Power equipment and principal continuous trust evaluation method based on quantum flower pollination algorithm optimized adabi-lstm model

Abstract

The application discloses a kind of power equipment and principal persistent trust evaluation method based on quantum flower pollination algorithm optimization AdaBi-LSTM model.The method first collects the information such as static information, interaction history, action attribute and environment attribute of power equipment and principal, balances data set by ADASYN method, carries out cleaning and normalization processing to data, forms the training set and test set of model.Then by adaptive bidirectional long short-term memory network, organically combines migration learning and bidirectional long short-term memory network, effectively trains the model, and constantly updates model parameters using quantum flower pollination algorithm, fully considers the correlation of historical information and future information of power equipment and principal, finally the model established is used as the prediction model of power equipment and principal behavior, realizes the accurate prediction of power equipment and principal behavior, real-time evaluates the trust degree of equipment and principal, adjusts the authority level of user, realizes accurate management and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of de-bordering security protection, and in particular to a power equipment and principal continuous trust evaluation method based on a quantum flower pollination algorithm optimized AdaBi-LSTM model. BACKGROUND

[0002] At present, in the power system, the existing border-based protection system and the method of "one-time authentication, one-time authorization, long-term validity" are difficult to cope with attack threats from internal and external devices and users with legal authority. However, the typical power network structure is becoming increasingly complex. A single power network can be composed of multiple intranets, remote devices, mobile office users and cloud services. As the boundaries of the power network become increasingly blurred, the traditional border-based protection strategy for power network security is no longer applicable. In this context, the industry has proposed to adopt a zero-trust architecture. The core idea of the zero-trust architecture is to not trust anyone, device or system inside and outside the network by default, and to reconstruct the trust basis of access control based on identity authentication and authorization. This means a security model that never trusts and always authenticates. In this mode, the problem of internal personnel violation or malicious attack can be well solved, providing protection for the realization of power Internet of Things "anytime, anywhere, anyone, anything" information connection and secure interaction. Assuming that the network is malicious, even the intranet of an enterprise is not an exception, continuous identity authentication and trust evaluation need to be studied, and by evaluating the trust of devices and users in real time, the permission level of users is adjusted to achieve accurate management and control.

[0003] Generally, trust evaluation methods can be roughly divided into two categories. One is to directly calculate the trust degree, which lacks dynamic nature and cannot achieve "continuous evaluation and never trust". The other only uses machine learning to predict the trust degree of devices and users, and the model prediction accuracy is not high without further optimization of parameters. SUMMARY

[0004] The present application proposes a power equipment and principal continuous trust evaluation method based on a quantum flower pollination algorithm optimized AdaBi-LSTM model. This method can help the system to realize continuous trust evaluation of devices and users, and enhance the dynamic adaptability of trust evaluation, the prediction accuracy of the model and the ability of the system to resist malicious attacks.

[0005] The application adopts a data set balancing method-adaptive synthetic sampling (ADASYN) and an AdaBi-LSTM model, finds the optimal or nearly optimal structure and parameters of the AdaBi-LSTM model by training the model and using a quantum flower pollination algorithm, realizes accurate prediction of power equipment and subject behaviors, and overcomes the defects of previous trust evaluation in adaptability, accuracy and algorithm efficiency.

[0006] To achieve the above object, the application comprises the following steps:

[0007] Step 1: collecting static information, interaction history, action attribute, environment attribute and other information of multiple devices and subjects in a power system;

[0008] Step 2: using the ADASYN method to process the imbalance of the data set;

[0009] Step 3: performing data cleaning and normalization processing on the data to obtain a data matrix, and dividing the data into a training set and a test set;

[0010] Step 4: constructing an AdaBi-LSTM trust evaluation model, searching for the optimal or nearly optimal structure and parameters of the AdaBi-LSTM model by using a quantum flower pollination algorithm, and obtaining the best device and subject trust degree prediction performance, including batch_size, layers and the number of units in each layer;

[0011] Step 5: taking the training set as the input of the AdaBi-LSTM model, training the model, verifying the prediction accuracy of the model, and achieving accurate prediction of the behaviors of the devices and subjects;

[0012] Step 6: according to the prediction result of the trust evaluation model, applying a similarity algorithm to calculate the similarity, and performing trust evaluation on the devices and subjects;

[0013] Step 7: dividing the trust level according to the trust evaluation result.

[0014] The step of using the ADASYN method to process the imbalance of the data set is:

[0015] Step 2.1: calculating the imbalance degree: d=m s / m l , wherein d s ∈(0,1], m l is a minority class sample, and m l is a majority class sample.

[0016] Step 2.2: calculating the number of samples to be synthesized: G=(m s) x beta, where beta is an element of [0, 1] is a parameter used to specify the desired level of balance after synthetic data generation. When beta = 1, G is equal to the difference between the number of majority class and minority class, at this time the majority class data and minority class data after synthetic data is exactly balanced.

[0017] Step 2.3: Calculate the K nearest neighbors for each sample belonging to the minority class using the Euclidean distance, calculate the proportion r i i i / K, i = 1,..., m s , where Delta i is the number of samples belonging to the majority class in the K neighbors, r i is an element of [0, 1].

[0018] Step 2.4: Get r i for each minority class sample in step 2.3, calculate the surrounding majority class of each minority class sample using the following formula:

[0019]

[0020] Where: is the density distribution,

[0021] Step 2.5: Calculate the number of synthetic samples for each minority class sample Where G is the total number of synthetic samples;

[0022] Step 2.6: Select 1 minority class sample from the K neighbors around each minority class sample to be synthesized, synthesize according to the following equation, repeat the synthesis until the number of synthesis required in step 2.5 is met:

[0023] s i = x i + (x zi - x i ) x lambda

[0024] Where: (x zi - x i ) is the difference vector in n-dimensional space, lambda is an element of [0, 1] is a random number.

[0025] The AdaBi-LSTM model structure satisfies:

[0026] ​​Temporal Distribution Characterization (TDC): Quantify the distribution of consecutive data in time series to divide it into K segments of the most dissimilar sequences. The assumption is that if the model can reduce the distribution difference of these K segments of the most dissimilar sequences, the model will have the strongest generalization ability. Therefore, the prediction effect will be better for unknown data.

[0027] To divide the time series into K segments of the most dissimilar sequences (corresponding to the maximum operation in the formal, while making K minimum), the temporal similarity quantization method characterizes this problem as an optimization problem:

[0028]

[0029]

[0030] Where: D is the time series data, n represents the number of time series data markers, d(·,·) represents the similarity measure function, Δ1, Δ2 and K0 are predefined parameters to avoid meaningless solutions. This optimization problem can be efficiently solved by dynamic programming algorithm.

[0031] Temporal Distribution Matching (TDM) constructs a transfer learning model for the above K time series to learn a model with temporal invariance. In particular, in order to not lose the temporal correlation in the transfer process, it is necessary to dynamically measure the importance of each time state in the Bi-LSTM unit. At this time, the importance of each time state to the entire training process in the transfer process can be dynamically learned.

[0032] The internal storage unit and gate mechanism of the Long Short-Term Memory Network LSTM can overcome the problem of gradient disappearance and explosion in traditional RNN, save the delay event in the time series, and extract it in subsequent training. The core calculation formula is as follows:

[0033] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0034] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0035] g t =tanh(W g ·[ht-1 , t ]+b g )

[0036] c t =f t *c t-1 +i t *g t

[0037] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0038] h t =o t *tanh(c t )

[0039] wherein: f t , i t , g t , o t are output values of the forget gate, the input gate, the update gate and the output gate respectively, the input of the four gates includes the output value h t-1 of the LSTM at t-1 moment and the input value x t at the current moment, W f , W i , W g , W o all represent weight matrices, b f , b i , b g , b o all represent bias vectors, c t is a memory cell, and sigma is a sigmoid activation function.

[0040] The unidirectional LSTM model uses previous information to predict subsequent information, while the bidirectional LSTM (Bi-LSTM) can simultaneously and comprehensively learn forward and backward time-related information to improve prediction accuracy. In power equipment and subject behavior prediction, considering that information about the past and future in the time series data of equipment and subject behavior can play an important role at the same time, the present application uses Bi-LSTM for prediction, and the Bi-LSTM model includes a forward LSTM layer and a backward LSTM layer. The horizontal direction calculates the forward LSTM hidden vector , and the vertical direction calculates the backward LSTM hidden vector at each time step t.

[0041] The quantum flower pollination algorithm satisfies:

[0042] In quantum space, the wave function The flower state of the quantum flower space is described, and the position of the i-th flower in the d-dimensional space is X i (t) = (x i1 , x i2 , …, x im ), a variable delta potential field is established to make the flower have a gathering state at a local point p (p1, p2, …, p d ) with a certain probability density appear at any point in space, so that the flower can search the entire feasible solution space, but will not diverge to infinity, the specific steps are as follows:

[0043] Step 4.1: initialization parameters: flower population size n, maximum iteration number N max , conversion probability p = 0.8;

[0044] Step 4.2: calculate the fitness value of the current population, record the global optimal value and the corresponding optimal solution;

[0045] Step 4.3: calculate the average optimal position avgbest of the flower and the contraction expansion factor β respectively,

[0046]

[0047]

[0048] Step 4.4: if the conversion probability p < rand, enter global search, update the solution according to the following formula, and handle the new solution out of bounds;

[0049]

[0050] Where: v = rand (0, 1), q is a local attraction factor, calculated by the formula q = αx (t) + (1-α)x best (α∈[0,1], x best is the current best position in the population), and L is the characteristic length of the delta potential well, which is an important variable in the flower evolution algorithm, calculated by the formula L (t) = 2β|avgbest-x (t) |;

[0051] Step 4.5: if the conversion probability p > rand, enter local search, update the solution according to the following formula, and handle the solution out of bounds;

[0052]

[0053] Where: q' = x bestL' is a new modified characteristic length of the delta potential well, which is calculated by the following formula:

[0054] L(t)' = 2 beta |avgbest - x best |

[0055] Step 4.6: Calculate the fitness value of the flower individual in step 4.4 and step 4.5, and record the optimal value and optimal position;

[0056] Step 4.7: judge the end condition, if it is satisfied, exit the program and output the optimal value and optimal solution, otherwise go to step 4.3.

[0057] Compared with the prior art, the method has the following advantages:

[0058] 1. The ADASYN technology is used to balance the data set, and imbalance is a problem that may occur in any type of data set, which is the main reason for the low efficiency of the prediction model. Therefore, the use of this technology can avoid the low efficiency of the prediction model.

[0059] 2. The adaptive bidirectional long short-term memory network used in the application solves the distribution shift problem in the machine learning field, that is, because the statistical characteristics of the time series can change over time, the distribution changes over time. In the application, the time series prediction time covariate shift (TCS) problem is established. An adaptive bidirectional long short-term memory network is proposed to solve the TCS problem. It is composed of two modules in turn. The first module is called time series similarity quantization, which aims to better represent the distribution information in the time series. The second module is time distribution matching, which aims to reduce the distribution mismatch in the time series, and learns an adaptive time series prediction model based on Bi-LSTM to better generalize and more accurately predict the behavior of power equipment and subjects, so as to more accurately and continuously assess the trust of power equipment and subjects.

[0060] 3. The quantum flower pollination algorithm is used to optimize the structure and parameters of AdaBi-LSTM to determine the optimal prediction model with the minimum training loss for target power equipment and subject behavior prediction, thereby improving the accuracy of the model. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 It is a structure diagram of the power equipment and subject continuous trust assessment method based on the quantum flower pollination algorithm optimized AdaBi-LSTM model.

[0062] Figure 2 It is a continuous trust assessment flowchart.

[0063] Figure 3 It is a self-adaptive bidirectional long short-term memory network model architecture diagram. Detailed Implementation

[0064] The following is in conjunction with the appendix Figure 1 Appendix Figure 2 Appendix Figure 3 The embodiments provide a detailed description of the method for evaluating the continuous trust of power equipment and entities based on the quantum flower pollination algorithm to optimize the AdaBi-LSTM model, but the implementation of the present invention is not limited thereto.

[0065] like Figure 1 As shown, the structure of the power equipment and subject continuous trust evaluation based on the quantum flower pollination algorithm-optimized AdaBi-LSTM model in this embodiment is mainly constructed using an adaptive bidirectional long short-term memory network optimized by the quantum flower pollination algorithm.

[0066] like Figure 2 As shown in this embodiment, the process of continuously assessing the trust of power equipment and entities is based on the AdaBi-LSTM model optimized by the quantum flower pollination algorithm.

[0067] like Figure 3 As shown, the AdaBi-LSTM model in this embodiment first establishes an adaptive bidirectional long short-term memory network model through a temporal similarity quantization module and a temporal distribution matching module. Then, the model is optimized using the quantum flower pollination algorithm, including batch size, layers, and the number of units per layer.

[0068] The entire ongoing trust assessment process is as follows:

[0069] (1) Collect static information, interaction history, action attributes, environmental attributes and other information of multiple devices and main entities in the power system;

[0070] (2) Use the ADASYN method to handle the imbalance of the dataset. The steps are as follows:

[0071] (2.1): Calculate the unbalance: d = m s / m l , where d∈(0,1], m s For minority class samples, m l This is the majority class sample.

[0072] (2.2): Calculate the number of samples to be synthesized: G = (m l -m s G is calculated as G = (0,1) × β, where β ∈ [0,1] is a parameter used to specify the required balance level after the synthetic data is generated. When β = 1, G equals the difference between the minority and majority classes, at which point the majority and minority class data are perfectly balanced after the synthetic data is generated.

[0073] (2.3): Calculate K nearest neighbors for each sample belonging to minority class using Euclidean distance, calculate the proportion r i is: r i i / K, i = 1,..., m s , where, Δ i is the number of samples belonging to the majority class in the K neighbors, r i ∈ [0, 1].

[0074] (2.4): Get r i for each minority class sample in (2.3)

[0075]

[0076] Where: is the density distribution,

[0077] (2.5): Calculate the number of synthetic samples for each minority class sample Where G is the total number of synthetic samples;

[0078] (2.6): Select 1 minority class sample from the K neighbors around each minority class sample to be synthesized, and synthesize according to the following equation, repeat the synthesis until the number of synthesis (2.5) is met:

[0079] s i = x i + (x zi -x i ) x lambda

[0080] Where: (x zi -x i ) is the difference vector in n-dimensional space, lambda ∈ [0, 1] is a random number.

[0081] (3) Data cleaning and normalization are performed to obtain a data matrix, and the data is divided into training set and test set. Data processing includes data cleaning and data normalization. Data cleaning includes deleting outliers in the data set and supplementing missing values. Data normalization can speed up the convergence of gradient descent and improve the performance of the prediction model. In many cases, gradient descent will become very complex non-standard data. Normalization can also make the deep learning model better extract high-level features from multiple historical output sequence data, thereby improving the performance of the proposed model. The present application normalizes each input sequence to the range (0, 1) by the following formula:

[0082]

[0083] Where:​ is the normalized value, x i is the original data, x max is the original data, x min is the maximum and minimum value of the original data, respectively.

[0084] (4) Construct an AdaBi-LSTM trust evaluation model, and use a quantum flower pollination algorithm to search for the optimal or near-optimal structure and parameters of the AdaBi-LSTM model to obtain the best device and subject trust degree prediction performance, including batch_size, layers, and the number of units in each layer;

[0085] The AdaBi-LSTM model includes:

[0086] Temporal Distribution Characterization (TDC): The continuous data distribution in the time series is quantified to divide it into K segments of the most dissimilar sequences. The assumption is that if the model can reduce the distribution difference of the K most dissimilar sequences, the model will have the strongest generalization ability. Therefore, the prediction effect on unknown data will be better.

[0087] To divide the time series into K segments of the most dissimilar sequences (corresponding to the maximum value operation in the formal, while making K minimum), the temporal similarity quantification method characterizes this problem as an optimization problem:

[0088]

[0089]

[0090] where: D is the time series data, n represents the number of time series data markers, d(·,·) represents the similarity measure function, Δ1, Δ2 and K0 are parameters defined in advance to avoid meaningless solutions. This optimization problem can be efficiently solved using a dynamic programming algorithm.

[0091] Temporal Distribution Matching (TDM) is a transfer learning model constructed for the above K time series to learn a model with temporal invariance. In particular, in order to not lose the temporal correlation in the transfer process, it is necessary to dynamically measure the importance of each time state in the Bi-LSTM unit. At this time, the importance of each time state to the entire training process during the transfer process can be dynamically learned.

[0092] Long Short-Term Memory (LSTM) networks overcome the vanishing and exploding gradient problems of traditional RNNs through their inherent storage units and gating mechanisms. LSTM stores delayed events in the time series and retrieves them during subsequent training. The core calculation formula is as follows:

[0093] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0094] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0095] g t =tanh(W g ·[h t-1 ,x t ]+b g )

[0096] c t =f t *c t-1 +i t *g t

[0097] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0098] h t =o t *tanh(c t )

[0099] Where: f t i t ,g t ,o t These are the output values ​​of the forget gate, input gate, update gate, and output gate, respectively. The inputs to the four gates include the LSTM's output value h at time t-1. t-1 and the input value x at the current time t W f W i W g W o Both represent multiplicative matrices, b f ,b i ,b g ,b o Both represent bias vectors, c tis a memory unit, and σ is a sigmoid activation function.

[0100] The unidirectional LSTM model uses previous information to predict subsequent information, while the bidirectional LSTM (Bi-LSTM) can simultaneously learn forward and backward time-related information to improve prediction accuracy. In power equipment and subject behavior prediction, considering that information about the past and future in equipment and subject behavior time series data can play an important role at the same time, the present application uses Bi-LSTM for prediction, and the Bi-LSTM model includes a forward LSTM layer and a backward LSTM layer. The horizontal direction calculates the forward LSTM hidden vector At the same time, the backward LSTM hidden vector is calculated at each time step t The vertical direction represents a one-way flow from the input layer to the hidden layer and then to the output layer. Then, the two hidden states are connected to calculate the final prediction of the Bi-LSTM.

[0101] The steps of the quantum flower pollination algorithm for optimizing model parameters are as follows:

[0102] Step 1: Initialize parameters: flower population size n, maximum number of iterations N max , conversion probability p = 0.8;

[0103] Step 2: Calculate the fitness value of the current population, record the global optimal value and the corresponding optimal solution;

[0104] Step 3: Calculate the average optimal position avgbest of the flower and the contraction expansion factor β respectively using the following formulas,

[0105]

[0106]

[0107] Step 4: If the conversion probability p < rand, enter global search, update the solution according to the following formula, and handle the out-of-bound processing of the new solution;

[0108]

[0109] where: v = rand(0, 1), q is a local attraction factor, calculated by the formula q = αx(t) + (1-α)x best (α ∈ [0, 1], x best is the current best position in the population), and L is the characteristic length of the δ potential well, which is an important variable in the flower evolution algorithm, calculated by the formula L(t) = 2β|avgbest-x(t)|;

[0110] Step 5: If the conversion probability p > rand, enter local search, update the solution according to the following formula, and handle the out-of-bound processing of the solution;

[0111]

[0112] wherein: q'=x best L' is the new modified characteristic length of the delta potential well, which is calculated by the following formula:

[0113] L(t)'=2beta|avgbest-x best |

[0114] Step 6: the fitness value of the flower individual in step 4 and step 5 is calculated, and the optimal value and optimal position are recorded;

[0115] Step 7: the end condition is judged, if it is satisfied, the program is exited and the optimal value and optimal solution are output, otherwise step 3 is turned.

[0116] (5) the training set is taken as the input of the AdaBi-LSTM model, the model is trained, the prediction accuracy (Acc) of the model is verified by the cross-validation method, so as to achieve accurate prediction of the device and the subject behavior, and the Acc formula of the accuracy is as follows:

[0117]

[0118]

[0119] Wherein, S(t+1) and respectively are the real behavior and the predicted behavior at t+1 moment (sampling point), N is the total length of data sequence, f(t) represents the function that the prediction at t+1 moment is inconsistent with the actual behavior, when S(t+1) and are the same, 1 is taken, and 0 is taken when they are different.

[0120] (6) according to the prediction result of the trust evaluation model, the similarity algorithm is applied to calculate the similarity, and the trust of the device and the subject is evaluated;

[0121] (7) according to the trust evaluation result, the trust level is divided, and is divided into trust, general trust, general untrust and untrust four levels.

[0122] The AdaBi-LSTM model finally tested and parameter-optimized will be used as the prediction model of the power device and the subject behavior. The preprocessed data about the static information, interaction history, action attribute and environment attribute of the device and the subject in the power system is taken as the input data of the AdaBi-LSTM model, the input data is learned through the AdaBi-LSTM model, and finally the behavior prediction result at the future moment is output.

[0123] Embodiment:

[0124] To verify the effectiveness of the proposed model, the trained network is used to experiment on untrusted behaviors to evaluate the effectiveness of the proposed method.

[0125] First, the power device and subject trust level is divided into 4 levels. The trust level from high to low is: {trusted, generally trusted, generally untrusted, untrusted}, and the corresponding value range is shown in Table 1:

[0126] Table 1: Power device and subject behavior trust level

[0127]

[0128] When the power device and subject behavior evaluation result is in the generally untrusted range, mark the power device and subject as an abnormal user; when the power device and subject behavior evaluation result is in the untrusted range, mark the power device and subject as a malicious user; when the evaluation result is in the trusted and generally trusted range, mark the user as a normal power device and subject, and respond to its attribute allocation requirements.

[0129] Then select the last 100 behaviors in the normal state as the current action to predict the next 20 actions, and then calculate the similarity value between each action in the predicted sequence and each action in the real sequence that occurs in the attack state.

[0130] Finally, in order to compare, the embodiment also predicts some sequences in the normal state, and the results are shown in Table 2.

[0131] Table 2: Similarity of predicted sequence and real sequence

[0132]

[0133] As can be seen from Table 2, when a normal behavior sequence appears, the similarity is close to 0, which is divided into the trusted level according to Table 2, and the permission allocation is responded to. As can be seen from the last row of the table, when a malicious behavior sequence appears, the similarity is almost close to 1, which is divided into the untrusted level according to the trust evaluation level table, and the abnormal behavior sequence can also be accurately divided into the generally untrusted level, and the permission allocation is not responded to. It can be clearly seen that the trust levels of normal sequences (first and second rows), abnormal sequences (third and fourth rows), and malicious sequences (last row) are obviously different and can accurately divide the trust level, which verifies that similarity is useful for evaluating power device and subject behavior.

[0134] The above merely expresses the preferred embodiments of the present application, which are described in more detail and in a more specific manner, but should not be understood as limiting the scope of the patent of the present application. It should be noted that, for those of ordinary skill in the art, several modifications, improvements and substitutions can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A power equipment and principal continuous trust evaluation method based on a quantum flower pollination algorithm optimized AdaBi-LSTM model, characterized in that, It comprises the following steps: Step 1, collect the static information, interaction history, action attribute, and environment attribute information of multiple devices and subjects in the power system; Step 2, use the ADASYN method to process the unbalance of the data set; Step 3, perform data cleaning and normalization processing to obtain a data matrix, and divide the data into a training set and a test set; Step 4, construct an AdaBi-LSTM trust evaluation model, use the quantum flower pollination algorithm to search for the optimal or near-optimal structure and parameters of the AdaBi-LSTM model to obtain the best device and subject trust prediction performance, including batch_size, layers, and the number of units per layer; Step 5, use the training set as the input of the AdaBi-LSTM model, train the model, and verify the prediction accuracy of the model to achieve accurate prediction of device and subject behavior; Step 6, according to the prediction results of the trust evaluation model, apply a similarity algorithm to calculate the similarity and perform trust evaluation on the device and subject; Step 7, divide the trust level according to the trust evaluation results; The core calculation formula of the AdaBi-LSTM trust evaluation model is: Time series similarity quantization TDC: quantize the continuous data distribution in the time series to divide it into K segments of the most dissimilar sequences; the assumption is that if the model can reduce the distribution difference of the K most dissimilar sequences, the model will have the strongest generalization ability; therefore, the prediction effect on unknown data will be better; To divide the time series into K segments of the most dissimilar sequences, the maximum value operation in the formal is used, and K is minimized; the time series similarity quantization method represents this problem as an optimization problem: wherein: is a time series data, denotes the number of time series data labeled segments, denotes a similarity measure function, and are predefined parameters to avoid meaningless solutions; this optimization problem is solved efficiently using a dynamic programming algorithm; Time series distribution matching TDM is to construct a transfer learning model for the above K time series to learn a model with time invariance; in order to not lose the time correlation in the transfer process, the importance of each time state in the Bi-LSTM unit must be dynamically measured; at this time, the importance of each time state to the entire training process in the transfer process can be dynamically learned; The core calculation formula of the long short-term memory network LSTM is as follows: ; ; ; ; wherein: are output values of a forget gate, an input gate, an update gate and an output gate, respectively, and the inputs of the four gates include the output value of the LSTM at the time t-1 and the input value at the current time t , , each represent a weight matrix, each represent a bias vector, is a memory cell, is a sigmoid activation function; In the power equipment and the main body behavior prediction, the Bi-LSTM is used to make the prediction, the Bi-LSTM model includes a forward LSTM layer and a backward LSTM layer; The horizontal direction calculates the forward LSTM hidden vector , and the backward LSTM hidden vector is calculated at each time step t, the vertical direction indicates one-way flow from the input layer to the hidden layer and then to the output layer; Then, two hidden states are connected to calculate the final prediction of the Bi-LSTM.

2. The power device and subject continuous trust evaluation method based on the quantum flower pollination algorithm optimized AdaBi-LSTM model according to claim 1, characterized in that, The step of using the ADASYN method to process the unbalance of the data set is: Step 2.1: Calculate the degree of imbalance: where , is the minority class sample, is the majority class sample; Step 2.2: Calculate the number of samples needed to be synthesized: wherein is a parameter for specifying the balance level required after the synthetic data is generated; when , equals the difference between the minority class and the majority class, at which time the majority class data and the minority class data after the synthetic data are exactly balanced; Step 2.3: Calculate the Euclidean distance for each sample belonging to the minority class. Calculate the proportion of the nearest neighbors. for: ,in, for The number of samples belonging to the majority class among the neighbors. ; Step 2.4: For each minority class sample obtained in step 2.3, calculate the surrounding majority class cases using the following formula: Step 2.4: For each minority class sample obtained in step 2.3, calculate the surrounding majority class cases using the following formula: ; wherein: is the density distribution, ; Step 2.5: Calculate the number of synthetic samples for each minority class sample , where is the total number of synthetic samples; Step 2.6: Synthesis of a minority class sample around each of the minority class samples to be synthesized select 1 minority class sample from the 6 neighbors, synthesize according to the following equation, repeat until the desired number of Step 2.5 syntheses is met: wherein: is a difference vector in the feature space, is a random number.

3. The power equipment and principal continuous trust evaluation method based on the quantum flower pollination algorithm optimized AdaBi-LSTM model according to claim 1, characterized in that, The quantum flower pollination algorithm step is: Step 4.1: Initialize parameters: flower population size , maximum number of iterations , transition probability = 0.8; Step 4.2: Calculate the fitness value of the current population and record the global optimal value and its corresponding optimal solution; Step 4.3: Calculate the average optimal position of the flower using the following formula and the expansion factor of contraction ​ Step 4.4: If the transition probability , go to global search, update the solution by the following equation, and handle the new solution out of bounds. where , is the local attraction factor, which is calculated by the formula , , is the current best position in the population, The characteristic length of the potential well is an important variable in the flower evolution algorithm, which is calculated by the formula ​​ Step 4.5: If transition probability , enter local search, update the solution according to the following formula, and perform boundary processing on the solution; wherein: , is the characteristic length of the new modification of the potential well, calculated from the following equation: Step 4.6: Calculate the fitness values of the flower individuals in steps 4.4 and 4.5, and record the optimal value and optimal position; Step 4.7: judge the end condition, if satisfied, exit the program and output the optimal value and optimal solution, otherwise go to step 4.3.

Images