High-resolution one-dimensional range profile anomaly detection method and device fused with multi-detector capability

Abstract

The invention discloses a high-resolution one-dimensional range profile anomaly detection method and device fusing multi-detector capability, and relates to the technical field of time sequence anomaly detection, and the method comprises the steps: obtaining a to-be-detected radar high-resolution one-dimensional range profile; inputting a to-be-detected radar high-resolution one-dimensional range profile into the trained neural network model for processing to obtain an anomaly detection result of the to-be-detected radar high-resolution one-dimensional range profile; wherein the trained neural network model is obtained by training the initial neural network model by taking preset category data as a training data set and taking preset classification loss as guidance; the preset classification loss comprises a regulatory factor, and the regulatory factor is used for representing the detection difficulty. According to the method, the results of a plurality of detectors do not need to be firstly calculated and then fused in the reasoning process, so that calculation resources during reasoning are saved.

Description

Technical Field

[0001] The present invention belongs to the technical field of time series anomaly detection, and particularly relates to a high-resolution one-dimensional range profile anomaly detection method and device that integrates the capabilities of multiple detectors. Background Art

[0002] A high-resolution range profile (HRRP) is obtained by transmitting a broadband radar signal and receiving the echo signal reflected by the target, and through processing techniques such as matched filtering and pulse compression, a high-resolution distribution image of the target in the range direction is obtained. The high-resolution range profile can reflect the structural characteristics and shape information of the target, so it has important applications in target recognition and classification. During the operation of the radar system, it will be affected by various noises and interferences, such as thermal noise, electronic interference, and multipath effects. These noises and interferences will cause abnormal points to appear in the HRRP data. At the same time, the radar data changes dynamically over time, and the movement of the target object, the change of environmental conditions, etc. will also affect the stability and consistency of the HRRP data.

[0003] HRRP data is usually high-dimensional, and the high-dimensional characteristics of the data make anomaly detection complex, and efficient algorithms need to be developed to process and analyze these data. Currently, there are few anomaly detection methods for high-resolution one-dimensional range profiles. For general time series anomaly detection methods, a single anomaly detection method usually performs well in specific scenarios, but when faced with complex and diverse anomaly patterns, it is often difficult to comprehensively capture all anomaly patterns and may perform poorly. For example, statistical-based methods may not be able to handle high-dimensional data, while machine learning-based methods may require a large amount of labeled data.

[0004] Therefore, there is an urgent need to provide a high-resolution one-dimensional range profile anomaly detection method to improve the defects existing in the above-mentioned existing solutions. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a high-resolution one-dimensional range profile anomaly detection method and device that integrates the capabilities of multiple detectors. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] In a first aspect, the present invention provides a high-resolution one-dimensional range profile anomaly detection method that integrates the capabilities of multiple detectors, including:

[0007] Obtain the radar high-resolution one-dimensional range profile to be detected;

[0008] Input the radar high-resolution one-dimensional range profile to be detected into a trained neural network model for processing to obtain the anomaly detection result of the radar high-resolution one-dimensional range profile to be detected;

[0009] Among them, the trained neural network model is obtained by training the initial neural network model with preset category data as the training data set and guided by a preset classification loss; the preset classification loss includes a regulation factor, and the regulation factor is used to characterize the detection difficulty.

[0010] In a second aspect, the present invention further provides a high-resolution one-dimensional range image anomaly detection device that fuses the capabilities of multiple detectors, which is used to implement the above-mentioned high-resolution one-dimensional range image anomaly detection method that fuses the capabilities of multiple detectors, and includes:

[0011] A data acquisition module, which is used to acquire the radar high-resolution one-dimensional range image to be detected;

[0012] A data anomaly recognition module, which is used to input the radar high-resolution one-dimensional range image to be detected into the trained neural network model for processing to obtain the anomaly detection result of the radar high-resolution one-dimensional range image to be detected;

[0013] Among them, the trained neural network model is obtained by training the initial neural network model with preset category data as the training data set and guided by a preset classification loss; the preset classification loss includes a regulation factor, and the regulation factor is used to characterize the detection difficulty.

[0014] Advantages of the present invention:

[0015] The high-resolution one-dimensional range image anomaly detection method and device that fuse the capabilities of multiple detectors provided by the present invention can not only combine the respective advantages to improve the overall detection performance, but also process diverse anomaly patterns, and improve the generalization ability of anomaly detection for different types of time series data by jointly using a variety of anomaly detection algorithms. Compared with directly using a single method and directly training in a deep neural network, this method improves the anomaly detection ability of the model. At the same time, compared with the scheme of directly integrating multiple detectors, the method proposed by the present invention does not need to calculate the results of multiple detectors first and then fuse them during the inference process, saving the computing resources during inference.

[0016] The following will further elaborate on the present invention in conjunction with the accompanying drawings and embodiments. Description of the Drawings

[0017] Figure 1 is a flowchart of a high-resolution one-dimensional range image anomaly detection method that fuses the capabilities of multiple detectors provided by an embodiment of the present invention;

[0018] Figure 2 is a schematic diagram of the training process of the neural network model provided by an embodiment of the present invention;

[0019] Figure 3It is a schematic diagram of a high-resolution one-dimensional range image anomaly detection device integrating the capabilities of multiple detectors provided by an embodiment of the present invention;

[0020] Figure 4 It is a schematic diagram of a statistical chart of anomaly detection evaluation indexes for a supervised model and an unsupervised algorithm provided by an embodiment of the present invention. Specific embodiments

[0021] The following further describes the present invention in detail with reference to specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0022] Existing time series anomaly detection methods can be roughly divided into three categories according to learning techniques: statistical methods, neural network methods, and basic model methods. Statistical methods rely on statistical hypotheses and detect situations that deviate from the expected data distribution as anomalies. Neural network methods learn the normal patterns of anomaly-free training data and then detect anomalies in new test data. Basic model methods utilize the knowledge of large models pre-trained on a large amount of time series or text data and apply it to time series anomaly detection tasks.

[0023] Time series anomaly detection is a complex task. A single method usually performs well in specific types of time series data or specific types of anomaly detection tasks, but may not work well in other scenarios. Many single methods rely on specific assumptions (such as data distribution, independence, etc.), and these assumptions may not hold in practical applications, thus affecting the detection effect. When facing different types of time series data, the generalization ability of a single method may be insufficient, making it difficult to handle diverse anomaly patterns. A single method may also be more sensitive to noise and interference in the data, with a higher computational complexity and longer training and inference times.

[0024] In view of this, considering that different methods are complementary in dealing with different types of abnormal data, by jointly using multiple methods, their respective advantages can be integrated to improve the overall detection performance. The present invention provides a high-resolution one-dimensional range image anomaly detection method integrating the capabilities of multiple detectors, which jointly uses multiple general time series anomaly detection methods to form a high-resolution one-dimensional range image anomaly detection method integrating the capabilities of multiple detectors.

[0025] Please refer to Figure 1 , Figure 1 It is a flowchart of a high-resolution one-dimensional range image anomaly detection method integrating the capabilities of multiple detectors provided by an embodiment of the present invention. A high-resolution one-dimensional range image anomaly detection method integrating the capabilities of multiple detectors provided by the present invention includes:

[0026] S101. Obtain the radar high-resolution one-dimensional range image to be detected.

[0027] Specifically, in this embodiment, considering that during the radar signal acquisition process, it may be affected by other electromagnetic wave interferences, environmental factors, target characteristics, radar parameters, etc., resulting in some abnormal samples in the high-resolution range profile dataset. If all the data is directly used to train the recognition model of the target recognition database, it may lead to a decline in recognition performance. Therefore, it is necessary to perform anomaly detection on the radar high-resolution one-dimensional range profile data.

[0028] S102. Input the radar high-resolution one-dimensional range profile to be detected into the trained neural network model for processing to obtain the anomaly detection result of the radar high-resolution one-dimensional range profile to be detected.

[0029] Among them, the trained neural network model is obtained by training the initial neural network model with preset category data as the training dataset and with a preset classification loss as the guidance; the preset classification loss includes a regulation factor, and the regulation factor is used to characterize the detection difficulty.

[0030] Specifically, in this embodiment, considering that the radar high-resolution range profile data is regarded as a time series, a single time series anomaly detection method usually has certain limitations. Some methods may be sensitive to specific types of anomalies and often difficult to comprehensively capture the anomaly features in the data. In view of this, the present invention provides a high-resolution one-dimensional range profile anomaly detection method that fuses the capabilities of multiple detectors. By fusing the results of multiple anomaly detection algorithms, it provides guidance for the training of the neural network anomaly detection model, uses the outputs of different detection algorithms as input features or auxiliary information of the neural network anomaly detection model, and enhances the detection ability of the neural network model for anomalies. This fusion method can capture the advantages and complementarities of each single algorithm. By jointly using multiple methods, it synthesizes the advantages of each single algorithm, improves the overall detection performance, and improves the generalization ability for different types of time series data. By combining multiple methods, it can reduce the bias of a single method and improve the detection accuracy. The fusion method can reduce the variance of the model by averaging the results of multiple models, improve the detection stability, and thus improve the robustness and detection accuracy of the neural network anomaly detection model. At the same time, the inference process does not need to calculate the results of multiple detectors first and then fuse them, saving computing resources.

[0031] In this embodiment, please refer to Figure 2 , Figure 2 which is a schematic diagram of the training process of the neural network model provided by the embodiment of the present invention. The training process of the trained neural network model includes:

[0032] Obtain data of multiple preset categories to construct a training dataset; among them, the training dataset includes multiple samples, and the samples include radar high-resolution one-dimensional range profiles.

[0033] Obtain the true labels of the samples in the training dataset; for example, the training dataset includes m high-resolution range image samples, and the true labels of the samples are Y i (i = 1, 2, …, m), and its elements are 0 or 1, where 0 indicates that the sample is normal and 1 indicates that the sample is abnormal;

[0034] Use the samples in the training dataset, the true labels of the samples, and a preset classification loss to train the initial neural network model to obtain a trained neural network model; optionally, anomaly detection can be regarded as a binary classification problem and trained using the preset classification loss.

[0035] In this embodiment, using the samples in the training dataset, the true labels of the samples, and a preset classification loss to train the initial neural network model to obtain a trained neural network model includes:

[0036] Input some samples in the training dataset into the j-th neural network model to be trained for training to obtain the predicted results of the classification output during the j-th training process;

[0037] According to the predicted results of the classification output during the j-th training process and the true labels of the samples used to train the j-th neural network model to be trained, calculate the classification loss, and multiply the classification loss by the adjustment factor as the preset classification loss for the j-th training process;

[0038] Perform backpropagation according to the preset classification loss of the j-th training process to update the network parameters of the j-th neural network model to be trained to obtain the (j + 1)-th neural network model to be trained; iterate in this way until the number of training times or the convergence degree meets the preset conditions to obtain a trained neural network model.

[0039] In this embodiment, the process of obtaining the adjustment factor includes:

[0040] Use multiple single-type anomaly detection methods to perform anomaly detection on the samples in the training dataset to obtain an anomaly detection result matrix, and its elements are 0 or 1; for example, there are n single-type anomaly detection methods, which can be existing anomaly detection methods in the prior art;

[0041] According to the anomaly detection result matrix, calculate the scores of the difficulty of anomaly detection of the samples in the training dataset, and the larger the value, the easier the sample is to be misjudged (abnormal samples are detected as normal, and normal samples are detected as abnormal), and the difficult-to-detect samples can be used to construct the adjustment factor according to the detection difficulty;

[0042] Calculate the adjustment factor according to the scores of the difficulty of anomaly detection of the samples in the training dataset.

[0043] It should be noted that each single type of anomaly detection method can detect one type of anomaly. By using multiple single types of anomaly detection methods, multiple types of anomalies can be detected. By fusing the results of multiple anomaly detection algorithms, guidance can be provided for the training of the neural network anomaly detection model. This method uses the outputs of different detection algorithms as input features or auxiliary information for the neural network detection model, enhancing the anomaly detection ability of the neural network model.

[0044] It should be noted that by using the method of majority voting, the scores of the difficulty levels of anomaly detection for the samples in the training dataset are calculated. That is, through the majority voting mechanism, the results of multiple algorithms are weighted, and the fused result is embedded in the calculation of the loss function during the neural network training process, thereby achieving the purpose of multiple algorithms collaborating to guide the training of the neural network model. In addition, by adjusting the parameters of the loss function, the requirements for different detection effects can be met.

[0045] In this embodiment, the expression of the anomaly detection result matrix is:

[0046] y ij (i = 1, 2, …, m; j = 1, 2, …, n);

[0047] Among them, y ij represents the anomaly detection result matrix, i represents the index of the sample in the training dataset, m represents the total number of samples in the training dataset, j represents the index of the preset anomaly detection method, and n represents the total number of preset anomaly detection methods.

[0048] In this embodiment, the expression of the scores of the difficulty levels of anomaly detection for the samples in the training dataset is:

[0049]

[0050] Among them, D represents the scores of the difficulty levels of anomaly detection for the samples in the training dataset, and Y i represents the true labels of the samples in the training dataset.

[0051] In this embodiment, the expression of the adjustment factor is:

[0052] H = α × (1 - D) γ + (1 - α);

[0053] Among them, H represents the adjustment factor, α represents the joint loss balance factor, and its value range is [0, 1]. When α = 0, it means that the current training only uses the original BCE loss, Loss = Loss BCE , and the adjustment factor does not play a role. When α = 1, it means that the current training only uses the acceptance of the difficulty level of detection, that is, the BCE loss guided by the adjustment factor, Loss = (1 - D) γ LossBCE When α takes values between 0 and 1, it means that the guidance of the adjustment factor is incorporated into the original loss. When using it, α should take values between 0 and 1, and the value range of D can be known from the formula to be between 0 and 1. To prevent the value of Loss from being too small when the value of D is too large (close to 1) and unable to effectively train the neural network. γ represents the difficulty sample attention degree index. When γ = 0, Loss = Loss BCE , the adjustment factor has no effect. The value of γ should be greater than 0. The larger the value of γ, the higher the attention degree of the model to difficult-to-detect samples, and the stronger the suppression of the loss of easy-to-detect samples. The smaller the value of γ, Loss approaches the original BCE loss, and the attention degree of the model to all samples approaches the same.

[0054] In this embodiment, the expression of the preset classification loss is:

[0055] Loss = H × Loss BCE ;

[0056] Loss = (α × (1 - D) γ + (1 - α)) × Loss BCE ;

[0057] Among them, Loss represents the preset classification loss, H represents the adjustment factor, and Loss BCE represents the classification loss, which can be the binary cross-entropy loss function, and its expression is:

[0058]

[0059] Among them, p i represents the probability that the p i th sample is predicted as the positive class, and Y i represents the true label of the sample in the training dataset; the parameters α and γ can be adjusted to meet different anomaly detection performance requirements.

[0060] In this embodiment, by introducing the adjustment factor to control the binary cross-entropy loss function, the purpose of increasing the loss value of difficult-to-detect samples is achieved, so that the network model can pay more attention to difficult-to-detect samples.

[0061] In summary, the high-resolution one-dimensional range image anomaly detection method that integrates the capabilities of multiple detectors provided by the present invention can not only combine their respective advantages to improve the overall detection performance by jointly using multiple anomaly detection algorithms, but also handle diverse anomaly patterns to enhance the generalization ability for anomaly detection of different types of time series data. Compared with directly using a single method and directly training in a deep neural network, this method improves the anomaly detection ability of the model. At the same time, compared with the scheme of directly integrating multiple detectors, this method does not need to calculate the results of multiple detectors first and then fuse them during the inference process, saving computational resources during inference.

[0062] Based on the same inventive concept, please refer to Figure 3 , Figure 3 FIG. is a schematic diagram of a high-resolution one-dimensional range image anomaly detection device that integrates the capabilities of multiple detectors provided by an embodiment of the present invention. The present invention also provides a high-resolution one-dimensional range image anomaly detection device that integrates the capabilities of multiple detectors, which is used to implement the high-resolution one-dimensional range image anomaly detection method provided by the above-mentioned embodiment of the present invention. For the embodiments of the method, please refer to the above, and details will not be repeated here; the device includes:

[0063] A data acquisition module, configured to acquire the radar high-resolution one-dimensional range image to be detected;

[0064] A data anomaly recognition module, configured to input the radar high-resolution one-dimensional range image to be detected into a trained neural network model for processing to obtain an anomaly detection result of the radar high-resolution one-dimensional range image to be detected;

[0065] Among them, the trained neural network model is obtained by training an initial neural network model with preset category data as the training data set and a preset classification loss as the guidance; the preset classification loss includes a regulation factor, and the regulation factor is used to characterize the detection difficulty.

[0066] In an optional embodiment of the present invention, the effect of the high-resolution one-dimensional range image anomaly detection method that integrates the capabilities of multiple detectors provided by the above-mentioned embodiment is verified through a simulation experiment. Specifically:

[0067] This example uses radar high-resolution one-dimensional range image data for verification. A data set is obtained, and the data set is first preprocessed. The data set has a total of 15,000 samples with a dimension of 4,096. Randomly select 1,000 samples from it to add anomalies, and the contamination degree is 6.7%. 30% of the data set is extracted as the training set, 14% as the validation set, and the remaining 56% as the test set.

[0068] The following starts from the LSTM (Long Short-Term Memory) model to analyze the impact of the fusion anomaly detection method on HRRP anomaly detection. In this embodiment, 9 general unsupervised anomaly detection methods are selected.

[0069] Please refer to Figure 4 , Figure 4 which is a schematic diagram of the statistical chart of the supervised model and the unsupervised algorithm anomaly detection evaluation index provided by the embodiment of the present invention. Table 1 is the statistical table of the anomaly detection evaluation index of the basic model and the fusion model. It can be seen from Table 1 that the basic model embedded with the adjustment factor has improved in all four evaluation indexes compared with the basic model without the embedded adjustment factor, and the improvement of the recall rate is the most significant, with an increase of 2.33%. This shows that the adjustment factor plays a positive role in the process of model training. Compared with the scheme of directly integrating multiple detectors, this scheme does not need to calculate the results of multiple detectors first and then fuse them during the inference process, saving the computing resources during inference.

[0070] Table 1 Statistical Table of Anomaly Detection Evaluation Indexes of Basic Model and Fusion Model

[0071]

[0072]

[0073] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "including", "comprising" or any other variant are intended to cover non-exclusive inclusion, so that an article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the article or device including the said element. "Connection" or "connected" and other similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The orientation or positional relationship indicated by "up", "down", "left", "right", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention.

[0074] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0075] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A high-resolution one-dimensional range image anomaly detection method that fuses the capabilities of multiple detectors, characterized in that Including: Obtain the high-resolution one-dimensional range image of the radar to be detected; Input the high-resolution one-dimensional range image of the radar to be detected into the trained neural network model for processing, and obtain the anomaly detection result of the high-resolution one-dimensional range image of the radar to be detected; Wherein, the trained neural network model is obtained by training the initial neural network model with preset category data as the training data set and a preset classification loss as the guidance; the preset classification loss includes a regulation factor, and the regulation factor is used to characterize the detection difficulty.

2. The high-resolution one-dimensional range image anomaly detection method integrating multi-detector capabilities according to claim 1, characterized in that The training process of the trained neural network model includes: Obtain data of multiple said preset categories to construct a training data set; wherein, the training data set includes multiple samples, and the samples include high-resolution one-dimensional range images of the radar; Obtain the true labels of the samples in the training data set; Use the samples in the training data set, the true labels of the samples, and a preset classification loss to train the initial neural network model to obtain the trained neural network model.

3. The high-resolution one-dimensional range profile anomaly detection method integrating multi-detector capabilities according to claim 2, characterized in that, Using the samples in the training data set, the true labels of the samples, and a preset classification loss to train the initial neural network model to obtain the trained neural network model, includes: Input a part of the samples in the training data set into the neural network model to be trained for the jth time for training, and obtain the prediction result of the classification output during the jth training process; According to the prediction result of the classification output during the jth training process and the true label of the sample for training the neural network model to be trained for the jth time, calculate the classification loss, and multiply the classification loss by the regulation factor as the preset classification loss for the jth training process; Perform backpropagation according to the preset classification loss of the jth training process to update the network parameters of the neural network model to be trained for the jth time, and obtain the neural network model to be trained for the (j + 1)th time; iterate in this way until the number of training times or the convergence degree meets the preset conditions, and obtain the trained neural network model.

4. The high-resolution one-dimensional range image anomaly detection method integrating multi-detector capabilities according to claim 3, wherein, The obtaining process of the regulation factor includes: Use multiple single-type anomaly detection methods to perform anomaly detection on the samples in the training data set to obtain an anomaly detection result matrix; According to the anomaly detection result matrix, calculate the scores of the anomaly detection difficulty of the samples in the training data set; According to the scores of the anomaly detection difficulty of the samples in the training data set, calculate and obtain the regulation factor.

5. The high-resolution one-dimensional range profile anomaly detection method integrating multi-detector capabilities according to claim 4, characterized in that The expression of the anomaly detection result matrix is: y ij (i = 1, 2, …, m; j = 1, 2, …, n); where y ij represents the anomaly detection result matrix, i represents the index of the samples in the training dataset, m represents the total number of samples in the training dataset, j represents the index of the preset anomaly detection method, and n represents the total number of the preset anomaly detection methods.

6. The high-resolution one-dimensional range image anomaly detection method integrating multi-detector capabilities according to claim 4, characterized in that, The expression of the scores of the anomaly detection difficulty of the samples in the training data set is: Among them, D represents the score of the difficulty of anomaly detection of samples in the training dataset, and Y i represents the true label of the samples in the training dataset.

7. The high-resolution one-dimensional range profile anomaly detection method integrating multi-detector capabilities according to claim 4, characterized in that The expression of the regulation factor is: H = α × (1 - D) γ + (1 - α); Wherein, H represents the regulation factor, α represents the joint loss balance factor, and γ represents the attention degree index of difficult and easy samples.

8. The high-resolution one-dimensional range profile anomaly detection method integrating multi-detector capabilities according to claim 4, characterized in that, The expression of the preset classification loss is: Loss=H×Loss BCE ; Among them, Loss represents the preset classification loss, H represents the adjustment factor, and Loss BCE represents the classification loss, and its expression is: Among them, p i represents the probability that the p i -th sample is predicted as the positive class, and Y i represents the true label of the samples in the training dataset.

9. A high-resolution one-dimensional range profile anomaly detection device integrating the capabilities of multiple detectors is used to implement the high-resolution one-dimensional range profile anomaly detection method integrating the capabilities of multiple detectors according to any one of claims 1 to 8, and is characterized in that, Including: A data acquisition module, used to obtain the high-resolution one-dimensional range image of the radar to be detected; A data anomaly recognition module, used to input the high-resolution one-dimensional range image of the radar to be detected into the trained neural network model for processing, and obtain the anomaly detection result of the high-resolution one-dimensional range image of the radar to be detected; Among them, the trained neural network model is obtained by training an initial neural network model with preset category data as the training data set and under the guidance of a preset classification loss; the preset classification loss includes a regulation factor, and the regulation factor is used to characterize the detection difficulty.

Images