Abnormity detection and maintenance decision optimization method and system for automatic test equipment

Abstract

The invention provides an anomaly detection and maintenance decision optimization method and system for automatic test equipment, and relates to the technical field of automation, and the method comprises the steps: deploying an edge calculation unit and a multi-level sensor network to collect equipment data, carrying out the fault diagnosis through a deep network model, and constructing a maintenance decision model in combination with equipment operation data. The model is input into a quantum enhancement dual-network system to calculate a reward value, and is used for training a maintenance strategy network under a federal reinforcement learning framework. And generating a maintenance resource configuration scheme through a quantum evolutionary algorithm, and generating a maintenance scheme through optimization of the multi-criterion evaluation model and the self-adaptive multi-target differential evolutionary algorithm. According to the scheme, a high-fidelity digital twin model and a multi-task deep transfer learning network are input for verification and updating, updated data are transmitted to a field control system through an industrial Internet of Things platform, and the updated data are recorded by using a block chain technology for updating a deep network model.

Description

Technical Field

[0001] The present invention relates to automation technology, and in particular to an abnormal detection and maintenance decision optimization method and system for automated test equipment. Background Art

[0002] Automated test equipment plays a crucial role in modern industrial production, and its operating status directly affects product quality and production efficiency. In order to ensure the stable and reliable operation of automated test equipment, effective abnormal detection and maintenance are required. Traditional abnormal detection methods mainly rely on manual experience and regular maintenance, which have problems such as low detection accuracy, high maintenance cost, and inability to detect potential faults in a timely manner. With the development of sensor technology, artificial intelligence, and Internet of Things technology, some new abnormal detection and maintenance methods have emerged, such as fault diagnosis and predictive maintenance based on machine learning.

[0003] The existing technologies mainly have the following defects and deficiencies: 1. Difficult to process complex multi-source heterogeneous data: Automated test equipment is usually equipped with various types of sensors, and the data collected by these sensors are diverse in type and format. Traditional methods are difficult to effectively fuse and analyze this data, resulting in low abnormal detection accuracy.

[0004] 2. Lack of adaptive maintenance decision-making ability: Traditional maintenance strategies are usually based on pre-set rules or experience and cannot be dynamically adjusted according to the actual operating status of the equipment, resulting in waste or shortage of maintenance resources.

[0005] 3. Difficult to achieve global optimization: The maintenance of automated test equipment involves multiple links and multiple goals, such as maintenance cost, downtime, equipment life, etc. Traditional methods are difficult to balance multiple goals and difficult to achieve global optimization. Summary of the Invention

[0006] Embodiments of the present invention provide an abnormal detection and maintenance decision optimization method and system for automated test equipment, which can solve the problems in the existing technologies.

[0007] In the first aspect of the embodiments of the present invention, there is provided an abnormal detection and maintenance decision optimization method for automated test equipment, including: Deploy edge computing units and multi-level sensor networks; use the self-organizing time synchronization protocol through the edge computing units to synchronize multi-source sensing data; use the empirical mode decomposition algorithm and variational mode decomposition technology to perform mode separation on the synchronized sensing data; construct a deep network model including a graph neural network module and a recursive variational autoencoder module, and the deep network model processes different sensor data through a cross-modal attention mechanism; match the output features of the deep network model with the device historical fault case database to generate a device state feature vector; establish a knowledge graph based on the device state feature vector to obtain a hierarchical fault diagnosis result; Receive the hierarchical fault diagnosis result, and construct a maintenance decision model in combination with device operation data; input the maintenance decision model into a quantum-enhanced dual network system to calculate a sequence of reward values; construct a dynamic probability graph model, input the sequence of reward values into the dynamic probability graph model, and combine the dynamic probability graph model with a Bayesian network and a hidden Markov model to generate a state transition matrix; train a maintenance policy network using a federated reinforcement learning framework; input the state transition matrix into the maintenance policy network, and input the output of the maintenance policy network into a quantum evolutionary algorithm to generate a multi-constraint maintenance resource allocation plan; Input the multi-constraint maintenance resource allocation plan into a multi-criteria evaluation model based on interval intuitionistic fuzzy sets and improved grey relational degrees to output evaluation result data; input the evaluation result data into an adaptive multi-objective differential evolution algorithm to generate optimized maintenance plan data; input the optimized maintenance plan data into a high-fidelity digital twin model to output integrated model data; input the integrated model data into a multi-task deep transfer learning network and a meta-learning network to generate maintenance decision model update data; transmit the maintenance decision model update data to the field control system through an industrial Internet of Things platform; record the maintenance decision model update data using a blockchain network, and the data of the blockchain network is used as the input of the device historical fault case database to update the deep network model.

[0008] The step of receiving the hierarchical fault diagnosis result and constructing a maintenance decision model in combination with device operation data includes: Receive the hierarchical fault diagnosis result, and extract fault type data, fault level data, and fault feature data from the hierarchical fault diagnosis result; input the fault type data, the fault level data, and the fault feature data into a multi-level adaptive tensor decomposition network, and the multi-level adaptive tensor decomposition network uses the high-order singular value decomposition method to perform non-linear dimensionality reduction and reconstruction on the fault data to generate a fault feature tensor; input the fault feature tensor into a fuzzy analytic hierarchy process network with a dynamic weight allocation mechanism to generate a fault comprehensive scoring matrix; Collect the real-time operation data of the device, and input the real-time operation data of the device and the fault comprehensive scoring matrix into a multi-scale variational mode decomposition device. The multi-scale variational mode decomposition device adaptively adjusts the decomposition parameters based on the fault comprehensive scoring matrix, and decomposes the real-time operation data of the device into multiple intrinsic mode components; input the intrinsic mode components into a multi-channel time-frequency analysis network with an attention mechanism. The multi-channel time-frequency analysis network includes a time-frequency energy channel, an instantaneous phase channel, and a modal correlation channel, and generates a multi-dimensional device operation feature tensor; Input the fault feature tensor and the multi-dimensional device operation feature tensor into a deep coupled feature fusion model. The deep coupled feature fusion model includes a spatial graph convolutional layer, a temporal graph convolutional layer, and a variational inference layer. The spatial graph convolutional layer constructs a feature space association graph, the temporal graph convolutional layer constructs a feature temporal association graph, and the variational inference layer generates a device state vector based on the dual association graph; input the device state vector into a multi-layer bidirectional neural network with a recursive attention mechanism, and generate a maintenance time series based on the device state vector; construct a maintenance decision model based on the maintenance time series.

[0009] Constructing the maintenance decision model based on the maintenance time series includes: Input the maintenance time series into an adaptive hybrid quantum optimization system. The adaptive hybrid quantum optimization system includes a quantum-behaved particle swarm module, a quantum simulated annealing module, a quantum genetic algorithm module, and a quantum entanglement enhancement module. The quantum-behaved particle swarm module uses a multi-particle entanglement coding method to perform quantum state coding on the maintenance time series to obtain a quantum coding sequence. The quantum simulated annealing module performs quantum tunneling state transfer on the quantum coding sequence to obtain a quantum transfer sequence. The quantum genetic algorithm module performs quantum chromosome crossover and quantum bit mutation operations on the quantum transfer sequence to obtain a quantum optimization sequence. The quantum entanglement enhancement module performs a quantum entanglement state operation on the quantum optimization sequence to generate a multi-dimensional maintenance decision sequence; Input the multi-dimensional maintenance decision sequence into an adaptive hierarchical cognitive computing network. The adaptive hierarchical cognitive computing network includes a maintenance strategy generation layer, a maintenance effect prediction layer, a knowledge reasoning layer, and a parameter optimization layer. The maintenance strategy generation layer uses a hierarchical recursive structure with a memory enhancement mechanism to transform the multi-dimensional maintenance decision sequence into a maintenance action space. The maintenance effect prediction layer uses a multi-head attention mechanism and a causal inference mechanism to predict the maintenance action space to obtain execution effect data. The knowledge reasoning layer inputs the maintenance action space and the execution effect data into a graph neural network to construct a maintenance knowledge graph. The parameter optimization layer inputs the maintenance action space, the execution effect data, and the maintenance knowledge graph into a meta-learning network to update the network parameters of the deep coupled feature fusion model; Construct an adaptive hybrid model based on the maintenance action space, the execution effect data, the maintenance knowledge graph, and the updated network parameters. The adaptive hybrid model includes an action mapping engine, a multi-modal fusion engine, and a policy evaluation engine. The action mapping engine uses a hierarchical attention mechanism to map the maintenance action space into a multi-level maintenance policy tensor. The multi-modal fusion engine fuses the features of the multi-level maintenance policy tensor, the execution effect data, and the maintenance knowledge graph to obtain a fused feature vector. The policy evaluation engine inputs the fused feature vector into a deep belief network to construct a maintenance decision model.

[0010] Input the maintenance decision model into a quantum-enhanced dual network system to calculate a sequence of reward values, including: Input the maintenance decision model into the quantum-enhanced dual network system. The quantum-enhanced dual network system includes a quantum policy network module and a hierarchical evaluation network module. The quantum policy network module uses a multi-particle quantum entanglement state encoder to map the maintenance decision model into a high-dimensional quantum feature space to obtain a sequence of quantum states. The hierarchical evaluation network module uses a dynamic graph neural network and a cross-scale multi-head attention mechanism to process the sequence of quantum states to obtain a spatio-temporal multi-dimensional value tensor. Input the spatio-temporal multi-dimensional value tensor into a recursive target decomposition system. The recursive target decomposition system includes an adaptive neural tensor decomposition module and a multi-scale temporal feature extraction module. The adaptive neural tensor decomposition module uses a tensor kernel decomposition method to recursively decompose the spatio-temporal multi-dimensional value tensor to obtain a device reliability dimension tensor, a maintenance cost dimension tensor, and a resource utilization dimension tensor. The multi-scale temporal feature extraction module uses a hierarchical deep temporal model to process the dimension tensors in sequence to obtain a set of maintenance decision features. Among them, failure evolution features and life prediction features are extracted from the device reliability dimension tensor, cost dynamic features and resource consumption features are extracted from the maintenance cost dimension tensor, and efficiency optimization features and load balancing features are extracted from the resource utilization dimension tensor. Calculate the dynamic weight coefficient matrix between the features in the set of maintenance decision features based on an improved non-dominated sorting genetic algorithm and an adaptive fuzzy analytic hierarchy process. The set of maintenance decision features and the dynamic weight coefficient matrix are input into a hybrid state evaluation system. The hybrid state evaluation system includes a conditional variational encoding module and a probability inference module. The conditional variational encoding module uses a conditional variational autoencoder to map the features into a non-linear state space to obtain a multi-modal state representation vector. The probability inference module uses an improved deep belief network to calculate a time-varying state transition probability tensor based on the multi-modal state representation vector and the dynamic weight coefficient matrix. Input the time-varying state transition probability tensor into a two-way evaluation system to calculate a sequence of reward values.

[0011] Inputting the time-varying state transition probability tensor into the bidirectional evaluation system to calculate a sequence of reward values, including: Inputting the time-varying state transition probability tensor into the bidirectional evaluation system, the bidirectional evaluation system includes a hierarchical action evaluation module and a state evaluation module. The hierarchical action evaluation module uses an improved bidirectional long short-term memory network and a cross-dimensional attention mechanism to evaluate and maintain actions to obtain multi-level action influence vectors. The state evaluation module uses a spatio-temporal graph convolutional network to evaluate state transitions to obtain state influence vectors. The multi-level action influence vectors and the state influence vectors are fused through an adaptive residual connection structure and a cross-modal feature fusion network to obtain a multi-dimensional reward vector; Constructing the meta-learning module, the meta-learning module includes a hierarchical gradient policy network, a dynamic programming network, and a conditional variational network. Among them, the hierarchical gradient policy network uses a hierarchical structure to calculate action gradients and state gradients. The dynamic programming network calculates the discount factor based on the Bellman equation. The conditional variational network uses an encoder-decoder structure to randomly sample state transitions to generate a conditional variational state value function; the multi-dimensional reward vector, the discount factor, and the conditional variational state value function are mapped to a unified feature space through a non-linear transformation to obtain a feature mapping matrix; the feature mapping matrix is decomposed to obtain a reward feature and a discount feature; the tensor product of the reward feature and the discount feature is calculated to obtain a reward-discount tensor, and then the reward-discount tensor and the conditional variational state value function are subjected to a tensor contraction operation to obtain an accumulated reward tensor; Inputting the accumulated reward tensor into the adversarial training network, the adversarial training network extracts features through a discriminator to obtain a discriminant score, and generates an optimized reward sample based on the discriminant score through a generator; a hierarchical knowledge distillation network is constructed using the discriminant score and the optimized reward sample, and the state transition features and action execution features in the pre-trained teacher network are transferred to the student network to obtain a knowledge transfer error; an adaptive loss function is constructed based on the discriminant score, the optimized reward sample, and the knowledge transfer error, and a distributed asynchronous parallel algorithm is used to optimize network parameters to generate a sequence of reward values including temporal dependence relationships and state transition relationships.

[0012] Inputting the state transition matrix into the maintenance strategy network and processing the output of the maintenance strategy network based on the quantum evolutionary algorithm to generate a multi-constraint maintenance resource allocation scheme, including: Input the state transition matrix into the maintenance policy network. The maintenance policy network includes a state encoding layer, a policy mapping layer, and an action generation layer. The state encoding layer uses a multi-head self-attention module to calculate the correlation weights between states for the state transition matrix. The multi-head self-attention module includes a query matrix, a key matrix, and a value matrix. Attention scores are obtained through the dot product operation between the query matrix and the key matrix, and the weighted features are obtained by multiplying the attention scores with the value matrix. At the same time, position encoding is introduced to retain the temporal information of state transitions. The weighted features, the correlation weights between states, and the position encoding are feature fused to obtain a state feature vector; Based on the state feature vector and the correlation weights between states, a value evaluation branch and a policy generation branch are constructed in the policy mapping layer. The value evaluation branch uses a fully connected layer with residual connection to calculate the state value function of the maintenance action. The residual connection includes a main channel and a bypass channel. The main channel performs a non-linear transformation on the state feature vector and the correlation weights between states. The bypass channel keeps the state feature vector and the correlation weights between states unchanged and superimposes them with the output of the main channel to obtain the state value function; The policy generation branch includes a policy network and a value network. The policy network outputs the probability distribution of the maintenance action based on the state feature vector and the correlation weights between states. The value network evaluates the expected return of the maintenance action based on the state feature vector and the correlation weights between states. The policy network is optimized by maximizing the expected return of entropy regularization to obtain the optimized probability distribution of the maintenance action; In the action generation layer, based on the correlation weights between states, the reparameterization sampling technique is used to convert the optimized probability distribution of the maintenance action into discrete maintenance policy parameters. Specifically, a temperature parameter is introduced into the optimized probability distribution of the maintenance action to control the randomness of sampling, and discrete action samples are generated based on the temperature parameter and the correlation weights between states to obtain the maintenance policy parameters, which are used as the output of the maintenance policy network; Based on the quantum evolutionary algorithm, the maintenance policy parameters are processed to generate a multi-constraint maintenance resource allocation scheme.

[0013] The processing of the maintenance policy parameters based on the quantum evolutionary algorithm to generate a multi-constraint maintenance resource allocation scheme includes: Encode the association weights between the maintenance policy parameters and the states as a qubit string; input the state value function and the association weights between the states into an adaptive angle calculation module. The adaptive angle calculation module uses a value difference function to calculate the difference between the current solution corresponding to the maintenance policy parameters and the historical optimal solution, sets an adaptive coefficient based on the difference according to a preset proportional coefficient, multiplies the adaptive coefficient by a preset reference rotation angle to obtain the actual rotation angle of the quantum rotation gate; uses the actual rotation angle to drive the quantum rotation gate to adjust the phase of the qubit string, constructs a policy gradient vector based on the state value function and the association weights between the states, designs a quantum interference operation matrix using the policy gradient vector, and adjusts the local phase of the qubit string through the quantum interference operation matrix to achieve population-directed evolution, obtaining the qubit string after the first evolution; Use the state value function as an evaluation index, and at the same time use the association weights between the states as weight factors to perform weighted calculation on each qubit in the qubit string after the first evolution to obtain a weighted fitness score; calculate the resource overrun degree, time conflict degree, and reliability deficiency degree in the maintenance plan corresponding to the qubit string after the first evolution. The resource overrun degree is the difference between the actual required maintenance resource amount and the preset resource capacity, the time conflict degree is the duration by which the actual execution time of the maintenance activity exceeds the preset time window, and the reliability deficiency degree is the difference between the actual reliability of the equipment after maintenance and the preset reliability target; calculate a penalty term based on the resource overrun degree, the time conflict degree, and the reliability deficiency degree; perform weighted combination of the weighted fitness score and the penalty term according to a preset ratio to obtain a comprehensive fitness evaluation result, and screen the qubit string after the first evolution based on the comprehensive fitness evaluation result to obtain the qubit string after the second evolution; Calculate the variance of the values of each qubit in the qubit string after the second evolution to obtain a gene diversity index, and perform weighted combination of the gene diversity index and the association weights between the states according to a preset weight to obtain a diversity evaluation index; when the diversity evaluation index is lower than a preset lower threshold, multiply the current quantum mutation probability by a preset growth rate greater than 1 to obtain a new quantum mutation probability; when the diversity evaluation index is higher than a preset upper threshold, multiply the current quantum mutation probability by a preset decay rate less than 1 to obtain a new quantum mutation probability; use the new quantum mutation probability to perform a flip operation on the qubits in the qubit string after the second evolution to achieve quantum evolution, decode the evolved qubit string to obtain multiple maintenance plans, and select the maintenance plan with the highest fitness and that simultaneously meets the resource capacity limit, time window constraint, and reliability requirements from the multiple maintenance plans as the multi-constraint maintenance resource allocation plan.

[0014] In the second aspect of the embodiments of the present invention, an abnormal detection and maintenance decision optimization system for an automated test device is provided, including: A first unit for deploying an edge computing unit and a multi-level sensor network; synchronizing multi-source sensing data through the edge computing unit using a self-organizing time synchronization protocol; performing modal separation on the synchronized sensing data using an empirical mode decomposition algorithm and a variational mode decomposition technique; constructing a deep network model including a graph neural network module and a recursive variational autoencoder module, and the deep network model processes different sensor data through a cross-modal attention mechanism; matching the output features of the deep network model with a device historical fault case database to generate a device state feature vector; establishing a knowledge graph based on the device state feature vector to obtain a hierarchical fault diagnosis result; A second unit for receiving the hierarchical fault diagnosis result, constructing a maintenance decision model in combination with device operation data; inputting the maintenance decision model into a quantum-enhanced dual network system to calculate a sequence of reward values; constructing a dynamic probability graph model, inputting the sequence of reward values into the dynamic probability graph model, and combining the dynamic probability graph model with a Bayesian network and a hidden Markov model to generate a state transition matrix; training a maintenance policy network using a federated reinforcement learning framework; inputting the state transition matrix into the maintenance policy network, and inputting the output of the maintenance policy network into a quantum evolutionary algorithm to generate a multi-constraint maintenance resource allocation scheme; A third unit for inputting the multi-constraint maintenance resource allocation scheme into a multi-criteria evaluation model based on interval intuitionistic fuzzy sets and improved grey relational degrees to output evaluation result data; inputting the evaluation result data into an adaptive multi-objective differential evolution algorithm to generate optimized maintenance scheme data; inputting the optimized maintenance scheme data into a high-fidelity digital twin model to output integrated model data; inputting the integrated model data into a multi-task deep transfer learning network and a meta-learning network to generate maintenance decision model update data; transmitting the maintenance decision model update data to a field control system through an industrial Internet of Things platform; recording the maintenance decision model update data using a blockchain network, and the data of the blockchain network is used as the input of the device historical fault case database to update the deep network model.

[0015] The beneficial effects of this application are as follows: 1. Improve the accuracy and efficiency of fault diagnosis: Through a multi-level sensor network and advanced signal processing technologies (empirical mode decomposition, variational mode decomposition), combined with deep learning models (graph neural network, recursive variational autoencoder, cross-modal attention mechanism) and knowledge graphs, it is possible to identify device faults more accurately and quickly, and achieve hierarchical fault diagnosis.

[0016] 2. Optimize maintenance decisions and reduce maintenance costs: Utilize advanced technologies such as quantum-enhanced dual-network systems, dynamic probabilistic graph models (combining Bayesian networks and hidden Markov models), federated reinforcement learning, and quantum evolutionary algorithms to generate and optimize multi-constraint maintenance resource allocation schemes, and finally formulate more scientific and economical maintenance strategies.

[0017] 3. Achieve continuous improvement and adaptability of maintenance decisions: Based on multi-criteria evaluation models such as interval intuitionistic fuzzy sets and improved grey relational degrees, adaptive multi-objective differential evolution algorithms, high-fidelity digital twin models, multi-task deep transfer learning, and meta-learning networks, a complete maintenance decision model update mechanism is constructed, and through the industrial Internet of Things platform and blockchain technology, closed-loop data flow and secure storage are realized, enabling maintenance decisions to be continuously optimized and improved according to actual situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic flowchart of the method for anomaly detection and maintenance decision optimization of the automated test equipment according to an embodiment of the present invention; Figure 2 is a schematic structural diagram of the system for anomaly detection and maintenance decision optimization of the automated test equipment according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0021] Figure 1 is a schematic flowchart of the method for anomaly detection and maintenance decision optimization of the automated test equipment according to an embodiment of the present invention, as Figure 1 shown, the method includes: S101. Deploy edge computing units and multi-level sensor networks; use the self-organizing time synchronization protocol through the edge computing units to synchronize multi-source sensing data; use the empirical mode decomposition algorithm and variational mode decomposition technology to perform mode separation on the synchronized sensing data; construct a deep network model including a graph neural network module and a recurrent variational autoencoder module, and the deep network model processes different sensor data through a cross-modal attention mechanism; match the output features of the deep network model with the device historical fault case database to generate a device status feature vector; establish a knowledge graph based on the device status feature vector to obtain a hierarchical fault diagnosis result; S102. Receive the hierarchical fault diagnosis result, and construct a maintenance decision model in combination with device operation data; input the maintenance decision model into a quantum-enhanced dual network system to calculate a reward value sequence; construct a dynamic probability graph model, input the reward value sequence into the dynamic probability graph model, and combine the dynamic probability graph model with a Bayesian network and a hidden Markov model to generate a state transition matrix; train a maintenance policy network using a federated reinforcement learning framework; input the state transition matrix into the maintenance policy network, and input the output of the maintenance policy network into a quantum evolutionary algorithm to generate a multi-constraint maintenance resource allocation plan; S103. Input the multi-constraint maintenance resource allocation plan into a multi-criteria evaluation model based on interval intuitionistic fuzzy sets and improved grey relational degrees to output evaluation result data; input the evaluation result data into an adaptive multi-objective differential evolution algorithm to generate optimized maintenance plan data; input the optimized maintenance plan data into a high-fidelity digital twin model to output integrated model data; input the integrated model data into a multi-task deep transfer learning network and a meta-learning network to generate maintenance decision model update data; transmit the maintenance decision model update data to the on-site control system through an industrial Internet of Things platform; record the maintenance decision model update data using a blockchain network, and the data of the blockchain network is used as the input of the device historical fault case database to update the deep network model.

[0022] In an alternative embodiment, the receiving the hierarchical fault diagnosis result and constructing a maintenance decision model in combination with device operation data includes: Receive the hierarchical fault diagnosis result, and extract fault type data, fault level data, and fault feature data from the hierarchical fault diagnosis result; input the fault type data, the fault level data, and the fault feature data into a multi-level adaptive tensor decomposition network, and the multi-level adaptive tensor decomposition network uses a high-order singular value decomposition method to perform non-linear dimensionality reduction and reconstruction on the fault data to generate a fault feature tensor; input the fault feature tensor into a fuzzy analytic hierarchy process network with a dynamic weight allocation mechanism to generate a fault comprehensive scoring matrix; The acquisition device collects real-time operation data of the device, and inputs the real-time operation data of the device and the fault comprehensive scoring matrix into a multi-scale variational mode decomposition device. The multi-scale variational mode decomposition device adaptively adjusts the decomposition parameters based on the fault comprehensive scoring matrix, and decomposes the real-time operation data of the device into multiple intrinsic mode components; the intrinsic mode components are input into a multi-channel time-frequency analysis network with an attention mechanism. The multi-channel time-frequency analysis network includes a time-frequency energy channel, an instantaneous phase channel, and a modal correlation channel, and generates a multi-dimensional device operation feature tensor; The fault feature tensor and the multi-dimensional device operation feature tensor are input into a deep coupled feature fusion model. The deep coupled feature fusion model includes a spatial graph convolutional layer, a temporal graph convolutional layer, and a variational inference layer. The spatial graph convolutional layer constructs a feature space correlation graph, the temporal graph convolutional layer constructs a feature temporal correlation graph, and the variational inference layer generates a device state vector based on the dual correlation graph; the device state vector is input into a multi-layer bidirectional neural network with a recursive attention mechanism, and a maintenance time series is generated based on the device state vector; a maintenance decision model is constructed based on the maintenance time series.

[0023] A method for constructing a maintenance decision model based on hierarchical fault diagnosis results and device operation data is as follows in specific implementation manners: First, receive the fault diagnosis results of the device. This diagnosis result adopts a hierarchical structure and can provide fault information with different granularities. For example, the hierarchical fault diagnosis results of a motor system may include: the first layer, motor system fault; the second layer, driver fault, mechanical fault; the third layer, driver overheating, bearing wear, etc. Three types of key data are extracted from the hierarchical fault diagnosis results: fault type, fault level, and fault features. Taking driver overheating as an example, the fault type is "overheating", the fault level is the third layer, and the fault features may include data such as temperature, current, and voltage. Assume the temperature is 80 °C, the current is 10 A, and the voltage is 220 V.

[0024] Next, input the extracted fault type, fault level, and fault feature data into a multi-level adaptive tensor decomposition network. This network processes the fault data using a method similar to high-order singular value decomposition. It can be understood as decomposing the original data into multiple core components and removing redundant information, thereby realizing data dimensionality reduction and reconstruction. For example, decompose and reconstruct the temperature, current, and voltage data of driver overheating to generate a more concise and representative fault feature tensor, such as [0.8, 0.5, 0.2].

[0025] Then, the generated fault feature tensor is input into a fuzzy analytic hierarchy process network with a dynamic weight assignment mechanism. This network dynamically adjusts its weights according to the importance of fault features. For example, for drive overheating, the temperature feature may be more important than the current and voltage features, so the network assigns a higher weight to the temperature. Finally, the network outputs a comprehensive fault scoring matrix, such as [0.7], representing the comprehensive score of drive overheating.

[0026] Meanwhile, the real-time operation data of the device is collected. These data can include various sensor data, such as temperature, pressure, vibration, etc. Suppose the real-time operation data of the motor system collected includes: motor speed 1500 rpm, vibration amplitude 0.5 mm, and ambient temperature 25 °C.

[0027] The collected real-time operation data of the device and the comprehensive fault scoring matrix are input into a multi-scale variational mode decomposition algorithm. This algorithm adaptively adjusts the decomposition parameters according to the comprehensive fault scoring matrix and decomposes the real-time operation data of the device into multiple intrinsic mode components. It can be understood as decomposing complex operation data into multiple simple signals with different characteristics. For example, the motor speed, vibration amplitude, and ambient temperature are decomposed into three intrinsic mode components: [0.9, 0.1, 0.2], [0.3, 0.8, 0.1], [0.2, 0.2, 0.9].

[0028] The decomposed intrinsic mode components are input into a multi-channel time-frequency analysis network with an attention mechanism. This network contains three channels: time-frequency energy channel, instantaneous phase channel, and modal correlation channel. Each channel extracts different aspects of features. For example, the time-frequency energy channel extracts the energy distribution of the signal, the instantaneous phase channel extracts the phase change of the signal, and the modal correlation channel analyzes the relationship between different modes. Finally, the network generates a multi-dimensional device operation feature tensor, such as [0.7, 0.6, 0.5].

[0029] The fault feature tensor and the multi-dimensional device operation feature tensor are input into a deep coupled feature fusion model. This model contains a spatial graph convolutional layer, a temporal graph convolutional layer, and a variational inference layer. The spatial graph convolutional layer constructs a feature space association graph to describe the spatial relationship between different features; the temporal graph convolutional layer constructs a feature temporal association graph to describe the time dependence between different features; the variational inference layer generates a device state vector based on the dual association graph, such as [0.8].

[0030] Finally, the device status vector is input into a multi-layer bidirectional neural network with a recursive attention mechanism. This network generates a maintenance time series based on the device status vector, such as [10, 20, 30], indicating that maintenance is required in the next 10 days, 20 days, and 30 days. Based on the generated maintenance time series, a maintenance decision model is constructed. For example, when the predicted maintenance time is less than a certain threshold, a maintenance alert is issued.

[0031] This application can achieve: 1. Improve the accuracy of maintenance decisions: By integrating the hierarchical fault diagnosis results and the real-time operation data of the device, the device status can be evaluated more comprehensively, thus improving the accuracy of maintenance decisions.

[0032] 2. Reduce maintenance costs: By predicting the maintenance time, unnecessary preventive maintenance can be avoided, thus reducing maintenance costs.

[0033] 3. Improve device reliability: By performing maintenance in a timely manner, the occurrence of device failures can be prevented, thus improving device reliability.

[0034] In an alternative embodiment, constructing the maintenance decision model based on the maintenance time series includes: Input the maintenance time series into an adaptive hybrid quantum optimization system, which includes a quantum behavior particle swarm module, a quantum simulated annealing module, a quantum genetic algorithm module, and a quantum entanglement enhancement module. The quantum behavior particle swarm module uses a multi-particle entanglement coding method to perform quantum state coding on the maintenance time series to obtain a quantum coding sequence. The quantum simulated annealing module performs quantum tunneling state transfer on the quantum coding sequence to obtain a quantum transfer sequence. The quantum genetic algorithm module performs quantum chromosome crossover and quantum bit mutation operations on the quantum transfer sequence to obtain a quantum optimization sequence. The quantum entanglement enhancement module performs quantum entanglement state operations on the quantum optimization sequence to generate a multi-dimensional maintenance decision sequence; Input the multi-dimensional maintenance decision sequence into an adaptive hierarchical cognitive computing network, which includes a maintenance strategy generation layer, a maintenance effect prediction layer, a knowledge reasoning layer, and a parameter optimization layer. The maintenance strategy generation layer uses a hierarchical recursive structure with a memory enhancement mechanism to transform the multi-dimensional maintenance decision sequence into a maintenance action space. The maintenance effect prediction layer uses a multi-head attention mechanism and a causal inference mechanism to predict the maintenance action space to obtain execution effect data. The knowledge reasoning layer inputs the maintenance action space and the execution effect data into a graph neural network to construct a maintenance knowledge graph. The parameter optimization layer inputs the maintenance action space, the execution effect data, and the maintenance knowledge graph into a meta-learning network to update the network parameters of the deep coupled feature fusion model; Construct an adaptive hybrid model based on the maintenance action space, the execution effect data, the maintenance knowledge graph, and the updated network parameters. The adaptive hybrid model includes an action mapping engine, a multimodal fusion engine, and a policy evaluation engine. The action mapping engine uses a hierarchical attention mechanism to map the maintenance action space into a multi-level maintenance policy tensor. The multimodal fusion engine fuses the features of the multi-level maintenance policy tensor, the execution effect data, and the maintenance knowledge graph to obtain a fused feature vector. The policy evaluation engine inputs the fused feature vector into a deep belief network to construct a maintenance decision model.

[0035] To construct a more intelligent maintenance decision model, this embodiment proposes a method for constructing an adaptive maintenance decision model based on maintenance time series. This method utilizes the advantages of quantum computing and cognitive computing to deeply analyze and process the maintenance time series, and finally generates more accurate and effective maintenance decisions.

[0036] First, preprocess the collected maintenance time series. For example, for the maintenance time series data of a certain device, including information such as the time of each maintenance, the maintenance type, the failure type, and the replaced parts. Suppose the maintenance records of this device for one year are collected, including data on 12 preventive maintenances and 3 breakdown repairs. Clean these data, remove invalid data and outliers, and perform normalization processing, convert the time into a time interval, and convert the maintenance type and failure type into digital codes.

[0037] Next, input the preprocessed maintenance time series into an adaptive hybrid quantum optimization system. This system includes a quantum-behaved particle swarm module, a quantum simulated annealing module, a quantum genetic algorithm module, and a quantum entanglement enhancement module. The quantum-behaved particle swarm module encodes the maintenance time series into a quantum state using a multi-particle entanglement encoding method. For example, encode information such as the time interval, maintenance type, and failure type of each maintenance into a quantum bit string. The quantum simulated annealing module uses the quantum tunneling effect to transfer the quantum state in different solution spaces to search for the optimal solution. The quantum genetic algorithm module performs crossover and mutation operations on the quantum state to simulate the biological evolution process and further optimize the solution space. The quantum entanglement enhancement module utilizes the quantum entanglement property to enhance the correlation between quantum states and improve the search efficiency. Through the collaborative work of these four modules, a multi-dimensional maintenance decision sequence is generated. For example, a probability distribution of different maintenance strategies for a future period of time can be generated.

[0038] Then, input the generated multi-dimensional maintenance decision sequence into the adaptive hierarchical cognitive computing network. This network includes a maintenance strategy generation layer, a maintenance effect prediction layer, a knowledge reasoning layer, and a parameter optimization layer. The maintenance strategy generation layer uses a hierarchical recursive structure with a memory enhancement mechanism to transform the multi-dimensional maintenance decision sequence into a maintenance action space. For example, it transforms a probability distribution into a specific maintenance strategy, such as performing a certain type of maintenance at a certain time point. The maintenance effect prediction layer uses a multi-head attention mechanism and a causal inference mechanism to predict the maintenance action space and obtain execution effect data. For example, it predicts the impact of different maintenance strategies on equipment life, performance, and cost. The knowledge reasoning layer inputs the maintenance action space and execution effect data into a graph neural network to construct a maintenance knowledge graph. For example, it represents the relationships between different maintenance strategies, execution effects, and equipment states as nodes and edges in the graph. The parameter optimization layer inputs the maintenance action space, execution effect data, and maintenance knowledge graph into a meta-learning network to update the network parameters of the deep coupled feature fusion model, enabling the model to better adapt to new data and environments.

[0039] Finally, construct an adaptive hybrid model based on the maintenance action space, execution effect data, maintenance knowledge graph, and updated network parameters. This model includes an action mapping engine, a multi-modal fusion engine, and a policy evaluation engine. The action mapping engine uses a hierarchical attention mechanism to map the maintenance action space into a multi-level maintenance strategy tensor. For example, it organizes different maintenance strategies according to time and type. The multi-modal fusion engine performs feature fusion on the multi-level maintenance strategy tensor, execution effect data, and maintenance knowledge graph to obtain a fused feature vector. For example, it integrates information from different data sources into a single vector. The policy evaluation engine inputs the fused feature vector into a deep belief network to construct a maintenance decision model and finally output the optimal maintenance strategy. For example, based on the current state of the equipment and future predictions, it decides when to perform what type of maintenance.

[0040] This application can achieve: 1. Improve the accuracy of maintenance decisions: Through the deep integration of quantum computing and cognitive computing, it can more comprehensively analyze maintenance time series data, mine hidden patterns, and thus generate more accurate maintenance strategies. For example, through quantum computing, it can more effectively search for the optimal solution, and through cognitive computing, it can better understand the operating state and maintenance requirements of the equipment.

[0041] 2. Enhance the adaptability of maintenance decisions: The adaptive hybrid model constructed by this method can dynamically adjust parameters according to new data and environments, thus better adapting to different equipment and scenarios. For example, when the operating state of the equipment changes, the model can automatically adjust the maintenance strategy to ensure the normal operation of the equipment.

[0042] 3. Optimize the configuration of maintenance resources: By predicting the execution effects of different maintenance strategies, the configuration of maintenance resources can be optimized, maintenance costs can be reduced, and the utilization rate of equipment can be improved. For example, by predicting the failure probability of equipment, unnecessary preventive maintenance can be avoided, thus reducing maintenance costs.

[0043] In an alternative embodiment, the maintenance decision-making model is input into a quantum-enhanced dual-network system, and a sequence of reward values is calculated, including: Input the maintenance decision-making model into the quantum-enhanced dual-network system. The quantum-enhanced dual-network system includes a quantum policy network module and a hierarchical evaluation network module. The quantum policy network module uses a multi-particle quantum entanglement state encoder to map the maintenance decision-making model to a high-dimensional quantum feature space to obtain a sequence of quantum states. The hierarchical evaluation network module uses a dynamic graph neural network and a cross-scale multi-head attention mechanism to process the sequence of quantum states to obtain a spatio-temporal multi-dimensional value tensor. Input the spatio-temporal multi-dimensional value tensor into a recursive objective decomposition system. The recursive objective decomposition system includes an adaptive neural tensor decomposition module and a multi-scale time-series feature extraction module. The adaptive neural tensor decomposition module uses a tensor kernel decomposition method to recursively decompose the spatio-temporal multi-dimensional value tensor to obtain a device reliability dimension tensor, a maintenance cost dimension tensor, and a resource utilization rate dimension tensor. The multi-scale time-series feature extraction module uses a hierarchical deep time-series model to process the dimension tensors in sequence to obtain a set of maintenance decision-making features. Among them, failure evolution features and life prediction features are extracted from the device reliability dimension tensor, cost dynamic features and resource consumption features are extracted from the maintenance cost dimension tensor, and efficiency optimization features and load balancing features are extracted from the resource utilization rate dimension tensor. Calculate the dynamic weight coefficient matrix between the features in the set of maintenance decision-making features based on an improved non-dominated sorting genetic algorithm and an adaptive fuzzy analytic hierarchy process method. The set of maintenance decision-making features and the dynamic weight coefficient matrix are input into a hybrid state evaluation system. The hybrid state evaluation system includes a conditional variational encoding module and a probability inference module. The conditional variational encoding module uses a conditional variational autoencoder to map the features to a non-linear state space to obtain a multi-modal state representation vector. The probability inference module uses an improved deep belief network to calculate a time-varying state transition probability tensor based on the multi-modal state representation vector and the dynamic weight coefficient matrix. Input the time-varying state transition probability tensor into a two-way evaluation system to calculate a sequence of reward values.

[0044] A device maintenance decision-making method based on quantum-enhanced dual networks and hybrid state evaluation aims to improve the efficiency and accuracy of device maintenance decision-making and reduce maintenance costs. The core of this method is to utilize the parallelism and high-speed processing capabilities of quantum computing, combined with the feature extraction and state evaluation capabilities of deep learning, to achieve precise evaluation of device states and formulation of optimal maintenance strategies.

[0045] First, construct a maintenance decision-making model. This model can preliminarily formulate a series of alternative maintenance strategies based on information such as the operating state of the device, historical maintenance records, and environmental factors. For example, for a wind turbine, the maintenance strategies can include regular inspections, condition monitoring, and fault prediction.

[0046] Then, input the maintenance decision-making model into the quantum-enhanced dual network system. This system includes a quantum policy network module and a hierarchical evaluation network module. The quantum policy network module uses a multi-particle quantum entanglement state encoder to map the maintenance decision-making model to a high-dimensional quantum feature space, obtaining a quantum state sequence. For example, different maintenance strategies are encoded into different quantum states, and using the characteristics of quantum superposition states, multiple strategies can be evaluated simultaneously. The hierarchical evaluation network module uses a dynamic graph neural network and a cross-scale multi-head attention mechanism to process the quantum state sequence, obtaining a spatio-temporal multi-dimensional value tensor. For example, the dynamic graph neural network can capture the dynamic changes of the device state over time, and the cross-scale multi-head attention mechanism can focus on key features at different time scales, thus more comprehensively evaluating the value of each maintenance strategy. Suppose there are three maintenance strategies, and after being processed by the quantum-enhanced dual network, a value tensor containing three dimensions of device reliability, maintenance cost, and resource utilization rate is obtained.

[0047] Next, input the spatio-temporal multi-dimensional value tensor into the recursive objective decomposition system. This system includes an adaptive neural tensor decomposition module and a multi-scale temporal feature extraction module. The adaptive neural tensor decomposition module uses a tensor kernel decomposition method to recursively decompose the spatio-temporal multi-dimensional value tensor, obtaining a device reliability dimension tensor, a maintenance cost dimension tensor, and a resource utilization rate dimension tensor. For example, the value tensor is decomposed into three independent tensors, respectively representing the values in different aspects. The multi-scale temporal feature extraction module uses a hierarchical deep temporal model to process these dimension tensors in turn, obtaining a set of maintenance decision-making features. For example, fault evolution features and remaining life prediction features are extracted from the device reliability dimension tensor; cost dynamic features and resource consumption features are extracted from the maintenance cost dimension tensor; efficiency optimization features and load balancing features are extracted from the resource utilization rate dimension tensor. Suppose the fault probability is extracted as 0.1 and the remaining life is 5 years from the device reliability dimension tensor; the annual maintenance cost is 10,000 yuan and the resource consumption is 100 units are extracted from the maintenance cost dimension tensor; the efficiency improvement is 20% and the load balancing degree is 90% are extracted from the resource utilization rate dimension tensor.

[0048] Subsequently, a dynamic weight coefficient matrix among the features in the maintenance decision feature set is calculated based on the improved non-dominated sorting genetic algorithm and the adaptive fuzzy analytic hierarchy process. For example, according to the actual operation condition of the device and the maintenance objective, the weights of different features are dynamically adjusted. Suppose the weight of the failure probability is 0.3, the weight of the remaining life is 0.2, the weight of the annual maintenance cost is 0.2, the weight of the resource consumption is 0.1, the weight of the efficiency improvement is 0.1, and the weight of the load balance degree is 0.1.

[0049] The maintenance decision feature set and the dynamic weight coefficient matrix are input into the hybrid state evaluation system. This system includes a conditional variational encoding module and a probabilistic inference module. The conditional variational encoding module uses a conditional variational autoencoder to map the features into a non-linear state space to obtain a multi-modal state representation vector. The probabilistic inference module uses an improved deep belief network to calculate a time-varying state transition probability tensor based on the multi-modal state representation vector and the dynamic weight coefficient matrix. For example, predict the state change probability of the device in the future period under different maintenance strategies.

[0050] Finally, the time-varying state transition probability tensor is input into the two-way evaluation system to calculate a reward value sequence. For example, according to indicators such as the long-term operation cost, reliability, and safety of the device, each maintenance strategy is comprehensively evaluated to obtain the final reward value.

[0051] This application can achieve: First, the accuracy of maintenance decision-making is improved. Through the quantum-enhanced dual network and the hybrid state evaluation system, the device state can be more accurately evaluated and the future development trend can be predicted, so as to formulate a more reasonable maintenance strategy.

[0052] Second, the maintenance cost is reduced. By optimizing the maintenance strategy, unnecessary maintenance work can be reduced, and the maintenance cost and resource consumption can be lowered.

[0053] Third, the reliability and safety of the device are improved. Through timely maintenance and preventive measures, the reliability and safety of the device can be improved, and losses caused by device failures can be avoided.

[0054] In an alternative embodiment, the inputting the time-varying state transition probability tensor into the two-way evaluation system to calculate a reward value sequence includes: Input the time-varying state transition probability tensor into a two-way evaluation system, which includes a hierarchical action evaluation module and a state evaluation module. The hierarchical action evaluation module uses an improved bidirectional long short-term memory network and a cross-dimensional attention mechanism to evaluate and maintain actions to obtain multi-level action influence vectors. The state evaluation module uses a spatio-temporal graph convolutional network to evaluate state transitions to obtain state influence vectors. The multi-level action influence vectors and the state influence vectors are fused through an adaptive residual connection structure and a cross-modal feature fusion network to obtain a multi-dimensional reward vector; Construct the meta-learning module, which includes a hierarchical gradient policy network, a dynamic programming network, and a conditional variational network. The hierarchical gradient policy network uses a hierarchical structure to calculate action gradients and state gradients. The dynamic programming network calculates the discount factor based on the Bellman equation. The conditional variational network uses an encoder-decoder structure to randomly sample state transitions to generate a conditional variational state value function. The multi-dimensional reward vector, the discount factor, and the conditional variational state value function are mapped to a unified feature space through a non-linear transformation to obtain a feature mapping matrix. The feature mapping matrix is decomposed to obtain a reward feature and a discount feature. Calculate the tensor product of the reward feature and the discount feature to obtain a reward-discount tensor, and then perform a tensor contraction operation on the reward-discount tensor and the conditional variational state value function to obtain an accumulated reward tensor; Input the accumulated reward tensor into an adversarial training network. The adversarial training network extracts features through a discriminator to obtain a discriminant score, and the generator generates an optimized reward sample based on the discriminant score. Use the discriminant score and the optimized reward sample to construct a hierarchical knowledge distillation network, and transfer the state transition features and action execution features in the pre-trained teacher network to the student network to obtain a knowledge transfer error. Based on the discriminant score, the optimized reward sample, and the knowledge transfer error, construct an adaptive loss function, and use a distributed asynchronous parallel algorithm to optimize the network parameters to generate a reward value sequence containing temporal dependence and state transition relationships.

[0055] To effectively evaluate the time-varying state transition probability tensor and generate a reward value sequence containing temporal dependence and state transition relationships, a method for generating a reward value sequence based on a two-way evaluation system and a meta-learning module is proposed.

[0056] First, preprocess the input time-varying state transition probability tensor. Assume the dimension of the state transition probability tensor is [time step, number of states, number of actions, number of states], where each element represents the probability of transitioning from a certain state to another state by performing a certain action at a specific time step. For example, a tensor with dimension [10, 5, 3, 5] represents a system with 10 time steps, 5 states, and 3 actions, recording the state transition probabilities at each time step. The preprocessing process includes data cleaning, such as removing noise and outliers, and data normalization, scaling all probability values in the tensor to between 0 and 1 to ensure the stability of subsequent calculations.

[0057] Next, input the preprocessed time-varying state transition probability tensor into a bidirectional evaluation system. This system consists of a hierarchical action evaluation module and a state evaluation module. The hierarchical action evaluation module uses an improved bidirectional long short-term memory network and a cross-dimensional attention mechanism to evaluate the impacts of actions at different levels and generate multi-level action impact vectors. For example, in a robot control scenario, a high-level action could be "grasp an object", and a low-level action could be "move the robotic arm". The cross-dimensional attention mechanism can capture the mutual influences between actions at different levels. The state evaluation module uses a spatio-temporal graph convolutional network to evaluate state transitions and generate state impact vectors. The spatio-temporal graph convolutional network can capture the dependencies of states in the time and space dimensions. Then, the multi-level action impact vectors and state impact vectors are fused through an adaptive residual connection structure and a cross-modal feature fusion network to obtain a multi-dimensional reward vector.

[0058] Subsequently, construct a meta-learning module. This module consists of a hierarchical gradient policy network, a dynamic programming network, and a conditional variational network. The hierarchical gradient policy network uses a hierarchical structure to calculate action gradients and state gradients for guiding policy optimization. The dynamic programming network calculates a discount factor based on the Bellman equation to balance the importance of current rewards and future rewards. The conditional variational network uses an encoder-decoder structure to perform random sampling on state transitions and generate a conditional variational state value function. Through a non-linear transformation, such as a multi-layer perceptron, the multi-dimensional reward vector, the discount factor, and the conditional variational state value function are mapped to a unified feature space to obtain a feature mapping matrix. Decompose the feature mapping matrix, such as performing singular value decomposition, to obtain reward features and discount features. Calculate the tensor product of the reward features and the discount features to obtain a reward-discount tensor, and then perform a tensor contraction operation, such as a tensor dot product, on the reward-discount tensor and the conditional variational state value function to obtain an accumulated reward tensor.

[0059] Finally, the cumulative reward tensor is input into the adversarial training network. This network consists of a discriminator and a generator. The discriminator extracts the features of the cumulative reward tensor and gives a discrimination score to evaluate the quality of the reward. The generator generates optimized reward samples based on the discrimination score to improve the reward generation strategy. Using the discrimination score and the optimized reward samples, a hierarchical knowledge distillation network is constructed to transfer the state transition features and action execution features in the pre-trained teacher network to the student network, and the knowledge transfer error is calculated to accelerate the training process. An adaptive loss function is constructed based on the discrimination score, the optimized reward samples, and the knowledge transfer error, and a distributed asynchronous parallel algorithm is used to optimize the network parameters, and finally a sequence of reward values including temporal dependence and state transition relationships is generated.

[0060] This application can achieve: 1. Improve the quality of the reward signal: Through the bidirectional evaluation system and the meta-learning module, the impacts of state transitions and actions can be evaluated more accurately, generating more informative reward signals, thereby improving the performance of the reinforcement learning algorithm.

[0061] 2. Accelerate model training: The meta-learning module can learn the general knowledge of the task, thus accelerating the learning speed of new tasks. The adversarial training and knowledge distillation techniques can further improve the training efficiency.

[0062] 3. Enhance the generalization ability of the model: By modeling the state transition probability tensor, the model can better understand the dynamic characteristics of the environment, thereby enhancing the generalization ability of the model and enabling it to adapt to different environments and tasks.

[0063] In an optional implementation manner, inputting the state transition matrix into the maintenance policy network and processing the output of the maintenance policy network based on the quantum evolutionary algorithm to generate a multi-constraint maintenance resource allocation scheme includes: Input the state transition matrix into the maintenance policy network. The maintenance policy network includes a state encoding layer, a policy mapping layer, and an action generation layer. The state encoding layer uses a multi-head self-attention module to calculate the correlation weights between states for the state transition matrix. The multi-head self-attention module includes a query matrix, a key matrix, and a value matrix. The attention score is obtained through the dot product operation of the query matrix and the key matrix, and the weighted feature is obtained by multiplying the attention score by the value matrix. At the same time, position encoding is introduced to retain the temporal information of the state transition. The weighted feature, the correlation weights between states, and the position encoding are fused to obtain a state feature vector; Based on the state feature vector and the association weights between the states, a value evaluation branch and a policy generation branch are constructed in the policy mapping layer. The value evaluation branch uses a fully connected layer with residual connections to calculate the state value function of the maintenance action. The residual connection includes a main channel and a bypass channel. The main channel performs a non-linear transformation on the state feature vector and the association weights between the states. The bypass channel keeps the state feature vector and the association weights between the states unchanged and superimposes them with the output of the main channel to obtain the state value function. The policy generation branch includes a policy network and a value network. The policy network outputs the probability distribution of the maintenance action based on the state feature vector and the association weights between the states. The value network evaluates the expected return of the maintenance action based on the state feature vector and the association weights between the states. The policy network is optimized by maximizing the entropy-regularized expected return to obtain the optimized probability distribution of the maintenance action. In the action generation layer, based on the association weights between the states, the optimized probability distribution of the maintenance action is converted into discrete maintenance policy parameters by using the reparameterization sampling technique. Specifically, a temperature parameter is introduced into the optimized probability distribution of the maintenance action to control the randomness of sampling, and discrete action samples are generated based on the temperature parameter and the association weights between the states to obtain the maintenance policy parameters, which are used as the output of the maintenance policy network. The maintenance policy parameters are processed based on the quantum evolutionary algorithm to generate a multi-constraint maintenance resource allocation scheme.

[0064] A multi-constraint maintenance resource allocation method based on the quantum evolutionary algorithm, the core of which is to use a maintenance policy network to learn the device state transition law and combine the quantum evolutionary algorithm to generate a maintenance resource allocation scheme that meets multiple constraint conditions. The specific implementation steps of this method are as follows: First, construct a device state transition matrix. This matrix describes the probability of the device transitioning between different states. For example, a device may have three states: normal, sub-healthy, and faulty. The state transition matrix is a 3x3 matrix, where each element represents the probability of the device transitioning from one state to another. For example, the element in the first row and second column of the matrix represents the probability of the device transitioning from the normal state to the sub-healthy state. Then, input the state transition matrix into the maintenance policy network. The maintenance policy network consists of a state encoding layer, a policy mapping layer, and an action generation layer.

[0065] In the state encoding layer, the multi-head self-attention mechanism is adopted to calculate the correlation weights between states. The multi-head self-attention mechanism includes a query matrix, a key matrix, and a value matrix. The attention scores are obtained by calculating the dot product of the query matrix and the key matrix, and then the weighted features are obtained by multiplying the attention scores with the value matrix. At the same time, position encoding is introduced to preserve the temporal information of state transitions. Finally, the weighted features, the correlation weights between states, and the position encoding are fused to obtain the state feature vector. For example, assuming that each state in the state transition matrix is represented by a three-dimensional vector, the position encoding can be a vector with the same dimension as the state vector, which is used to represent the position of the state in the sequence.

[0066] Next, in the policy mapping layer, a value evaluation branch and a policy generation branch are constructed based on the state feature vector and the correlation weights between states. The value evaluation branch uses a fully connected layer with residual connections to calculate the state value function of the maintenance action. The residual connection includes a main channel and a bypass channel. The main channel performs a non-linear transformation on the state feature vector and the correlation weights between states, while the bypass channel keeps the state feature vector and the correlation weights between states unchanged and superimposes the output of the main channel with the output of the bypass channel to obtain the state value function. The policy generation branch includes a policy network and a value network. The policy network outputs the probability distribution of the maintenance action based on the state feature vector and the correlation weights between states. The value network evaluates the expected return of the maintenance action based on the state feature vector and the correlation weights between states. The policy network is optimized by maximizing the entropy-regularized expected return to obtain the optimized probability distribution of the maintenance action.

[0067] In the action generation layer, based on the correlation weights between states, the reparameterization sampling technique is adopted to convert the optimized probability distribution of the maintenance action into discrete maintenance policy parameters. Specifically, the randomness of sampling is controlled by introducing a temperature parameter into the optimized probability distribution of the maintenance action, and discrete action samples are generated based on the temperature parameter and the correlation weights between states to obtain the maintenance policy parameters, which are used as the output of the maintenance policy network. For example, assuming that the maintenance actions are "repair" and "replace", the maintenance policy parameters can be a value representing the probability of selecting the "repair" action.

[0068] Finally, the maintenance policy parameters are input into the quantum evolutionary algorithm for processing to generate a multi-constrained maintenance resource allocation scheme. The quantum evolutionary algorithm uses the characteristics of the superposition state and entanglement state of quantum bits for search and optimization, and can effectively handle multi-constrained optimization problems. For example, constraint conditions such as maintenance cost, time, and reliability can be incorporated into the fitness function of the quantum evolutionary algorithm, so as to generate a maintenance resource allocation scheme that meets multiple constraint conditions.

[0069] This application can achieve: 1. Improve maintenance efficiency: By learning the law of equipment state transition and optimizing maintenance strategies, maintenance resources can be allocated more effectively, thus improving maintenance efficiency.

[0070] 2. Reduce maintenance costs: By optimizing maintenance strategies, unnecessary maintenance operations can be reduced, thus reducing maintenance costs.

[0071] 3. Enhance equipment reliability: By more effective maintenance strategies, equipment failures can be detected and handled in a timely manner, thus enhancing equipment reliability.

[0072] In an optional implementation manner, processing the maintenance strategy parameters based on the quantum evolutionary algorithm to generate a multi-constraint maintenance resource allocation scheme includes: Encoding the maintenance strategy parameters and the correlation weights between states into a quantum bit string; inputting the state value function and the correlation weights between states into an adaptive angle calculation module, the adaptive angle calculation module calculates the difference between the current solution corresponding to the maintenance strategy parameters and the historical optimal solution by using a value difference function, sets an adaptive coefficient based on the difference according to a preset proportional coefficient, and multiplies the adaptive coefficient by a preset reference rotation angle to obtain the actual rotation angle of the quantum rotation gate; driving the quantum rotation gate with the actual rotation angle to adjust the phase of the quantum bit string, and constructing a policy gradient vector based on the state value function and the correlation weights between states, designing a quantum interference operation matrix by using the policy gradient vector, and adjusting the local phase of the quantum bit string through the quantum interference operation matrix to achieve population-directed evolution and obtain the quantum bit string after the first evolution; Taking the state value function as an evaluation index, and at the same time taking the correlation weights between states as weight factors, performing weighted calculation on each quantum bit in the quantum bit string after the first evolution to obtain a weighted fitness score; calculating the resource overrun degree, time conflict degree, and reliability deficiency degree in the maintenance plan corresponding to the quantum bit string after the first evolution, where the resource overrun degree is the difference between the actual required maintenance resource amount and the preset resource capacity, the time conflict degree is the duration by which the actual execution time of the maintenance activity exceeds the preset time window, and the reliability deficiency degree is the difference between the actual reliability of the equipment after maintenance and the preset reliability target; calculating a penalty term based on the resource overrun degree, the time conflict degree, and the reliability deficiency degree; performing weighted combination on the weighted fitness score and the penalty term according to a preset ratio to obtain a comprehensive fitness evaluation result, and screening the quantum bit string after the first evolution based on the comprehensive fitness evaluation result to obtain the quantum bit string after the second evolution; Calculate the variance of the values of each qubit in the second evolved qubit string to obtain the gene diversity index, and perform a weighted combination of the gene diversity index and the correlation weight between states according to a preset weight to obtain the diversity evaluation index; when the diversity evaluation index is lower than the preset lower threshold, multiply the current quantum mutation probability by a preset growth rate greater than 1 to obtain a new quantum mutation probability; when the diversity evaluation index is higher than the preset upper threshold, multiply the current quantum mutation probability by a preset decay rate less than 1 to obtain a new quantum mutation probability; perform a flipping operation on the qubits in the second evolved qubit string using the new quantum mutation probability to achieve quantum evolution, decode the evolved qubit string to obtain multiple maintenance plans, and select the maintenance plan with the highest fitness that simultaneously meets the resource capacity limit, time window constraint, and reliability requirement from the multiple maintenance plans as the multi-constraint maintenance resource allocation plan.

[0073] A multi-constraint maintenance resource allocation method based on a quantum evolutionary algorithm is used to generate maintenance plans that meet resource capacity limits, time window constraints, and reliability requirements. This method encodes maintenance strategy parameters, state value functions, and correlation weights between states into qubit strings, and uses the parallelism and probability of quantum computing, combined with adaptive parameter adjustment and quantum mutation strategies, to efficiently search for optimal maintenance plans.

[0074] First, encode the maintenance strategy parameters. Convert each maintenance strategy parameter, such as maintenance time, number of maintenance personnel, spare part replacement type, etc., and their correlation weights with the device state into a string of qubits. The state of each qubit represents the value or weight of a parameter. For example, the "0" state of a qubit represents replacing spare part type A, and the "1" state represents replacing spare part type B. The correlation weight is represented by the superposition state of the qubit, and the greater the weight, the higher the probability that the qubit is in the corresponding state.

[0075] Next, perform adaptive angle calculation. Input the state value functions of the device, such as reliability, availability, performance indicators, etc., and the correlation weights between states into the adaptive angle calculation module. This module calculates the value difference between the solution corresponding to the current maintenance strategy parameters and the historical optimal solution. According to this difference, adjust the adaptive coefficient according to a preset proportional coefficient. Multiply the adaptive coefficient by a preset reference rotation angle to obtain the actual rotation angle of the quantum rotation gate. For example, if the current solution is much worse than the historical optimal solution, the adaptive coefficient will increase, resulting in a larger rotation angle, thereby increasing the search space.

[0076] Then, the first quantum evolution is carried out. The actually calculated rotation angle is used to drive the quantum rotation gate to adjust the phase of the encoded qubit string. At the same time, based on the state value function and the correlation weight between states, a policy gradient vector is constructed. This vector indicates the direction in which the qubit string should evolve to improve the quality of the solution. According to the policy gradient vector, a quantum interference operation matrix is designed to adjust the local phase of the qubit string, realizing the directional evolution of the population.

[0077] The evolved qubit string is evaluated. The state value function is used as the evaluation index, and the correlation weight between states is used as the weight factor to perform weighted calculations on each qubit to obtain the weighted fitness score. At the same time, the resource overrun degree, time conflict degree, and reliability deficiency degree of the maintenance plan are calculated. For example, if a certain plan requires more resources than the preset resource capacity, a resource overrun penalty will be generated. The weighted fitness score and the penalty term are weighted and combined according to a preset ratio to obtain the comprehensive fitness evaluation result. Based on the comprehensive fitness evaluation result, the qubit strings with higher fitness are selected to complete the second evolution.

[0078] Subsequently, diversity evaluation and quantum mutation are carried out. The variance of the values of each qubit in the qubit string after the second evolution is calculated to obtain the gene diversity index. The gene diversity index and the correlation weight between states are weighted and combined to obtain the diversity evaluation index. If the diversity evaluation index is lower than the preset lower threshold, the quantum mutation probability is increased; if it is higher than the upper threshold, the quantum mutation probability is decreased. The qubit string is flipped using the new quantum mutation probability to realize quantum evolution.

[0079] Finally, decoding and scheme selection are carried out. The evolved qubit string is decoded to obtain multiple maintenance plans. For example, the "0" or "1" state of each qubit is translated back to the corresponding maintenance strategy parameter value. The plan with the highest fitness and that simultaneously meets the resource capacity limit, time window constraint, and reliability requirement is selected as the final multi-constraint maintenance resource allocation plan.

[0080] For example, assume there are two maintenance strategy parameters: maintenance time and spare part type. There are two options for maintenance time: morning and afternoon; there are two options for spare part type: A and B. Encode these two parameters into two qubits. Assume the current solution is "morning, spare part type A" and the historical optimal solution is "afternoon, spare part type B". By calculating the value difference, an adaptive rotation angle can be obtained, such as 30 degrees. Then, perform rotation and interference operations on the qubit string. Assume that after the first evolution, four qubit strings are obtained, representing four solutions: "morning, A", "morning, B", "afternoon, A", and "afternoon, B". Evaluate these four solutions. Assume that the comprehensive fitness of the "afternoon, B" solution is the highest. Then, calculate the gene diversity index and adjust the quantum mutation probability according to the diversity evaluation index. Finally, perform mutation and decoding on the qubit string, and ultimately select the "afternoon, B" solution as the optimal solution.

[0081] This application can achieve: 1. Improve the solution efficiency: Utilizing the parallelism of quantum computing, multiple maintenance solutions can be evaluated simultaneously, thereby accelerating the search speed. Especially when facing large-scale complex problems, the advantage is more significant.

[0082] 2. Enhance the solution quality: Through adaptive parameter adjustment and quantum mutation strategies, the solution space can be better explored, jumping out of local optimal solutions to find better maintenance solutions, thereby improving the reliability and availability of the equipment.

[0083] 3. Strengthen the feasibility of the solution: This method considers multiple constraint conditions such as resource capacity limitations, time window constraints, and reliability requirements during the solution process, ensuring the feasibility of the final solution and enabling it to be directly applied to actual maintenance work.

[0084] Figure 2 It is a schematic structural diagram of an abnormal detection and maintenance decision optimization system for an automated test device according to an embodiment of the present invention. As Figure 2 shown, the system includes: A first unit for deploying an edge computing unit and a multi-level sensor network; synchronously processing multi-source sensing data through the edge computing unit using a self-organizing time synchronization protocol; performing modal separation on the synchronized sensing data using an empirical mode decomposition algorithm and a variational mode decomposition technique; constructing a deep network model including a graph neural network module and a recurrent variational autoencoder module, and the deep network model processes different sensor data through a cross-modal attention mechanism; matching the output features of the deep network model with the device historical fault case database to generate a device state feature vector; establishing a knowledge graph based on the device state feature vector to obtain a hierarchical fault diagnosis result; A second unit, configured to receive the hierarchical fault diagnosis result, construct a maintenance decision model by combining device operation data; input the maintenance decision model into a quantum-enhanced dual-network system to calculate a sequence of reward values; construct a dynamic probability graph model, input the sequence of reward values into the dynamic probability graph model, and combine the dynamic probability graph model with a Bayesian network and a hidden Markov model to generate a state transition matrix; train a maintenance policy network using a federated reinforcement learning framework; input the state transition matrix into the maintenance policy network, and input the output of the maintenance policy network into a quantum evolutionary algorithm to generate a multi-constraint maintenance resource allocation scheme; A third unit, configured to input the multi-constraint maintenance resource allocation scheme into a multi-criteria evaluation model based on interval intuitionistic fuzzy sets and improved grey relational degrees to output evaluation result data; input the evaluation result data into an adaptive multi-objective differential evolution algorithm to generate optimized maintenance scheme data; input the optimized maintenance scheme data into a high-fidelity digital twin model to output integrated model data; input the integrated model data into a multi-task deep transfer learning network and a meta-learning network to generate maintenance decision model update data; transmit the maintenance decision model update data to a field control system through an industrial Internet of Things platform; record the maintenance decision model update data using a blockchain network, and the data of the blockchain network is used as the input of the device historical fault case database to update the deep network model.

[0085] The present invention may be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for performing various aspects of the present invention.

[0086] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Anomaly detection and maintenance decision optimization method for automated test equipment, characterized in that: include: Deploy edge computing units and multi-layer sensor networks; Synchronously processing multi-source sensor data using a self-organizing time synchronization protocol through the edge computing unit; The synchronized sensor data is modally separated using the empirical mode decomposition algorithm and the variational mode decomposition technique; a deep network model including a graph neural network module and a recursive variational autoencoder module is constructed, and the deep network model processes different sensor data through a cross-modal attention mechanism; the output features of the deep network model are matched with the equipment historical fault case database to generate an equipment state feature vector; a knowledge graph is established based on the equipment state feature vector to obtain a hierarchical fault diagnosis result; Receiving the hierarchical fault diagnosis results, and building a maintenance decision model in combination with the equipment operation data; inputting the maintenance decision model into the quantum enhanced dual network system, and calculating a reward value sequence; Constructing a dynamic probability graph model, inputting the reward value sequence into the dynamic probability graph model, and combining the dynamic probability graph model with a Bayesian network and a hidden Markov model to generate a state transfer matrix; A maintenance strategy network is trained using a federated reinforcement learning framework; the state transfer matrix is ​​input into the maintenance strategy network, and the output of the maintenance strategy network is input into a quantum evolutionary algorithm to generate a multi-constraint maintenance resource configuration plan; Inputting the multi-constraint maintenance resource allocation scheme into a multi-criteria evaluation model based on interval intuitionistic fuzzy sets and improved grey relational degree, and outputting evaluation result data; Input the evaluation result data into an adaptive multi-objective differential evolution algorithm to generate optimized maintenance plan data; input the optimized maintenance plan data into a high-fidelity digital twin model to output integrated model data; input the integrated model data into a multi-task deep transfer learning network and a meta-learning network to generate maintenance decision model update data; transmit the maintenance decision model update data to the field control system through the industrial Internet of Things platform; A blockchain network is used to record and maintain decision model update data, and the data of the blockchain network is used as input to the equipment historical failure case database to update the deep network model.

2. The method according to claim 1, characterized in that The receiving the hierarchical fault diagnosis result and building a maintenance decision model in combination with the equipment operation data includes: Receive hierarchical fault diagnosis results, extract fault type data, fault level data and fault feature data from the hierarchical fault diagnosis results; input the fault type data, the fault level data and the fault feature data into a multi-level adaptive tensor decomposition network, the multi-level adaptive tensor decomposition network uses a high-order singular value decomposition method to perform nonlinear dimensionality reduction and reconstruction on the fault data to generate a fault feature tensor; input the fault feature tensor into a fuzzy hierarchical analysis network with a dynamic weight allocation mechanism to generate a comprehensive fault scoring matrix; Collecting real-time operation data of the equipment, inputting the real-time operation data of the equipment and the comprehensive fault score matrix into a multi-scale variational modal decomposer, wherein the multi-scale variational modal decomposer adaptively adjusts decomposition parameters based on the comprehensive fault score matrix to decompose the real-time operation data of the equipment into multiple intrinsic modal components; inputting the intrinsic modal components into a multi-channel time-frequency analysis network with an attention mechanism, wherein the multi-channel time-frequency analysis network includes a time-frequency energy channel, an instantaneous phase channel, and a modal correlation channel to generate a multi-dimensional equipment operation feature tensor; The fault feature tensor and the multi-dimensional equipment operation feature tensor are input into a deep coupling feature fusion model, wherein the deep coupling feature fusion model comprises a spatial graph convolution layer, a temporal graph convolution layer and a variational inference layer, wherein the spatial graph convolution layer constructs a feature space association graph, the temporal graph convolution layer constructs a feature temporal association graph, and the variational inference layer generates a device state vector based on the dual association graph; the device state vector is input into a multi-layer bidirectional neural network with a recursive attention mechanism, and a maintenance time series is generated based on the device state vector; and a maintenance decision model is constructed based on the maintenance time series.

3. The method according to claim 2, characterized in that The constructing of a maintenance decision model based on the maintenance time series includes: Input the maintenance time sequence into an adaptive hybrid quantum optimization system, wherein the adaptive hybrid quantum optimization system comprises a quantum behavior particle swarm module, a quantum simulated annealing module, a quantum genetic algorithm module and a quantum entanglement enhancement module, wherein the quantum behavior particle swarm module uses a multi-particle entanglement encoding method to perform quantum state encoding on the maintenance time sequence to obtain a quantum coding sequence, the quantum simulated annealing module performs quantum tunneling state transfer on the quantum coding sequence to obtain a quantum transfer sequence, the quantum genetic algorithm module performs quantum chromosome crossover and quantum bit mutation operations on the quantum transfer sequence to obtain a quantum optimization sequence, and the quantum entanglement enhancement module performs quantum entanglement state operations on the quantum optimization sequence to generate a multidimensional maintenance decision sequence; Input the multidimensional maintenance decision sequence into an adaptive hierarchical cognitive computing network, the adaptive hierarchical cognitive computing network includes a maintenance strategy generation layer, a maintenance effect prediction layer, a knowledge reasoning layer and a parameter optimization layer, the maintenance strategy generation layer uses a hierarchical recursive structure with a memory enhancement mechanism to convert the multidimensional maintenance decision sequence into a maintenance action space, the maintenance effect prediction layer uses a multi-head attention mechanism and a causal inference mechanism to predict the maintenance action space to obtain execution effect data, the knowledge reasoning layer inputs the maintenance action space and the execution effect data into a graph neural network to construct a maintenance knowledge graph, and the parameter optimization layer inputs the maintenance action space, the execution effect data and the maintenance knowledge graph into a meta-learning network to update the network parameters of the deep coupling feature fusion model; An adaptive hybrid model is constructed based on the maintenance action space, the execution effect data, the maintenance knowledge graph and the updated network parameters. The adaptive hybrid model includes an action mapping engine, a multimodal fusion engine and a strategy evaluation engine. The action mapping engine uses a hierarchical attention mechanism to map the maintenance action space into a multi-level maintenance strategy tensor. The multimodal fusion engine performs feature fusion on the multi-level maintenance strategy tensor, the execution effect data and the maintenance knowledge graph to obtain a fused feature vector. The strategy evaluation engine inputs the fused feature vector into a deep belief network to construct a maintenance decision model.

4. The method according to claim 1, characterized in that: The maintenance decision model is input into the quantum enhanced dual network system to calculate the reward value sequence, including: Input the maintenance decision model into the quantum enhanced dual network system, the quantum enhanced dual network system includes a quantum strategy network module and a hierarchical evaluation network module, the quantum strategy network module uses a multi-particle quantum entangled state encoder to map the maintenance decision model to a high-dimensional quantum feature space to obtain a quantum state sequence, and the hierarchical evaluation network module uses a dynamic graph neural network and a cross-scale multi-head attention mechanism to process the quantum state sequence to obtain a spatiotemporal multi-dimensional value tensor; The spatiotemporal multidimensional value tensor is input into a recursive target decomposition system, wherein the recursive target decomposition system includes an adaptive neural tensor decomposition module and a multi-scale time series feature extraction module, wherein the adaptive neural tensor decomposition module uses a tensor kernel decomposition method to recursively decompose the spatiotemporal multidimensional value tensor to obtain an equipment reliability dimension tensor, a maintenance cost dimension tensor and a resource utilization dimension tensor, and the multi-scale time series feature extraction module uses a hierarchical deep time series model to sequentially process the dimension tensors to obtain a maintenance decision feature set, wherein fault evolution features and life prediction features are extracted from the equipment reliability dimension tensor, cost dynamic features and resource consumption features are extracted from the maintenance cost dimension tensor, and performance optimization features and load balancing features are extracted from the resource utilization dimension tensor; The dynamic weight coefficient matrix between the features in the maintenance decision feature set is calculated based on an improved non-dominated sorting genetic algorithm and an adaptive fuzzy hierarchical analysis method; the maintenance decision feature set and the dynamic weight coefficient matrix are input into a hybrid state evaluation system, and the hybrid state evaluation system comprises a conditional variational encoding module and a probabilistic reasoning module, the conditional variational encoding module uses a conditional variational autoencoder to map the features into a nonlinear state space to obtain a multimodal state representation vector, and the probabilistic reasoning module uses an improved deep belief network to calculate a time-varying state transition probability tensor based on the multimodal state representation vector and the dynamic weight coefficient matrix; the time-varying state transition probability tensor is input into a bidirectional evaluation system to calculate a reward value sequence.

5. The method according to claim 4, characterized in that The step of inputting the time-varying state transition probability tensor into a bidirectional evaluation system and calculating a reward value sequence comprises: The time-varying state transition probability tensor is input into a bidirectional evaluation system, wherein the bidirectional evaluation system comprises a hierarchical action evaluation module and a state evaluation module, wherein the hierarchical action evaluation module uses an improved bidirectional long short-term memory network and a cross-dimensional attention mechanism to evaluate the maintenance action to obtain a multi-level action influence vector, and the state evaluation module uses a spatiotemporal graph convolutional network to evaluate the state transition to obtain a state influence vector, and the multi-level action influence vector and the state influence vector are fused through an adaptive residual connection structure and a cross-modal feature fusion network to obtain a multi-dimensional reward vector; Constructing the meta-learning module, the meta-learning module includes a hierarchical gradient policy network, a dynamic programming network and a conditional variational network, wherein the hierarchical gradient policy network uses a hierarchical structure to calculate action gradients and state gradients, the dynamic programming network calculates discount factors based on the Bellman equation, and the conditional variational network uses an encoder-decoder structure to randomly sample state transitions to generate a conditional variational state value function; mapping the multidimensional reward vector, the discount factor and the conditional variational state value function to a unified feature space through nonlinear transformation to obtain a feature mapping matrix; decomposing the feature mapping matrix to obtain reward features and discount features; calculating the tensor product of the reward feature and the discount feature to obtain a reward-discount tensor, and then performing a tensor contraction operation on the reward-discount tensor and the conditional variational state value function to obtain a cumulative reward tensor; The accumulated reward tensor is input into an adversarial training network, which extracts features through a discriminator to obtain a discriminant score, and generates an optimized reward sample based on the discriminant score through a generator; a hierarchical knowledge distillation network is constructed using the discriminant score and the optimized reward sample, and the state transition features and action execution features in the pre-trained teacher network are transferred to the student network to obtain a knowledge transfer error; an adaptive loss function is constructed based on the discriminant score, the optimized reward sample and the knowledge transfer error, and a distributed asynchronous parallel algorithm is used to optimize network parameters to generate a reward value sequence containing timing dependencies and state transfer relationships.

6. The method according to claim 1, characterized in that The step of inputting the state transfer matrix into the maintenance strategy network, processing the output of the maintenance strategy network based on a quantum evolutionary algorithm, and generating a multi-constraint maintenance resource configuration scheme includes: The state transfer matrix is ​​input into the maintenance strategy network, the maintenance strategy network includes a state encoding layer, a strategy mapping layer and an action generation layer, the state encoding layer uses a multi-head self-attention module to calculate the association weights between states for the state transfer matrix, the multi-head self-attention module includes a query matrix, a key matrix and a value matrix, the attention score is obtained by the dot product operation of the query matrix and the key matrix, the attention score is multiplied by the value matrix to obtain the weighted feature, and the position code is introduced to retain the timing information of the state transfer, and the weighted feature, the association weights between the states and the position code are feature fused to obtain the state feature vector; Based on the state feature vector and the association weight between the states, a value evaluation branch and a strategy generation branch are constructed in the strategy mapping layer. The value evaluation branch uses a fully connected layer with residual connection to calculate the state value function of the maintenance action. The residual connection includes a trunk channel and a bypass channel. The trunk channel performs a nonlinear transformation on the state feature vector and the association weight between the states. The bypass channel keeps the association weight between the state feature vector and the states unchanged and superimposes it with the output of the trunk channel to obtain the state value function. The strategy generation branch includes a strategy network and a value network. The strategy network outputs the probability distribution of the maintenance action based on the state feature vector and the association weight between the states. The value network evaluates the expected return of the maintenance action based on the state feature vector and the association weight between the states. The optimized probability distribution of the maintenance action is obtained by optimizing the strategy network through the expected return of maximizing entropy regularization. In the action generation layer, based on the association weights between the states, the optimized maintenance action probability distribution is converted into discrete maintenance strategy parameters by using a reparameterized sampling technique, specifically by introducing a temperature parameter into the optimized maintenance action probability distribution to control the randomness of sampling, and generating discrete action samples based on the temperature parameter and the association weights between the states to obtain the maintenance strategy parameters as the output of the maintenance strategy network; The maintenance strategy parameters are processed based on a quantum evolutionary algorithm to generate a multi-constraint maintenance resource configuration plan.

7. The method according to claim 6, characterized in that The process of processing the maintenance strategy parameters based on the quantum evolutionary algorithm to generate a multi-constraint maintenance resource configuration scheme includes: The maintenance strategy parameter and the association weight between the states are encoded into a quantum bit string; the state value function and the association weight between the states are input into an adaptive angle calculation module, the adaptive angle calculation module uses a value difference function to calculate the difference between the current solution and the historical optimal solution corresponding to the maintenance strategy parameter, and sets an adaptive coefficient according to a preset proportional coefficient based on the difference, and multiplies the adaptive coefficient with a preset reference rotation angle to obtain an actual rotation angle of the quantum revolving gate; the actual rotation angle is used to drive the quantum revolving gate to adjust the phase of the quantum bit string, and a policy gradient vector is constructed based on the state value function and the association weight between the states, and a quantum interference operation matrix is ​​designed using the policy gradient vector, and the local phase of the quantum bit string is adjusted by the quantum interference operation matrix to realize population directed evolution, so as to obtain a quantum bit string of the first evolution; The state value function is used as an evaluation index, and the association weight between the states is used as a weight factor, and each quantum bit in the quantum bit string of the first evolution is weightedly calculated to obtain a weighted fitness score; the resource excess degree, time conflict degree and reliability deficiency degree in the maintenance plan corresponding to the quantum bit string of the first evolution are calculated, the resource excess degree is the difference between the actual required maintenance resource amount and the preset resource capacity, the time conflict degree is the duration of the actual execution time of the maintenance activity exceeding the preset time window, and the reliability deficiency degree is the difference between the actual reliability of the equipment after maintenance and the preset reliability target; the penalty term is calculated based on the resource excess degree, the time conflict degree and the reliability deficiency degree; the weighted fitness score and the penalty term are weightedly combined according to a preset ratio to obtain a comprehensive fitness evaluation result, and the quantum bit string of the first evolution is screened based on the comprehensive fitness evaluation result to obtain the quantum bit string of the second evolution; The variance of the value of each quantum bit in the quantum bit string of the second evolution is calculated to obtain a gene diversity index, and the association weight between the genetic diversity index and the state is weighted and combined according to the preset weight to obtain a diversity evaluation index; when the diversity evaluation index is lower than the preset lower limit threshold, the current quantum mutation probability is multiplied by a preset growth rate greater than 1 to obtain a new quantum mutation probability; when the diversity evaluation index is higher than the preset upper limit threshold, the current quantum mutation probability is multiplied by a preset decay rate less than 1 to obtain a new quantum mutation probability; the new quantum mutation probability is used to perform a flip operation on the quantum bits in the quantum bit string of the second evolution to realize quantum evolution, and the evolved quantum bit string is decoded to obtain multiple maintenance schemes, and the maintenance scheme with the highest fitness and satisfying the resource capacity limitation, time window constraint and reliability requirements is selected from the multiple maintenance schemes based on the comprehensive fitness evaluation result as the multi-constraint maintenance resource configuration scheme.

8. An abnormality detection and maintenance decision optimization system for automated test equipment, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to deploy edge computing units and multi-level sensor networks; The edge computing unit uses a self-organizing time synchronization protocol to synchronize multi-source sensor data; uses an empirical mode decomposition algorithm and a variational mode decomposition technique to perform modal separation on the synchronized sensor data; constructs a deep network model including a graph neural network module and a recursive variational autoencoder module, and the deep network model processes different sensor data through a cross-modal attention mechanism; matches the output features of the deep network model with a historical equipment fault case database to generate an equipment state feature vector; establishes a knowledge graph based on the equipment state feature vector to obtain a hierarchical fault diagnosis result; The second unit is used to receive the hierarchical fault diagnosis result, build a maintenance decision model in combination with the equipment operation data; input the maintenance decision model into the quantum enhanced dual network system, and calculate the reward value sequence; Constructing a dynamic probability graph model, inputting the reward value sequence into the dynamic probability graph model, and combining the dynamic probability graph model with a Bayesian network and a hidden Markov model to generate a state transfer matrix; A maintenance strategy network is trained using a federated reinforcement learning framework; the state transfer matrix is ​​input into the maintenance strategy network, and the output of the maintenance strategy network is input into a quantum evolutionary algorithm to generate a multi-constraint maintenance resource configuration plan; The third unit is used to input the multi-constraint maintenance resource configuration plan into a multi-criteria evaluation model based on interval intuitionistic fuzzy sets and improved grey relational degree, and output evaluation result data; Input the evaluation result data into an adaptive multi-objective differential evolution algorithm to generate optimized maintenance plan data; input the optimized maintenance plan data into a high-fidelity digital twin model to output integrated model data; input the integrated model data into a multi-task deep transfer learning network and a meta-learning network to generate maintenance decision model update data; transmit the maintenance decision model update data to the field control system through the industrial Internet of Things platform; A blockchain network is used to record and maintain decision model update data, and the data of the blockchain network is used as input to the equipment historical failure case database to update the deep network model.

Images