Bi-LSTM-based light quantum arrival time sequence noise suppression method

Abstract

The invention provides a quantum ranging noise suppression method based on Bi-LSTM (Bidirectional Long Short Term Memory) time sequence prediction and reconstruction. Firstly, entangled photon arrival time sequence signals in the quantum ranging system are preprocessed, specifically, data cleaning, denoising and standardization processing are included, and the data quality is improved and made to meet the model input requirement; then, segmenting the preprocessed data into time sequence fragments with fixed lengths, and converting the time sequence fragments into a three-dimensional tensor format to adapt to an input structure of a Bi-LSTM network; secondly, sending the processed data into a Bi-LSTM self-encoding network, and performing primary processing on the grouped input data by using a full connection layer; thirdly, inputting the primarily processed data into a Bi-LSTM (Bidirectional Long Short Term Memory) layer, capturing a bidirectional dependency relationship of a time sequence by using a bidirectional LSTM layer, extracting key features and compressing the key features into low-dimensional features for representation; then, in a decoding stage, the encoded data are input into a Bi-LSTM layer; and finally, enabling the data to pass through a full connection layer to obtain reconstruction of an original input time sequence by the model.

Description

Technical Field

[0001] The invention belongs to the field of quantum signal processing and signal processing, and particularly relates to noise suppression of photon arrival time in a dynamic environment, specifically a noise suppression method based on sequence prediction and reconstruction. Background Art

[0002] In modern positioning and navigation technology, accurate measurement of photon arrival time is one of the key factors to achieve high-precision positioning. However, various noise factors in dynamic environments (such as environmental scattering, dynamic object interference, multipath effects in signal propagation, etc.) will have a significant impact on the measurement of photon arrival time, resulting in reduced positioning accuracy and insufficient system robustness. For example, in a quantum positioning system, the dynamic changes in the scattering environment will lead to an increase in the photon loss rate of the light source signal light time pulse sequence, thereby reducing positioning accuracy.

[0003] In addition, traditional noise suppression methods have limitations when dealing with complex noise in dynamic environments. For example, methods based on geometric and semantic constraints do not work well when dealing with non-rigid objects and low-dynamic moving targets. However, noise suppression methods based on deep learning, such as Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (Bi-LSTM), perform well in processing time series data and can effectively capture long-term dependencies in time series, thereby better suppressing noise.

[0004] In recent years, Bi-LSTM networks have been widely used in many fields, such as self-interference suppression in communication systems, noise suppression of electrocardiogram signals, and real-time noise suppression of speech signals. These studies have shown that Bi-LSTM networks can significantly improve the system's noise resistance and positioning accuracy by learning noise characteristics and reconstructing signals. However, the existing technology still has shortcomings in positioning applications in dynamic environments. For example, most dynamic SLAM methods are mainly designed for indoor RGBD environments and lack versatility. In addition, when dealing with complex noise in dynamic environments, traditional methods often cannot effectively suppress the impact of noise on the arrival time of photons, which leads to a decrease in positioning accuracy and robustness.

[0005] In order to solve the above problems and to deepen the understanding of time series data and improve the quality of its reconstruction, the present invention proposes a new noise suppression method based on time series prediction and reconstruction. Summary of the invention

[0006] The purpose of the present invention is to provide a noise suppression method for denoising a photon arrival time series. The method improves the noise suppression effect in the entangled photon arrival time series by using a time series reconstruction technology of an autoencoder based on a bidirectional long short-term memory network (Bi-directional Long Short-Term Memory, Bi-LSTM).

[0007] The technical solution adopted by the present invention is: a noise suppression method based on the combination of an autoencoder and a Bi-LSTM network, which specifically includes the following steps:

[0008] Step 1: Use the quantum distance measurement system to obtain the arrival time series signal of the entangled photons containing noise, and save the obtained signal photons and idle photons arrival time series locally through TDC technology, which are recorded as CH1 and CH2 respectively;

[0009] Step 2: Preprocess the time series data in CH1 and CH2, including removing outliers, standardizing, and obtaining eigenvalues, to improve the efficiency and accuracy of subsequent model processing;

[0010] Step 3: Group the preprocessed entangled photon arrival time series into data to form an input format suitable for deep learning models. Grouped input data X input It is usually a three-dimensional tensor with a shape of (number of samples, time steps, number of features). The processed input data X input Satisfy the formula:

[0011] X input =Reshape(X p ,(S,T,F)) (1)

[0012] The Reshape function is used to reshape the data into the required three-dimensional tensor, X p is the preprocessed data, S, T, F are the number of samples, time steps, and number of features, respectively, X input is the processed input data;

[0013] Step 4: The input data X processed by the above steps input The data is sent to the Bi-LSTM autoencoder network, and a fully connected layer with 32 neurons is used to perform preliminary processing on the input data. The Gaussian Error Linear Unit (GeLU) is selected as the activation function to enhance the model's ability to process nonlinear relationships. Output data of the fully connected layer fc1 Satisfy the formula:

[0014] GeLU(x)=x×Φ(x) (2)

[0015] Output fc1 =GeLU(W fc1 ×X input +b fc1 ) (3)

[0016] Among them, the GeLU function is an activation function used to enhance the model's ability to handle nonlinear relationships. x is the input of the GeLU function, Φ(x) is the cumulative distribution function of the standard normal distribution, and W fc1 is the weight matrix of the fully connected layer, X input is the data input to the fully connected layer, b fc1 is the bias term, Output fc1 It is the output data after being processed by the fully connected layer;

[0017] This step aims to map the input data into a low-dimensional space while preserving the key features of the time series;

[0018] Step 5: Output the data processed by the fully connected layer fc1 The data is passed to a Bi-LSTM layer containing 64 neurons for data processing. The Bi-LSTM layer can capture the forward and reverse dependencies of the time series at the same time through a bidirectional processing mechanism, thereby extracting the deep features of the time series. This layer also applies the GeLU activation function to maintain consistency and efficiency in processing. After being processed by the Bi-LSTM layer, the data is further compressed into a low-dimensional latent space to form a compact feature representation of the time series;

[0019] Step 6: The data processing enters the decoder stage. The decoder consists of multiple LSTM layers. Its structure is symmetrical with the encoder and is responsible for gradually decoding the low-dimensional features back to the shape of the original time series.

[0020] Step 7: After entering the decoder stage, the feature data extracted in the encoding stage is output bi-lstm The output data of the Bi-LSTM layer is: bi-lstm-dec Satisfy the formula:

[0021] Output bi-lstm-dec =GeLU(W bi-lstm-dec ×Output bi-lstm +b bi-lstm-dec ) (4)

[0022] Among them, W bi-lstm-decis the weight matrix of the fully connected layer, Output bi-lstm is the feature data extracted in the encoding stage, b bi-lstm-dec is the bias term, Output bi-lstm-dec is the output data of the Bi-LSTM layer;

[0023] Step 8: Output the data obtained in the previous step bi-lstm-dec The input is sent to a fully connected layer with 100 neurons, which uses the GeLU activation function to complete the reconstruction process with high quality. The output data of the fully connected layer satisfies the formula:

[0024] Output fc2 =GeLU(W fc2 ×Output bi-lstm-dec +b fc2 ) (5)

[0025] Among them, W fc2 is the weight matrix of the fully connected layer, Output bi-lstm-dec is the output data of the Bi-LSTM layer and is also the data input to the fully connected layer. fc2 is the bias term, Output fc2 It is the output data after being processed by the fully connected layer;

[0026] Step 9: Finally, the output of the decoder is mapped back to the feature dimension of the original input data to obtain the reconstructed time series data.

[0027] The step 2 includes the following steps:

[0028] Step 2 (a) The time series X after outliers are removed original Split into fixed-length sequences, each containing N time steps. The processed time series data matrix X processed Satisfy the formula:

[0029] X processed =Split(X original ,N) (6)

[0030] Among them, the Split function divides the original data into multiple sequences of length N;

[0031] Step 2 (ii) Perform minimum-maximum normalization on the features of each time step to make them have a uniform scale. The standardized time series data matrix X norm Satisfy the formula:

[0032]

[0033] Among them, X processedis the processed time series data matrix, X norm is the standardized time series data matrix;

[0034] Step 2 (3): According to the coincidence counting algorithm, the standardized time series data matrix X norm Calculate and get the eigenvalue F of each time step photon . Satisfies the formula:

[0035] F photon =CoincidenceCount(X norm ) (8)

[0036] Among them, the CoincidenceCount function calculates the standardized data according to the coincidence counting algorithm to obtain the eigenvalue of each time step, F photon is the eigenvalue at each time step;

[0037] Step 2 (iv) The calculated eigenvalue F photon Vectorize to form the feature vector X for each time step f The feature vector X at each time step f Satisfy the formula:

[0038] X f = Vectorize(F photon ) (9)

[0039] The Vectorize function is used to convert the feature values ​​calculated by the coincidence counting algorithm into a one-dimensional array so that these data can be used as input to the autoencoder network.

[0040] The step five includes the following steps:

[0041] The core of Bi-LSTM is the LSTM unit. Each LSTM unit performs the following operations at each time step:

[0042] Step 5 (a) Calculate the activation value of the input gate to determine how much of the current input information is written into the cell state. The activation value of the input gate satisfies:

[0043]

[0044] Step 5 (ii) calculates the activation value of the forget gate to determine which information in the cell state needs to be forgotten. The activation value of the forget gate satisfies:

[0045]

[0046] Step 5 (iii) Combine the operations of the input gate and the forget gate to update the cell state. Satisfy the formula:

[0047]

[0048] Step 5 (iv) Calculate the activation value of the output gate to determine how much information in the cell state is output. The activation value of the output gate satisfies:

[0049]

[0050] Among them, f h is the activation function of the system state, f s is the activation function of the internal state, usually the tanh function, which can transform the input vector to between (-1,1). Its purpose is to limit the vector value of each layer of the neural network, so as to avoid the large difference between the values ​​inside the vector, which will have a negative impact on the final output of the neural network; g is the gate unit, which is updated with the time step. In essence, it is a feedforward neural network with the sigmoid function (denoted as σ in the figure) as the activation function. The sigmoid function can be understood as a set of weights that control the output within the interval of (0,1); X t represents the tth sequence value in the sequence data, W is the weight matrix of the network, b is the bias, and the subscripts i, f, and o represent the input gate, forget gate, and output gate respectively;

[0051] Bi-LSTM captures the bidirectional dependency information in the sequence through the LSTM layers in the forward and backward directions. The specific calculation process is:

[0052] Step 5 (V) At time t, Bi-LSTM can use the information at time t+1 and time t-1 at the same time. In the Bi-LSTM network, the output y of each node t The calculation formula is as follows:

[0053] h t = f(W t [s t-1 ; x] + b t ) (14)

[0054] h' t = f(W t ' [s' t-1 ; x] + b' t ) (15)

[0055] y t = g(W y [h t ; h' t ] + b y ) (16)

[0056] Among them, h tis the hidden state of the forward LSTM unit at time t, h t ' is the hidden layer state of the reverse LSTM unit at time t. t and W t ' corresponds to the weight matrix of the forward LSTM and the reverse LSTM at time t. s t-1 is the hidden state of the forward LSTM unit at time t-1, s t ' -1 is the hidden state of the reverse LSTM unit at time t-1. x is the input data at time t. y t is the output of each node of the Bi-LSTM network at time t. f is the activation function in the LSTM unit, which is usually used to calculate the hidden state. g is used to calculate the output y t The activation function. t and b t ' is the bias term of the forward LSTM and the reverse LSTM at time t, W y is the value corresponding to the calculated output y t The weight matrix when b y Yes and W y The bias term used together to adjust the output y t Baseline value of

[0057] Step 5 (6), the output of the Bi-LSTM layer satisfies the formula:

[0058] Output bi-lstm =GeLU(W bi-lstm ×Output fc1 +b bi-lstm ) (17)

[0059] Among them, W bi-lstm is the weight matrix of the fully connected layer, Output fc1 is the output data of the fully connected layer and the input data of the Bi-LSTM layer. bi-lstm is the bias term, Output bi-lstm It is the output data of the Bi-LSTM layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is the optical path diagram of the photon distance measurement of the present invention

[0061] Figure 2 This is the LSTM network structure diagram of the present invention;

[0062] Figure 3 This is a Bi-LSTM network structure diagram of the present invention;

[0063] Figure 4This is the Bi-LSTM autoencoding network structure diagram of the present invention. Specific implementation plan

[0064] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0065] Step 1: Use the quantum distance measurement system to obtain the arrival time series signal of the entangled photons containing noise, and save the obtained signal photons and idle photons arrival time series locally through TDC technology, which are recorded as CH1 and CH2 respectively;

[0066] Step 2: Preprocess the time series data in CH1 and CH2, including removing outliers, standardizing, and obtaining eigenvalues, to improve the efficiency and accuracy of subsequent model processing;

[0067] Step 3: Group the preprocessed entangled photon arrival time series into data to form an input format suitable for deep learning models. Grouped input data X input It is usually a three-dimensional tensor with a shape of (number of samples, time steps, number of features). The processed input data X input Satisfy the formula:

[0068] X input =Reshape(X p, (S,T,F)) (18)

[0069] The Reshape function is used to reshape the data into the required three-dimensional tensor, X p is the preprocessed data, S, T, F are the number of samples, time steps, and number of features, respectively, X input is the processed input data.

[0070] In this experiment, we finally formed the input data X input It is a time series data containing 64 samples, each sample has 100 time steps, and each time step has 1 feature, so the shape of the input data is (64, 100, 1);

[0071] Step 4: The input data X processed by the above steps input The data is sent to the Bi-LSTM autoencoder network, and a fully connected layer with 32 neurons is used to perform preliminary processing on the input data. The Gaussian Error Linear Unit (GeLU) is selected as the activation function to enhance the model's ability to process nonlinear relationships. Output data of the fully connected layer fc1 Satisfy the formula:

[0072] GeLU(x)=x×Φ(x) (19)

[0073] Output fc1 =GeLU(W fc1 ×X input +b fc1 ) (20)

[0074] Among them, the GeLU function is an activation function used to enhance the model's ability to handle nonlinear relationships. x is the input of the GeLU function, Φ(x) is the cumulative distribution function of the standard normal distribution, and W fc1 is the weight matrix of the fully connected layer, X input is the data input to the fully connected layer, b fc1 is the bias term, Output fc1 It is the output data after being processed by the fully connected layer;

[0075] This step aims to map the input data into a low-dimensional space while retaining the key features of the time series. The shape of the data after preliminary processing is (64, 100, 32);

[0076] Step 5: Output the data processed by the fully connected layer fc1 Pass in a Bi-LSTM layer with 64 neurons. The Bi-LSTM layer can capture the forward and reverse dependencies of the time series at the same time through a bidirectional processing mechanism, thereby extracting the deep features of the time series. This layer also applies the GeLU activation function to maintain consistency and efficiency of processing;

[0077] Step 6: After being processed by the Bi-LSTM layer, the data is further compressed into a low-dimensional latent space to form a compact feature representation of the time series. The shape of the encoder output data is (64, 64). The data processing process enters the decoder stage. The decoder consists of multiple LSTM layers, and its structure is symmetrical with the encoder. It is responsible for gradually decoding the low-dimensional features back to the shape of the original time series;

[0078] Step 7: After entering the decoder stage, the feature data extracted in the encoding stage is output bi-lstm The output data of the Bi-LSTM layer is: bi-lstm-dec Satisfy the formula:

[0079] Output bi-lstm-dec =GeLU(W bi-lstm-dec ×Output bi-lstm +b bi-lstm-dec ) (twenty one)

[0080] Among them, W bi-lstm-dec is the weight matrix of the fully connected layer, Output bi-lstm is the feature data extracted in the encoding stage, b bi-lstm-dec is the bias term, Output bi-lstm-dec is the output data of the Bi-LSTM layer;

[0081] Step 8: Output the data obtained in the previous step bi-lstm-dec Input a fully connected layer with 100 neurons, which uses the GeLU activation function to complete the reconstruction process with high quality. The output data of the fully connected layer satisfies the formula:

[0082] Output fc2 =GeLU(W fc2 ×Output bi-lstm-dec +b fc2 ) (twenty two)

[0083] Among them, W fc2 is the weight matrix of the fully connected layer, Output bi-lstm-dec is the output data of the Bi-LSTM layer and is also the data input to the fully connected layer. fc2 is the bias term, Output fc2 It is the output data after being processed by the fully connected layer;

[0084] Step 9: Finally, the output of the decoder is mapped back to the feature dimension of the original input data, thereby obtaining the reconstructed time series data in the form of (64, 100).

[0085] The step 2 includes the following steps:

[0086] Step 2 (a) The time series X after outliers are removed original Split into fixed-length sequences, each containing 100 time steps. The processed time series data matrix X processed Satisfy the formula:

[0087] X processed =Split(X original ,100) (23)

[0088] Among them, the Split function splits the original data into multiple sequences of length 100;

[0089] Step 2 (ii) Perform minimum-maximum normalization on the features of each time step to make them have a uniform scale. The standardized time series data matrix X norm Satisfy the formula:

[0090]

[0091] Among them, X processed is the processed time series data matrix, X norm is the standardized time series data matrix;

[0092] Step 2 (3): According to the coincidence counting algorithm, the standardized time series data matrix X norm Calculate and get the eigenvalue F of each time step photon . Satisfies the formula:

[0093] F photon =CoincidenceCount(X norm ) (25)

[0094] Among them, the CoincidenceCount function calculates the standardized data according to the coincidence counting algorithm to obtain the eigenvalue of each time step, F photon is the eigenvalue at each time step;

[0095] Step 2 (iv) The calculated eigenvalue F photon Vectorize to form the feature vector X for each time step f The feature vector X at each time step f Satisfy the formula:

[0096] X f = Vectorize(F photon ) (26)

[0097] The Vectorize function is used to convert the feature values ​​calculated by the coincidence counting algorithm into a one-dimensional array so that these data can be used as input to the autoencoder network.

[0098] The step five includes the following steps:

[0099] The core of Bi-LSTM is the LSTM unit. Each LSTM unit performs the following operations at each time step:

[0100] Step 5 (a) Calculate the activation value of the input gate to determine how much of the current input information is written into the cell state. The activation value of the input gate satisfies:

[0101]

[0102] Step 5 (ii) calculates the activation value of the forget gate to determine which information in the cell state needs to be forgotten. The activation value of the forget gate satisfies:

[0103]

[0104] Step 5 (iii), combine the operation of input gate and forget gate to update the cell state. Satisfy the formula:

[0105]

[0106] Step 5 (iv) Calculate the activation value of the output gate to determine how much information in the cell state is output. The activation value of the output gate satisfies:

[0107]

[0108] Among them, f h is the activation function of the system state, f s is the activation function of the internal state, usually the tanh function, which can transform the input vector to between (-1,1). Its purpose is to limit the vector value of each layer of the neural network, so as to avoid the large difference between the values ​​inside the vector, which will have a negative impact on the final output of the neural network; g is the gate unit, which is updated with the time step. In essence, it is a feedforward neural network with the sigmoid function (denoted as σ in the figure) as the activation function. The sigmoid function can be understood as a set of weights that control the output within the interval of (0,1); X t represents the tth sequence value in the sequence data, W is the weight matrix of the network, b is the bias, and the subscripts i, f, and o represent the input gate, forget gate, and output gate respectively;

[0109] Bi-LSTM captures the bidirectional dependency information in the sequence through the LSTM layers in the forward and backward directions. The specific calculation process is:

[0110] Step 5 (V) At time t, Bi-LSTM can use the information at time t+1 and time t-1 at the same time. In the Bi-LSTM network, the output y of each node t The calculation formula is as follows:

[0111] h t = f(W t [s t-1 ; x] + b t ) (31)

[0112] h' t = f(W t ' [s' t-1 ; x] + b' t ) (32)

[0113] y t = g(W y [h t ; h' t ] + b y ) (33)

[0114] Among them, h t is the hidden state of the forward LSTM unit at time t, h t ' is the hidden layer state of the reverse LSTM unit at time t. t and W t ' corresponds to the weight matrix of the forward LSTM and the reverse LSTM at time t. s t-1 is the hidden state of the forward LSTM unit at time t-1, s t ' -1 is the hidden state of the reverse LSTM unit at time t-1. x is the input data at time t. y t is the output of each node of the Bi-LSTM network at time t. f is the activation function in the LSTM unit, which is usually used to calculate the hidden state. g is used to calculate the output y t The activation function. t and b t ' is the bias term of the forward LSTM and the reverse LSTM at time t, W y is the value corresponding to the calculated output y t The weight matrix when b y Yes and W y The bias term used together to adjust the output y t Baseline value of

[0115] Step 5 (6), the output of the Bi-LSTM layer satisfies the formula:

[0116] Output bi-lstm =GeLU(W bi-lstm ×Output fc1 +b bi-lstm ) (34)

[0117] Among them, W bi-lstm is the weight matrix of the fully connected layer, Output fc1 is the output data of the fully connected layer and the input data of the Bi-LSTM layer. bi-lstm is the bias term, Output bi-lstm It is the output data of the Bi-LSTM layer.

Claims

1. A method for suppressing noise in the time series of light quantum arrival based on Bi-LSTM, characterized in that The following steps are involved: Step 1: Use the quantum distance measurement system to obtain the arrival time series signal of the entangled photons containing noise, and save the obtained signal photons and idle photons arrival time series locally through TDC technology, which are recorded as CH1 and CH2 respectively; Step 2: Preprocess the time series data in CH1 and CH2, including removing outliers, standardizing, and obtaining eigenvalues, to improve the efficiency and accuracy of subsequent model processing; Step 3: Group the preprocessed entangled photon arrival time series into data to form an input format suitable for deep learning models. Grouped input data X input It is usually a three-dimensional tensor with a shape of (number of samples, time steps, number of features). The processed input data X input Satisfy the formula: X input =Reshape(X p ,(S,T,F)) (1) The Reshape function is used to reshape the data into the required three-dimensional tensor, X p is the preprocessed data, S, T, F are the number of samples, time steps, and number of features, respectively, X input is the processed input data. In this experiment, we finally formed the input data X input It is a time series data containing 64 samples, each sample has 100 time steps, and each time step has 1 feature, so the shape of the input data is (64, 100, 1); Step 4: The input data X processed by the above steps input The data is sent to the Bi-LSTM autoencoder network, and a fully connected layer with 32 neurons is used to perform preliminary processing on the input data. The Gaussian Error Linear Unit (GeLU) is selected as the activation function to enhance the model's ability to process nonlinear relationships. Output data of the fully connected layer fc1 Satisfy the formula: GeLU(x)=x×Φ(x) (2) Output fc1 =GeLU(W fc1 ×X input +b fc1 ) (3) Among them, the GeLU function is an activation function used to enhance the model's ability to handle nonlinear relationships. x is the input of the GeLU function, Φ(x) is the cumulative distribution function of the standard normal distribution, and W fc1 is the weight matrix of the fully connected layer, X input is the data input to the fully connected layer, b fc1 is the bias term, Output fc1 It is the output data after being processed by the fully connected layer; This step aims to map the input data into a low-dimensional space while retaining the key features of the time series. The shape of the data after preliminary processing is (64, 100, 32); Step 5: Output the data processed by the fully connected layer fc1 Pass in a Bi-LSTM layer with 64 neurons. The Bi-LSTM layer can capture the forward and reverse dependencies of the time series at the same time through a bidirectional processing mechanism, thereby extracting the deep features of the time series. This layer also applies the GeLU activation function to maintain consistency and efficiency of processing; Step 6: After being processed by the Bi-LSTM layer, the data is further compressed into a low-dimensional latent space to form a compact feature representation of the time series. The shape of the encoder output data is (64, 64). The data processing process enters the decoder stage. The decoder consists of multiple LSTM layers, and its structure is symmetrical with the encoder. It is responsible for gradually decoding the low-dimensional features back to the shape of the original time series; Step 7: After entering the decoder stage, the feature data extracted in the encoding stage is output bi-lstm The output data of the Bi-LSTM layer is: bi-lstm-dec Satisfy the formula: Output bi-lstm-dec =GeLU(W bi-lstm-dec ×Output bi-lstm +b bi-lstm-dec ) (4) Among them, W bi-lstm-dec is the weight matrix of the fully connected layer, Output bi-lstm is the feature data extracted in the encoding stage, b bi-lstm-dec is the bias term, Output bi-lstm-dec is the output data of the Bi-LSTM layer; Step 8: Output the data obtained in the previous step bi-lstm-dec Input a fully connected layer with 100 neurons, which uses the GeLU activation function to complete the reconstruction process with high quality. The output data of the fully connected layer satisfies the formula: Output fc2 =GeLU(W fc2 ×Output bi-lstm-dec +b fc2 ) (5) Among them, W fc2 is the weight matrix of the fully connected layer, Output bi-lstm-dec is the output data of the Bi-LSTM layer and is also the data input to the fully connected layer. fc2 is the bias term, Output fc2 It is the output data after being processed by the fully connected layer; Step 9: Finally, the output of the decoder is mapped back to the feature dimension of the original input data, thereby obtaining the reconstructed time series data in the form of (64, 100).

2. According to the method of claim 1, the method is characterized by: The step 2 includes the following steps: Step 2 (a) The time series X after outliers are removed original Split into fixed-length sequences, each containing 100 time steps. The processed time series data matrix X processed Satisfy the formula: X processed =Split(X original ,100) (6) Among them, the Split function splits the original data into multiple sequences of length 100; Step 2 (ii) Perform minimum-maximum normalization on the features of each time step to make them have a uniform scale. The standardized time series data matrix X norm Satisfy the formula: Among them, X processed is the processed time series data matrix, X norm is the standardized time series data matrix; Step 2 (3): According to the coincidence counting algorithm, the standardized time series data matrix X norm Calculate and get the eigenvalue F of each time step photon . Satisfies the formula: F photon =CoincidenceCount(X norm ) (8) Among them, the CoincidenceCount function calculates the standardized data according to the coincidence counting algorithm to obtain the eigenvalue of each time step, F photon is the eigenvalue at each time step; Step 2 (iv) The calculated eigenvalue F photon Vectorize to form the feature vector X for each time step f The feature vector X at each time step f Satisfy the formula: X f =Vectorize(F photon ) (9) The Vectorize function is used to convert the feature values ​​calculated by the coincidence counting algorithm into a one-dimensional array so that these data can be used as input to the autoencoder network.

3. According to claim 1, a method for suppressing noise in a light quantum arrival time series based on Bi-LSTM is characterized in that: The step five includes the following steps: The core of Bi-LSTM is the LSTM unit. Each LSTM unit performs the following operations at each time step: Step 5 (a) Calculate the activation value of the input gate to determine how much of the current input information is written into the cell state. The activation value of the input gate satisfies: Step 5 (ii) calculates the activation value of the forget gate to determine which information in the cell state needs to be forgotten. The activation value of the forget gate satisfies: Step 5 (iii), combine the operation of input gate and forget gate to update the cell state. Satisfy the formula: Step 5 (iv) Calculate the activation value of the output gate to determine how much information in the cell state is output. The activation value of the output gate satisfies: Among them, f h is the activation function of the system state, f s is the activation function of the internal state, usually the tanh function, which can transform the input vector to between (-1,1). Its purpose is to limit the vector value of each layer of the neural network, so as to avoid the large difference between the values ​​inside the vector, which will have a negative impact on the final output of the neural network; g is the gate unit, which is updated with the time step. In essence, it is a feedforward neural network with the sigmoid function (denoted as σ in the figure) as the activation function. The sigmoid function can be understood as a set of weights that control the output within the interval of (0,1); X t represents the tth sequence value in the sequence data, W is the weight matrix of the network, b is the bias, and the subscripts i, f, and o represent the input gate, forget gate, and output gate respectively; Bi-LSTM captures the bidirectional dependency information in the sequence through the LSTM layers in the forward and backward directions. The specific calculation process is: Step 5 (V) At time t, Bi-LSTM can use the information at time t+1 and time t-1 at the same time. In the Bi-LSTM network, the output y of each node t The calculation formula is as follows: h t = f(W t [s t-1 ; x] + b t ) (14) h' t = f(W t ' [s' t-1 ; x] + b' t ) (15) y t = g(W y [h t ; h' t ] + b y ) (16) Among them, h t is the hidden state of the forward LSTM unit at time t, h t ' is the hidden layer state of the reverse LSTM unit at time t. t and W t ' corresponds to the weight matrix of the forward LSTM and the reverse LSTM at time t. s t-1 is the hidden state of the forward LSTM unit at time t-1, s t ' -1 is the hidden state of the reverse LSTM unit at time t-1. x is the input data at time t. y t is the output of each node of the Bi-LSTM network at time t. f is the activation function in the LSTM unit, which is usually used to calculate the hidden state. g is used to calculate the output y t The activation function. t and b t ' is the bias term of the forward LSTM and the reverse LSTM at time t, W y is the value corresponding to the calculated output y t The weight matrix when b y Yes and W y The bias term used together to adjust the output y t Baseline value of Step 5 (6), the output of the Bi-LSTM layer satisfies the formula: Output bi-lstm =GeLU(W bi-lstm ×Output fc1 +b bi-lstm ) (17) Among them, W bi-lstm is the weight matrix of the fully connected layer, Output fc1 is the output data of the fully connected layer and the input data of the Bi-LSTM layer. bi-lstm is the bias term, Output bi-lstm It is the output data of the Bi-LSTM layer.

Images