CN116680639A - Deep-learning-based anomaly detection method for sensor data of deep-sea submersible - Google Patents

Abstract

The present invention relates to an abnormality detection method for deep-sea submersible sensor data based on deep learning, comprising: preprocessing the sensor data of the deep-sea submersible to be detected and inputting it into a trained abnormality detection model to perform multi-category abnormality detection; abnormality detection The model includes a diffusion model, a multi-head self-attention mechanism layer, and a fully connected layer; the diffusion model outputs the extracted feature information, and assigns weights to each input item through the self-attention mechanism layer, and selects the The information that is more critical to the current task goal is classified through the fully connected layer to determine whether there is an abnormality. The present invention pays better attention to the important features of the data; the attention mechanism can help the model to better pay attention to the important features in the data, thereby improving the accuracy and performance of the model.

Description

Anomaly detection method for deep-sea submersible sensor data based on deep learning

technical field

The invention relates to a method for abnormality detection of sensor data of a deep-sea submersible based on deep learning, and belongs to the technical field of sensor equipment detection.

Background technique

Manned submersible is an important equipment for ocean exploration and marine scientific research. Its sensor system can collect seabed environmental data and submersible status data in real time, which plays an important role in the operation monitoring and fault diagnosis of submersibles. However, due to the complex and changeable marine environment, the sensor data is affected by noise and interference, and there are data anomalies, which require real-time anomaly detection and positioning.

As a machine learning technique based on artificial neural networks, deep learning can automatically learn feature representation from data and perform anomaly detection without human intervention. Deep learning has made significant progress in image recognition, speech recognition, natural language processing and other fields, and is gradually being applied to marine science and technology. In recent years, deep learning technology has also gained widespread attention and application in the abnormal detection of submersible sensor data. Its advantage is that it can automatically learn feature representation and discriminant criteria, has strong modeling ability for nonlinear and complex data, and can improve the accuracy and efficiency of anomaly detection.

Although the manned submersible sensor data anomaly detection algorithm based on deep learning has strong modeling ability and high-precision detection effect, there are still some defects and deficiencies, including the following aspects:

1. Insufficient data volume: Anomaly detection algorithms based on deep learning require a large amount of data for training and verification. However, in practical applications, due to the high cost of sensor data acquisition, the data volume is usually small, which will affect the performance and generality of the algorithm. ability.

2. Difficulty in obtaining labels: Data sets that need to be labeled by deep learning algorithms usually require manual labeling, which is costly and difficult to obtain complete and accurate label data sets, which also affects the performance of the algorithm.

3. Dependence on hardware: Deep learning algorithms require a large amount of computing resources and high-performance hardware devices, which increases the cost and complexity of the algorithm.

4. Model overfitting: Due to the strong fitting ability of the deep learning model, if the distribution of the training data set and the test data set are inconsistent, it may lead to overfitting of the model and affect the performance and generalization ability of the algorithm.

Therefore, the anomaly detection method for sensor data of manned submersibles based on deep learning still needs further optimization and improvement to improve its performance and reliability, and enhance its feasibility and sustainability in practical applications.

Diffusion models can be used in anomaly detection to detect abnormal behavior in information dissemination. Specifically, the diffusion model can use statistical analysis methods based on propagation behavior to monitor abnormal nodes and events in the process of information dissemination, so as to discover potential abnormal behaviors. For example, in social networks, diffusion models can be used to detect abnormal behaviors such as false information, phishing, and zombie fans, helping social platforms improve information security and credibility. In the financial field, diffusion models can be used to detect financial crimes such as abnormal transactions and money laundering. Overall, diffusion models have broad application prospects in anomaly detection.

However, due to the complex mathematical description of the diffusion model, the curse of dimensionality may occur when dealing with high-dimensional data, which will greatly increase the amount of calculation and affect the accuracy and performance of the model. Diffusion models may have local optimal solutions when dealing with nonlinear problems, which will affect the optimization and training effects of the model.

The smoothing of the data by the diffusion model may lead to the loss of information, especially for some abnormal points with obvious characteristics, which may be smoothed out, which will affect the detection effect.

Diffusion models usually require a lot of computing resources and time, especially when dealing with large-scale data, it may be necessary to use distributed computing and other technologies to speed up the calculation, which will also increase the computational cost and complexity.

To sum up, although the diffusion model is widely used in many fields, it is also necessary to select the appropriate model and algorithm according to the specific situation in practical application to solve the problem and avoid the influence caused by the inadequacy.

Contents of the invention

Aiming at the deficiencies in the prior art, the present invention provides a method for abnormal detection of sensor data of deep-sea submersibles based on deep learning;

Based on deep learning, the present invention proposes a new algorithm combining the attention mechanism and the diffusion model, thereby optimizing the performance of the diffusion model and improving the accuracy of the diffusion model for data processing.

The invention utilizes a multi-head self-attention mechanism to generate a random step size, thereby simulating the change process of sensor data. First define an initial time step, and then introduce a multi-head self-attention mechanism. Specifically, for each node, a random step value is calculated according to the weight and diffusion coefficient of its neighbor nodes to update the value of the node. At the same time, considering the association relationship between different nodes in the sliding window, the step size can be adjusted according to the edge weight between nodes, so as to better simulate the data change process.

The present invention combines the multi-head attention mechanism with the reverse diffusion network in the diffusion model, compares the predicted denoised data matrix with the corresponding time step original data matrix, and adjusts the diffusion coefficient to remove noise according to the weight of its adjacent nodes , for anomaly detection.

Technical scheme of the present invention is:

A method for anomaly detection of deep-sea submersible sensor data based on deep learning, comprising:

Preprocess the sensor data of the deep-sea submersible to be detected and input it into the trained anomaly detection model for multi-category anomaly detection;

The anomaly detection model includes a diffusion model, a multi-head self-attention mechanism layer, and a fully connected layer; the diffusion model outputs the extracted feature information, and passes through the self-attention mechanism layer. The self-attention mechanism layer assigns weights to each input item. Through the weight The size selects the information that is more critical to the current task goal from many feature information, and classifies it through the fully connected layer to determine whether there is an abnormality.

Preferably according to the present invention, pretreatment comprises:

Select the sensor data, remove the interference data before and after the sensor enters the water, and form the initial data matrix X _s (n), n∈R;

Perform normalization processing on each point in the initial data matrix X _s (n) and obtain the matrix X _{normal (n)} , and the normalization processing is shown in formula (I):

Slice the normalized sensor data X _normal(n) to form an m×6×6 matrix X′ _s ={x ₀ , x ₁ ,...,x _m }, where x _m is 6x6 matrix.

Preferably according to the present invention, the training process of the abnormality detection model includes:

Build a data set: collect the sensing data collected by six sensors, namely, the fuel tank pressure sensor of the deep-sea submersible, the VP2 fuel tank temperature sensor, the 10LPM compensator displacement sensor, the 15LPM compensator displacement sensor, the VP1 fuel tank temperature sensor, and the 24V current detection sensor. Carry out the preprocessing to obtain the training data set, when the deep-sea submersible breaks down, the corresponding sensor data is marked as abnormal, otherwise it is marked as normal;

Input the training data set to the anomaly detection model for training, including:

Use the optimized data set X _'s to train the diffusion model. The diffusion model is divided into forward diffusion process and reverse diffusion process: the forward diffusion process is to continuously add Gaussian noise to the sliced data matrix X _'s , and the reverse diffusion process The process is to continuously denoise, expecting to finally restore it to the original data matrix X′ _s ;

In the training process of each round of forward diffusion, a random time step t is selected for the training sample, t∈[0, T], and the Gaussian noise corresponding to the time step t is applied to the optimized data matrix. The time step is converted into the corresponding time step embedding E(t); Gaussian noise is continuously added to the optimized data set X' _s to form a noise matrix X _noise (t);

In the back diffusion process, in each round of training, the noise matrix X _noise (t) and the encoded time step embedding E(t) are used as input to train the Unet network in the diffusion model; each round After training, the Unet network will predict the noise to be removed from the noise matrix to form a new matrix X _rnoise (t), and add the prediction result of the Unet network and the corresponding time step in the forward diffusion process to E(t) Gaussian noise for comparison, calculate the predicted loss rate Loss(t), and pass this result through the multi-head self-attention mechanism layer to randomly generate the time step embedding E(t+1) in the next round of training;

The trained diffusion model will output the extracted features, pass the feature information through the multi-head self-attention mechanism layer, assign weights to each input item, and select the information that is more critical to the current task goal from many feature information through the weight size. And classify through the fully connected layer, and judge whether there is a deep-sea submersible or not according to whether there is an abnormality in the classification result.

Preferably according to the present invention, the process of multi-category anomaly detection includes:

1) Generate a random step size, including:

Define the initial time step t ₀ ;

Perform phase encoding on the initial time step t ₀ to obtain the initial time step embedding E(0); the value of the analog quantity within a time period is represented by a pulse time, and the pulse sequence obtained by connecting all time periods represents the entire time The change of analog quantity in the process;

In the process of training the Unet network, the loss rate Loss(t) value generated by each iteration will be used as negative feedback, and the time step embedding E(t) will be used as input through the multi-head self-attention mechanism layer to generate the next round of training Iterated time step embedding E(t+1);

2) Forward diffusion, including:

During each round of training, the noise corresponding to the time step embedding E(t) is added to the data matrix X′ _s , and the noise matrix X _noise (t) corresponding to the time step t is obtained after each round of training. After the training, the final noise matrix X _noise (T) is obtained;

For the data matrix X _noise (t), in each time step, q(X _noise (t)|X _noise (t-1)) is expressed as a subject to the mean Normal distribution with variance (1-α _t )I, I is the identity matrix: as shown in formula (II): q(X _noise (t)|X _noise (t-1))=m(X _noise (t );/>

α _t is used as a hyperparameter to adjust the amount of noise added according to the time step t;

By formula (III), namely:

The formula (IV) is derived:

3) Backdiffusion, including:

In each round of training, the time is stepped into the UNet network embedded in E(t) and the noise matrix X _noise (T) input into the diffusion model, and the predicted and removed noise is obtained through network training, and the matrix X _rnoise (t) is obtained, Comparing this prediction result with the noise added by E(t) under the corresponding time step embedding, the loss Loss(t) is calculated;

For the matrix X _rnoise (t) generated after each round of training, there is formula (V) under the time step embedding E(t):

In formula (V), ε _θ (X _rnoise (t), t) represents the noise removed by prediction, and ε is the Gaussian noise corresponding to the time step embedding E(t);

After the back-diffusion process is over, the final reduction forms the matrix X _rnoise (0);

4) After the diffusion model training is over, the vector feature matrix of the original data matrix X _'s will be output

X _feature ＝{x _feature0 ，x _feature1 ，...，x _featurem }, where x _featurem is a 6×6 matrix, and the self-attention value h _m is calculated separately for the feature result x _featurem of each slice, and each node represents A data sample, each edge represents the relationship between two nodes, for each node, update the value of the node according to the weight of its neighbor nodes, increase the weight of the node that has a greater impact on the task goal, and vice versa reduce the weight , and then calculate the multi-head self-attention mechanism MultiHead(X _feature ) through formula (VI), and finally get the feature that has a great influence on detecting whether there is an abnormal phenomenon in the deep-sea submersible sensor:

MultiHead(X _feature )=Concat(h ₀ , h ₁ , . . . , h _m ) (VI)

The classification of data is realized through the fully connected layer, and the data corresponding to the abnormal state is marked as 1, and the data corresponding to the normal state is marked as 0.

Preferably, according to the present invention, the difference between the predicted noise and the actual noise is calculated according to the cross-entropy loss function, which is used as a loss Loss to evaluate the performance of the abnormality detection model, measure the pros and cons of the prediction results of the abnormality detection model, and use it as a negative feedback item ; Loss Loss is shown in formula (VII):

In formula (VII), y _i represents the label of the sample, the flag of abnormal data is 1, and the flag of normal data is 0, and p _i represents the probability that the sample is predicted to be abnormal data.

A computer device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the abnormality detection method based on deep-sea submersible sensor data are realized.

A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the deep-learning-based abnormality detection method for deep-sea submersible sensor data are realized.

The beneficial effects of the present invention are:

The present invention uses the graph attention mechanism to calculate the weights between nodes, which can bring the following benefits:

1. Better focus on important features of data: The attention mechanism can help the model to better focus on important features in the data, thereby improving the accuracy and performance of the model.

2. Strengthen the discrimination ability of the model: the attention mechanism can make the model pay more attention to the differences and differences between different features, so that the model has stronger discrimination and discrimination ability.

3. Improve the robustness of the model: the attention mechanism can make the model more flexible and robust, and can better adapt to different data distributions and noise interference, thereby improving the generalization ability and robustness of the model.

4. Stronger interpretability: The attention mechanism can make the model more interpretable, so as to better understand the decision-making process and results of the model, and help to further optimize the performance and effect of the model.

5. Can handle more complex data: The attention mechanism can make the diffusion model more suitable for processing more complex high-dimensional data, thus expanding the application range and potential of the diffusion model.

Description of drawings

Fig. 1 is the schematic flow chart of the anomaly detection method of the deep-sea submersible sensor data based on the present invention:

Figure 2 is a schematic diagram of the network structure of the diffusion model:

Fig. 3 is a schematic diagram of the network structure of the anomaly detection model of the present invention.

Detailed ways

The present invention will be further limited below in conjunction with the accompanying drawings and embodiments, but not limited thereto.

Example 1

A method for anomaly detection of deep-sea submersible sensor data based on deep learning, including:

Preprocess the sensor data of the deep-sea submersible and input it into the trained anomaly detection model for multi-classification anomaly detection; the sensor data of the deep-sea submersible includes the fuel tank pressure sensor of the deep-sea submersible, the VP2 fuel tank temperature sensor, and the displacement of the 10LPM compensator Sensor, 15LPM compensator displacement sensor, VP1 fuel tank temperature sensor, 24V current detection sensor, the sensing data collected by these 6 sensors;

As shown in Figure 3, the anomaly detection model includes a diffusion model, a multi-head self-attention mechanism layer, and a fully-connected layer; the diffusion model outputs the extracted feature information, and passes through the self-attention mechanism layer, and the self-attention mechanism layer is for each input Items are assigned weights, and the information that is more critical to the current task goal is selected from a large number of feature information through the weight, and is classified through the fully connected layer to determine whether there is an abnormality.

It is an effective method to introduce the attention mechanism into the diffusion model in the present invention, which can make full use of the relationship between nodes, improve the accuracy and efficiency of the model, further improve the accuracy and performance of the diffusion model, and contribute to better application in various data analysis and processing tasks. Compared with the traditional Euler method or implicit Euler method, the present invention can better handle complex graph data, and has better applicability and universality.

The invention proposes to introduce an attention mechanism into the diffusion model to generate a random step size, which can make the model pay more attention to important nodes and edges, thereby generating a more realistic and accurate random step size, which helps to better simulate the change process of sensor data , the attention mechanism can make the model more flexible and robust, and can better adapt to different data distributions and noise interference, thereby improving the generalization ability and robustness of the model, which can further improve the accuracy of the model and help further optimization Model performance and effects.

Example 2

According to the anomaly detection method of a deep-sea submersible sensor data based on deep learning described in embodiment 1, the difference is that:

preprocessing, including:

Select the sensor data, remove the interference data before and after the sensor enters the water, and form the initial data matrix X _s (n), n∈R; the sensor data includes: VP1 fuel tank temperature in the sensor, VP2 fuel tank temperature, 24V current, fuel tank pressure , 10LPM compensator displacement, 15LPM compensator displacement data.

Perform normalization processing on each point in the initial data matrix X _s (n) and obtain the matrix X _{normal (n)} , and the normalization processing is shown in formula (I):

Slice the normalized sensor data X _normal(n) to form an m×6×6 matrix form X′ _s ={x ₀ , x ₁ ,...,x _m }, where x _m is 6x6 matrix.

The training process of the anomaly detection model includes:

Build a data set: collect the sensing data collected by six sensors, namely, the fuel tank pressure sensor of the deep-sea submersible, the VP2 fuel tank temperature sensor, the 10LPM compensator displacement sensor, the 15LPM compensator displacement sensor, the VP1 fuel tank temperature sensor, and the 24V current detection sensor. Carry out the preprocessing to obtain the training data set, when the deep-sea submersible breaks down, the corresponding sensor data is marked as abnormal, otherwise it is marked as normal;

Input the training data set to the anomaly detection model for training, including:

Use the optimized data set X _'s to train the diffusion model. The diffusion model is divided into forward diffusion process and reverse diffusion process: the forward diffusion process is to continuously add Gaussian noise to the sliced data matrix X _'s , and the reverse diffusion process The process is to continuously denoise, expecting to finally restore it to the original data matrix X′ _s ;

In the training process of each round of forward diffusion, a random time step t is selected for the training sample, t∈[0, T], and the Gaussian noise corresponding to the time step t is applied to the optimized data matrix. The time step is converted into the corresponding time step embedding E(t); Gaussian noise is continuously added to the optimized data set X′ _s to form a noise matrix X _noise (t); after the forward diffusion process is completed, the noise matrix is finally formed X _noise (T);

In the back diffusion process, in each round of training, the noise matrix X _noise (t) and the encoded time step embedding E(t) are used as input to train the Unet network in the diffusion model; each round After training, the Unet network will predict the noise to be removed from the noise matrix to form a new matrix X _rnoise (t), and add the prediction result of the Unet network and the corresponding time step in the forward diffusion process to E(t) Gaussian noise for comparison, calculate the predicted loss rate Loss(t), and pass this result through the multi-head self-attention mechanism layer to randomly generate the time step embedding E(t+1) in the next round of training;

The trained diffusion model will output the extracted features, pass the feature information through the multi-head self-attention mechanism layer, assign weights to each input item, and select the information that is more critical to the current task goal from many feature information through the weight size. And classify through the fully connected layer, and judge whether there is a deep-sea submersible or not according to whether there is an abnormality in the classification result.

As shown in Figure 1, the process of multi-category anomaly detection includes:

1) Generate a random step size, including:

Define the initial time step t ₀ ;

Perform phase encoding on the initial time step t ₀ to obtain the initial time step embedding E(0); the value of the analog quantity within a time period is represented by a pulse time, and the pulse sequence obtained by connecting all time periods represents the entire time The change of analog quantity in the process;

In the process of training the Unet network, the loss rate Loss(t) value generated by each iteration will be used as negative feedback, and the time step embedding E(t) will be used as input through the multi-head self-attention mechanism layer to generate the next round of training Iterated time step embedding E(t+1);

2) As shown in the solid line part of Figure 2, forward diffusion includes:

During each round of training, the noise corresponding to the time step embedding E(t) is added to the data matrix X′ _s , and the noise matrix X _noise (t) corresponding to the time step t is obtained after each round of training. After the training, the final noise matrix X _noise (T) is obtained;

For the data matrix X _noise (t), in each time step, q(X _noise (t)|X _noise (t-1)) is expressed as a subject to the mean Normal distribution with variance (1-α _t )I, I is the identity matrix: as shown in formula (II):

α _t is used as a hyperparameter to adjust the amount of noise added according to the time step t;

By formula (III), namely:

The formula (IV) is derived:

3) As shown in the dotted line part of Figure 2, backdiffusion includes:

In each round of training, the time is stepped into the UNet network embedded in E(t) and the noise matrix X _noise (T) input into the diffusion model, and the predicted and removed noise is obtained through network training, and the matrix X _rnoise (t) is obtained, Comparing this prediction result with the noise added by E(t) under the corresponding time step embedding, the loss Loss(t) is calculated;

For the matrix X _rnoise (t) generated after each round of training, there is formula (V) under the time step embedding E(t):

In formula (V), ε _θ (X _rnoise (t), t) represents the noise removed by prediction, and ε is the Gaussian noise corresponding to the time step embedding E(t);

After the back-diffusion process is over, the final reduction forms the matrix X _rnoise (0);

4) After the diffusion model training is over, the vector feature matrix of the original data matrix X _'s will be output

X _feature ＝{x _feature0 ，x _feature1 ，...，x _featurem }, where x _featurem is a 6×6 matrix, and the self-attention value h _m is calculated separately for the feature result x _featurem of each slice, and each node represents A data sample, each edge represents the relationship between two nodes, for each node, update the value of the node according to the weight of its neighbor nodes, increase the weight of the node that has a greater impact on the task goal, and vice versa reduce the weight , and then calculate the multi-head self-attention mechanism MultiHead(X _feature ) through formula (VI), and finally get the feature that has a great influence on detecting whether there is an abnormal phenomenon in the deep-sea submersible sensor:

MultiHead(X _feature )=Concat(h ₀ , h ₁ , . . . , h _m ) (VI)

The classification of data is realized through the fully connected layer, and the data corresponding to the abnormal state is marked as 1, and the data corresponding to the normal state is marked as 0.

By combining the diffusion model and the graph attention mechanism, a more accurate random step size can be generated, which helps to better simulate the change process of sensor data, thereby improving the performance and effect of anomaly detection.

According to the cross-entropy loss function, the difference between the predicted noise and the actual noise is calculated as the loss Loss, which is used to evaluate the performance of the anomaly detection model, measure the pros and cons of the prediction results of the anomaly detection model, and use it as a negative feedback item; the loss Loss is as follows: VII) as shown:

In formula (VII), y _i represents the label of the sample, the flag of abnormal data is 1, and the flag of normal data is 0, and p _i represents the probability that the sample is predicted to be abnormal data.

Example 3

A computer device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the deep learning-based deep-sea submersible sensor data abnormality detection method described in Embodiment 1 or 2 are realized.

Example 4

A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method for abnormality detection of deep-sea submersible sensor data based on deep learning described in Embodiment 1 or 2 are realized.

Claims (7)

1. A method for anomaly detection of deep-sea submersible sensor data based on deep learning, characterized in that, comprising: Preprocess the sensor data of the deep-sea submersible to be detected and input it into the trained anomaly detection model for multi-category anomaly detection; The anomaly detection model includes a diffusion model, a multi-head self-attention mechanism layer, and a fully connected layer; the diffusion model outputs the extracted feature information, and passes through the self-attention mechanism layer. The self-attention mechanism layer assigns weights to each input item. Through the weight The size selects the information that is more critical to the current task goal from many feature information, and classifies it through the fully connected layer to determine whether there is an abnormality. 2. a kind of anomaly detection method based on the deep-sea submersible sensor data of deep learning according to claim 1, is characterized in that, preprocessing comprises: Select the sensor data, remove the interference data before and after the sensor enters the water, and form the initial data matrix X _s (n),n∈R; Perform normalization processing on each point in the initial data matrix X _s (n) and obtain the matrix X _{normal (n)} , and the normalization processing is shown in formula (I): Slice the normalized sensor data X _normal(n) to form an m×6×6 matrix X′ _s ={x ₀ ,x ₁ ,…,x _m }, where x _m is 6× 6 matrix. 3. the abnormal detection method of a kind of deep-sea submersible sensor data based on deep learning according to claim 1, is characterized in that, the training process of abnormal detection model comprises: Build a data set: collect the sensing data collected by six sensors, namely, the fuel tank pressure sensor of the deep-sea submersible, the VP2 fuel tank temperature sensor, the 10LPM compensator displacement sensor, the 15LPM compensator displacement sensor, the VP1 fuel tank temperature sensor, and the 24V current detection sensor. Carry out the preprocessing to obtain the training data set, when the deep-sea submersible breaks down, the corresponding sensor data is marked as abnormal, otherwise it is marked as normal; Input the training data set to the anomaly detection model for training, including: Use the optimized data set x _'s to train the diffusion model. The diffusion model is divided into forward diffusion process and reverse diffusion process: the forward diffusion process is to continuously add Gaussian noise to the sliced data matrix x _'s , and the reverse diffusion process The process is to continuously denoise, expecting to finally restore it to the original data matrix x′ _s ; In the training process of each round of forward diffusion, a random time step t,t∈[0,T] is selected for the training sample, and the Gaussian noise corresponding to the time step t is applied to the optimized data matrix. The time step is converted into the corresponding time step embedding E(t); Gaussian noise is continuously added to the optimized data set X' _s to form a noise matrix x _noise (t); In the back diffusion process, in each round of training, the noise matrix n _noise (t) and the encoded time step embedding E(t) are used as input to train the Ut network in the diffusion model; each round After training, the Unet network will predict the noise to be removed from the noise matrix to form a new matrix X _rnoise (t), and add the prediction result of the Unet network to the corresponding time step in the forward diffusion process)(t) Gaussian noise for comparison, calculate the predicted loss rate Loss(t), and pass this result through the multi-head self-attention mechanism layer to randomly generate the time step embedding E(t+1) in the next round of training; The trained diffusion model will output the extracted features, pass the feature information through the multi-head self-attention mechanism layer, assign weights to each input item, and select the information that is more critical to the current task goal from many feature information through the weight size. And classify through the fully connected layer, and judge whether there is a deep-sea submersible or not according to whether there is an abnormality in the classification result. 4. a kind of abnormal detection method based on the deep-sea submersible sensor data of deep sea according to claim 1, it is characterized in that, the process of multi-category abnormal detection comprises: 1) Generate a random step size, including: Define the initial time step t ₀ ; Perform phase encoding on the initial time step t ₀ to obtain the initial time step embedding E(0); the value of the analog quantity within a time period is represented by a pulse time, and the pulse sequence obtained by connecting all time periods represents the entire time The change of analog quantity in the process; In the process of training the Unet network, the loss rate Loss(t) value generated by each iteration will be used as negative feedback, and the time step embedding E(t) will be used as input through the multi-head self-attention mechanism layer to generate the next round of training Iterated time step embedding E(t+1); 2) Forward diffusion, including: During each round of training, the noise corresponding to the time step embedding E(t) is added to the data matrix X′ _s , and the noise matrix X _noise (t) corresponding to the time step t is obtained after each round of training. After the training, the final noise matrix X _noise (T) is obtained; For the data matrix X _noise (t), in each time step, q(X _noise (t)|X _noise (t-1)) is expressed as a subject to the mean Normal distribution with variance (1-α _t )I, I is the identity matrix: as shown in formula (II): α _t is used as a hyperparameter to adjust the amount of noise added according to the time step t; By formula (III), namely: The formula (IV) is derived: 3) Backdiffusion, including: In each round of training, the time is stepped into the UNet network embedded in E(t) and the noise matrix X _noise (T) input into the diffusion model, and the predicted and removed noise is obtained through network training, and the matrix X _rnoise (t) is obtained, Comparing this prediction result with the noise added by E(t) under the corresponding time step embedding, the loss Loss(t) is calculated; For the matrix X _rnoise (t) generated after each round of training, there is formula (V) under the time step embedding E(t): In formula (V), ε _θ (X _rnoise (t), t) represents the noise removed by prediction, and ε is the Gaussian noise corresponding to the time step embedding E(t); After the back-diffusion process is over, the final reduction forms the matrix X _rnoise (0); 4) After the diffusion model training is over, the vector feature matrix X _feature of the original data matrix X′ _s will be output = {x _feature0 , x _feature1 ,…, x _featurem }, where x _featurem is a 6×6 matrix, for each The feature result x _featurem of the slice calculates the self-attention value h _m separately, each node represents a data sample, and each edge represents the association relationship between two nodes. For each node, update the node according to the weight of its neighbor nodes value, increase the weight of the node that has a greater impact on the task target, and vice versa, reduce the weight, and then calculate the multi-head self-attention mechanism MuktiHead(X _feature Q through formula (VI), and finally get the detection of whether there is an abnormality in the deep-sea submersible sensor Characteristics of a phenomenon with greater influence: MultiHead(x _feature )=Concat(h ₀ , h ₁ ,...,h _m ) (VI) The classification of data is realized through the fully connected layer, and the data corresponding to the abnormal state is marked as 1, and the data corresponding to the normal state is marked as 0. 5. according to claim 1-4 arbitrary described a kind of abnormal detection method based on deep-sea submersible sensor data, it is characterized in that, according to cross-entropy loss function calculation prediction noise and the gap of actual noise, as loss Loss , used to evaluate the performance of the anomaly detection model, measure the pros and cons of the prediction results of the anomaly detection model, and use it as a negative feedback item; the loss Loss is shown in formula (VII): In formula (VII), y _i represents the label of the sample, the flag of abnormal data is 1, and the flag of normal data is 0, and p _i represents the probability that the sample is predicted to be abnormal data. 6. A computer device, comprising a memory and a processor, the memory stores a computer program, wherein when the processor executes the computer program, the computer program based on deep learning according to any one of claims 1-5 is realized. Steps of a method for anomaly detection of bathyscaphe sensor data. 7. A computer-readable storage medium, on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the deep-sea submersible sensor data based on deep learning described in any one of claims 1-5 is realized The steps of the anomaly detection method.