An engineering equipment operation state diagnosis system based on deep learning

Abstract

The application discloses a kind of based on deep learning's engineering equipment operating condition diagnosis system, comprising: operating condition acquisition and working condition construction module, for obtaining operating condition sequence and working condition vector;Topology graph construction module, for constructing device topology graph and establishing sensor mapping;Improved OmniAnomaly variation recursive network module, for output structure latent vector, degradation latent vector, reconstructed operating condition, health degree and failure advance amount;Working condition safety manifold mapping module, for applying safety manifold and time monotonicity constraint to degradation latent vector;Topological risk diffusion module, for generating node risk distribution;Joint loss training module, for constructing joint loss and training diagnosis model;Operating condition diagnosis and risk assessment module, for output abnormal score, abnormal time period and risk level.The application improves equipment anomaly identification and risk assessment capability.

Description

Technical Field

[0001] This invention relates to the field of engineering equipment condition monitoring and intelligent diagnosis technology, and in particular to an engineering equipment operation condition diagnosis system based on deep learning. Background Technology

[0002] During long-term continuous operation, engineering equipment typically relies on sensors for online monitoring, including those for vibration, temperature, current, and voltage. Traditional condition diagnosis often combines threshold alarms, empirical rules, and mechanistic models, manually setting operating ranges, fault characteristics, and threshold parameters to determine abnormal operating states. However, in complex operating environments characterized by frequent fluctuations in operating conditions, frequent load switching, and significant changes in environmental conditions, these methods are highly sensitive to parameter settings, struggle to adapt to slow degradation processes over long timescales, exhibit delayed responses to early-stage, hidden faults, and fail to provide quantitative health status and fault prediction information. Diagnostic results often rely on empirical interpretation, resulting in limited automation.

[0003] With the development of data-driven methods, time-series anomaly detection techniques based on models such as principal component analysis, autoencoders, recurrent neural networks, and variational recurrent networks are increasingly being used for the diagnosis of operational status of engineering equipment. Existing methods typically use multivariate operational state sequences as input, measuring the degree of anomaly by reconstructing errors or predicting residuals. Some works introduce variational inference and latent vector modeling of time-dependent structures to improve the ability to characterize nonlinear time series. However, existing models often treat latent states as a single whole, failing to distinguish between structure-related and degradation-related behaviors. They do not construct degradation safety manifold constraints through operating condition vectors, lack explicit modeling of operating condition changes such as load variations and start-stop fluctuations, making them prone to false alarms in scenarios with sudden changes in operating conditions. They also struggle to ensure that degradation-related latent quantities exhibit monotonically degrading characteristics in the time dimension, thus affecting the stability of health assessment and fault lead estimation.

[0004] In engineering equipment with multiple components and measurement points, there is a correspondence between the sensor installation location and the component topology. Operational risks often spread between nodes along the component connections. Existing time-series anomaly detection methods mainly make overall anomaly judgments in the sample feature space, rarely combining the equipment topology map to construct a node-level risk distribution model. They do not utilize the discrete diffusion process of degenerate latent vectors on the topology map to characterize the evolution of node risks, making it difficult to achieve precise localization of critical nodes and high-risk areas. At the same time, existing training strategies mostly use a single loss based on reconstruction error or a multi-factor loss with fixed weights, failing to dynamically adjust the weights of safety manifold deviation, node risk diffusion deviation, health deviation, and fault lead time deviation on the time axis. This fails to highlight the role of samples close to the fault moment in the optimization process, and provides insufficient support for identifying abnormal time periods and classifying operational risk levels.

[0005] Therefore, how to provide a deep learning-based system for diagnosing the operational status of engineering equipment is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based engineering equipment operation status diagnosis system. This invention utilizes an improved OmniAnomaly variational recurrent network, operating condition safety manifold mapping, and topological risk diffusion method to model the structural latent vector and degradation latent vector of sensor operation status sequences, outputting health status, fault lead time, node risk distribution, and anomaly score. This enables precise diagnosis of operation status and risk level classification of abnormal time periods, and has the advantages of adapting to complex operating conditions, identifying slow degradation, and locating high-risk nodes with high accuracy.

[0007] According to an embodiment of the present invention, an engineering equipment operation status diagnosis system based on deep learning includes: The operation status acquisition and working condition construction module is used to acquire sensor operation status signals to generate operation status sequences, and generate working condition vector sequences based on load, speed, environmental parameters and maintenance status. The topology graph construction module is used to construct the device topology graph based on the component connection relationships and sensor installation locations, and to establish the mapping relationship between sensors and topology nodes; An improved OmniAnomaly variational recursive network module is used to receive the running state sequence and output the structural latent vector, the degenerate latent vector, the reconstructed running state sequence, the health sequence, and the fault advance sequence; The working condition safety manifold mapping module is used to receive the working condition vector sequence and output the degenerate safety manifold parameters, and apply safety manifold constraints and time monotonic degradation constraints to the degenerate potential vectors to generate a degenerate potential vector sequence. The topology risk diffusion module is used to receive the degraded potential vector sequence and the device topology map, and perform discrete risk diffusion on the device topology map to generate a node risk distribution sequence; The joint loss training module is used to construct a joint loss function based on the running state sequence, the reconstructed running state sequence, the degraded potential vector sequence, the degraded safe manifold parameters, the node risk distribution sequence, the health sequence, and the failure advance sequence. The loss weight is dynamically adjusted according to the failure advance sequence to train the improved OmniAnomaly variational recursive network module and the topology risk diffusion module to obtain the running state diagnosis model. The Operation Status Diagnosis and Risk Assessment module is used to input new operation status sequences and operating condition vector sequences into the operation status diagnosis model within the monitoring period, and output degradation potential vector sequences, node risk distribution sequences, health sequences, failure lead sequences, and anomaly score sequences. Based on the anomaly score sequences, it determines the abnormal time periods and operation risk levels.

[0008] Optionally, modules can be integrated using the following methods: Collect sensor operating status sequences, construct operating condition vector sequences, and build a topology map based on component connection relationships; An improved OmniAnomaly variational recurrent network is constructed based on the running state sequence, which encodes the running state sequence as structural latent vectors and degenerate latent vectors; Input the sequence of operating condition vectors into the operating condition safety manifold mapping subnetwork to generate a degenerate safety manifold, and apply safety manifold constraints and time monotonic degradation constraints to the degenerate potential vectors; Input the topology graph and the degenerate potential vector into the topology risk diffusion module, and discretize the risk diffusion on the topology graph to obtain the node risk distribution; Input the structural latent vector and the degradation latent vector into the decoder, reconstruct the running state sequence, and output the health status and failure advance. A joint loss function is constructed based on reconstruction error, safety manifold deviation, node risk distribution, health status and fault lead time. The loss weight is dynamically adjusted according to the fault lead time. An improved OmniAnomaly variational recurrent network and topology risk diffusion module are trained to obtain the operational status diagnosis model. The new operating status sequence is input into the operating status diagnostic model, which outputs a degradation potential vector and anomaly score. The abnormal time period and operating risk level are determined based on the anomaly score.

[0009] Optionally, the improved OmniAnomaly variational recurrent network includes: The running state encoding unit receives the running state sequence and generates a potential state vector through a recursive structure. The subspace decomposition unit divides the potential state vector into structural potential vectors and degenerate potential vectors; The running state decoding unit receives the structural latent vector and the degenerate latent vector and reconstructs the running state sequence; The health and fault lead output unit generates health and fault lead based on the degradation potential vector.

[0010] Optionally, the construction of the improved OmniAnomaly variational recurrent network specifically includes: The operating state sequence is arranged in the order of sampling time to obtain an operating state vector sequence that corresponds one-to-one with the sampling time. Each operating state vector includes vibration monitoring value, temperature monitoring value, current monitoring value and voltage monitoring value. The vibration monitoring value, temperature monitoring value, current monitoring value and voltage monitoring value are scaled according to a preset normalization method. The scaling results are spliced ​​in a fixed order to form an operating state vector, which is used to represent the operating state at the sampling time point. The sequence of running state vectors is input into the running state encoding unit. At each sampling time point, the running state encoding unit receives the current running state vector and the internal state of the previous sampling time point. Through recursive calculation, the new internal state and the potential state vector corresponding to the sampling time point are obtained. The internal state is used to record the comprehensive information of the running state vectors of the previous sampling time points. The potential state vectors of all sampling time points are arranged in the order of sampling time to form a potential state vector sequence. The potential state vector sequence is input into the subspace decomposition unit. At each sampling time point, the subspace decomposition unit splits the potential state vector into a structural potential vector and a degenerate potential vector according to a preset subspace allocation scheme. The structural potential vector is used to represent the components related to the equipment structure in the operating state, and the degenerate potential vector is used to represent the components that accumulate and change over time in the operating state. The structural potential vectors at all sampling time points are arranged in the order of sampling time to form a structural potential vector sequence, and the degenerate potential vectors at all sampling time points are arranged in the order of sampling time to form a degenerate potential vector sequence.

[0011] Optionally, the processing of the working condition safety manifold mapping sub-network specifically includes: The operating condition vector sequence is input into the operating condition safety manifold mapping subnetwork in the order of sampling time. The operating condition safety manifold mapping subnetwork includes an input layer, a hidden layer, and an output layer. The input layer receives load, speed, ambient temperature, ambient humidity, and maintenance status parameters from the operating condition vector component by component. The hidden layer performs linear weighting, bias superposition, and nonlinear activation operations on the input vector. The output layer outputs the manifold center vector and manifold scale vector with the same dimension as the degenerate latent vector at each sampling time. The manifold center vector sequence and manifold scale vector sequence corresponding to each sampling time are registered as the degenerate safety manifold parameter sequence. At each sampling time, the degenerate latent vector and the corresponding degenerate safe manifold parameter are read. For each dimension of the degenerate latent vector, the difference between the degenerate latent vector component and the manifold center vector component is calculated. The safe manifold deviation of the dimension is obtained according to the ratio of the difference to the manifold scale vector component. The safe manifold deviation of all dimensions is compared with a preset threshold. When the safe manifold deviation of any dimension is greater than the preset threshold, the magnitude of the degenerate latent vector component is scaled along the direction pointing to the manifold center vector component in that dimension, so that the degenerate latent vector falls into the safe range defined by the degenerate safe manifold parameter. The adjusted degenerate latent vectors are arranged in the order of sampling time to obtain the degenerate latent vector sequence that satisfies the safe manifold constraint. Apply time-monotonically degenerate constraints to the degenerate latent vector sequence that satisfies the safe manifold constraint. Starting from the first sampling time after the start time, read the current sampling time degenerate latent vector and the previous sampling time degenerate latent vector sequentially. Compare the components of the current sampling time degenerate latent vector with those of the previous sampling time in the dimension of the degeneracy index. When the component of the current sampling time degenerate latent vector is less than that of the previous sampling time degenerate latent vector, replace the component of the current sampling time degenerate latent vector with that of the previous sampling time degenerate latent vector. Update the corresponding degenerate latent vectors of all sampling times in the order of sampling time to obtain a degenerate latent vector sequence that simultaneously satisfies the safe manifold constraint and the time-monotonically degenerate constraint.

[0012] Optionally, the generation of the node risk distribution specifically includes: At each sampling time, the node index and component connection relationship are read from the topology graph. Based on the mapping relationship between the sensor installation position and the node index, the components of each dimension in the degradation potential vector are assigned to the corresponding node to form a node degradation vector that corresponds one-to-one with the node index. Each node degradation vector is composed of the degradation potential vector components assigned to the node arranged in a preset order. When a node is associated with more than one sensor, the degradation potential vector components corresponding to the associated sensors are summed according to a preset aggregation method to obtain the node degradation vector components. The node degradation vector is input into the topology risk diffusion module, which includes a risk initialization unit, a diffusion coefficient storage unit, and a risk iteration unit. The risk initialization unit calculates the initial risk intensity of the node based on the node degradation vector. The initial risk intensity of the node is obtained by multiplying the components of the node degradation vector by preset weights in each dimension and summing them. The diffusion coefficient storage unit registers the diffusion coefficient for each edge according to the edge weights in the topology graph. The risk iteration unit reads the current node risk intensity, edge weights, and diffusion coefficients at each diffusion step. For each node, it traverses the edges connected to the node, multiplies the node risk intensity by the corresponding edge weights and diffusion coefficients to obtain the risk increment, adds the risk increment to the risk intensity of the connected nodes, and retains the updated node risk intensity. After completing one round of diffusion step update, the new node risk intensity is registered according to the node index. The new node risk intensities are arranged in the order of diffusion step to form a node risk intensity iteration sequence. When the number of diffusion steps reaches the preset number of diffusion steps, the node risk intensity corresponding to the last diffusion step is selected from the node risk intensity iteration sequence. The node risk distribution at the current sampling time is obtained by arranging them according to the node index. The node risk distribution at each sampling time is arranged in the order of sampling time to form a node risk distribution sequence. The node risk distribution sequence is used to represent the risk diffusion status of the engineering equipment in the time axis and topology node dimensions.

[0013] Optionally, the generation of the health status and failure advance rate specifically includes: The structural latent vector sequence and the degenerate latent vector sequence are input into the running state decoding unit in the order of sampling time. At each sampling time, the running state decoding unit receives the current structural latent vector, the current degenerate latent vector, and the internal state of the previous sampling time. The unit performs a linear weighting operation on the current structural latent vector and the current degenerate latent vector, and then weights and superimposes the linear weighting result with the internal state of the previous sampling time and inputs it into the recursive activation function to generate a new internal state. At the output layer, a linear transformation and a nonlinear activation operation are performed based on the new internal state to obtain the reconstructed running state vector corresponding to the current sampling time. The reconstructed running state vectors corresponding to all sampling times are arranged in the order of sampling time to form a reconstructed running state sequence. The degraded potential vector sequence is input into the health mapping unit in the health and fault advance output unit. The health mapping unit receives the current degraded potential vector at each sampling time, performs linear weighting and bias superposition operations on each component of the degraded potential vector to obtain an intermediate health representation, and then inputs a nonlinear activation function that limits the output range to the intermediate health representation. The output is a health value between a preset minimum health value and a preset maximum health value. The health values ​​corresponding to all sampling times are arranged in the order of sampling time to form a health sequence. The degraded potential vector sequence is input into the fault advance mapping unit in the health and fault advance output unit. The fault advance mapping unit receives the current degraded potential vector and the sampling time interval parameter at each sampling time. It performs linear weighting and nonlinear activation operations on the current degraded potential vector to obtain the failure proximity representation. Based on the failure proximity representation and the sampling time interval parameter, it infers the predicted fault time and calculates the number of sampling time intervals between the current sampling time and the predicted fault time as the fault advance. The fault advance corresponding to all sampling times is arranged in the order of sampling time to form the fault advance sequence.

[0014] Optionally, the generation of the operational status diagnostic model specifically includes: Based on the running state sequence and the reconstructed running state sequence, a reconstruction error term is constructed. At each sampling time, the square of the difference in each feature dimension is calculated based on the numerical difference between the running state vector and the reconstructed running state vector in each feature dimension. The squares of the differences in each feature dimension are added together to obtain the reconstruction error for that sampling time. The reconstruction errors corresponding to all sampling times are arranged in the order of sampling time to form a reconstruction error sequence. A safety manifold deviation term is constructed based on the degenerate latent vector sequence and the degenerate safety manifold parameter sequence. At each sampling time, the difference between the component of the degenerate latent vector in each dimension and the component of the manifold center vector in the corresponding dimension at the same sampling time is calculated. The safety manifold deviation degree in each dimension is obtained by the ratio of the difference to the component of the manifold scale vector in the corresponding dimension. The safety manifold deviation degree is weighted and summed in all dimensions to obtain the safety manifold deviation at that sampling time. The safety manifold deviations corresponding to all sampling times are arranged in the order of sampling time to form a safety manifold deviation sequence. Based on the node risk distribution sequence, a node risk diffusion deviation term is constructed. At each sampling time, the node weighted risk intensity is obtained by multiplying the risk intensity of each node in the node risk distribution by a preset node coefficient. The node risk diffusion deviation at that sampling time is obtained by performing summation and squaring operations on the node weighted risk intensity at all nodes. The node risk diffusion deviations corresponding to all sampling times are arranged in the order of sampling time to form a node risk diffusion deviation sequence. Based on the health status sequence and the fault lead time sequence, construct health status deviation items and fault lead time deviation items. At each sampling time, calculate the square of the health status difference based on the difference between the health status and the preset health reference value as the health status deviation for that sampling time. Calculate the square of the fault lead time difference based on the difference between the fault lead time and the preset target lead time as the fault lead time deviation for that sampling time. Arrange the health status deviation and fault lead time deviation corresponding to all sampling times in the order of sampling time to form the health status deviation sequence and the fault lead time deviation sequence. At each sampling time, a weight coefficient is determined based on the fault advance sequence. The smaller the fault advance, the larger the weight coefficient. The reconstruction error, safety manifold deviation, node risk diffusion deviation, health deviation, and fault advance deviation at that sampling time are multiplied by their respective weight coefficients and summed to obtain the joint loss for that sampling time. The joint losses of all sampling times are summed to obtain the joint loss function. Based on the joint loss function, the parameters of the improved OmniAnomaly variational recursive network and the parameters of the topology risk diffusion module are optimized and trained to obtain the operational status diagnostic model.

[0015] Optionally, the determination of the abnormal time period and operational risk level specifically includes: During the monitoring period, sensor operating status sequences and operating condition vector sequences are collected. The sensor operating status sequences are arranged in order of sampling time to form a new operating status sequence, and the operating condition vectors are arranged in order of sampling time to form a new operating condition vector sequence. The new operating status sequences are input into an improved OmniAnomaly variational recursive network to obtain structural latent vector sequences and degenerate latent vector sequences. The new operating condition vector sequences are input into the operating condition safety manifold mapping subnetwork to obtain a degenerate safety manifold parameter sequence. The degenerate latent vector sequences and degenerate safety manifold parameter sequences are input into a topology risk diffusion module to obtain a node risk distribution sequence. The degenerate latent vector sequences are input into a health and fault lead output unit to obtain a health sequence and a fault lead sequence. At each sampling time, an anomaly score is calculated based on the degenerate latent vector, degenerate safety manifold parameter, node risk distribution, health and fault lead. The anomaly scores corresponding to all sampling times are arranged in order of sampling time to form an anomaly score sequence. At each sampling time, the abnormal score is compared with the preset abnormal score threshold. When the abnormal score is greater than the preset abnormal score threshold, the corresponding sampling time is marked as an abnormal sampling time. The abnormal sampling times with consecutive sampling times are merged into an abnormal time period in the order of sampling time. All abnormal time periods are arranged in the order of their start sampling time to form an abnormal time period set. For each abnormal time period in the abnormal time period set, the maximum abnormal score within that time period is selected from the abnormal score sequence as the abnormal intensity index for that time period; the maximum node risk intensity within that time period is selected from the node risk distribution sequence as the risk diffusion index for that time period; the minimum health value within that time period is selected from the health value sequence as the health level index for that time period; and the minimum failure lead time within that time period is selected from the failure lead time sequence as the remaining safe time index for that time period. The abnormal intensity index, risk diffusion index, health level index, and remaining safe time index are matched with the preset operational risk classification standard to determine the operational risk level corresponding to each abnormal time period, and the abnormal time period and operational risk level are output.

[0016] The beneficial effects of this invention are: Compared with existing time-series anomaly detection methods based on a single potential state, this invention uses an improved OmniAnomaly variational recurrent network to encode the operating state sequence into structural potential vectors and degradation potential vectors. It also combines the operating condition vector sequence to construct a degradation safety manifold and a time-monotonic degradation constraint, so that degradation-related information exhibits a stable and monotonic evolution pattern on the time axis. This effectively suppresses false alarms caused by operating condition switching and start-stop fluctuations. At the same time, it outputs a health sequence and a fault advance sequence, enabling early identification and quantitative prediction of slow degradation processes under complex operating conditions, thereby improving the accuracy and stability of engineering equipment condition assessment.

[0017] By introducing a topology graph construction module and a topology risk diffusion module, the degraded potential vector sequence is mapped to a node degradation vector. Discrete risk diffusion is performed on the equipment topology graph to generate a node risk distribution sequence, so that the operational risk can be uniformly represented in the time dimension and the topology node dimension. This not only distinguishes between overall anomalies and local anomalies, but also locates high-risk nodes and key components, providing a refined basis for maintenance strategy formulation and maintenance resource allocation, and reducing blind maintenance and accidental downtime.

[0018] The joint loss training module, based on the reconstruction error term, safety manifold deviation term, node risk diffusion deviation term, health deviation term, and failure lead time deviation term, introduces a dynamic weight adjustment mechanism based on the failure lead time sequence. This mechanism assigns higher weights to samples closer to the predicted failure time during optimization, making the operational status diagnostic model more sensitive to high-risk periods. Combined with the anomaly scoring sequence, anomaly time periods, and operational risk levels output by the operational status diagnostic and risk assessment module, this invention can complete anomaly detection, health assessment, risk classification, and spatial location tasks within a unified framework, reducing reliance on manual experience and improving the automation level of engineering equipment operation safety management. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based engineering equipment operation status diagnosis system proposed in this invention; Figure 2 This is a schematic diagram of the improved OmniAnomaly variational recursive network module structure of an engineering equipment operation status diagnosis system based on deep learning proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-2 A deep learning-based system for diagnosing the operational status of engineering equipment includes: The operation status acquisition and working condition construction module is used to acquire sensor operation status signals to generate operation status sequences, and generate working condition vector sequences based on load, speed, environmental parameters and maintenance status. The topology graph construction module is used to construct the device topology graph based on the component connection relationships and sensor installation locations, and to establish the mapping relationship between sensors and topology nodes; An improved OmniAnomaly variational recursive network module is used to receive the running state sequence and output the structural latent vector, the degenerate latent vector, the reconstructed running state sequence, the health sequence, and the fault advance sequence; The working condition safety manifold mapping module is used to receive the working condition vector sequence and output the degenerate safety manifold parameters, and apply safety manifold constraints and time monotonic degradation constraints to the degenerate potential vectors to generate a degenerate potential vector sequence. The topology risk diffusion module is used to receive the degraded potential vector sequence and the device topology map, and perform discrete risk diffusion on the device topology map to generate a node risk distribution sequence; The joint loss training module is used to construct a joint loss function based on the running state sequence, the reconstructed running state sequence, the degraded potential vector sequence, the degraded safe manifold parameters, the node risk distribution sequence, the health sequence, and the failure advance sequence. The loss weight is dynamically adjusted according to the failure advance sequence to train the improved OmniAnomaly variational recursive network module and the topology risk diffusion module to obtain the running state diagnosis model. The Operation Status Diagnosis and Risk Assessment module is used to input new operation status sequences and operating condition vector sequences into the operation status diagnosis model within the monitoring period, and output degradation potential vector sequences, node risk distribution sequences, health sequences, failure lead sequences, and anomaly score sequences. Based on the anomaly score sequences, it determines the abnormal time periods and operation risk levels.

[0022] In this embodiment, the modules are interconnected using the following method: Collect sensor operating status sequences, construct operating condition vector sequences, and build a topology map based on component connection relationships; An improved OmniAnomaly variational recurrent network is constructed based on the running state sequence, which encodes the running state sequence as structural latent vectors and degenerate latent vectors; Input the sequence of operating condition vectors into the operating condition safety manifold mapping subnetwork to generate a degenerate safety manifold, and apply safety manifold constraints and time monotonic degradation constraints to the degenerate potential vectors; Input the topology graph and the degenerate potential vector into the topology risk diffusion module, and discretize the risk diffusion on the topology graph to obtain the node risk distribution; Input the structural latent vector and the degradation latent vector into the decoder, reconstruct the running state sequence, and output the health status and failure advance. A joint loss function is constructed based on reconstruction error, safety manifold deviation, node risk distribution, health status and fault lead time. The loss weight is dynamically adjusted according to the fault lead time. An improved OmniAnomaly variational recurrent network and topology risk diffusion module are trained to obtain the operational status diagnosis model. The new operating status sequence is input into the operating status diagnostic model, which outputs a degradation potential vector and anomaly score. The abnormal time period and operating risk level are determined based on the anomaly score.

[0023] In this embodiment, the improved OmniAnomaly variational recursive network includes: The running state encoding unit receives the running state sequence and generates a potential state vector through a recursive structure. The subspace decomposition unit divides the potential state vector into structural potential vectors and degenerate potential vectors; The running state decoding unit receives the structural latent vector and the degenerate latent vector and reconstructs the running state sequence; The health and fault lead output unit generates health and fault lead based on the degradation potential vector.

[0024] This invention, by setting up an operating state encoding unit, a subspace decomposition unit, an operating state decoding unit, and a health and fault lead output unit in an improved OmniAnomaly variational recursive network, maps the original operating state sequence into structural latent vectors and degradation latent vectors and establishes a stable latent space representation. This enables the model to simultaneously capture the structural change characteristics and degradation trend change characteristics of the equipment in the recursive structure. The structural latent vector improves the accuracy of operating state reconstruction, and the degradation latent vector enhances the ability to separate degradation patterns. Thus, at the decoding end, an accurate reconstructed operating state sequence is obtained, and the health and fault lead are output simultaneously. This allows operating state monitoring, health assessment, and fault time prediction to be collaboratively modeled and numerically unified within a unified framework, significantly improving the coherence of equipment state understanding and the foresight of fault prediction.

[0025] In this embodiment, the construction of the improved OmniAnomaly variational recurrent network specifically includes: The operating state sequence is arranged in the order of sampling time to obtain an operating state vector sequence that corresponds one-to-one with the sampling time. Each operating state vector includes vibration monitoring value, temperature monitoring value, current monitoring value and voltage monitoring value. The vibration monitoring value, temperature monitoring value, current monitoring value and voltage monitoring value are scaled according to a preset normalization method. The scaling results are spliced ​​in a fixed order to form an operating state vector, which is used to represent the operating state at the sampling time point. The sequence of running state vectors is input into the running state encoding unit. At each sampling time point, the running state encoding unit receives the current running state vector and the internal state of the previous sampling time point. Through recursive calculation, the new internal state and the potential state vector corresponding to the sampling time point are obtained. The internal state is used to record the comprehensive information of the running state vectors of the previous sampling time points. The potential state vectors of all sampling time points are arranged in the order of sampling time to form a potential state vector sequence. The potential state vector sequence is input into the subspace decomposition unit. At each sampling time point, the subspace decomposition unit splits the potential state vector into a structural potential vector and a degenerate potential vector according to a preset subspace allocation scheme. The structural potential vector is used to represent the components related to the equipment structure in the operating state, and the degenerate potential vector is used to represent the components that accumulate and change over time in the operating state. The structural potential vectors at all sampling time points are arranged in the order of sampling time to form a structural potential vector sequence, and the degenerate potential vectors at all sampling time points are arranged in the order of sampling time to form a degenerate potential vector sequence.

[0026] This invention explicitly designs the construction process of the improved OmniAnomaly variational recurrent network, ensuring that operational state data receives a unified time and feature scale representation before entering the network. By normalizing and concatenating vibration, temperature, current, and voltage monitoring values ​​to form a consistent operational state vector, the recurrent structure can stably receive and process cross-dimensional, multi-dimensional monitoring signals. The operational state encoding unit accumulates internal states over time and outputs a latent state vector, enabling the model to simultaneously capture instantaneous operational states and time-dependent patterns. The subspace decomposition unit, based on a preset subspace allocation scheme, splits the latent state vector into structural latent vectors and degradation latent vectors, achieving explicit spatial separation of structural change features and degradation trend features. This allows subsequent diagnostic models to perform reconstruction, health assessment, and fault prediction based on a clearer latent representation. This network construction method makes equipment state representation more stable, degradation pattern extraction more accurate, and significantly improves model interpretability and diagnostic accuracy.

[0027] In this embodiment, the processing of the working condition safety manifold mapping subnetwork specifically includes: The operating condition vector sequence is input into the operating condition safety manifold mapping subnetwork in the order of sampling time. The operating condition safety manifold mapping subnetwork includes an input layer, a hidden layer, and an output layer. The input layer receives load, speed, ambient temperature, ambient humidity, and maintenance status parameters from the operating condition vector component by component. The hidden layer performs linear weighting, bias superposition, and nonlinear activation operations on the input vector. The output layer outputs the manifold center vector and manifold scale vector with the same dimension as the degenerate latent vector at each sampling time. The manifold center vector sequence and manifold scale vector sequence corresponding to each sampling time are registered as the degenerate safety manifold parameter sequence. At each sampling time, the degenerate latent vector and the corresponding degenerate safe manifold parameter are read. For each dimension of the degenerate latent vector, the difference between the degenerate latent vector component and the manifold center vector component is calculated. The safe manifold deviation of the dimension is obtained according to the ratio of the difference to the manifold scale vector component. The safe manifold deviation of all dimensions is compared with a preset threshold. When the safe manifold deviation of any dimension is greater than the preset threshold, the magnitude of the degenerate latent vector component is scaled along the direction pointing to the manifold center vector component in that dimension, so that the degenerate latent vector falls into the safe range defined by the degenerate safe manifold parameter. The adjusted degenerate latent vectors are arranged in the order of sampling time to obtain the degenerate latent vector sequence that satisfies the safe manifold constraint. Apply time-monotonically degenerate constraints to the degenerate latent vector sequence that satisfies the safe manifold constraint. Starting from the first sampling time after the start time, read the current sampling time degenerate latent vector and the previous sampling time degenerate latent vector sequentially. Compare the components of the current sampling time degenerate latent vector with those of the previous sampling time in the dimension of the degeneracy index. When the component of the current sampling time degenerate latent vector is less than that of the previous sampling time degenerate latent vector, replace the component of the current sampling time degenerate latent vector with that of the previous sampling time degenerate latent vector. Update the corresponding degenerate latent vectors of all sampling times in the order of sampling time to obtain a degenerate latent vector sequence that simultaneously satisfies the safe manifold constraint and the time-monotonically degenerate constraint.

[0028] In this embodiment, the generation of the node risk distribution specifically includes: At each sampling time, the node index and component connection relationship are read from the topology graph. Based on the mapping relationship between the sensor installation position and the node index, the components of each dimension in the degradation potential vector are assigned to the corresponding node to form a node degradation vector that corresponds one-to-one with the node index. Each node degradation vector is composed of the degradation potential vector components assigned to the node arranged in a preset order. When a node is associated with more than one sensor, the degradation potential vector components corresponding to the associated sensors are summed according to a preset aggregation method to obtain the node degradation vector components. The node degradation vector is input into the topology risk diffusion module, which includes a risk initialization unit, a diffusion coefficient storage unit, and a risk iteration unit. The risk initialization unit calculates the initial risk intensity of the node based on the node degradation vector. The initial risk intensity of the node is obtained by multiplying the components of the node degradation vector by preset weights in each dimension and summing them. The diffusion coefficient storage unit registers the diffusion coefficient for each edge according to the edge weights in the topology graph. The risk iteration unit reads the current node risk intensity, edge weights, and diffusion coefficients at each diffusion step. For each node, it traverses the edges connected to the node, multiplies the node risk intensity by the corresponding edge weights and diffusion coefficients to obtain the risk increment, adds the risk increment to the risk intensity of the connected nodes, and retains the updated node risk intensity. After completing one round of diffusion step update, the new node risk intensity is registered according to the node index. The new node risk intensities are arranged in the order of diffusion step to form a node risk intensity iteration sequence. When the number of diffusion steps reaches the preset number of diffusion steps, the node risk intensity corresponding to the last diffusion step is selected from the node risk intensity iteration sequence. The node risk distribution at the current sampling time is obtained by arranging them according to the node index. The node risk distribution at each sampling time is arranged in the order of sampling time to form a node risk distribution sequence. The node risk distribution sequence is used to represent the risk diffusion status of the engineering equipment in the time axis and topology node dimensions.

[0029] In this embodiment, the generation of health status and failure advance specifically includes: The structural latent vector sequence and the degenerate latent vector sequence are input into the running state decoding unit in the order of sampling time. At each sampling time, the running state decoding unit receives the current structural latent vector, the current degenerate latent vector, and the internal state of the previous sampling time. The unit performs a linear weighting operation on the current structural latent vector and the current degenerate latent vector, and then weights and superimposes the linear weighting result with the internal state of the previous sampling time and inputs it into the recursive activation function to generate a new internal state. At the output layer, a linear transformation and a nonlinear activation operation are performed based on the new internal state to obtain the reconstructed running state vector corresponding to the current sampling time. The reconstructed running state vectors corresponding to all sampling times are arranged in the order of sampling time to form a reconstructed running state sequence. The degraded potential vector sequence is input into the health mapping unit in the health and fault advance output unit. The health mapping unit receives the current degraded potential vector at each sampling time, performs linear weighting and bias superposition operations on each component of the degraded potential vector to obtain an intermediate health representation, and then inputs a nonlinear activation function that limits the output range to the intermediate health representation. The output is a health value between a preset minimum health value and a preset maximum health value. The health values ​​corresponding to all sampling times are arranged in the order of sampling time to form a health sequence. The degraded potential vector sequence is input into the fault advance mapping unit in the health and fault advance output unit. The fault advance mapping unit receives the current degraded potential vector and the sampling time interval parameter at each sampling time. It performs linear weighting and nonlinear activation operations on the current degraded potential vector to obtain the failure proximity representation. Based on the failure proximity representation and the sampling time interval parameter, it infers the predicted fault time and calculates the number of sampling time intervals between the current sampling time and the predicted fault time as the fault advance. The fault advance corresponding to all sampling times is arranged in the order of sampling time to form the fault advance sequence.

[0030] In this embodiment, the generation of the operational status diagnostic model specifically includes: Based on the running state sequence and the reconstructed running state sequence, a reconstruction error term is constructed. At each sampling time, the square of the difference in each feature dimension is calculated based on the numerical difference between the running state vector and the reconstructed running state vector in each feature dimension. The squares of the differences in each feature dimension are added together to obtain the reconstruction error for that sampling time. The reconstruction errors corresponding to all sampling times are arranged in the order of sampling time to form a reconstruction error sequence. A safety manifold deviation term is constructed based on the degenerate latent vector sequence and the degenerate safety manifold parameter sequence. At each sampling time, the difference between the component of the degenerate latent vector in each dimension and the component of the manifold center vector in the corresponding dimension at the same sampling time is calculated. The safety manifold deviation degree in each dimension is obtained by the ratio of the difference to the component of the manifold scale vector in the corresponding dimension. The safety manifold deviation degree is weighted and summed in all dimensions to obtain the safety manifold deviation at that sampling time. The safety manifold deviations corresponding to all sampling times are arranged in the order of sampling time to form a safety manifold deviation sequence. Based on the node risk distribution sequence, a node risk diffusion deviation term is constructed. At each sampling time, the node weighted risk intensity is obtained by multiplying the risk intensity of each node in the node risk distribution by a preset node coefficient. The node risk diffusion deviation at that sampling time is obtained by performing summation and squaring operations on the node weighted risk intensity at all nodes. The node risk diffusion deviations corresponding to all sampling times are arranged in the order of sampling time to form a node risk diffusion deviation sequence. Based on the health status sequence and the fault lead time sequence, construct health status deviation items and fault lead time deviation items. At each sampling time, calculate the square of the health status difference based on the difference between the health status and the preset health reference value as the health status deviation for that sampling time. Calculate the square of the fault lead time difference based on the difference between the fault lead time and the preset target lead time as the fault lead time deviation for that sampling time. Arrange the health status deviation and fault lead time deviation corresponding to all sampling times in the order of sampling time to form the health status deviation sequence and the fault lead time deviation sequence. At each sampling time, a weight coefficient is determined based on the fault advance sequence. The smaller the fault advance, the larger the weight coefficient. The reconstruction error, safety manifold deviation, node risk diffusion deviation, health deviation, and fault advance deviation at that sampling time are multiplied by their respective weight coefficients and summed to obtain the joint loss for that sampling time. The joint losses of all sampling times are summed to obtain the joint loss function. Based on the joint loss function, the parameters of the improved OmniAnomaly variational recursive network and the parameters of the topology risk diffusion module are optimized and trained to obtain the operational status diagnostic model.

[0031] This invention constructs a reconstruction error term, a safety manifold deviation term, a node risk diffusion deviation term, a health deviation term, and a failure lead time deviation term during the generation of the operational status diagnostic model. It also introduces dynamic weighting coefficients based on failure lead time. This unified approach incorporates the operational status reconstruction accuracy, the degree of fit between the degradation potential vector and the operational condition safety manifold, the diffusion intensity of node risk in the topology, and the deviations between health and failure lead time from the reference target into a joint loss function. This allows data samples near the failure time to receive higher optimization priority during training. Thus, under the same optimization framework, it simultaneously constrains the time series reconstruction quality, degradation trajectory morphology, and spatial risk distribution morphology, effectively improving the model's comprehensive sensitivity to early degradation, locally high-risk nodes, and near-failure periods. This significantly reduces the probability of false alarms and false negatives, enhancing the accuracy and stability of the operational status diagnostic results for engineering equipment.

[0032] In this embodiment, the determination of the abnormal time period and the operational risk level specifically includes: During the monitoring period, sensor operating status sequences and operating condition vector sequences are collected. The sensor operating status sequences are arranged in order of sampling time to form a new operating status sequence, and the operating condition vectors are arranged in order of sampling time to form a new operating condition vector sequence. The new operating status sequences are input into an improved OmniAnomaly variational recursive network to obtain structural latent vector sequences and degenerate latent vector sequences. The new operating condition vector sequences are input into the operating condition safety manifold mapping subnetwork to obtain a degenerate safety manifold parameter sequence. The degenerate latent vector sequences and degenerate safety manifold parameter sequences are input into a topology risk diffusion module to obtain a node risk distribution sequence. The degenerate latent vector sequences are input into a health and fault lead output unit to obtain a health sequence and a fault lead sequence. At each sampling time, an anomaly score is calculated based on the degenerate latent vector, degenerate safety manifold parameter, node risk distribution, health and fault lead. The anomaly scores corresponding to all sampling times are arranged in order of sampling time to form an anomaly score sequence. At each sampling time, the abnormal score is compared with the preset abnormal score threshold. When the abnormal score is greater than the preset abnormal score threshold, the corresponding sampling time is marked as an abnormal sampling time. The abnormal sampling times with consecutive sampling times are merged into an abnormal time period in the order of sampling time. All abnormal time periods are arranged in the order of their start sampling time to form an abnormal time period set. For each abnormal time period in the abnormal time period set, the maximum abnormal score within that time period is selected from the abnormal score sequence as the abnormal intensity index for that time period; the maximum node risk intensity within that time period is selected from the node risk distribution sequence as the risk diffusion index for that time period; the minimum health value within that time period is selected from the health value sequence as the health level index for that time period; and the minimum failure lead time within that time period is selected from the failure lead time sequence as the remaining safe time index for that time period. The abnormal intensity index, risk diffusion index, health level index, and remaining safe time index are matched with the preset operational risk classification standard to determine the operational risk level corresponding to each abnormal time period, and the abnormal time period and operational risk level are output.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to the operational status monitoring and health management of a large-scale engineering equipment. This equipment includes various types of sensors such as vibration, temperature, pressure, current, and voltage, distributed across different functional components. These components have clear connections, and the equipment is characterized by a large number of sensors, frequent fluctuations in operating conditions, and a slow degradation trend over long-term operation. Traditional methods relying on threshold rules and experience-based judgments are prone to false alarms and missed alarms during complex operating condition transitions and early stages of weak degradation. This makes it difficult to achieve continuous assessment of equipment health and predict fault advances, and it is also impossible to identify high-risk nodes based on the equipment topology. This invention achieves end-to-end intelligent diagnosis of the operational status of the engineering equipment through an operational status acquisition and operating condition construction module, an improved OmniAnomaly variational recursive network module, an operating condition safety manifold mapping module, a topology risk diffusion module, a joint loss training module, and an operational status diagnosis and risk assessment module.

[0034] In the application process, the operating status signals collected by dozens of sensors installed on the equipment are first uniformly sampled to construct an operating status sequence. A working condition vector sequence is generated based on load, ambient temperature, operating mode, and start / stop status. A topology graph is established based on the connection relationships of equipment components, mapping sensors to topology nodes. Then, the operating status sequence is input into an improved OmniAnomaly variational recursive network to obtain structural latent vectors and degenerate latent vectors. The structural latent vectors are used to reconstruct the operating status sequence, while the degenerate latent vectors characterize the equipment degradation trend. Next, the working condition vector sequence is input into a working condition safety manifold mapping subnetwork. The degenerate safety manifold is constructed from the manifold center vector and the manifold scale vector. Safety manifold constraints are applied to the degenerate latent vectors to make their distribution more stable under the influence of working conditions. Simultaneously, time-monotonic degradation constraints are incorporated to ensure that the degenerate latent vectors degenerate along time... The dimensions exhibit continuous monotonic changes, reducing misjudgments caused by fluctuations in operating conditions. To identify potentially high-risk components in the equipment, the degradation potential vector sequence is input into the topology risk diffusion module, allowing degradation features to propagate across the equipment topology map and form a node risk distribution. Subsequently, a joint loss function is constructed based on reconstruction error, safety manifold deviation, node risk diffusion deviation, health deviation, and fault lead time deviation. The loss weights are dynamically adjusted based on the fault lead time, giving higher weights to samples close to degradation extremes during training. This enables the model to more sensitively capture early degradation signs. Finally, the new operating state sequence is input into the trained operating state diagnostic model, outputting degradation potential vectors, node risk distribution, health, fault lead time, and anomaly score. The anomaly score determines the abnormal time period, and a quantitative risk assessment of the equipment status is performed based on the trends of high-risk nodes and health changes.

[0035] To verify the advantages of this invention, a comparison was made between the traditional threshold-based method, the reconstruction error method based on autoencoders, and the diagnostic method proposed in this invention. The comparison focused on the differences among the three methods in terms of anomaly detection accuracy, degradation trend identification accuracy, high-risk node localization capability, and fault lead prediction error. The comparison results are shown in Table 1. Table 1. Performance Comparison of Three Methods in Engineering Equipment Operation Status Diagnosis Tasks

[0036] Experimental results show that the traditional threshold method has a high false alarm rate in scenarios with frequent operating condition changes. The autoencoder method can alleviate the false alarm problem to a certain extent, but it is not sensitive enough to the slow degradation process. This invention introduces the operating condition safety manifold and time monotonic degradation constraint to make the degradation potential vector more robust to changes in operating conditions. The topological risk diffusion module identifies the risk propagation path on the equipment structure, improves the accuracy of high-risk node location, and the joint loss training mechanism makes the model more sensitive to samples in the later stage of degradation, thereby significantly improving the accuracy of fault advance prediction.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based engineering equipment operation state diagnosis system, characterized by, include: The operation status acquisition and working condition construction module is used to acquire sensor operation status signals to generate operation status sequences, and generate working condition vector sequences based on load, speed, environmental parameters and maintenance status. The topology graph construction module is used to construct the device topology graph based on the component connection relationships and sensor installation locations, and to establish the mapping relationship between sensors and topology nodes; An improved OmniAnomaly variational recursive network module is used to receive the running state sequence and output the structural latent vector, the degenerate latent vector, the reconstructed running state sequence, the health sequence, and the fault advance sequence; The working condition safety manifold mapping module is used to receive the working condition vector sequence and output the degenerate safety manifold parameters, and apply safety manifold constraints and time monotonic degradation constraints to the degenerate potential vectors to generate a degenerate potential vector sequence. The topology risk diffusion module is used to receive the degraded potential vector sequence and the device topology map, and perform discrete risk diffusion on the device topology map to generate a node risk distribution sequence; The joint loss training module is used to construct a joint loss function based on the running state sequence, the reconstructed running state sequence, the degraded potential vector sequence, the degraded safe manifold parameters, the node risk distribution sequence, the health sequence, and the failure advance sequence. The loss weight is dynamically adjusted according to the failure advance sequence to train the improved OmniAnomaly variational recursive network module and the topology risk diffusion module to obtain the running state diagnosis model. The Operation Status Diagnosis and Risk Assessment module is used to input new operation status sequences and operating condition vector sequences into the operation status diagnosis model within the monitoring period, and output degradation potential vector sequences, node risk distribution sequences, health sequences, failure lead sequences, and anomaly score sequences. Based on the anomaly score sequences, it determines the abnormal time periods and operation risk levels. 2.The deep learning-based engineering equipment operation state diagnosis system according to claim 1, characterized in that, The modules are connected in the following way: Collect sensor operating status sequences, construct operating condition vector sequences, and build a topology map based on component connection relationships; An improved OmniAnomaly variational recurrent network is constructed based on the running state sequence, which encodes the running state sequence as structural latent vectors and degenerate latent vectors; Input the sequence of operating condition vectors into the operating condition safety manifold mapping subnetwork to generate a degenerate safety manifold, and apply safety manifold constraints and time monotonic degradation constraints to the degenerate potential vectors; Input the topology graph and the degenerate potential vector into the topology risk diffusion module, and discretize the risk diffusion on the topology graph to obtain the node risk distribution; Input the structural latent vector and the degradation latent vector into the decoder, reconstruct the running state sequence, and output the health status and failure advance. A joint loss function is constructed based on reconstruction error, safety manifold deviation, node risk distribution, health status and fault lead time. The loss weight is dynamically adjusted according to the fault lead time. An improved OmniAnomaly variational recurrent network and topology risk diffusion module are trained to obtain the operational status diagnosis model. The new operating status sequence is input into the operating status diagnostic model, which outputs a degradation potential vector and anomaly score. The abnormal time period and operating risk level are determined based on the anomaly score. 3.The deep learning-based engineering equipment operation state diagnosis system according to claim 2, characterized in that, The improved OmniAnomaly variational recurrent network includes: The running state encoding unit receives the running state sequence and generates a potential state vector through a recursive structure. The subspace decomposition unit divides the potential state vector into structural potential vectors and degenerate potential vectors; The running state decoding unit receives the structural latent vector and the degenerate latent vector and reconstructs the running state sequence; The health and fault lead output unit generates health and fault lead based on the degradation potential vector.

4. The engineering equipment operation status diagnosis system based on deep learning according to claim 2, characterized in that, The construction of the improved OmniAnomaly variational recurrent network specifically includes: Arrange the running state sequence according to the sampling time order to obtain a running state vector sequence that corresponds one-to-one with the sampling time. Each running state vector includes a preset number of running state features. The sequence of running state vectors is input into the running state encoding unit. At each sampling time point, the running state encoding unit receives the current running state vector and the internal state of the previous sampling time point, and outputs the potential state vector corresponding to the sampling time point to form a sequence of potential state vectors. The potential state vector sequence is input into the subspace decomposition unit. At each sampling time point, the subspace decomposition unit divides the potential state vector into structural potential vector and degenerate potential vector, resulting in a structural potential vector sequence and a degenerate potential vector sequence.

5. The engineering equipment operation status diagnosis system based on deep learning according to claim 2, characterized in that, The processing of the working condition safety manifold mapping sub-network specifically includes: The working condition vector sequence is input into the working condition safety manifold mapping subnetwork in the order of sampling time. The working condition safety manifold mapping subnetwork outputs the manifold center vector and manifold scale vector at each sampling time. The manifold center vector and manifold scale vector constitute the degenerate safety manifold parameter sequence. At each sampling time, the degenerate latent vector and the corresponding degenerate safe manifold parameters are read. The safe manifold deviation is calculated based on the distance between the degenerate latent vector and the manifold center vector and the manifold scale vector. When the safe manifold deviation exceeds a preset threshold, the degenerate latent vector is adjusted and mapped to the degenerate safe manifold range to generate a sequence of degenerate latent vectors that satisfy the safe manifold constraints. A time-monotonic degradation constraint is established on the degenerate latent vector sequence. At each sampling time, the current degenerate latent vector is compared with the degenerate latent vector at the previous sampling time. In the dimension of degradation index, the value of the current degenerate latent vector is restricted to be no less than the value of the degenerate latent vector at the previous sampling time, thus obtaining a degenerate latent vector sequence that satisfies the safe manifold constraint and the time-monotonic degradation constraint.

6. The engineering equipment operation status diagnosis system based on deep learning according to claim 2, characterized in that, The generation of the node risk distribution specifically includes: At each sampling time, the node index and component connection relationship are read from the topology graph, and the components of each dimension in the degradation potential vector are assigned to the nodes according to the sensor-node mapping relationship to form a node degradation vector that corresponds one-to-one with the node index. The node degradation vector is input into the topology risk diffusion module, which includes a risk initialization unit, a diffusion coefficient storage unit, and a risk iteration unit. The risk initialization unit calculates the initial risk intensity of the node based on the node degradation vector and the edge weights in the topology graph. The diffusion coefficient storage unit registers the diffusion coefficient for each edge in the topology graph. At each diffusion step, the risk iteration unit updates the node risk intensity based on the current node risk intensity, edge weights, and diffusion coefficients. The current node risk intensity is then distributed to adjacent nodes according to the edge weights and diffusion coefficients and added to the risk intensity of adjacent nodes to form an iterative sequence of node risk intensity arranged according to the diffusion step. When the number of diffusion steps reaches the preset number of diffusion steps, the node risk intensity obtained from the last iteration is selected from the node risk intensity iteration sequence, and the node risk distribution at the current sampling time is obtained by arranging them according to the node index. The node risk distributions corresponding to all sampling times are arranged in chronological order to form a node risk distribution sequence.

7. The engineering equipment operation status diagnosis system based on deep learning according to claim 2, characterized in that, The generation of the health status and failure advance rate specifically includes: The structural latent vector sequence and the degenerate latent vector sequence are input into the running state decoding unit in the order of sampling time. At each sampling time, the running state decoding unit receives the corresponding structural latent vector, degenerate latent vector and internal state of the previous sampling time, generates a new internal state through recursive calculation, and generates a reconstructed running state vector corresponding to the sampling time at the output layer. All reconstructed running state vectors are arranged in the order of sampling time to form a reconstructed running state sequence. The degraded potential vector sequence is input into the health mapping unit in the health and fault advance output unit. The health mapping unit receives the corresponding degraded potential vector at each sampling time, performs linear transformation and nonlinear activation operation on each component of the degraded potential vector, and outputs a health value within a preset value range. All health values ​​are arranged in the order of sampling time to form a health sequence. The degraded potential vector sequence is input into the fault advance mapping unit in the health and fault advance output unit. The fault advance mapping unit receives the corresponding degraded potential vector and sampling time interval parameter at each sampling time, calculates the predicted fault time based on the degraded potential vector, and calculates the fault advance based on the number of sampling time intervals between the predicted fault time and the current sampling time. All fault advances are arranged in the order of sampling time to form a fault advance sequence.

8. The engineering equipment operation status diagnosis system based on deep learning according to claim 2, characterized in that, The generation of the operational status diagnostic model specifically includes: A reconstruction error term is constructed based on the running state sequence and the reconstructed running state sequence. The reconstruction error term represents the sum of squared differences between the running state vector and the reconstructed running state vector in each feature dimension at each sampling time. A safe manifold deviation term is constructed based on the degenerate latent vector sequence and the degenerate safe manifold parameter sequence. The safe manifold deviation term represents the degree of deviation between the degenerate latent vector and the manifold center vector in the degenerate safe manifold at each sampling time, which is obtained by weighting the deviation by the manifold scale vector. A node risk diffusion bias term is constructed based on the node risk distribution sequence. The node risk diffusion bias term represents the weighted aggregation result of the risk intensity of all nodes in the node risk distribution in the time axis and the topological node dimension at each sampling time. Based on the health status sequence and the fault lead time sequence, a health status deviation term and a fault lead time deviation term are constructed. The health status deviation term represents the squared difference between the health status and the preset health reference value at each sampling time, and the fault lead time deviation term represents the squared difference between the fault lead time and the preset target lead time at each sampling time. The reconstruction error term, safety manifold deviation term, node risk diffusion deviation term, health deviation term, and fault lead time deviation term are weighted and summed according to preset baseline weights to form a joint loss function. The weights of each deviation term are dynamically adjusted on the time axis according to the fault lead time sequence, so that the weights of the deviation terms corresponding to the sampling time close to the predicted fault time are increased. The joint loss function is used to train the parameters of the improved OmniAnomaly variational recursive network and the parameters of the topology risk diffusion module to obtain the operational status diagnosis model.

9. The deep learning-based engineering equipment operation status diagnosis system according to claim 2, characterized in that, The determination of the abnormal time period and operational risk level specifically includes: During the monitoring period, sensor operating status sequences and operating condition vector sequences are collected. The new operating status sequences and operating condition vector sequences are input into the operating status diagnosis model to obtain the degradation potential vector sequence, node risk distribution sequence, health sequence, fault lead time sequence and anomaly score sequence arranged according to sampling time. Based on the comparison results between the abnormal scoring sequence and the preset abnormal scoring threshold, mark whether each sampling time is an abnormal sampling time, and merge the abnormal sampling times with consecutive sampling times into an abnormal time period. Within each abnormal time period, the abnormal score sequence, node risk distribution sequence, health sequence, and fault lead time sequence are read and matched with the preset operational risk classification standard to determine the operational risk level corresponding to the abnormal time period, and the abnormal time period and operational risk level are output.

Images