Quantum communication network equipment quality evaluation method and system

Abstract

The invention relates to a quantum communication network equipment quality evaluation method and system, in particular to the field of quantum communication, and the method comprises the steps: achieving millisecond-level data collection and self-adaptive cleaning through a multi-mode sensor array, and generating a high-quality time series data flow with a confidence label; on the basis of quantum annealing optimization feature selection, dynamically extracting a feature subset with the most discriminative power; a hybrid model is trained under a federated learning framework, and the generalization ability of the model is improved while privacy is protected through encryption gradient aggregation; and finally, in combination with real-time reasoning and an adjacent node collaborative voting mechanism, accurate sensing of the running state of the quantum communication equipment and rapid positioning of an abnormal root are realized, and the accuracy, the real-time performance and the robustness of quality evaluation are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of quantum communication, and more specifically, to a method and system for evaluating the quality of quantum communication network equipment. Background Technology

[0002] With the rapid development of quantum communication technology, quantum communication networks have gradually expanded from controlled laboratory environments to practical application scenarios, including complex outdoor environments such as deserts, polar regions, high mountains, and urban high-temperature areas. Under these extreme natural conditions, quantum communication network equipment, such as quantum key distribution terminals, single-photon detectors, and quantum light sources, needs to be exposed to severe environmental fluctuations for extended periods. External environmental factors manifest as sudden temperature changes caused by the alternation of day and night, ranging from minus 30 degrees Celsius to 50 degrees Celsius, accompanied by significant humidity fluctuations, strong winds and dust storms, abnormal electromagnetic interference, and other complex factors. Internal factors involve limitations in the thermal management design of the equipment itself, mismatches in the thermal expansion coefficients of optical component materials, and insufficient sealing performance. This complex interplay of internal and external factors causes the performance indicators of quantum communication links, such as qubit error rate, photon count rate, and key generation rate, to exhibit strong non-stationarity and random fluctuation characteristics. Traditional network operation and maintenance and quality assessment methods based on the assumption of a stable laboratory environment face severe challenges in such dynamic real-world scenarios, making it difficult to accurately reflect the actual operating status and long-term reliability of the equipment.

[0003] Currently, quality evaluation methods for quantum communication network equipment mainly rely on periodic manual inspections, offline data analysis, or simple alarm mechanisms based on fixed thresholds. Existing technical solutions typically collect equipment performance parameters and environmental data at specific time intervals, such as every minute or longer, and use static statistical methods, such as calculating historical averages or setting empirical thresholds, to identify anomalies. However, this approach has significant drawbacks: its data sampling frequency is far lower than the actual rate of change of environmental disturbances, making it unable to capture transient interference signals on the order of microseconds or milliseconds. This results in a delayed response of the system to rapidly changing environmental factors. More importantly, existing methods lack the ability to deeply analyze temporal correlations, making it difficult to effectively distinguish between performance fluctuations caused by brief external environmental shocks and performance degradation caused by inherent defects or gradual failures in the equipment itself. This lack of discriminatory ability can easily lead to misjudgments in the quality evaluation system, potentially misreporting normal environmental adaptive responses as equipment failures, or conversely, masking the true trend of equipment performance degradation. Therefore, there is an urgent need to develop an adaptive quality evaluation technology and system that can adapt to dynamic changes in extreme environments, possess high real-time performance, and intelligently identify the root causes of anomalies. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a method and system for evaluating the quality of quantum communication network equipment, thereby resolving the issues raised in the background section.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for evaluating the quality of quantum communication network equipment, specifically including the following steps: Step S1: Deploy a multimodal sensor array at the equipment node of the quantum communication network, synchronously collect environmental parameter sequences and equipment performance index sequences at millisecond frequency as raw data, perform real-time alignment and standardization processing on the collected raw data, and use an adaptive dual-threshold anomaly detection algorithm to clean the data, add confidence labels to the marked data points, generate a preprocessed time series data stream with confidence labels, and calculate its statistical characteristics including volatility. Step S2: Construct a feature selection cost function based on the idea of ​​quantum annealing, dynamically adjust the weight coefficients of the feature selection cost function according to the statistical characteristics of the preprocessed time series data stream with confidence labels, solve for the minimum value of the feature selection cost function to obtain the dynamic optimal feature subset, and establish a feature mapping relationship. Step S3: Under the federated learning framework, each edge node locally trains the LSTM-Transformer hybrid model structure using the dynamic optimal feature subset, and globally aggregates and updates the encrypted model gradients through a contribution-based dynamic weighted aggregation algorithm, thereby generating an optimized global quality evaluation model. Step S4: Each edge node loads the optimized global quality evaluation model, applies the dynamic optimal feature subset to the real-time collected data for feature extraction and millisecond-level inference, and when a performance anomaly is detected, performs lightweight collaborative voting with neighboring nodes, determines the root cause of the anomaly based on the majority consensus principle, and outputs the quality evaluation result. In a preferred embodiment, the specific process of deploying a multimodal sensor array at the quantum communication network equipment node in step S1, and synchronously collecting environmental parameter sequences and equipment performance index sequences as raw data at a millisecond-level frequency, is as follows: A1. The multimodal sensor array includes a temperature sensor, a humidity sensor, an electromagnetic field strength sensor, a quantum bit error rate detection unit, and a photon count rate detection unit, wherein the millisecond frequency is a sampling interval of one millisecond to ten milliseconds; A2. The temperature sensor, humidity sensor, electromagnetic field strength sensor, quantum bit error rate detection unit and photon count rate detection unit in the multimodal sensor array are driven synchronously to trigger data acquisition operations via a high-precision clock signal at a millisecond frequency. A3. Obtain the original environmental parameter sequence and the original equipment performance index sequence with timestamp alignment. Use the original environmental parameter sequence and the original equipment performance index sequence together as the original data. The original environmental parameter sequence is a sequence of environmental parameter data composed of temperature, humidity and electromagnetic field strength data collected at each sampling time in chronological order. The original equipment performance index sequence is a sequence of equipment performance index data composed of quantum bit error rate and photon count rate data collected at each sampling time in chronological order. The specific process of real-time alignment and standardization of the collected raw data is as follows: B1. Based on the hardware timestamp carried by each data point in the original environmental parameter sequence and the original equipment performance index sequence, the environmental parameter data and equipment performance index data collected at the same sampling time are paired to form a parallel data stream aligned in the time dimension. B2. Standardization processing adopts the Z-score normalization method. The mean and standard deviation of the aligned environmental parameter sequence and the equipment performance index sequence are calculated respectively. Then, the mean of the environmental parameter sequence is subtracted from each data point in the environmental parameter sequence and divided by the standard deviation of the environmental parameter sequence. The mean of the equipment performance index sequence is subtracted from each data point in the equipment performance index sequence and divided by the standard deviation of the equipment performance index sequence to obtain the standardized environmental parameter sequence and the standardized equipment performance index sequence.

[0006] In a preferred embodiment, the specific operation of using an adaptive dual-threshold anomaly detection algorithm for data cleaning and assigning confidence labels to the marked data points is as follows: C1. Set a hard threshold and a soft threshold. The hard threshold is determined based on the physical measurement range limits of various sensors in the multimodal sensor array. The soft threshold is dynamically calculated based on the data within a sliding window of a preset width, and is the arithmetic mean of all data points within the sliding window plus twice the standard deviation of all data points within the sliding window. C2. During data cleaning, first determine whether the current data point in the standardized environmental parameter sequence or standardized equipment performance index sequence exceeds the upper or lower limit of the hard threshold, and determine the data point that exceeds the upper or lower limit as an invalid data point. For data points that do not exceed the hard threshold range, it is further determined whether their values ​​are greater than the soft threshold. Data points that are greater than the soft threshold are marked as potential outliers. C3. For data points that are determined to be invalid in C2, assign them a confidence label value of zero; C4. For data points marked as potential outliers in C2, their confidence labels are calculated according to the following rules: First, calculate the absolute difference between the value of the data point and the arithmetic mean of all data points in the sliding window. Then, divide this absolute difference by the product of a preset sensitivity adjustment coefficient and the standard deviation of all data points in the sliding window. Finally, take the negative power of the exponential function with the natural constant as the base of the calculated quotient. The result is the confidence label value of the data point. C5. For data points that pass the hard and soft threshold detection in C2 and are not marked as potential outliers, assign them a confidence label value of one. Finally, generate a preprocessed time series data stream with confidence labels, consisting of a standardized environmental parameter sequence, a standardized equipment performance index sequence, and corresponding confidence labels. The calculation of its statistical characteristics, including volatility, is specifically as follows: The relative fluctuation entropy of a preprocessed time-series data stream with confidence labels is calculated. The calculation process for this relative fluctuation entropy includes: The numerical range of the preprocessed time series data stream with confidence labels is divided into MA consecutive intervals; the probability of a data point falling into each interval is calculated, and the mean and standard deviation of the data points in each interval are calculated respectively. Finally, the probability of each interval is multiplied by the logarithm of that probability, then multiplied by the ratio of the standard deviation to the mean of that interval, and the sum of the results for all intervals is taken as a negative number to obtain the value of the relative fluctuation entropy.

[0007] In a preferred embodiment, the specific operation of constructing the feature selection cost function based on the quantum annealing concept in step S2 is as follows: Define a feature selection vector, which is a binary sequence where each element takes the value of zero or one, where zero indicates that the corresponding feature is not selected and one indicates that the corresponding feature is selected; The feature selection cost function adopts a dynamic conditional structure complexity function, and the specific calculation process of this function is as follows: First, the time axis is divided into multiple consecutive time intervals, the total number of which is determined by the interval division in the preprocessed time series data stream with confidence labels; For each time interval, the first component is calculated, which is the context importance weight of the time interval multiplied by the conditional probability that the data falls into the time interval under the selected feature subset conditions, multiplied by the base-2 logarithm of the conditional probability, and the summation of the calculation results for all time intervals is taken as the average, and finally the negative value is taken. Secondly, the second component is calculated, which is a balance coefficient multiplied by the feature dynamic sensitivity weight for each time interval, multiplied by the prior probability that the data belongs to the time interval under the condition of a selected feature subset, and then multiplied by the entropy measure of the dynamic complexity of the feature set under the time interval. The results of the calculation over all time intervals are summed and averaged. Finally, the first component and the second component are added together to obtain the dynamic conditional structure complexity function value; wherein, the context importance weight is calculated by the sigmoid function based on the confidence labels in the preprocessed time series data stream, and a confidence amplification factor is introduced to adjust the influence of the confidence labels on the weights.

[0008] In a preferred embodiment, the specific process of dynamically adjusting the weight coefficients of the feature selection cost function based on the statistical characteristics of the preprocessed time-series data stream with confidence labels is as follows: Statistical characteristics, including relative fluctuation entropy and average confidence, are extracted from the preprocessed time-series data stream with confidence labels. The average confidence is the arithmetic mean of all confidence label values. Based on the environmental parameter sequence and equipment performance index sequence in the preprocessed time-series data stream, the mutual information matrix between features is calculated, and the spectral norm of the matrix is ​​further calculated. The weight coefficients in the feature selection cost function are dynamically adjusted, including the balance coefficient, the feature dynamic sensitivity weight, and the context importance weight. The adjustment rule for the balance coefficient is that it is directly proportional to the value of the relative fluctuation entropy; when the value of the relative fluctuation entropy increases, the value of the balance coefficient increases accordingly. The adjustment rule for the confidence amplification factor of context importance weight is that it is related to the average confidence score through a hyperbolic tangent function; the value of the confidence amplification factor increases when the average confidence score increases. The adjustment rule for the feature dynamic sensitivity weight is that it is inversely proportional to the spectral norm of the mutual information matrix between features; the value of the feature dynamic sensitivity weight decreases as the spectral norm increases. The adjustment process is executed periodically based on a sliding window mechanism, with the size of the sliding window preset according to the data sampling frequency; The specific process of finding the minimum value of the feature selection cost function to obtain the dynamically optimal feature subset and establishing the feature mapping relationship is as follows: The minimum value of the feature selection cost function is solved using a quantum annealing optimization framework. The specific process includes encoding operations, Hamiltonian construction operations, evolution and cooling operations, and result reading operations. The encoding operation encodes the feature selection binary vector and the dynamically adjusted weight coefficients into the quantum state of the qubit; The Hamiltonian construction operation sets the initial Hamiltonian to an easily prepared ground state, and the target Hamiltonian is constructed as a quantum version of the feature selection cost function, with its ground state corresponding to the minimum value of the feature selection cost function. The evolution and cooling operations simulate the adiabatic evolution process of a quantum system and utilize the quantum tunneling effect to evolve the system state from the initial Hamiltonian to the target Hamiltonian. The result reading operation decodes the optimal feature selection binary vector and the corresponding weight coefficient configuration from the final quantum state after the evolution ends; The subset of features indicated by the binary vector of the optimal feature selection is taken as the dynamic optimal feature subset, and the adjustment path of the weight coefficients during quantum annealing is recorded. Based on this path, a mapping relationship from the original feature space to the dynamic optimal feature subset is established.

[0009] In a preferred embodiment, the specific operation of each edge node locally training the LSTM-Transformer hybrid model structure using the dynamic optimal feature subset in step S3 is as follows: First, the LSTM-Transformer hybrid model structure consists of cascaded Long Short-Term Memory (LSTM) network units and Transformer Encoder (LEE) units. The LSTM network units are used to process the preprocessed temporal data stream with confidence labels generated in step S1 from the dynamic optimal feature subset. Each LSTM network unit is composed of multiple LSTM sub-units. Each LSTM sub-unit receives the input features of the current time step and the hidden state of the previous time step, and outputs the hidden state of the current time step. The hidden states of all time steps form a hidden state sequence, and the number of time steps is determined by the length of the preprocessed temporal data stream. The hidden state sequence is passed as input to the transformer encoder unit, which consists of a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer calculates the attention weight between each hidden state and other hidden states to generate a weighted feature representation. The feedforward neural network layer performs a non-linear transformation on the features at each position and outputs a global context feature sequence. The global context feature sequence is aggregated into a fixed-length feature vector through a global average pooling layer. Finally, the feature vector is mapped to the quality evaluation result through a fully connected layer. During local training, weighted cross-entropy loss is used as the objective function. The weights of the weighted cross-entropy loss are determined by the confidence labels generated in step S1, i.e., the loss of each sample is multiplied by the confidence label value of that sample. Local training is optimized using the stochastic gradient descent algorithm. In each training iteration, the gradient of the weighted cross-entropy loss function with respect to all parameters of the model is calculated. This gradient is a vector and is called the local loss gradient. Throughout the local training process, the local loss gradient generated in each iteration is recorded to form a local loss gradient sequence. The learning rate is dynamically adjusted according to the training progress.

[0010] In a preferred embodiment, the specific process of globally aggregating and updating the encrypted model gradient using a contribution-based dynamic weighted aggregation algorithm is as follows: After completing local training, each edge node uses homomorphic encryption technology to encrypt the final gradient in its local loss gradient sequence to obtain the encrypted gradient, and uploads the encrypted gradient to the aggregation server. The aggregation server calculates the contribution index of each edge node. The contribution index is a multi-dimensional contribution evaluation index, which is calculated based on performance improvement, data quality, and training fluctuation. The performance improvement is measured by the relative reduction of the local model loss to the global model loss of the previous round. The relative reduction is the difference between the global model loss of the previous round and the local model loss of the current round, divided by the global model loss of the previous round. The data quality item is measured by calculating the average confidence level of the local data, where the average confidence level is the arithmetic mean of all confidence level label values; The training fluctuation term is measured by calculating the coefficient of variation of the local loss gradient, which is the standard deviation of the local loss gradient sequence divided by the mean. The relative reduction of the performance improvement item, the average confidence of the data quality item, and the coefficient of variation of the training fluctuation item are weighted and combined. The first weight coefficient, the second weight coefficient, and the third weight coefficient are used in the weighting combination. The first weight coefficient corresponds to the performance improvement item, the second weight coefficient corresponds to the data quality item, and the third weight coefficient corresponds to the training fluctuation item. The values ​​of the first weight coefficient, the second weight coefficient, and the third weight coefficient are in the range of zero to one, and the sum of the first weight coefficient, the second weight coefficient, and the third weight coefficient is one. The weighted combination result is the first weight coefficient multiplied by the relative reduction, plus the second weight coefficient multiplied by the average confidence label value, minus the third weight coefficient multiplied by the coefficient of variation, to obtain the multidimensional contribution evaluation index value of each node. Based on the multi-dimensional contribution evaluation index values, the softmax function is used to calculate the aggregate weight of each node. The temperature parameter in the softmax function is used to control the smoothness of the weight distribution. The aggregation server calculates the weighted average gradient in the encrypted state, that is, it multiplies the encrypted gradient of each node by its aggregate weight and sums them to obtain the encrypted weighted average gradient. Then, the encrypted weighted average gradient is decrypted to obtain the decrypted gradient. The global model parameters are updated using the decrypted gradient. The update formula is the global model parameters minus the learning rate multiplied by the decrypted gradient. The above aggregation and update process is repeated until the rate of change of the global model loss is less than a preset threshold, thereby generating an optimized global quality evaluation model.

[0011] In a preferred embodiment, in step S4, after each edge node loads the optimized global quality assessment model, the specific steps for feature extraction and millisecond-level inference of the real-time collected data are as follows: First, each edge node extracts features from the real-time data stream using the dynamic optimal feature subset generated in step S2, selecting only the feature dimension corresponding to the dynamic optimal feature subset to generate a real-time feature vector. Secondly, the real-time feature vector is input into the optimized global quality assessment model generated in step S3 for millisecond-level inference. The inference process is optimized to a millisecond-level response and outputs a quality score, which is a value between zero and one. Then, an anomaly threshold is set. The anomaly threshold is a preset fixed value, ranging from zero to one. The quality score is compared with the anomaly threshold. When the quality score is lower than the anomaly threshold, a performance anomaly flag is triggered. The performance anomaly flag is a binary variable, with a value of one indicating an anomaly and a value of zero indicating normal. Meanwhile, the global quality assessment model outputs an anomaly root cause hypothesis and a confidence score, which is a value between zero and one. The confidence score is calculated based on the information entropy of the probability distribution output by the model. When the performance anomaly flag is triggered, a lightweight collaborative voting mechanism with neighboring nodes is initiated. Neighboring nodes are determined based on network topology or geographical proximity, including nodes whose hop distance is less than or equal to the preset maximum distance. Edge nodes broadcast query requests to neighboring nodes and receive anomaly information triples returned by neighboring nodes, including anomaly flags, anomaly root cause hypothesis, and confidence level; the communication process of broadcasting query requests and receiving anomaly information triples adopts a lightweight coding protocol.

[0012] In a preferred embodiment, the specific steps of the lightweight collaborative voting and anomaly root cause determination are as follows: After receiving the anomaly information triple from neighboring nodes, the edge node initiating the vote calculates a consensus score for each possible anomaly root cause type. The consensus score is calculated as follows: First, obtain the set of neighboring nodes of the node, and denote the size of the set of neighboring nodes as the number of neighboring nodes; For each neighboring node, its individual contribution to the root cause type is calculated. The individual contribution is the product of the indicator function value, the distance decay weight, and the prediction confidence of the neighboring node. The indicator function value is 1 if the anomaly root cause hypothesis of the neighboring node is consistent with the currently calculated root cause type, and 0 otherwise; the distance decay weight is calculated by the exponential decay function, the input of which is the product of the negative distance decay factor and the logical distance between nodes, which is represented by the number of network hops; the prediction confidence directly uses the confidence value provided by the neighboring node. Then, sum the individual contribution values ​​of all neighboring nodes and divide by the number of neighboring nodes to obtain the collaborative consistency score for this root cause type. After calculating the scores for all root cause types, the root cause type with the highest consensus score is selected as the final judgment result. If multiple root cause types have the same score, the type with the highest number of votes is selected first according to the majority consensus principle. Meanwhile, the maximum collaborative consistency score is used as the confidence level for judgment; Finally, the quality evaluation results are output, including anomaly indicators, anomaly root cause determination results, and determination confidence levels; Each node sends the results to the central server for aggregation, generating a global quality report and recording it for model optimization.

[0013] This application also provides a quality evaluation system for quantum communication network equipment, specifically including: The data acquisition and preprocessing module is deployed at the equipment node of the quantum communication network. It is used to synchronously acquire environmental parameter sequences and equipment performance index sequences at millisecond frequency through a multimodal sensor array as raw data. The acquired raw data is aligned and standardized in real time, and an adaptive dual-threshold anomaly detection algorithm is used for data cleaning. Confidence labels are added to the marked data points, generating a preprocessed time series data stream with confidence labels and calculating its statistical characteristics. The feature selection optimization module connects to the data acquisition and preprocessing module. It is used to construct a feature selection cost function based on the quantum annealing concept, dynamically adjust the weight coefficients of the cost function according to the statistical characteristics of the preprocessed time series data stream with confidence labels, solve for the minimum value of the cost function to obtain the dynamically optimal feature subset, and establish a feature mapping relationship. The model training and aggregation module connects to the feature selection and optimization module. Under the federated learning framework, each edge node locally trains the LSTM-Transformer hybrid model structure using a dynamically optimal feature subset. The encrypted model gradient is globally aggregated and updated through a contribution-based dynamic weighted aggregation algorithm, thereby generating an optimized global quality evaluation model. The real-time evaluation and collaborative decision-making module connects to the model training and aggregation module. It is used to load the optimized global quality evaluation model on each edge node. It applies the dynamic optimal feature subset to extract features and perform millisecond-level inference on the real-time collected data. When a performance anomaly is detected, it performs lightweight collaborative voting with neighboring nodes, determines the root cause of the anomaly based on the majority consensus principle, and outputs the quality evaluation result.

[0014] The beneficial effects of this invention are as follows: it achieves millisecond-level data acquisition and adaptive cleaning through a multimodal sensor array, generating a high-quality time-series data stream with confidence labels; it optimizes feature selection based on quantum annealing, dynamically extracting the most discriminative feature subset; it trains a hybrid model under a federated learning framework, protecting privacy while improving the model's generalization ability through encrypted gradient aggregation; and finally, it combines real-time inference with a neighboring node collaborative voting mechanism to achieve accurate perception of the operating status of quantum communication equipment and rapid location of anomalies, significantly improving the accuracy, real-time performance, and robustness of quality evaluation. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1The method for evaluating the quality of quantum communication network equipment, as shown, specifically includes the following steps: Step S1: Deploy a multimodal sensor array at the equipment node of the quantum communication network, synchronously collect environmental parameter sequences and equipment performance index sequences at millisecond frequency as raw data, perform real-time alignment and standardization processing on the collected raw data, and use an adaptive dual-threshold anomaly detection algorithm to clean the data, add confidence labels to the marked data points, generate a preprocessed time series data stream with confidence labels, and calculate its statistical characteristics including volatility. Step S2: Construct a feature selection cost function based on the idea of ​​quantum annealing. Dynamically adjust the weight coefficients of the feature selection cost function according to the statistical characteristics of the preprocessed time series data stream with confidence labels. Solve for the minimum value of the feature selection cost function to obtain the dynamic optimal feature subset. This dynamic optimal feature subset is used for training the subsequent quality assessment model and establishing the feature mapping relationship. Step S3: Under the federated learning framework, each edge node locally trains the LSTM-Transformer hybrid model structure using the dynamic optimal feature subset, and globally aggregates and updates the encrypted model gradients through a contribution-based dynamic weighted aggregation algorithm, thereby generating an optimized global quality evaluation model. Step S4: Each edge node loads the optimized global quality evaluation model, applies the dynamic optimal feature subset to extract features and perform millisecond-level inference on the real-time collected data, and when a performance anomaly is detected, performs lightweight collaborative voting with neighboring nodes, determines the root cause of the anomaly based on the majority consensus principle, and outputs the quality evaluation result.

[0020] In this embodiment, the specific process of deploying a multimodal sensor array at the quantum communication network equipment node in step S1, and synchronously collecting environmental parameter sequences and equipment performance index sequences as raw data at a millisecond-level frequency, is as follows: A1. The multimodal sensor array includes a temperature sensor, a humidity sensor, an electromagnetic field strength sensor, a quantum bit error rate detection unit, and a photon count rate detection unit, with a sampling interval of one to ten milliseconds at a millisecond frequency. A2. The temperature sensor, humidity sensor, electromagnetic field strength sensor, quantum bit error rate detection unit and photon count rate detection unit in the multimodal sensor array are driven synchronously to trigger data acquisition operations via a high-precision clock signal at a millisecond frequency. A3. Obtain the original environmental parameter sequence and the original equipment performance index sequence with timestamp alignment. Use the original environmental parameter sequence and the original equipment performance index sequence together as the original data. The original environmental parameter sequence is a sequence of environmental parameter data composed of temperature, humidity and electromagnetic field strength data collected at each sampling time in chronological order. The original equipment performance index sequence is a sequence of equipment performance index data composed of quantum bit error rate and photon count rate data collected at each sampling time in chronological order. The specific process of real-time alignment and standardization of the collected raw data is as follows: B1. Based on the hardware timestamp carried by each data point in the original environmental parameter sequence and the original equipment performance index sequence, the environmental parameter data and equipment performance index data collected at the same sampling time are paired to form a parallel data stream that is strictly aligned in the time dimension. B2. The standardization process uses the Z-score normalization method to calculate the mean and standard deviation of the aligned environmental parameter sequence and the equipment performance index sequence, respectively. Then, for each data point in the environmental parameter sequence, the mean of the environmental parameter sequence is subtracted and divided by the standard deviation of the environmental parameter sequence. For each data point in the equipment performance index sequence, the mean of the equipment performance index sequence is subtracted and divided by the standard deviation of the equipment performance index sequence to obtain the standardized environmental parameter sequence and the standardized equipment performance index sequence. The data cleaning process employs an adaptive dual-threshold anomaly detection algorithm. The specific steps for assigning confidence labels to the marked data points are as follows: C1. Set a hard threshold and a soft threshold. The hard threshold is determined based on the physical measurement range limits of various sensors in the multimodal sensor array. The soft threshold is dynamically calculated based on the data within a preset width sliding window, which is the arithmetic mean of all data points within the sliding window plus twice the standard deviation of all data points within the sliding window. C2. During data cleaning, first determine whether the current data point in the standardized environmental parameter sequence or standardized equipment performance index sequence exceeds the upper or lower limit of the hard threshold, and determine the data point that exceeds the upper or lower limit as an invalid data point. For data points that do not exceed the hard threshold, further determine whether their values ​​are greater than the soft threshold, and mark data points that are greater than the soft threshold as potential outliers; C3. For data points that are determined to be invalid in C2, assign them a confidence label value of zero; C4. For data points marked as potential outliers in C2, their confidence labels are calculated according to the following rules: First, calculate the absolute difference between the value of the data point and the arithmetic mean of all data points in the sliding window. Then, divide this absolute difference by the product of a preset sensitivity adjustment coefficient and the standard deviation of all data points in the sliding window. Finally, take the negative power of the exponential function with the natural constant as the base of the calculated quotient. The result is the confidence label value of the data point. C5. For data points that pass the hard and soft threshold detection in C2 and are not marked as potential outliers, assign them a confidence label value of one. Finally, generate a preprocessed time series data stream with confidence labels, consisting of a standardized environmental parameter sequence, a standardized equipment performance index sequence, and corresponding confidence labels. The specific calculation of its statistical properties, including volatility, is as follows: The relative fluctuation entropy of a preprocessed time-series data stream with confidence labels is calculated. The calculation process for this relative fluctuation entropy includes: The numerical range of the preprocessed time series data stream with confidence labels is divided into MA consecutive intervals; the probability of a data point falling into each interval is calculated, and the mean and standard deviation of the data points in each interval are calculated respectively. Finally, the probability of each interval is multiplied by its logarithm, then multiplied by the ratio of the standard deviation to the mean of that interval. The sum of the results for all intervals is then taken as a negative number to obtain the relative volatility entropy. This relative volatility entropy is used to characterize the randomness and instability of the preprocessed time-series data stream. The formula for calculating the relative volatility entropy is: ; in, Representing relative fluctuation entropy, it is a dimensionless scalar value, and its calculation result is a non-negative real number. The larger the value, the more complex the data fluctuation characteristics. Replacing traditional variance or standard deviation, it can more profoundly reveal the randomness and instability of the sequence. The larger the value, the more complex, violent, and unpredictable the fluctuation of the preprocessed time series data stream. MA represents the number of intervals (the total number of continuous intervals into which the numerical range of the preprocessed time series data stream is divided; its value is usually determined according to the amount of data and the desired granularity, for example, 10, 20, 50). It determines the granularity of probability distribution estimation. If MA is too large, it is easily affected by noise; if it is too small, it cannot capture the details of fluctuation. It is a key parameter for controlling the accuracy of calculation. This represents the interval index (i.e., the loop variable for summation, representing the index of the interval). (A number of intervals, traversing from 1 to MA), is used to calculate the contribution of each interval to the total entropy value. This represents the interval probability, i.e., the probability that a data point falls within the interval. The empirical probability within a given interval is calculated using the formula: the number of data points within that interval divided by the total number of data points, with the value set to satisfy... and It reflects the distribution of data across the entire numerical range, embodies the concept of information entropy, and measures the degree of dispersion (uncertainty) in the data distribution. This represents the standard deviation within the interval, i.e., the first... The standard deviation of all data points within an interval, a non-negative real number, if there is only one point in the interval, then A value of 0 is used to measure the dispersion or fluctuation of the data within that interval. The larger the value, the more volatile the data is, even within the same range. Represents the mean within the interval, the th The arithmetic mean of all data points within an interval, a real number. To avoid division by zero, empty intervals can be processed or the division of MA can be ensured so that each interval contains data; represents the center position of the data in that interval, and is related to in the formula. Composition of coefficient of variation ( This is used to normalize the volatility of different numerical level ranges, making them comparable.

[0021] In this embodiment, it is specifically necessary to explain the following steps in step S2: the specific operation of constructing the feature selection cost function based on the quantum annealing concept is as follows: Define a feature selection vector, which is a binary sequence where each element takes the value of zero or one, where zero indicates that the corresponding feature is not selected and one indicates that the corresponding feature is selected; The feature selection cost function adopts a dynamic conditional structural complexity function, and the specific calculation process of this function is as follows: First, the time axis is divided into multiple consecutive time intervals, the total number of which is determined by the interval division in the preprocessed time series data stream with confidence labels; For each time interval, the first component is calculated, which is the context importance weight of the time interval multiplied by the conditional probability that the data falls into the time interval under the selected feature subset conditions, multiplied by the base-2 logarithm of the conditional probability, and the summation of the calculation results for all time intervals is taken as the average, and finally the negative value is taken. Secondly, the second component is calculated, which is a balance coefficient multiplied by the feature dynamic sensitivity weight for each time interval, multiplied by the prior probability that the data belongs to the time interval under the condition of a selected feature subset, and then multiplied by the entropy measure of the dynamic complexity of the feature set under the time interval. The results of the calculation over all time intervals are summed and averaged. Finally, the first component and the second component are added together to obtain the dynamic conditional structure complexity function value. The context importance weight is calculated using the sigmoid function based on the confidence labels in the preprocessed time-series data stream, and a confidence amplification factor is introduced to adjust the influence of the confidence labels on the weights. The value of the confidence amplification factor is determined according to the following rule: when the average confidence level is high, the confidence amplification factor is increased accordingly to strengthen the contribution of the high-confidence interval. The feature dynamic sensitivity weight is calculated using a non-linear increasing function based on the gradient norm of each feature in the corresponding time interval. The entropy measure uses the Shannon entropy function to quantify the uncertainty of dynamic feature values. The formula for calculating the dynamic conditional structure complexity function is: in, The DCSC (Discretionary Cost Function) represents the complexity of the dynamic conditional structure and is the core indicator of the feature selection cost function. It measures the macroscopic discreteness and microscopic volatility of the data distribution under a selected feature subset. The smaller the value, the better the feature subset; the higher the value, the more complex and unstable the dynamic behavior of the system represented by that feature subset is. The goal of feature selection is to find the feature subset that minimizes the DCSC value, that is, the feature combination that best simplifies the system representation and captures the core dynamics. The feature selection vector is a set, where each element... It is a binary variable ( (It is a feature index) The value can be 0 or 1, indicating the first or second digit. The vector represents whether each feature is selected (1 for selected, 0 for unselected). This vector is the optimization objective, and the optimal value is determined by minimizing the DCSC (Discretionary Value Score). MA represents the total number of time intervals into which the preprocessed time-series data stream is divided (corresponding to the number of intervals in step S1). MA is a positive integer used to characterize the time resolution, and its value is usually determined based on the amount of data and the desired granularity, for example, 10, 20, or 50. In the formula, MA is the upper limit of the summation. The interval index representing the time interval (i.e., the summation loop variable, representing the first...) There are several time intervals, traversing from 1 to MA (where MA is an integer variable). In the DCSC formula... Used to iterate through each interval to calculate the contribution. Indicates the first The contextual importance weights for each time interval are calculated using a nonlinear transformation (such as the sigmoid function) based on the confidence labels output in step S1. Their purpose is to emphasize the influence of high-confidence data intervals in the evaluation; the higher the confidence level of a time period, the greater its data distribution pattern contributes to the feature selection cost function. This represents the currently selected subset of features (i.e., the feature selection vector). Under the condition that the data falls into the first The conditional probability of each time interval is a probability value ranging from [0,1]. It is calculated as follows: based on the M time intervals divided in step S1, the probability is obtained by statistically analyzing the frequency of data points falling into each interval under the selected feature subset (i.e., ...). It reflects the ability of a feature subset to represent the temporal distribution pattern of data. It is a direct manifestation of the concept of information entropy and is used to measure the "macroscopic discreteness" of data. This represents the balance coefficient, an adjustable parameter greater than zero, used to weigh the importance of the two terms in the formula. A key regulator: when data fluctuates drastically (high RFE in step S1), it can be increased. Increase the weight of the second component (dynamic fluctuation term). Indicates the first The feature dynamic sensitivity weights are calculated based on the gradient norm of each feature within that time interval, for example, using an exponential function for amplification. Their function is to quantify the sensitivity of a feature to changes within that time period; the more sensitive a feature is, the more its fluctuations contribute to the complexity. This represents the prior probability, i.e., the probability within the selected feature subset (i.e., the feature selection vector). Under the condition that the data belongs to the first The probability of a time interval is set in the formula as... equal, Entropy measures representing dynamic complexity, i.e., the first... Dynamic eigenvalues ​​calculated based on selected features over a given time interval. Shannon entropy, where It is the relative fluctuation entropy of the characteristics within this interval. The entropy is used to measure the "micro-volatility" of data within each time interval. Even if the macro-distribution is similar, the drastic fluctuations of data within the interval will lead to an increase in this entropy value. The function of this formula is: The DCSC function considers both the macro-dispersion (information entropy part) and micro-volatility (dynamic characteristics part) of the data distribution. It can more profoundly reveal the ability of feature subsets to represent the complexity of the system in the time series evolution and avoid the neglect of nonlinear time series dependence by traditional methods. The weight coefficients of the feature selection cost function are dynamically adjusted based on the statistical characteristics of the preprocessed time-series data stream with confidence labels. The specific process is as follows: Statistical characteristics, including relative fluctuation entropy and average confidence, are extracted from the preprocessed time series data stream with confidence labels. The average confidence is the arithmetic mean of all confidence label values. The mutual information matrix between features is calculated based on the environmental parameter sequence and equipment performance index sequence in the preprocessed time series data stream, and the spectral norm of the matrix is ​​further calculated. The weight coefficients in the feature selection cost function are dynamically adjusted, including the balance coefficient, the feature dynamic sensitivity weight, and the context importance weight. The adjustment rule for the balance coefficient is that it is directly proportional to the value of the relative fluctuation entropy; when the value of the relative fluctuation entropy increases, the value of the balance coefficient increases accordingly. The adjustment rule for the confidence amplification factor of context importance weight is that it is related to the average confidence score through a hyperbolic tangent function; the value of the confidence amplification factor increases when the average confidence score increases. The adjustment rule for the feature dynamic sensitivity weight is that it is inversely proportional to the spectral norm of the mutual information matrix between features; the value of the feature dynamic sensitivity weight decreases as the spectral norm increases. The adjustment process is executed periodically based on a sliding window mechanism, with the size of the sliding window preset according to the data sampling frequency; The specific process of finding the minimum value of the feature selection cost function to obtain the dynamically optimal feature subset and establishing the feature mapping relationship is as follows: The minimum value of the feature selection cost function is solved using a quantum annealing optimization framework. The specific process includes encoding operations, Hamiltonian construction operations, evolution and cooling operations, and result reading operations. The encoding operation encodes the feature-selected binary vector and the dynamically adjusted weight coefficients into the quantum state of the qubit; The Hamiltonian construction operation sets the initial Hamiltonian to an easily prepared ground state, and the target Hamiltonian is constructed as a quantum version of the feature selection cost function, with its ground state corresponding to the minimum value of the feature selection cost function. The evolution and cooling operation simulates the adiabatic evolution process of a quantum system and utilizes the quantum tunneling effect to make the system state evolve from the initial Hamiltonian to the target Hamiltonian, avoiding local optima and gradually converging to the global optimum. The result reading operation decodes the optimal feature selection binary vector and the corresponding weight coefficient configuration from the final quantum state after the evolution ends; The subset of features indicated by the binary vector of the optimal feature selection is taken as the dynamic optimal feature subset, and the adjustment path of the weight coefficients during quantum annealing is recorded. Based on this path, a mapping relationship from the original feature space to the dynamic optimal feature subset is established. The dynamic optimal feature subset is directly output to the training steps of the subsequent quality evaluation model.

[0022] In this embodiment, it is specifically necessary to explain how the weight coefficients of the feature selection cost function are dynamically adjusted based on the statistical characteristics of the preprocessed time-series data stream with confidence labels. The specific process is as follows: Statistical characteristics, including relative fluctuation entropy and average confidence, are extracted from the preprocessed time series data stream with confidence labels. The average confidence is the arithmetic mean of all confidence label values. The mutual information matrix between features is calculated based on the environmental parameter sequence and equipment performance index sequence in the preprocessed time series data stream, and the spectral norm of the matrix is ​​further calculated. The weight coefficients in the feature selection cost function are dynamically adjusted, including the balance coefficient, the feature dynamic sensitivity weight, and the context importance weight. The adjustment rule for the balance coefficient is that it is directly proportional to the value of the relative fluctuation entropy; when the value of the relative fluctuation entropy increases, the value of the balance coefficient increases accordingly. The adjustment rule for the confidence amplification factor of context importance weight is that it is related to the average confidence score through a hyperbolic tangent function; the value of the confidence amplification factor increases when the average confidence score increases. The adjustment rule for the feature dynamic sensitivity weight is that it is inversely proportional to the spectral norm of the mutual information matrix between features; the value of the feature dynamic sensitivity weight decreases as the spectral norm increases. The adjustment process is executed periodically based on a sliding window mechanism, with the size of the sliding window preset according to the data sampling frequency; The specific process of finding the minimum value of the feature selection cost function to obtain the dynamically optimal feature subset and establishing the feature mapping relationship is as follows: The minimum value of the feature selection cost function is solved using a quantum annealing optimization framework. The specific process includes encoding operations, Hamiltonian construction operations, evolution and cooling operations, and result reading operations. The encoding operation encodes the feature-selected binary vector and the dynamically adjusted weight coefficients into the quantum state of the qubit; The Hamiltonian construction operation sets the initial Hamiltonian to an easily prepared ground state, and the target Hamiltonian is constructed as a quantum version of the feature selection cost function, with its ground state corresponding to the minimum value of the feature selection cost function. The evolution and cooling operation simulates the adiabatic evolution process of a quantum system and utilizes the quantum tunneling effect to make the system state evolve from the initial Hamiltonian to the target Hamiltonian, avoiding local optima and gradually converging to the global optimum. The result reading operation decodes the optimal feature selection binary vector and the corresponding weight coefficient configuration from the final quantum state after the evolution ends; The subset of features indicated by the binary vector of the optimal feature selection is taken as the dynamic optimal feature subset, and the adjustment path of the weight coefficients during quantum annealing is recorded. Based on this path, a mapping relationship from the original feature space to the dynamic optimal feature subset is established. The dynamic optimal feature subset is directly output to the training steps of the subsequent quality assessment model. The specific process of globally aggregating and updating the gradient of the encrypted model using a contribution-based dynamic weighted aggregation algorithm is as follows: After completing local training, each edge node uses homomorphic encryption technology to encrypt the final gradient in its local loss gradient sequence to obtain the encrypted gradient, and then uploads the encrypted gradient to the aggregation server. The aggregation server calculates the contribution index of each edge node. The contribution index is a multi-dimensional contribution evaluation index, which is calculated based on performance improvement, data quality, and training fluctuation. The performance improvement is measured by the relative reduction between the local model loss and the global model loss of the previous round. The relative reduction is the difference between the global model loss of the previous round and the local model loss of the current round, divided by the global model loss of the previous round. The larger the relative reduction, the more significant the performance improvement. Data quality is measured by calculating the average confidence level of local data, where the average confidence level is the arithmetic mean of all confidence level labels. A higher average confidence level indicates better data quality. Training variability is measured by calculating the coefficient of variation of the local loss gradient. The coefficient of variation is the standard deviation of the local loss gradient sequence divided by the mean. The smaller the coefficient of variation, the more stable the training process. The relative reduction of the performance improvement item, the average confidence of the data quality item, and the coefficient of variation of the training fluctuation item are weighted and combined. The first weight coefficient, the second weight coefficient, and the third weight coefficient are used in the weighting combination. The first weight coefficient corresponds to the performance improvement item, the second weight coefficient corresponds to the data quality item, and the third weight coefficient corresponds to the training fluctuation item. The values ​​of the first weight coefficient, the second weight coefficient, and the third weight coefficient are in the range of zero to one, and the sum of the first weight coefficient, the second weight coefficient, and the third weight coefficient is one. The weighted combination result is obtained by multiplying the first weight coefficient by the relative reduction, adding the second weight coefficient by the average confidence label value, and subtracting the third weight coefficient by the coefficient of variation, thus obtaining the multi-dimensional contribution evaluation index value for each node. The formula is as follows: ; in, Indicates the first The multi-dimensional contribution evaluation index for each edge node is a scalar value used to quantify the degree of contribution of a node to the global model. The larger the value, the higher the contribution. The calculation formula comprehensively considers three dimensions: performance improvement, data quality, and training stability. Let represent the first, second, and third weighting coefficients, respectively, used to balance the relative importance of the three dimensions in the formula. They satisfy ... That is, the sum of the weights is 1. Specifically: Used for weighted performance improvement items Used for weighted data quality items, The values ​​of the first, second, and third weighting coefficients, used for weighted training of the fluctuation term, are preset by the user and are usually adjusted according to actual needs, ranging from 0 to 1. This represents the global model loss in the previous round (i.e., the loss at time t). The global model loss after one round of aggregation is a scalar used as a benchmark to evaluate the performance improvement of the node's local model. The loss value is typically derived from the global model evaluation in federated learning. Indicates the first The local model loss of each edge node in the current round (round t) is a scalar value calculated through local training of the node and used for comparison with the global loss. Indicates the first The average confidence level of local data from each edge node ranges from 0 to 1, with higher values ​​indicating better data quality. Indicates the first The coefficient of variation of the local loss gradient of each edge node is a scalar used to measure the stability of the node training process. It is calculated as the absolute value of the standard deviation of the local loss gradient sequence divided by the mean. The smaller the coefficient of variation, the more stable the training process; conversely, the greater the fluctuation. The function of this formula is to overcome the limitations of traditional federated learning aggregation algorithms that only consider the amount of data or simple accuracy. By evaluating the node contribution from multiple dimensions, it avoids the negative impact of low-quality or malicious nodes on the global model. Finally, this index is used to calculate the aggregation weight of each node to achieve dynamic weighted aggregation. Based on the multi-dimensional contribution evaluation index values, the softmax function is used to calculate the aggregate weight of each node. The temperature parameter in the softmax function is used to control the smoothness of the weight distribution; the larger the temperature parameter value, the smoother the weight distribution. The formula for calculating the aggregate weight is as follows: ; in, Indicates the first The aggregate weights of the edge nodes are used to weight the encryption gradient of each node during global aggregation. A larger value indicates a higher contribution from that node. The weights are calculated using the softmax function, based on the node's contribution index, and their values ​​range from 0 to 1. The sum of the weights of all nodes is 1. Indicates the first The multi-dimensional contribution evaluation index value of each edge node This represents the temperature parameter, which controls the smoothness of the weight distribution in the softmax function. The larger the value, the more evenly the weight distribution; the smaller the value, the more concentrated the weight is on high-contribution nodes. Its value range is typically a positive decimal, such as between 0.1 and 1.0. exp represents an exponential function, used in the softmax formula to convert the contribution index to positive numbers, ensuring the weights are non-negative. Indicates the summation index, represents the node number, and is related to... Similarly, the value range is: integers, from 1 to... , The total number of edge nodes represents the number of nodes participating in federated learning. It is used to calculate the summation range, which is a positive integer, such as 10 or 100, depending on the network size. The aggregation server calculates the weighted average gradient in encrypted mode, which is obtained by multiplying the encrypted gradient of each node by its aggregation weight and then summing the results. The formula is: ; in, This represents an encryption function, specifically a homomorphic encryption operation, which encrypts the model gradients to ensure data privacy. For example... Indicates the first An encrypted version of the local gradient of each edge node. Indicates the first The local model gradients of each edge node, i.e., the gradient vectors of the model parameters after local training, are used for global aggregation. The global model gradient, represented by the weighted average of the encrypted gradients of all nodes and then decrypted, is used to update the global model parameters. The encrypted weighted average gradient is then decrypted to obtain the decrypted gradient. The decrypted gradient is used to update the global model parameters, with the update formula being the global model parameters minus the learning rate multiplied by the decrypted gradient. The formula is: ; in, Represents the global model parameters after the t-th iteration, and represents the optimized model state, where t is the iteration number (a positive integer). Indicates the t-th The global model parameters before the first iteration are used to represent the historical states in the parameter update formula. The learning rate is a positive real number (usually small, such as 0.01), used to control the step size of gradient descent and affecting the model's convergence speed. The aggregation and update process is repeated until the rate of change of the global model loss is less than a preset threshold (e.g., ...). This is used to determine whether the model has converged. Training stops when the global loss rate of change is less than this threshold or when the maximum number of training epochs is reached (e.g., 1000, used to set an upper limit on the number of training epochs to prevent infinite loops), thereby generating an optimized global quality assessment model.

[0023] In this embodiment, it is specifically necessary to explain the following steps in step S4: after each edge node loads the optimized global quality assessment model, the specific steps for feature extraction and millisecond-level inference of the real-time collected data are as follows: First, each edge node extracts features from the real-time data stream using the dynamic optimal feature subset generated in step S2, selecting only the feature dimension corresponding to the dynamic optimal feature subset to generate a real-time feature vector. Secondly, the real-time feature vector is input into the optimized global quality assessment model generated in step S3 for millisecond-level inference. The inference process is optimized to a millisecond-level response to meet the real-time requirements, and a quality score is output, which is a value between zero and one. Then, set an anomaly threshold, which is a preset fixed value with a range between zero and one, usually set to 0.5; compare the quality score with the anomaly threshold, and when the quality score is lower than the anomaly threshold, trigger a performance anomaly flag, which is a binary variable with a value of one indicating an anomaly and a value of zero indicating normal. Simultaneously, the global quality assessment model outputs an anomaly root cause hypothesis, which is a discrete value taken from a predefined set of anomaly root cause types, containing multiple possible anomaly cause categories; and outputs a confidence score, which is a value between zero and one, representing the reliability of the model's anomaly root cause hypothesis, calculated based on the information entropy or maximum probability value of the model's output probability distribution; when a performance anomaly flag is triggered, a lightweight collaborative voting mechanism with neighboring nodes is initiated, with neighboring nodes determined based on network topology or geographical proximity, including nodes whose hop distance is less than or equal to a preset maximum distance; Edge nodes broadcast query requests to neighboring nodes and receive anomaly information triples returned by neighboring nodes, including anomaly flags, anomaly root cause hypothesis, and confidence level; the communication process of broadcasting query requests and receiving anomaly information triples adopts a lightweight coding protocol. The specific steps for lightweight collaborative voting and anomaly root cause determination are as follows: After receiving the anomaly information triples from neighboring nodes, the edge node initiating the vote calculates a consensus score for each possible anomaly root cause type. The consensus score is calculated as follows: First, obtain the set of neighboring nodes of the node, and denote the size of the set of neighboring nodes as the number of neighboring nodes; For each neighboring node, calculate its individual contribution to the root cause type. The individual contribution is the product of the indicator function value, the distance decay weight, and the prediction confidence of the neighboring node. The indicator function value is 1 if the anomaly root cause hypothesis of the neighboring node is consistent with the currently calculated root cause type, and 0 otherwise; the distance decay weight is calculated by the exponential decay function, the input of which is the product of the negative distance decay factor and the logical distance between nodes, which is represented by the number of network hops; the prediction confidence directly uses the confidence value provided by the neighboring node. Then, sum the individual contribution values ​​of all neighboring nodes and divide by the number of neighboring nodes to obtain the collaborative consistency score for this root cause type. After calculating the scores for all root cause types, the root cause type with the highest consensus score is selected as the final judgment result. If multiple root cause types have the same score, the type with the highest number of votes is selected first according to the majority consensus principle. Meanwhile, the maximum collaborative consistency score is used as the confidence level for judgment; Finally, the quality evaluation results are output, including anomaly indicators, anomaly root cause determination results, and determination confidence levels; Each node sends the results to the central server for aggregation, generates a global quality report, and records it for model optimization. The model iterative optimization includes feedback to step S3.

[0024] Example 2 This embodiment provides, for example Figure 2 The quantum communication network equipment quality evaluation system shown here specifically includes: The data acquisition and preprocessing module is deployed at the equipment node of the quantum communication network. It is used to synchronously acquire environmental parameter sequences and equipment performance index sequences at millisecond frequency through a multimodal sensor array as raw data. The acquired raw data is aligned and standardized in real time, and an adaptive dual-threshold anomaly detection algorithm is used for data cleaning. Confidence labels are added to the marked data points, generating a preprocessed time series data stream with confidence labels and calculating its statistical characteristics. The feature selection optimization module connects to the data acquisition and preprocessing module. It is used to construct a feature selection cost function based on the quantum annealing concept, dynamically adjust the weight coefficients of the cost function according to the statistical characteristics of the preprocessed time series data stream with confidence labels, solve for the minimum value of the cost function to obtain the dynamically optimal feature subset, and establish a feature mapping relationship. The model training and aggregation module connects to the feature selection and optimization module. Under the federated learning framework, each edge node locally trains the LSTM-Transformer hybrid model structure using a dynamically optimal feature subset. The encrypted model gradient is globally aggregated and updated through a contribution-based dynamic weighted aggregation algorithm, thereby generating an optimized global quality evaluation model. The real-time evaluation and collaborative decision-making module connects to the model training and aggregation module. It is used to load the optimized global quality evaluation model on each edge node. It applies the dynamic optimal feature subset to extract features and perform millisecond-level inference on the real-time collected data. When a performance anomaly is detected, it performs lightweight collaborative voting with neighboring nodes, determines the root cause of the anomaly based on the majority consensus principle, and outputs the quality evaluation result.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for evaluating the quality of quantum communication network equipment, characterized in that, Specifically, the following steps are included: Step S1: Deploy a multimodal sensor array at the equipment node of the quantum communication network, synchronously collect environmental parameter sequences and equipment performance index sequences at millisecond frequency as raw data, perform real-time alignment and standardization processing on the collected raw data, and use an adaptive dual-threshold anomaly detection algorithm to clean the data, add confidence labels to the marked data points, generate a preprocessed time series data stream with confidence labels, and calculate its statistical characteristics including volatility. Step S2: Construct a feature selection cost function based on the idea of ​​quantum annealing, dynamically adjust the weight coefficients of the feature selection cost function according to the statistical characteristics of the preprocessed time series data stream with confidence labels, solve for the minimum value of the feature selection cost function to obtain the dynamic optimal feature subset, and establish a feature mapping relationship. Step S3: Under the federated learning framework, each edge node locally trains the LSTM-Transformer hybrid model structure using the dynamic optimal feature subset, and globally aggregates and updates the encrypted model gradients through a contribution-based dynamic weighted aggregation algorithm, thereby generating an optimized global quality evaluation model. Step S4: Each edge node loads the optimized global quality evaluation model, applies the dynamic optimal feature subset to the real-time collected data for feature extraction and millisecond-level inference, and when a performance anomaly is detected, performs lightweight collaborative voting with neighboring nodes, determines the root cause of the anomaly based on the majority consensus principle, and outputs the quality evaluation result.

2. The method for evaluating the quality of quantum communication network equipment according to claim 1, characterized in that: In step S1, the specific process of deploying a multimodal sensor array at the quantum communication network equipment node and synchronously collecting environmental parameter sequences and equipment performance index sequences as raw data at a millisecond-level frequency is as follows: A1. The multimodal sensor array includes a temperature sensor, a humidity sensor, an electromagnetic field strength sensor, a quantum bit error rate detection unit, and a photon count rate detection unit, wherein the millisecond frequency is a sampling interval of one millisecond to ten milliseconds; A2. The temperature sensor, humidity sensor, electromagnetic field strength sensor, quantum bit error rate detection unit and photon count rate detection unit in the multimodal sensor array are driven synchronously to trigger data acquisition operations via a high-precision clock signal at a millisecond frequency. A3. Obtain the original environmental parameter sequence and the original equipment performance index sequence with timestamp alignment. Use the original environmental parameter sequence and the original equipment performance index sequence together as the original data. The original environmental parameter sequence is a sequence of environmental parameter data composed of temperature, humidity and electromagnetic field strength data collected at each sampling time in chronological order. The original equipment performance index sequence is a sequence of equipment performance index data composed of quantum bit error rate and photon count rate data collected at each sampling time in chronological order. The specific process of real-time alignment and standardization of the collected raw data is as follows: B1. Based on the hardware timestamp carried by each data point in the original environmental parameter sequence and the original equipment performance index sequence, the environmental parameter data and equipment performance index data collected at the same sampling time are paired to form a parallel data stream aligned in the time dimension. B2. Standardization processing adopts the Z-score normalization method. The mean and standard deviation of the aligned environmental parameter sequence and the equipment performance index sequence are calculated respectively. Then, the mean of the environmental parameter sequence is subtracted from each data point in the environmental parameter sequence and divided by the standard deviation of the environmental parameter sequence. The mean of the equipment performance index sequence is subtracted from each data point in the equipment performance index sequence and divided by the standard deviation of the equipment performance index sequence to obtain the standardized environmental parameter sequence and the standardized equipment performance index sequence.

3. The method for evaluating the quality of quantum communication network equipment according to claim 2, characterized in that: The specific steps for data cleaning using the adaptive dual-threshold anomaly detection algorithm, and for assigning confidence labels to the marked data points, are as follows: C1. Set a hard threshold and a soft threshold. The hard threshold is determined based on the physical measurement range limits of various sensors in the multimodal sensor array. The soft threshold is dynamically calculated based on the data within a sliding window of a preset width, and is the arithmetic mean of all data points within the sliding window plus twice the standard deviation of all data points within the sliding window. C2. During data cleaning, first determine whether the current data point in the standardized environmental parameter sequence or standardized equipment performance index sequence exceeds the upper or lower limit of the hard threshold, and determine the data point that exceeds the upper or lower limit as an invalid data point. For data points that do not exceed the hard threshold range, it is further determined whether their values ​​are greater than the soft threshold. Data points that are greater than the soft threshold are marked as potential outliers. C3. For data points that are determined to be invalid in C2, assign them a confidence label value of zero; C4. For data points marked as potential outliers in C2, their confidence labels are calculated according to the following rules: First, calculate the absolute difference between the value of the data point and the arithmetic mean of all data points in the sliding window. Then, divide this absolute difference by the product of a preset sensitivity adjustment coefficient and the standard deviation of all data points in the sliding window. Finally, take the negative power of the exponential function with the natural constant as the base of the calculated quotient. The result is the confidence label value of the data point. C5. For data points that pass the hard and soft threshold detection in C2 and are not marked as potential outliers, assign them a confidence label value of one. Finally, generate a preprocessed time series data stream with confidence labels, consisting of a standardized environmental parameter sequence, a standardized equipment performance index sequence, and corresponding confidence labels. The calculation of its statistical characteristics, including volatility, is specifically as follows: The relative fluctuation entropy of a preprocessed time-series data stream with confidence labels is calculated. The calculation process for this relative fluctuation entropy includes: The numerical range of the preprocessed time series data stream with confidence labels is divided into MA consecutive intervals; Calculate the probability of a data point falling within each interval, and then calculate the mean and standard deviation of the data points within each interval. Finally, the probability of each interval is multiplied by the logarithm of that probability, then multiplied by the ratio of the standard deviation to the mean of that interval, and the sum of the results for all intervals is taken as a negative number to obtain the value of the relative fluctuation entropy.

4. The method for evaluating the quality of quantum communication network equipment according to claim 3, characterized in that: In step S2, the specific operation of constructing the feature selection cost function based on the quantum annealing concept is as follows: Define a feature selection vector, which is a binary sequence where each element takes the value of zero or one, where zero indicates that the corresponding feature is not selected and one indicates that the corresponding feature is selected; The feature selection cost function adopts a dynamic conditional structural complexity function, and the specific calculation process of this function is as follows: First, the time axis is divided into multiple consecutive time intervals, the total number of which is determined by the interval division in the preprocessed time series data stream with confidence labels; For each time interval, the first component is calculated, which is the context importance weight of the time interval multiplied by the conditional probability that the data falls into the time interval under the selected feature subset conditions, multiplied by the base-2 logarithm of the conditional probability, and the summation of the calculation results for all time intervals is taken as the average, and finally the negative value is taken. Secondly, the second component is calculated, which is a balance coefficient multiplied by the feature dynamic sensitivity weight for each time interval, multiplied by the prior probability that the data belongs to the time interval under the condition of a selected feature subset, and then multiplied by the entropy measure of the dynamic complexity of the feature set under the time interval. The results of the calculation over all time intervals are summed and averaged. Finally, the first component and the second component are added together to obtain the dynamic conditional structure complexity function value; wherein, the context importance weight is calculated by the sigmoid function based on the confidence labels in the preprocessed time series data stream, and a confidence amplification factor is introduced to adjust the influence of the confidence labels on the weights.

5. The method for evaluating the quality of quantum communication network equipment according to claim 4, characterized in that: The specific process of dynamically adjusting the weight coefficients of the feature selection cost function based on the statistical characteristics of the preprocessed time-series data stream with confidence labels is as follows: Statistical characteristics, including relative fluctuation entropy and average confidence, are extracted from the preprocessed time-series data stream with confidence labels. The average confidence is the arithmetic mean of all confidence label values. Based on the environmental parameter sequence and equipment performance index sequence in the preprocessed time-series data stream, the mutual information matrix between features is calculated, and the spectral norm of the matrix is ​​further calculated. The weight coefficients in the feature selection cost function are dynamically adjusted, including the balance coefficient, the feature dynamic sensitivity weight, and the context importance weight. The adjustment rule for the balance coefficient is that it is directly proportional to the value of the relative fluctuation entropy; when the value of the relative fluctuation entropy increases, the value of the balance coefficient increases accordingly. The adjustment rule for the confidence amplification factor of context importance weight is that it is related to the average confidence score through a hyperbolic tangent function; the value of the confidence amplification factor increases when the average confidence score increases. The adjustment rule for the feature dynamic sensitivity weight is that it is inversely proportional to the spectral norm of the mutual information matrix between features; the value of the feature dynamic sensitivity weight decreases as the spectral norm increases. The adjustment process is executed periodically based on a sliding window mechanism, with the size of the sliding window preset according to the data sampling frequency; The specific process of finding the minimum value of the feature selection cost function to obtain the dynamically optimal feature subset and establishing the feature mapping relationship is as follows: The minimum value of the feature selection cost function is solved using a quantum annealing optimization framework. The specific process includes encoding operations, Hamiltonian construction operations, evolution and cooling operations, and result reading operations. The encoding operation encodes the feature selection binary vector and the dynamically adjusted weight coefficients into the quantum state of the qubit; The Hamiltonian construction operation sets the initial Hamiltonian to an easily prepared ground state, and the target Hamiltonian is constructed as a quantum version of the feature selection cost function, with its ground state corresponding to the minimum value of the feature selection cost function. The evolution and cooling operations simulate the adiabatic evolution process of a quantum system and utilize the quantum tunneling effect to evolve the system state from the initial Hamiltonian to the target Hamiltonian. The result reading operation decodes the optimal feature selection binary vector and the corresponding weight coefficient configuration from the final quantum state after the evolution ends; The subset of features indicated by the binary vector of the optimal feature selection is taken as the dynamic optimal feature subset, and the adjustment path of the weight coefficients during quantum annealing is recorded. Based on this path, a mapping relationship from the original feature space to the dynamic optimal feature subset is established.

6. The method for evaluating the quality of quantum communication network equipment according to claim 5, characterized in that: In step S3, the specific operation of each edge node locally training the LSTM-Transformer hybrid model structure using the dynamically optimal feature subset is as follows: First, the LSTM-Transformer hybrid model structure consists of cascaded Long Short-Term Memory (LSTM) network units and Transformer Encoder (LEE) units. The LSTM network units are used to process the preprocessed temporal data stream with confidence labels generated in step S1 from the dynamic optimal feature subset. Each LSTM network unit is composed of multiple LSTM sub-units. Each LSTM sub-unit receives the input features of the current time step and the hidden state of the previous time step, and outputs the hidden state of the current time step. The hidden states of all time steps form a hidden state sequence, and the number of time steps is determined by the length of the preprocessed temporal data stream. The hidden state sequence is passed as input to the transformer encoder unit, which consists of a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer calculates the attention weight between each hidden state and other hidden states to generate a weighted feature representation. The feedforward neural network layer performs a non-linear transformation on the features at each position and outputs a global context feature sequence. The global context feature sequence is aggregated into a fixed-length feature vector through a global average pooling layer, and finally the feature vector is mapped to the quality evaluation result through a fully connected layer. During local training, weighted cross-entropy loss is used as the objective function. The weights of the weighted cross-entropy loss are determined by the confidence labels generated in step S1, i.e., the loss of each sample is multiplied by the confidence label value of that sample. Local training is optimized using the stochastic gradient descent algorithm. In each training iteration, the gradient of the weighted cross-entropy loss function with respect to all parameters of the model is calculated. This gradient is a vector and is called the local loss gradient. Throughout the local training process, the local loss gradient generated in each iteration is recorded to form a local loss gradient sequence. The learning rate is dynamically adjusted according to the training progress.

7. The method for evaluating the quality of quantum communication network equipment according to claim 6, characterized in that: The specific process of globally aggregating and updating the encrypted model gradient using a contribution-based dynamic weighted aggregation algorithm is as follows: After completing local training, each edge node uses homomorphic encryption technology to encrypt the final gradient in its local loss gradient sequence to obtain the encrypted gradient, and then uploads the encrypted gradient to the aggregation server. The aggregation server calculates the contribution index of each edge node. The contribution index is a multi-dimensional contribution evaluation index, which is calculated based on performance improvement items, data quality items, and training fluctuation items. The performance improvement is measured by the relative reduction of the local model loss to the global model loss of the previous round. The relative reduction is the difference between the global model loss of the previous round and the local model loss of the current round, divided by the global model loss of the previous round. The data quality item is measured by calculating the average confidence level of the local data, where the average confidence level is the arithmetic mean of all confidence level label values; The training fluctuation term is measured by calculating the coefficient of variation of the local loss gradient, which is the standard deviation of the local loss gradient sequence divided by the mean. The relative reduction of the performance improvement item, the average confidence of the data quality item, and the coefficient of variation of the training fluctuation item are weighted and combined. The first weight coefficient, the second weight coefficient, and the third weight coefficient are used in the weighting combination. The first weight coefficient corresponds to the performance improvement item, the second weight coefficient corresponds to the data quality item, and the third weight coefficient corresponds to the training fluctuation item. The values ​​of the first weight coefficient, the second weight coefficient, and the third weight coefficient are in the range of zero to one, and the sum of the first weight coefficient, the second weight coefficient, and the third weight coefficient is one. The weighted combination result is the first weight coefficient multiplied by the relative reduction, plus the second weight coefficient multiplied by the average confidence label value, minus the third weight coefficient multiplied by the coefficient of variation, to obtain the multidimensional contribution evaluation index value of each node. Based on the multi-dimensional contribution evaluation index values, the softmax function is used to calculate the aggregate weight of each node. The temperature parameter in the softmax function is used to control the smoothness of the weight distribution. The aggregation server calculates the weighted average gradient in the encrypted state, that is, it multiplies the encrypted gradient of each node by its aggregate weight and sums them to obtain the encrypted weighted average gradient. Then, the encrypted weighted average gradient is decrypted to obtain the decrypted gradient. The global model parameters are updated using the decrypted gradient. The update formula is the global model parameters minus the learning rate multiplied by the decrypted gradient. The above aggregation and update process is repeated until the rate of change of the global model loss is less than a preset threshold, thereby generating an optimized global quality evaluation model.

8. The method for evaluating the quality of quantum communication network equipment according to claim 7, characterized in that: In step S4, after each edge node loads the optimized global quality assessment model, the specific steps for feature extraction and millisecond-level inference of the real-time collected data are as follows: First, each edge node extracts features from the real-time data stream using the dynamic optimal feature subset generated in step S2, selecting only the feature dimension corresponding to the dynamic optimal feature subset to generate a real-time feature vector. Secondly, the real-time feature vector is input into the optimized global quality assessment model generated in step S3 for millisecond-level inference. The inference process is optimized to a millisecond-level response and outputs a quality score, which is a value between zero and one. Then, an anomaly threshold is set. The anomaly threshold is a preset fixed value, ranging from zero to one. The quality score is compared with the anomaly threshold. When the quality score is lower than the anomaly threshold, a performance anomaly flag is triggered. The performance anomaly flag is a binary variable, with a value of one indicating an anomaly and a value of zero indicating normal. At the same time, the global quality assessment model outputs an anomaly root cause hypothesis; It also outputs a confidence score, which is a value between zero and one, calculated based on the information entropy of the model's output probability distribution. When a performance anomaly flag is triggered, a lightweight collaborative voting mechanism with neighboring nodes is initiated. Neighboring nodes are determined based on network topology or geographical proximity, including nodes whose hop distance is less than or equal to a preset maximum distance. Edge nodes broadcast query requests to neighboring nodes and receive anomaly information triples returned by neighboring nodes, including anomaly flags, anomaly root cause hypothesis, and confidence level; the communication process of broadcasting query requests and receiving anomaly information triples adopts a lightweight coding protocol.

9. The method for evaluating the quality of quantum communication network equipment according to claim 8, characterized in that: The specific steps for the lightweight collaborative voting and anomaly root cause determination are as follows: After receiving the anomaly information triple from neighboring nodes, the edge node initiating the vote calculates a consensus score for each possible anomaly root cause type. The consensus score is calculated as follows: First, obtain the set of neighboring nodes of the node, and denote the size of the set of neighboring nodes as the number of neighboring nodes; For each neighboring node, its individual contribution to the root cause type is calculated. The individual contribution is the product of the indicator function value, the distance decay weight, and the prediction confidence of the neighboring node. The indicator function value is 1 if the anomaly root cause hypothesis of the neighboring node is consistent with the currently calculated root cause type, and 0 otherwise; the distance decay weight is calculated by the exponential decay function, the input of which is the product of the negative distance decay factor and the logical distance between nodes, which is represented by the number of network hops; the prediction confidence directly uses the confidence value provided by the neighboring node. Then, sum the individual contribution values ​​of all neighboring nodes and divide by the number of neighboring nodes to obtain the collaborative consistency score for this root cause type. After calculating the scores for all root cause types, the root cause type with the highest consensus score is selected as the final judgment result. If multiple root cause types have the same score, the type with the highest number of votes is selected first according to the majority consensus principle. Meanwhile, the maximum collaborative consistency score is used as the confidence level for judgment; Finally, the quality evaluation results are output, including anomaly indicators, anomaly root cause determination results, and determination confidence levels; Each node sends the results to the central server for aggregation, generating a global quality report and recording it for model optimization.

10. A quality evaluation system for quantum communication network equipment is applied to the quality evaluation method for quantum communication network equipment as described in any one of claims 1-9, characterized in that: Specifically, it includes: The data acquisition and preprocessing module is deployed at the equipment node of the quantum communication network. It is used to synchronously acquire environmental parameter sequences and equipment performance index sequences at millisecond frequency through a multimodal sensor array as raw data. The acquired raw data is aligned and standardized in real time, and an adaptive dual-threshold anomaly detection algorithm is used for data cleaning. Confidence labels are added to the marked data points, generating a preprocessed time series data stream with confidence labels and calculating its statistical characteristics. The feature selection optimization module connects to the data acquisition and preprocessing module. It is used to construct a feature selection cost function based on the quantum annealing concept, dynamically adjust the weight coefficients of the cost function according to the statistical characteristics of the preprocessed time series data stream with confidence labels, solve for the minimum value of the cost function to obtain the dynamically optimal feature subset, and establish a feature mapping relationship. The model training and aggregation module connects to the feature selection and optimization module. Under the federated learning framework, each edge node locally trains the LSTM-Transformer hybrid model structure using a dynamically optimal feature subset. The encrypted model gradient is globally aggregated and updated through a contribution-based dynamic weighted aggregation algorithm, thereby generating an optimized global quality evaluation model. The real-time evaluation and collaborative decision-making module connects to the model training and aggregation module. It is used to load the optimized global quality evaluation model on each edge node. It applies the dynamic optimal feature subset to extract features and perform millisecond-level inference on the real-time collected data. When a performance anomaly is detected, it performs lightweight collaborative voting with neighboring nodes, determines the root cause of the anomaly based on the majority consensus principle, and outputs the quality evaluation result.

Images