method of claim 3, wherein the decoy jamming bomb discrimination model is generated based on the following equation: ###0001### where, P is the probability of the decoy jamming bomb discrimination model, and P is the probability of the decoy jamming bomb discrimination model. the adjusting the single-layer fully connected classification head in the preset network according to the spatio-temporal embedding vector comprises: performing clustering screening on the spatio-temporal embedding vector to obtain a first spatio-temporal embedding vector, and obtaining a label corresponding to the first spatio-temporal embedding vector, wherein the label comprises a first label for representing a real target and a second label for representing a decoy jamming projectile; adjusting the single-layer fully connected classification head in the preset network according to the first spatio-temporal embedding vector and the label corresponding to the first spatio-temporal embedding vector, while keeping parameters of the first encoder unchanged; taking vectors in the spatio-temporal embedding vector except the first spatio-temporal embedding vector as second spatio-temporal embedding vectors, and performing weak augmentation processing on original images corresponding to the second spatio-temporal embedding vectors to obtain a pseudo-label candidate set; screening a weakly augmented image and a pseudo-label corresponding to the weakly augmented image from the pseudo-label candidate set according to a preset decreasing threshold sequence; performing strong augmentation processing on the weakly augmented image to obtain a strongly augmented image; adjusting the single-layer fully connected classification head in the preset network according to the pseudo-label corresponding to the weakly augmented image and the strongly augmented image, while keeping parameters of the first encoder unchanged.

5. The method of claim 4, wherein the decoy jamming bomb discrimination model is generated by: the obtaining the decoy sequence according to limit values of physical parameters of the decoy jamming projectile comprises: constructing a parameter generator according to the limit values of the physical parameters of the decoy jamming projectile; obtaining the decoy sequence according to parameters output by the parameter generator.

6. The method of claim 5, wherein the decoy jamming bomb discrimination model is generated based on the following equation: ###0001### where, P is the probability of the decoy jamming bomb discrimination model, and P is the probability of the decoy jamming bomb discrimination model. the obtaining the intermediate discriminant model according to the initial discriminant model and the decoy sequence comprises: unfreezing a second-to-last layer of the first encoder in the initial discriminant model and the single-layer fully connected classification head in the initial discriminant model to construct an adversarial discriminator; determining a confusion probability of the decoy sequence according to the adversarial discriminator; training the adversarial discriminator according to the confusion probability, continuing to train the adversarial discriminator by using the decoy sequence, the time-series consistent radiation sequence, and the attitude sequence after a first preset condition is reached, and constructing the intermediate discriminant model according to the first encoder and the trained adversarial discriminator after a second preset condition is reached.

7. The method of claim 6, wherein the decoy jamming bomb discrimination model is generated by, the training the intermediate discriminant model by using the decoy sequence, the time-series consistent radiation sequence, and the attitude sequence to obtain a target discriminant model comprises: inserting the decoy sequence into the time-series consistent radiation sequence and the attitude sequence according to a time axis to obtain a synthetic sequence; adjusting last two layers of a first encoder in the intermediate discriminant model according to the synthetic sequence to obtain a second encoder; adjust a single-layer full-connection classification head of an adversarial discriminator in the second encoder and the intermediate discriminator model according to the original image corresponding to the first spatio-temporal embedding vector; construct the target discriminator model according to the adjusted second encoder and the adjusted adversarial discriminator.

8. The method of claim 1, wherein the decoy drone discrimination model is generated by: Further comprising: decompose a three-dimensional convolution kernel of the target discriminator model into an intra-frame two-dimensional convolution kernel and an inter-frame one-dimensional convolution kernel; prune an output channel of the target discriminator model according to a contribution degree of a reverse propagation gradient corresponding to a difficult example sequence to the output channel, wherein the difficult example sequence is a bait sequence with a confidence in the middle; perform inference delay time and anti-bait precision detection on the target discriminator model; deploy the target discriminator model in a case where the inference delay time is less than a preset time and the anti-bait precision is greater than a preset precision. 9.A device for generating a decoy jamming bomb discrimination model, comprising a processor and a memory, wherein, The memory stores computer program instructions capable of being executed by the processor, and the processor executes the computer program instructions to implement the steps of the method in any one of claims 1 to 8.

Images