Multi-ignition signal and anomaly detection method based on frame difference and deep learning

Abstract

The invention belongs to the technical field of image recognition and intelligent detection, and particularly discloses a multi-ignition signal and anomaly detection method based on frame difference and deep learning, comprising the following steps: S1, video image acquisition; s2, target detection positioning and ROI region extraction are carried out; s3, performing frame difference method and gray Top-K analysis on each ROI region to obtain an image with a suspected ignition region; s4, ignition signal discrimination and ignition frequency statistics are carried out by adopting the trained convolutional neural network model; and S5, ignition sequence modeling and abnormity identification are carried out. According to the multi-ignition signal and anomaly detection method based on the frame difference and deep learning, non-contact measurement, multi-target positioning, accurate detection and counting of ignition signals and automatic discrimination of abnormal signals are realized, and the method has the advantages of high detection precision, high reliability, high automation degree and wide application range; the method is suitable for multi-ignition signal detection and anomaly detection in a complex scene.

Description

Technical Field

[0001] The present invention relates to the field of image recognition and intelligent detection technology, and in particular to a multi-ignition signal and anomaly detection method based on frame difference and deep learning. Background Art

[0002] Igniters are widely used in a variety of fields, including aircraft engines, missile launch systems, and gas ignition equipment. Their operating status impacts the safety, reliability, and operational efficiency of these devices. Accurately detecting the ignition status and sequence of multiple igniters is a critical requirement, especially in multi-igniter systems.

[0003] In existing technologies, ignition signal detection methods primarily rely on direct acquisition of electrical signals. Typical implementations involve using components such as current probes, Hall effect current sensors, or sampling resistors to convert the ignition signal's current into a voltage signal. The igniter's operating status is then determined based on characteristics such as the signal's waveform, amplitude, and duration. However, these traditional methods suffer from the following deficiencies:

[0004] (1) The detection method is contact-based and must be directly connected to the device being tested, which not only causes interference but may also damage the circuit;

[0005] (2) When the number of igniters is large and dense, complex wiring and interfaces are required, resulting in cumbersome construction and high maintenance costs;

[0006] (3) The detection process is highly complex and there is a risk of misoperation;

[0007] (4) The detection scheme lacks an intuitive display of the ignition status, cannot count the number of ignitions, and is difficult to quickly and automatically identify the ignition sequence and abnormal status.

[0008] Computer vision can be used to achieve non-contact detection of igniters. Existing image detection methods mainly include frame difference method and image threshold method. However, these methods still have the following drawbacks:

[0009] (1) Sensitive to factors such as illumination changes, background interference, and noise, resulting in a high false detection rate;

[0010] (2) Without the introduction of target detection or target positioning algorithms, it is impossible to accurately identify and separate the multiple igniters in the video, resulting in the inability to independently identify and count multiple points;

[0011] (3) There is a lack of a modeling mechanism based on time series. Usually, only the changes in a single frame of image are focused, which makes it impossible to effectively establish the order of ignition signals and thus cannot identify abnormal ignition behaviors such as skipping and repetition.

[0012] (4) There is a lack of mechanisms for recording, extracting, and analyzing abnormal events, making it difficult to provide effective video evidence for abnormal situations.

[0013] In summary, existing ignition status detection technologies have yet to simultaneously meet the requirements for non-contact, multi-target positioning and detection, high-accuracy ignition identification, abnormal sequence recognition, and automatic recording. Therefore, a comprehensive solution that can overcome these shortcomings and achieve detection and abnormality identification of multiple ignition signals is urgently needed. Summary of the Invention

[0014] The purpose of the present invention is to provide a multi-ignition signal and anomaly detection method based on frame difference and deep learning, which overcomes the problems of lack of multi-target positioning capability and inability to realize abnormal sequence recognition in the existing technology, realizes non-contact measurement, multi-target positioning, accurate detection and counting of ignition signals and automatic discrimination of abnormal signals, has the advantages of high detection accuracy, strong reliability, high degree of automation and wide adaptability, and is suitable for multi-ignition signal detection and anomaly detection in complex scenarios.

[0015] To achieve the above objectives, the present invention provides a multi-ignition signal and anomaly detection method based on frame difference and deep learning, comprising the following steps:

[0016] S1, acquiring continuous image frames containing multiple ignition positions and performing grayscale preprocessing;

[0017] S2. Analyze the image frame using the target detection model, identify the ignition locations, and construct the ROI region based on the detection results;

[0018] S3. For each ROI region, extract the image content of the corresponding ROI region in the current frame and the previous frame, select the first K pixels with the highest grayscale values, and calculate the grayscale difference. When the grayscale difference of all selected pixels is greater than a preset threshold, it is determined that there is a suspected ignition signal in the ROI region, and an image of the suspected ignition region is obtained;

[0019] S4. Input the image of the suspected ignition area obtained in S3 into the trained convolutional neural network model to perform a binary classification task to further determine whether it is an ignition signal. For the ROI area determined to be an ignition signal, record and update its ignition count;

[0020] S5. Generate an ignition sequence for the identified ignition position and classify the ignition sequence using the trained long short-term memory network model. If the classification label is repeated or skipped, it is determined that there is an abnormality, and the image frame containing the ignition signal is automatically saved to generate a video clip marking the abnormal position.

[0021] Preferably, S2 is specifically:

[0022] The target detection model is used to analyze the video frame images, identify the ignition positions, obtain the detection frame of each ignition position, eliminate the repeated detection of the same ignition position through the center point deduplication mechanism, and construct the ROI area based on the center coordinates of the ignition position. After detecting and locating all ignition positions, they are sorted and numbered according to the spatial position of each igniter.

[0023] Preferably, in S2, the detection frame is defined as follows:

[0024] B i =(x 1,i ,y 1,i ,x 2,i ,y 2,i ,s i ),s i ∈[0,1];

[0025] Among them, (x 1,i ,y 1,i ) and (x 2,i ,y 2,i ) represent the upper left corner and lower right corner coordinates of the detection frame of the i-th ignition position, s i To test confidence;

[0026] The formula for calculating the center coordinates is:

[0027]

[0028] Based on the center coordinates, a rectangular area with fixed width and height is constructed. The center point deduplication mechanism is that when the Euclidean distance between the center coordinates of any two ignition positions is less than the set threshold, only the one with higher confidence is retained. The calculation formula of the Euclidean distance is:

[0029]

[0030] Preferably, in S3, a double loop traversal mechanism is used to extract the image content of the corresponding ROI area in the current frame and the previous frame for each ROI area, specifically:

[0031] Outer loop: traverse each frame of the video sequence;

[0032] Inner loop: Analyze each ROI area in the current frame image one by one.

[0033] Preferably, in S3, the first K pixels with the highest grayscale values ​​are selected, which are:

[0034]

[0035] in, and They represent the top K pixel sets with the highest grayscale values ​​in the grayscale image of the i-th ROI region in the current t-th frame and the t-1-th frame of the previous frame, respectively. and Represent the grayscale images of the i-th ROI region in the current frame and the previous frame respectively;

[0036] The grayscale difference calculation formula is as follows:

[0037]

[0038] in, Indicates the grayscale difference between the jth maximum grayscale value in the i ROI region in the current frame and the previous frame.

[0039] Preferably, in S3, K is preferably 5.

[0040] Preferably, in S4, the trained convolutional neural network model includes an input layer, two convolutional layers, two pooling layers, two fully connected layers and an output layer, which is used to determine whether the ROI area contains an ignition signal.

[0041] Preferably, in S4, an activation flag mechanism for recording the ignition status of each ROI region in the previous frame is introduced into the counting logic of the ignition number to determine whether repeated counting occurs in consecutive frames.

[0042] Preferably, in S5, the long short-term memory network model is trained by a cross entropy loss function to classify the firing sequence including the current firing number, and outputs a category label corresponding to the current firing sequence, wherein the category label includes normal sequence, repeated sequence, and skipped sequence;

[0043] The normal order satisfies a strictly increasing relationship or a cyclic increasing relationship, that is:

[0044] Strictly increasing: (d t-1 =d t-2 +1) and (d t =d t-1 +1);

[0045] Cycle increment: If d t-2 =8, then [8,9,1] belongs to normal cyclic increment;

[0046] There are two consecutive numbers in the repeating sequence that are exactly the same, namely:

[0047] d t-2 =d t-1 or d t-1 =d t ;

[0048] Skip when the numbering is not incremented continuously, that is:

[0049] d t-1 ≠d t-2 +1 or d t ≠d t-1 +1, and does not meet the cyclic increment rule.

[0050] Preferably, in S5, when the ignition sequence is determined to be an abnormal ignition behavior, the image frames in the adjacent time windows before and after the abnormal event are automatically saved, and a marking box is drawn on the abnormal ignition ROI area, and the frame sequence is exported as a video file.

[0051] Therefore, the present invention adopts the above-mentioned multi-ignition signal and anomaly detection method based on frame difference and deep learning, and the beneficial effects are as follows:

[0052] (1) Non-contact measurement: This invention adopts pure image and video detection method, which does not require physical electrical connection, has high safety, low equipment maintenance cost, is suitable for complex environment application and has a wide range of adaptability;

[0053] (2) Fast and accurate multi-target positioning: This invention combines the YOLO target detection model and the center point deduplication mechanism to achieve fast and accurate multi-target positioning with high detection accuracy and strong reliability;

[0054] (3) Accurate detection and counting of ignition signals: This invention combines frame difference analysis with LeNet convolutional neural network classification to significantly reduce the false detection rate and missed detection rate;

[0055] (4) Automatic identification of abnormal signals: The present invention introduces the abnormal sequence recognition function through the LSTM model to predict and identify abnormal ignition signal sequences, which can accurately identify ignition abnormal events such as repetition and skipping, with a high degree of automation;

[0056] (5) Automatically save abnormal videos: The present invention automatically saves abnormal video clips through the LSTM model, marks the location where the abnormality occurs, and improves the visualization of abnormality analysis.

[0057] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of the overall process of an embodiment of a multi-ignition signal and anomaly detection method based on frame difference and deep learning of the present invention;

[0059] Figure 2 2 is a flow chart of S2 of an embodiment of a method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to the present invention;

[0060] Figure 33 is a flow chart of S3 of an embodiment of a method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to the present invention;

[0061] Figure 4 4 is a flow chart of S4 of an embodiment of a method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to the present invention;

[0062] Figure 5 This is a flow chart of S5 of an embodiment of a multi-ignition signal and anomaly detection method based on frame difference and deep learning of the present invention. DETAILED DESCRIPTION

[0063] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0064] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0065] Example 1

[0066] like Figure 1 As shown, a multi-point ignition signal and anomaly detection method based on frame difference and deep learning includes the following steps:

[0067] S1. Video image acquisition. Specifically, first, use an industrial camera to capture a video image sequence of multiple igniters. The video resolution and frame rate are set according to actual needs. If the captured video image is a color image, it needs to be converted to grayscale. The ignition count is displayed at each ignition position in the captured video.

[0068] S2, target detection and positioning and ROI area extraction, specifically, Figure 2 As shown, the target detection model (You Only Look Once, YOLO) is used to analyze the video frame image to identify each ignition position. In this embodiment, the YOLOv8 target detection model is used to perform target detection on the input image to identify multiple ignition positions in the video frame and output the detection frame coordinates of each ignition position, which are defined as follows:

[0069] B i =(x 1,i ,y 1,i ,x 2,i ,y 2,i ,s i ),s i ∈[0,1]

[0070] Among them, (x 1,i ,y 1,i ) and (x 2,i ,y 2,i) represent the upper left corner and lower right corner coordinates of the detection frame of the i-th ignition position, s i Confidence of detection.

[0071] Deduplication is performed by calculating the center coordinates of the detection frame. The center coordinate calculation formula is:

[0072]

[0073] When the Euclidean distance between the center coordinates of any two ignition positions is less than the set threshold, only the one with higher confidence is retained. The calculation formula of the Euclidean distance is:

[0074]

[0075] After deduplication, when the number of detected ignition positions is equal to the number of igniters, the positions of all igniters have been found. The ignition positions are sorted according to their spatial positions in the video frame image, numbered in sequence, and a fixed-size region of interest (ROI area) is constructed based on the center point of the ignition position.

[0076] S3, frame difference method and grayscale Top-K analysis, specifically, Figure 3 As shown in the figure, a double loop traversal mechanism is used to process the entire video image sequence, and detailed analysis is performed frame by frame and ROI by ROI. Specifically:

[0077] Outer loop: traverse each frame of the video sequence;

[0078] Inner loop: Analyze each ROI area in the current frame image one by one.

[0079] For each ROI region, extract the top K pixels with the highest grayscale values ​​in the corresponding ROI region in the current frame and the previous frame, and record them as:

[0080]

[0081] in, and Respectively represent the top K pixel sets with the highest grayscale values ​​in the grayscale image of the i-th ROI region in the current frame (t-th frame) and the previous frame (t-1-th frame); and Represent the grayscale images of the i-th ROI region in the current frame and the previous frame respectively.

[0082] Calculate the grayscale difference of the corresponding area:

[0083]

[0084] in, Indicates the grayscale difference between the jth maximum grayscale value in the i ROI region in the current frame and the previous frame.

[0085] If all differences satisfy:

[0086]

[0087] Among them, T s is the preset threshold, i.e. the grayscale difference threshold, which is set to an empirical value of 5.

[0088] The ROI area is determined to be a suspected ignition signal, and an image of the suspected ignition area is obtained.

[0089] S4, ignition signal identification and ignition number statistics, specifically, Figure 4 As shown in the figure, the image of the suspected ignition area is scaled and normalized, and then input into the convolutional neural network model (LeNet) for binary classification task detection to further determine whether it is an ignition signal.

[0090] If the above conditions are not met, it is considered that no ignition signal is generated in the current frame of the ROI area, and the next ROI area is processed. After all ROI areas are traversed, the next frame image is entered for the same processing.

[0091] In this embodiment, the LeNet convolutional neural network model is used to classify and identify images of areas with suspected ignition signals. The images of areas with suspected ignition signals are scaled to 32×32 pixels, normalized, and then input into the LeNet convolutional neural network model. The specific structure of the LeNet convolutional neural network model is as follows:

[0092] Input layer: 1×32×32 pixels;

[0093] The first convolutional layer (Conv1): 6 5×5 convolution kernels;

[0094] The first pooling layer (Pool1): 2×2 maximum pooling;

[0095] The second convolutional layer (Conv2): 16 5×5 convolution kernels;

[0096] Second pooling layer (Pool2): 2×2 maximum pooling;

[0097] First fully connected layer (FC1): 120 dimensions;

[0098] Second fully connected layer (FC2): 84 dimensions;

[0099] Softmax output layer: outputs the probabilities of two categories (ignition signal and non-ignition signal).

[0100] LeNet convolutional neural network model output formula:

[0101]

[0102] Among them, x (i) It is the ROI area image after scaling and normalization; To output the probability vectors of two categories, ignition signal or non-ignition signal, the Softmax function is used to normalize them to between 0 and 1.

[0103] Classification judgment formula:

[0104]

[0105] in, is the classification result label, and the category with a greater probability is taken as the classification result, 0 represents a non-ignition signal, and 1 represents an ignition signal.

[0106] When the model outputs an ignition signal, an "active flag" mechanism is further introduced to record the ignition status of the previous frame in each ROI area to prevent repeated counting of consecutive frames. The specific method is as follows:

[0107] If an ignition signal is detected in the ROI area in the current frame, and the flag bit of the ROI area in the previous frame is in the "inactive" state, the ignition count of the corresponding number of the ROI area is increased by one, and the corresponding flag bit of the current frame is changed to the "active" state;

[0108] If an ignition signal is detected in the ROI area of ​​the current frame, but the corresponding ROI area flag bit in the previous frame is in the "active" state, it is considered that the same ignition signal continues in consecutive frames and is no longer counted;

[0109] If no ignition signal is detected in the ROI area of ​​the current frame, the flag bit of the corresponding ROI area is set to the "inactive" state.

[0110] S5, ignition sequence modeling and anomaly identification, specifically, Figure 5 As shown, the ignition sequence under normal working conditions cycles from left to right in the ROI numbering. When ignition sequence anomalies such as number jumps and duplications occur in actual operation, traditional methods are difficult to effectively identify. In this embodiment, to detect and identify anomalies in the ignition sequence of multiple ignition signals, the ROI region numbers corresponding to the ignition signals are first extracted from the entire video to form a complete ignition sequence.

[0111] At the beginning of the video analysis, starting from the first frame, the ignition position is detected frame by frame. When a certain ROI area is determined to contain the ignition position, the number of the current ROI area is recorded. Every time a new ignition position is detected, the corresponding ROI area number is added to the end of the ignition sequence, recorded as:

[0112] S={d1,d2,d3,...,d i ,...d t}

[0113] Where S is the ignition sequence of the entire video; t is the total number of ignition positions; d i is the ROI area number corresponding to the i-th ignition position.

[0114] Perform sliding window processing on the firing order sequence S to construct sequence classification samples:

[0115] Define a sliding window of length L and a step size of 1 to form a sample set X t :

[0116] X t =[d t-L+1 ,d t-L+2 ,...,d t ]

[0117] For each sequence sample X of length L=3 t =[d t-2 ,d t-1 ,d t ], define the following three firing order tags:

[0118] (1) Normal order:

[0119] Satisfy a strict increasing relationship or a cyclic increasing relationship, that is:

[0120] Strictly increasing: (d t-1 =d t-2 +1) and (d t =d t-1 +1);

[0121] Cycle increment: If d t-2 =8, then [8,9,1] belongs to normal cyclic increment.

[0122] (2) Repeat order:

[0123] There are two consecutive numbers in the sequence that are exactly the same: d t-2 =d t-1 or d t-1 =d t .

[0124] (3) Skip:

[0125] The numbering is not incremented continuously: d t-1 ≠d t-2 +1 or d t ≠d t-1+1, and does not meet the cyclic increment rule.

[0126] A classification model of the firing sequence is constructed using a trained long short-term memory network model (LSTM). In this embodiment, the trained LSTM network model adopts a bidirectional long short-term memory network (Bi-LSTM):

[0127] The model input is a firing order sequence X with a length of L=3 t , after vector encoding through the embedding layer (Embedding) to capture the sequence structure and order characteristics, the model can be expressed as:

[0128] h t =Bi-LSTM(E(X t );θ)

[0129] Among them, E(·) is the Embedding layer, which converts the number into a real vector representation of fixed dimension; h t is the output hidden state vector of Bi-LSTM; θ is the network parameter pre-trained by sequence classification task.

[0130] In the output layer, the Softmax function is used to calculate the category probability:

[0131]

[0132] in, is the category probability vector output by the model, with three categories in total (normal order, repeated order, skipped); W, b are the weight matrix and bias vector of the output layer, respectively, obtained through training.

[0133] The predicted category is the category with the highest probability:

[0134]

[0135] The category labels are defined as:

[0136] Normal order: C t =0;

[0137] Repeat order: C t =1;

[0138] Skip: C t =2.

[0139] When the ignition sequence is judged as abnormal ignition behavior, the image frames in the adjacent time windows before and after the abnormal event are automatically saved, and a red mark box is drawn on the abnormal ignition ROI area, and the frame sequence is exported as a video file.

[0140] Whenever a new ignition signal is detected, the latest ignition number is combined with the previous two historical numbers to form a new sliding window sequence X t , and input the above classification model for classification judgment:

[0141] If the classification result C t =0, which means normal sequence, the system continues to operate normally;

[0142] If the classification result C t =1 or C t =2, the system immediately determines it as an abnormal ignition signal, marks the position of the abnormal ignition ROI area, saves the relevant frames before and after the abnormality, and outputs them.

[0143] Therefore, the present invention adopts the above-mentioned multi-ignition signal and anomaly detection method based on frame difference and deep learning to realize non-contact measurement, multi-target positioning, precise detection and counting of ignition signals, and automatic discrimination of abnormal signals. It has the advantages of high detection accuracy, strong reliability, high degree of automation and wide adaptability, and is suitable for multi-ignition signal detection and anomaly detection in complex scenarios.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-fire signal and anomaly detection method based on frame difference and deep learning, characterized in that: The following steps are involved: S1, acquiring continuous image frames containing multiple ignition positions and performing grayscale preprocessing; S2. Analyze the image frame using the target detection model, identify the ignition locations, and construct the ROI region based on the detection results; S3. For each ROI region, extract the image content of the corresponding ROI region in the current frame and the previous frame, select the first K pixels with the highest grayscale values, and calculate the grayscale difference. When the grayscale difference of all selected pixels is greater than a preset threshold, it is determined that there is a suspected ignition signal in the ROI region, and an image of the suspected ignition region is obtained; S4. Input the image of the suspected ignition area obtained in S3 into the trained convolutional neural network model to perform a binary classification task to further determine whether it is an ignition signal. For the ROI area determined to be an ignition signal, record and update its ignition count; S5. Generate an ignition sequence for the identified ignition position and classify the ignition sequence using the trained long short-term memory network model. If the classification label is repeated or skipped, it is determined that there is an abnormality, and the image frame containing the ignition signal is automatically saved to generate a video clip marking the abnormal position.

2. The method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to claim 1, characterized in that: S2 is specifically: The target detection model is used to analyze the video frame images, identify the ignition positions, obtain the detection frame of each ignition position, eliminate the repeated detection of the same ignition position through the center point deduplication mechanism, and construct the ROI area based on the center coordinates of the ignition position. After detecting and locating all ignition positions, they are sorted and numbered according to the spatial position of each igniter.

3. The method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to claim 2, characterized in that: In S2, the detection box is defined as follows: B i =(x 1,i ,y 1,i ,x 2,i ,y 2,i ,s i ),s i ∈[0,1]; Among them, (x 1,i ,y 1,i ) and (x 2,i ,y 2,i ) represent the upper left corner and lower right corner coordinates of the detection frame of the i-th ignition position, s i To test confidence; The formula for calculating the center coordinates is: Based on the center coordinates, a rectangular area with fixed width and height is constructed. The center point deduplication mechanism is that when the Euclidean distance between the center coordinates of any two ignition positions is less than the set threshold, only the one with higher confidence is retained. The calculation formula of the Euclidean distance is:

4. The method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to claim 1, wherein: In S3, a double loop traversal mechanism is used to extract the image content of the corresponding ROI area in the current frame and the previous frame for each ROI area, specifically: Outer loop: traverse each frame of the video sequence; Inner loop: Analyze each ROI area in the current frame image one by one.

5. The method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to claim 1, wherein: In S3, the first K pixels with the highest grayscale values ​​are selected, which are: in, and They represent the top K pixel sets with the highest grayscale values ​​in the grayscale image of the i-th ROI region in the current t-th frame and the t-1-th frame of the previous frame, respectively. and Represent the grayscale images of the i-th ROI region in the current frame and the previous frame respectively; The grayscale difference calculation formula is as follows: in, Indicates the grayscale difference between the jth maximum grayscale value in the i ROI region in the current frame and the previous frame.

6. The method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to claim 5, characterized in that: In S3, K is preferably 5.

7. The method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to claim 1, characterized in that: In S4, the trained convolutional neural network model includes an input layer, two convolutional layers, two pooling layers, two fully connected layers and an output layer.

8. The method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to claim 1, wherein: In S4, an activation flag mechanism for recording the ignition status of each ROI region in the previous frame is introduced into the counting logic of the ignition times to determine whether repeated counting occurs in consecutive frames.

9. The method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to claim 1, characterized in that: In S5, the long short-term memory network model is trained using a cross entropy loss function to classify the firing sequence including the current firing number, and outputs a category label corresponding to the current firing sequence, wherein the category label includes normal sequence, repeated sequence, and skipped sequence; The normal order satisfies a strictly increasing relationship or a cyclic increasing relationship, that is: Strictly increasing: (d t-1 =d t-2 +1) and (d t =d t-1 +1); Cycle increment: If d t-2 =8, then [8,9,1] belongs to normal cyclic increment; There are two consecutive numbers in the repeating sequence that are exactly the same, namely: d t-2 =d t-1 or d t-1 =d t ; Skip when the numbering is not incremented continuously, that is: d t-1 ≠d t-2 +1 or d t ≠d t-1 +1, and does not meet the cyclic increment rule.

10. The method for detecting multiple ignition signals and anomalies based on frame difference and deep learning according to claim 1, characterized in that: In S5, when the ignition sequence is determined to be an abnormal ignition behavior, the image frames in the adjacent time windows before and after the abnormal event are automatically saved, and a marking box is drawn on the abnormal ignition ROI area, and the frame sequence is exported as a video file.

Images