Production automation equipment fault diagnosis and detection system

Abstract

The invention discloses a fault diagnosis and detection system for production automation equipment. The fault diagnosis and detection system comprises a data sensing layer which is used for carrying out multi-mode signal acquisition and real-time preprocessing; the feature extraction layer is used for constructing a recursive block convolution module, capturing transient impact features in four time steps by using an L1-layer gating convolution unit, associating a 16-time-step cross-block periodic degradation mode with an L2-layer sparse attention mechanism, aggregating multi-sensor spatial-temporal features by using an L3-layer global context node, and performing multi-scale feature extraction; the causal reasoning layer is used for establishing a physical constraint driven causal graph engine and outputting a fault propagation path with probability weight; the state modeling layer is used for constructing a continuous health evolution model by adopting a Shenchang differential equation, embedding a physical constraint loss function, and performing equipment full life cycle health state prediction and residual service life estimation in combination with a three-stage memory fusion mechanism of LSTM short-term memory, differentiable neural dictionary medium-term memory and knowledge graph long-term memory; and the decision support layer is used for generating a personalized maintenance work order.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment fault detection, and in particular to a production automation equipment fault diagnosis and detection system. BACKGROUND

[0002] At present, production automation equipment fault diagnosis and detection has become a core demand to improve the competitiveness of manufacturing industry. With the evolution of industrial equipment towards high precision, high integration and high load rate, the traditional maintenance mode based on manual inspection and threshold alarm cannot meet the needs of modern production lines. Unplanned downtime caused by equipment failure can cause tens of thousands of yuan of production capacity loss per hour, and excessive maintenance significantly increases spare parts inventory costs and operation and maintenance manpower investment. Intelligent diagnosis technology based on artificial intelligence and multi-modal sensor fusion has emerged as the times require.

[0003] In the prior art, the running state of power plant equipment is presented by multi-dimensional sensor data stream. These data have strong time sequence correlation. LSTM effectively analyzes the nonlinear dynamics of the evolution of equipment health state over time. However, LSTM controls the retention proportion of cell state through the forgetting gate. However, in long sequence modeling, the model tends to optimize recent loss through gradient descent, causing the forgetting gate to gradually close the early information channel. Therefore, a production automation equipment fault diagnosis and detection system is proposed. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art and to provide a production automation equipment fault diagnosis and detection system.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0006] A production automation equipment fault diagnosis and detection system comprises:

[0007] The data perception layer adopts an industrial Internet of Things sensor network and an edge computing node, combines wavelet denoising and moving average filtering technology to perform multi-modal signal acquisition and real-time preprocessing, and synchronizes data through IEEE 1588 protocol to generate structured time-frequency domain features and working condition labels.

[0008] The feature extraction layer constructs a recursive block convolution module, uses L1 layer gated convolution units to capture transient shock features within 4 time steps, L2 layer sparse attention mechanism to associate 16 time step cross-block periodic degradation patterns, L3 layer global context nodes to aggregate multi-sensor spatio-temporal features, and simultaneously generates redundant connections based on device working conditions through a dynamic sparse mask generator to perform multi-scale feature extraction.

[0009] Causal reasoning layer: build a physical constraint driven causal graph engine, learn the directed acyclic graph topology of fundamental cause variables, observed variables and disturbance variables through structural equation model, integrate counterfactual reasoning module to calculate intervention effect, generate counterfactual samples, and output fault propagation path with probability weight;

[0010] State modeling layer: adopt neural ordinary differential equation to construct continuous health evolution model, embed physical constraint loss function, combine three-level memory fusion mechanism of LSTM short-term memory, differentiable neural dictionary medium-term memory and knowledge graph long-term memory to perform device full life cycle health state prediction and remaining useful life estimation;

[0011] Decision support layer: integrate reinforcement learning engine to build state-action-reward three-dimensional decision space with device health index, working condition parameter and maintenance cost, generate personalized maintenance work order, and equip with human-computer interaction interface.

[0012] The above technical scheme further comprises:

[0013] Further, the data perception layer performs multi-modal signal acquisition and real-time preprocessing to generate structured time-frequency domain features and working condition labels, including the following steps:

[0014] Data acquisition: deploy industrial Internet of Things sensors, and use Modbus TCP / IP or OPC UA protocol to transmit raw signals to edge computing nodes;

[0015] Wavelet threshold denoising: use db4 wavelet basis to perform 4-layer wavelet decomposition on the vibration signal, and suppress noise through a soft threshold function, wherein the soft threshold function is expressed as Where λ is a threshold coefficient, σ j is the standard deviation of the jth layer noise, which is estimated by σ j =median(|cD j | / 0.6745), and the signal is reconstructed by inverse wavelet transform;

[0016] Sliding average filtering: exponential weighted moving average is used for slow varying signals;

[0017] IEEE 1588 time synchronization: use path delay calculation through IEEE 1588 master-slave clock architecture and hardware timestamp technology;

[0018] Working condition label generation: read device operating parameters from PLC, and encode one-hot working condition labels after fuzzy logic classification;

[0019] Structured data output: perform short-time Fourier transform on the vibration signal to generate time-frequency spectrogram, and combine time domain statistical features, frequency domain features and working condition labels to form a structured data package, which provides input for the feature extraction layer and the causal reasoning layer.

[0020] Further, the recursive block convolution module is constructed, the L1 layer gated convolution unit is used to capture transient impact features within 4 time steps, the L2 layer sparse attention mechanism is used to associate 16 time step cross-block periodic degradation modes, the L3 layer global context node is used to aggregate multi-sensor spatio-temporal features, and meanwhile, a dynamic sparse mask generator is used to cut redundant connections based on device working conditions, so as to perform multi-scale feature extraction, including the following steps:

[0021] L1 layer transient feature capture: a gated convolution unit is used to perform local feature extraction on 4 time steps of sensor data;

[0022] L2 layer cross-block periodic association: 16 time steps are divided into 4 data blocks, and a sparse self-attention mechanism is used to establish cross-block association, which is represented as Wherein, Q / K / V are query / key / value matrices, d k is a key vector dimension, Mask(·) is a dynamically generated sparse mask, and the mask generation formula is Mask=I(σ(MLP(op_status))>τ), wherein I(·) is an indicator function, τ is a working condition perception threshold, and op_status is a device operating parameter;

[0023] L3 layer global context aggregation: a global context node is constructed, and a graph attention network is used to aggregate multi-sensor spatio-temporal features, which is represented as Wherein, is the i-th node feature output by the L2 layer, W a is a shared weight matrix, and the attention coefficient α ij is generated by the constraint of device physical topology (such as bearing-gear transmission chain);

[0024] Dynamic feature re-labeling: the final feature is re-labeled by a dynamic sparse mask: Wherein, Conv 1×1 is a 1x1 convolution to realize channel compression, and Mask is dynamically generated by a working condition perception generator, and the calculation formula is Mask=σ(W m ·op_status+b m )⊙base_mask, wherein base_mask is a static topology mask, W m and b m are learnable parameters.

[0025] Further, the physical constraint driven causal graph engine is established, the directed acyclic graph topology of root cause variables, observed variables and disturbance variables is learned by a structural equation model, including the following steps:

[0026] Variable definition and assumption space construction: define the set of root cause variables, the set of observed variables and the set of environmental interference variables based on the physical mechanism of the device, and assume causal sufficiency to exclude unobserved confounding variables;

[0027] Structural equation model construction: for each observed variable o j Establish a structural equation, assuming that the observed variables are linearly combined by the root cause variables, environmental interference variables and error terms, represented as Wherein, and are the root cause parent node and the environmental interference parent node of o j , β ji and γ jk are path coefficients, representing the strength of causal effect, ε j ~ Laplace(0, b j ) is a non-Gaussian error term, and the error term independence constraint is imposed

[0028] Causal discovery based on PC algorithm: PC algorithm is used to perform causal discovery, conditional independence is tested by chi-square goodness-of-fit test, redundant edges are removed step by step and the undirected skeleton is determined, and then the directed acyclic graph is generated by using V-structure rule and Meek propagation orientation rule;

[0029] Causal graph output and verification: the model fitting degree is evaluated by AIC criterion, and the optimal causal structure is selected, so as to form an interpretable causal graph model.

[0030] Further, the specific steps of the counterfactual reasoning module for calculating the intervention effect are:

[0031] Counterfactual world construction: hard intervention operation is used to forcibly assign values to target variables, and the original causal chain is cut off by reconstructing the structural equation model;

[0032] Counterfactual sample generation: generate counterfactual samples that meet the physical constraints using a generative adversarial network, and force the counterfactual samples to meet the data distribution and causal mechanism double constraints through a counterfactual consistency loss function;

[0033] Intervention effect quantification: calculate the state change before and after the intervention of the total effect, and decompose the natural direct effect and the total indirect effect to distinguish the direct and indirect influence paths;

[0034] Fault propagation path probability update: update the posterior distribution of path coefficients based on Bayesian inference, dynamically adjust the causal propagation probability with weight, and thus construct the causal evidence chain from virtual intervention to real decision.

[0035] Further, the specific steps of the output fault propagation path with probability weight are:

[0036] Path decomposition: decompose the causal graph into all possible paths for any observed variable o j , the probability distribution is decomposed as where n j is the number of parent nodes of o j ;

[0037] Path strength quantification: calculate the contribution degree of each causal path c i →o j ;

[0038] Bayesian belief propagation: perform probabilistic inference on the causal graph, denoted as Calculate the posterior probability through the message passing algorithm;

[0039] Output the weighted causal graph: for each directed edge c i →o j , calculate its propagation probability, generate a set of weighted fault propagation paths Π={(c i →o j →...→o k ),w i→k}, where the weight w i→k is the product of all edge probabilities on the path.

[0040] Further, the state modeling layer performs the specific steps of device full life cycle health state prediction and remaining useful life estimation:

[0041] Continuous health state modeling: build a continuous dynamics model centered on the physical degradation law of the device, denoted as where h(t)∈R d is the device health state vector, x(t) is the multi-modal observation data, f θ (·) is a learnable neural network parameterized function, which uses a parameterized neural network to describe the nonlinear evolution process of the health state vector over time, and realizes high-precision state advance through a Dormand-Prince numerical solver, while converting the Paris crack propagation law into a regularization term embedded in the loss function;

[0042] Three-level memory fusion mechanism: establish a three-level memory fusion architecture, the short-term memory unit uses an LSTM network to capture transient degradation features in minute-level time series data, the medium-term memory stores event-level memory through a differentiable neural dictionary, and the long-term memory correlates device full life cycle data through a knowledge graph;

[0043] Memory fusion and state update: feature fusion through cross-time scale attention mechanism, integration of continuous evolution state and discrete memory information through feedforward network, generation of health index fused with physical law and data-driven, and mapping of health index to probability distribution of remaining useful life by Weibull proportional risk model.

[0044] Further, the decision support layer generates a specific step of personalized maintenance work order:

[0045] State space modeling: the health index of the equipment, the working condition parameters and the maintenance cost are quantitatively fused through a multi-dimensional state vector modeling, and the Min-Max normalization processing is adopted to eliminate the dimension difference;

[0046] Action space design: a discrete-continuous hybrid action space is designed, the discrete action set is combined, and the continuous parameter action is parameterized;

[0047] Reward function construction: a composite reward function is constructed, a multi-objective optimization mechanism of equipment availability reward item, maintenance cost penalty item (combined with the budget proportion of the expected downtime loss) and operation compliance reward item is adopted, and the agent is guided to learn the optimal strategy;

[0048] Reinforcement learning: a deep Q network architecture is adopted to process high-dimensional state input, a priority experience replay mechanism is used to store key state transition samples, and a target network soft update and an epsilon-greedy exploration strategy decay are adopted to balance exploration and utilization;

[0049] Maintenance work order generation: according to the current state, the action combination maximizing the Q value is selected, a structured work order containing specific operation instructions, priority label and safety notes is output, and the decision basis is displayed through a visual interface and a manual confirmation link is set.

[0050] The present application has the following beneficial effects:

[0051] In the present application, the neural ordinary differential equation realizes continuous modeling of the health state of the equipment through an ordinary differential equation solver, decouples the discrete time step into a continuous evolution process, and can accurately capture slow physical processes. Combined with a three-level memory fusion mechanism, the LSTM processes the minute-level vibration sequence, the differentiable neural dictionary stores the daily event features, and the knowledge graph precipitates the monthly maintenance experience, forming a "transient-event-history" full life cycle memory chain. In addition, the recursive block convolution module decouples the time series feature extraction into three dimensions: the L1 layer captures transient impact and avoids short-term noise from hiding early fault signs; the L2 layer excavates periodic degradation patterns; and the L3 layer constructs a device-level health portrait. The recursive block convolution module has transient perception and monthly trend analysis capabilities. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1A system block diagram of a production automation equipment fault diagnosis detection system is provided. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0054] Please refer to Figure 1 The present application is a production automation equipment fault diagnosis detection system, which comprises:

[0055] The data perception layer: industrial Internet of Things sensor network and edge computing nodes are adopted, combined with wavelet denoising and moving average filtering technology, to perform multi-modal signal acquisition and real-time preprocessing, and data synchronization is performed through IEEE 1588 protocol, to generate structured time-frequency domain features and working condition labels.

[0056] The feature extraction layer: a recursive block convolution module is constructed, L1 layer gate convolution units are used to capture transient impact features within 4 time steps, L2 layer sparse attention mechanism is used to associate 16 time step cross-block periodic degradation modes, L3 layer global context nodes are used to aggregate multi-sensor spatio-temporal features, and at the same time, a dynamic sparse mask generator is used to generate redundant connections based on device working conditions, to perform multi-scale feature extraction.

[0057] The causal reasoning layer: a physically constrained causal graph engine is established, a directed acyclic graph topology of root cause variables, observed variables and interference variables is learned through a structural equation model, an intervention effect is calculated by integrating a counterfactual reasoning module, counterfactual samples are generated, and a fault propagation path with probability weight is output.

[0058] The state modeling layer: a neural ordinary differential equation is used to construct a continuous health evolution model, a physical constraint loss function such as Paris crack propagation law is embedded, a three-level memory fusion mechanism of LSTM short-term memory, differentiable neural dictionary medium-term memory and knowledge graph long-term memory is combined, and device full life cycle health state prediction and remaining useful life estimation are performed.

[0059] The decision support layer: a reinforcement learning engine is integrated, a state-action-reward three-dimensional decision space is constructed by using device health index, working condition parameters and maintenance cost, personalized maintenance work orders containing lubrication period adjustment, spare parts replacement and other operations are generated, and a man-machine interaction interface is provided to support artificial coverage confirmation.

[0060] In one embodiment, the data perception layer performs multi-modal signal acquisition and real-time preprocessing, generating structured time-frequency domain features and working condition labels, including the following steps:

[0061] Data acquisition: Deploy industrial Internet of Things sensors such as vibration accelerometers, infrared thermometers, and current clamp meters, etc. Use Modbus TCP / IP or OPC UA protocol to transmit raw signals to edge computing nodes, realize multi-modal physical quantity acquisition of key parts of equipment (such as bearing seat, motor winding);

[0062] Wavelet threshold denoising: Use db4 wavelet basis to decompose the vibration signal into 4 layers, and suppress noise through a soft threshold function, which is expressed as Where λ is the threshold coefficient, σ j is the standard deviation of the jth layer noise, which is estimated by σ j = median(|cD j | / 0.6745), and then the signal is reconstructed by inverse wavelet transform, which improves the signal-to-noise ratio of the original vibration signal from 12 dB to 25 dB while preserving the 120 Hz feature frequency amplitude;

[0063] Exponential weighted moving average filtering: Exponential weighted moving average filtering is used for slow-changing signals such as temperature and current;

[0064] IEEE 1588 time synchronization: IEEE 1588 master-slave clock architecture and hardware timestamp technology are used to calculate the path delay;

[0065] Working condition label generation: Read the load rate, oil temperature and other equipment operating parameters from PLC, and encode them into one-hot working condition labels (such as high load + normal oil temperature represented as [0, 0, 1; 1, 0, 0]) after fuzzy logic classification;

[0066] Structured data output: Perform short-time Fourier transform on the vibration signal to generate time-frequency spectrogram, and combine time domain statistical features, frequency domain features and working condition labels to form a structured data package, providing input for the feature extraction layer and causal reasoning layer.

[0067] In one embodiment, the recursive block convolution module is constructed, which uses L1 layer gated convolution unit to capture transient shock features within 4 time steps, L2 layer sparse attention mechanism to associate 16 time step cross-block periodic degradation patterns, L3 layer global context node to aggregate multi-sensor spatio-temporal features, and simultaneously generate dynamic sparse mask generator based on device working condition to cut redundant connections, for multi-scale feature extraction, including the following steps:

[0068] L1 layer transient feature capture: Gated Convolutional Unit (GCU) is used to extract local features from 4 time steps of sensor data;

[0069] L2 layer cross-block periodic association: 16 time steps are divided into 4 data blocks, and cross-block association is established through sparse self-attention mechanism, which is represented as wherein Q / K / V are query / key / value matrices, d k is the key vector dimension, Mask(·) is a dynamically generated sparse mask, and the mask generation formula is Mask = I(σ(MLP(op_status)) > τ), wherein I(·) is an indicator function, τ is a working condition perception threshold (such as activating the cross-block connection when the load rate is greater than 80%), and op_status is a device operating parameter (rotational speed, temperature, etc.), and the mechanism selectively retains the periodic degradation mode related to the current working condition (such as gear meshing frequency harmonics);

[0070] L3 layer global context aggregation: a global context node (GCN) is constructed, and multi-sensor spatio-temporal features are aggregated through a graph attention network (GAT), which is represented as wherein is the i-th node feature output by the L2 layer, W a is a shared weight matrix, and the attention coefficient α ij is generated by the device physical topology (such as a bearing-gear transmission chain) constraint;

[0071] Dynamic feature re-labeling: the final feature is re-labeled through a dynamic sparse mask: wherein Conv 1×1 is a 1x1 convolution to realize channel compression, and Mask is dynamically generated by a working condition perception generator, and the calculation formula is Mask = σ(W m ·op_status + b m )⊙base_mask, wherein base_mask is a static topology mask (such as a bearing node and a temperature sensor forced association), W m and b m are learnable parameters.

[0072] In one embodiment, the establishment of the physical constraint driven causal graph engine learns the directed acyclic graph topology of the root cause variable, the observed variable and the disturbance variable through a structural equation model, including the following steps:

[0073] Variable definition and hypothesis space construction: based on the device physical mechanism, a set of root cause variables (such as bearing wear δ, gear engagement stiffness k) and a set of observed variables (such as vibration energy Ev, temperature rise rate ) and environmental interference variables (e.g. load shock Fimp), and assumes causal sufficiency to exclude unobserved confounders, assuming all common cause variables are observed, i.e. there are no unobserved confounders, thus ensuring identifiability of causal discovery;

[0074] Structural equation modeling (SEM) construction: for each observed variable o j Establish structural equation, assume that the observed variables are linearly combined by the root cause variables, environmental interference variables and error terms, represented as Wherein, and are the root cause parent node and environmental interference parent node of o j , β ji and γ jk are path coefficients, representing causal effect strength, ε j ~ Laplace (0, b j ) is a non-Gaussian error term, and the error term independence constraint ensures model identifiability;

[0075] Causal discovery based on PC algorithm: PC algorithm is used to perform causal discovery, conditional independence is tested by chi-square goodness-of-fit test, redundant edges are removed step by step and the undirected skeleton is determined, and then the directed acyclic graph is generated by using V-structure rule (such as identifying x→z←y mode) and Meek propagation orientation rule (such as avoiding directional propagation of directed ring);

[0076] Causal graph output and verification: the model fitting degree is evaluated by AIC criterion (AIC = -2lnL + 2k) to select the optimal causal structure, thus forming an interpretable causal graph model.

[0077] In one embodiment, the counterfactual reasoning module calculates the specific steps of the intervention effect;

[0078] Counterfactual world construction: hard intervention operation is used to assign values to target variables, and the original causal chain is cut off by reconstructing the structural equation model;

[0079] Counterfactual sample generation: generate counterfactual samples that meet physical constraints using a generative adversarial network, and force the counterfactual samples to meet the dual constraints of data distribution and causal mechanism through a counterfactual consistency loss function (including a prediction consistency term and a gradient penalty term);

[0080] Intervention effect quantification: calculate the total effect (TE) to quantify the state change before and after the intervention, and decompose the natural direct effect (NDE) and the total indirect effect (TIE) to distinguish direct and indirect influence paths;

[0081] Fault propagation path probability updating: updating the path coefficient posterior distribution based on Bayesian inference, dynamically adjusting the weighted causal propagation probability, thereby constructing the causal evidence chain from virtual intervention to real decision, and realizing the explainable root cause tracing of equipment failure.

[0082] In one embodiment, the specific steps of outputting the fault propagation path with probability weight are:

[0083] Path decomposition: decomposing the probability distribution of any observed variable o j into where n j is the number of parent nodes of o j .

[0084] Path strength quantification: calculating the contribution degree of each causal path c i →o j .

[0085] Bayesian belief propagation: performing probability inference on the causal graph, denoted as Calculate the posterior probability through the message passing algorithm (such as Sum-Product);

[0086] Output weighted causal graph: for each directed edge c i →o j , calculate its propagation probability, generate a set of weighted fault propagation paths Π={(c i →o j →...→o k ),w i→k}, where the weight w i→k is the product of the probabilities of all edges on the path.

[0087] In one embodiment, the state modeling layer performs the following specific steps for device full life cycle health state prediction and remaining useful life estimation:

[0088] Continuous health state modeling: construct a continuous dynamics model centered on the physical degradation law of the device, denoted as where h(t)∈R d is the device health state vector (including crack length a, wear δ, etc. Physical parameters), x(t) is multi-modal observation data (vibration, temperature, current), f θ (·) is a learnable neural network parameterized function, which uses a parameterized neural network to describe the nonlinear evolution process of the health state vector over time, and implements high-precision state advancement through the Dormand-Prince numerical solver, while converting the Paris crack propagation law into a regularization term embedded loss function to ensure that the model prediction conforms to the material fatigue mechanism;

[0089] Three-level memory fusion mechanism: Establish a three-level memory fusion architecture, the short-term memory unit uses LSTM network to capture the transient degradation characteristics in the minute-level time series data (such as the periodic shift of vibration spectrum), the medium-term memory stores event-level memory (such as lubrication records, overload impact events) through differentiable neural dictionary (DND), and the long-term memory associates device life cycle data (including historical maintenance cases, failure modes of similar devices) through knowledge graph;

[0090] Memory fusion and state update: Feature fusion is performed through cross-time scale attention mechanism, and continuous evolution state and discrete memory information are integrated through feedforward network to generate health index (HI) that combines physical law and data-driven, and Weibull proportional risk model is used to map health index to probability distribution of remaining useful life (RUL).

[0091] In one embodiment, the decision support layer generates specific steps of the personalized maintenance work order:

[0092] State space modeling: The device health index, working condition parameters and maintenance cost are quantitatively fused through multi-dimensional state vector modeling, and Min-Max normalization processing is used to eliminate dimension difference;

[0093] Action space design: A discrete-continuous hybrid action space including lubrication period adjustment, spare parts replacement and other operations is designed, the discrete action set is combined, and the continuous parameter action is parameterized;

[0094] Reward function construction: A composite reward function is constructed, a multi-objective optimization mechanism of device availability reward item (weighted calculation based on the difference between health index and emergency threshold), maintenance cost penalty item (budget proportion combined with actual cost and expected downtime loss) and operation compliance reward item (statistical number of safe procedure compliant actions) is used to guide the agent to learn the optimal strategy;

[0095] Reinforcement learning: Deep Q network (DQN) architecture is used to process high-dimensional state input, priority experience replay mechanism is used to store key state transition samples, and target network soft update and ε-greedy exploration strategy decay are used to balance exploration and utilization;

[0096] Maintenance work order generation: According to the current state, the action combination that maximizes the Q value is selected, and a structured work order (such as JSON format) containing specific operation instructions, priority label and safety notes is output, at the same time, the decision basis (Q value distribution, causal reasoning result) is displayed through the visualization interface and the human confirmation link is set to ensure that the key operation needs to be verified by the engineer again.

[0097] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A production automation equipment failure diagnostic detection system characterized by, Comprise: Data perception layer: adopt industrial internet of things sensor network and edge computing node, combine wavelet denoising and moving average filtering technology for multi-modal signal acquisition and real-time preprocessing, and generate structured time-frequency domain features and working condition labels through IEEE 1588 protocol data synchronization; Feature extraction layer: construct recursive block convolution module, use L1 layer gated convolution unit to capture transient impact features within 4 time steps, L2 layer sparse attention mechanism to associate 16 time step cross-block periodic degradation mode, L3 layer global context node to aggregate multi-sensor spatio-temporal features, and through dynamic sparse mask generator to generate redundant connections based on device working condition, for multi-scale feature extraction; Causal reasoning layer: establish a physically constrained causal graph engine, learn the directed acyclic graph topology of the root cause variable, observed variable and interference variable through structural equation model, integrate counterfactual reasoning module to calculate intervention effect, generate counterfactual samples, and output fault propagation path with probability weight; State modeling layer: use neural ordinary differential equation to construct continuous health evolution model, embed physical constraint loss function, combine LSTMs short-term memory, differentiable neural dictionary medium-term memory and knowledge graph long-term memory three-level memory fusion mechanism, for device full life cycle health state prediction and remaining useful life estimation; Decision support layer: integrate reinforcement learning engine to construct state-action-reward three-dimensional decision space with device health index, working condition parameters and maintenance cost, generate personalized maintenance work order, and equip with human-computer interaction interface.

2. The production automation equipment failure diagnostic detection system according to claim 1, wherein, The data perception layer carries out multi-modal signal acquisition and real-time preprocessing, generates structured time-frequency domain features and working condition labels, including the following steps: Data acquisition: deploy industrial internet of things sensors, use Modbus TCP / IP or OPC UA protocol to transmit raw signals to edge computing nodes; Wavelet threshold denoising: the vibration signal is decomposed by db4 wavelet base to 4 layers, the noise is suppressed by soft threshold function, and the soft threshold function is expressed as Wherein λ is a threshold coefficient, σ j is the standard deviation of the jth layer noise, which is estimated by σ j =median(|cD j | / 0.6745), and the signal is reconstructed by inverse wavelet transform. Moving average filtering: exponential weighted moving average for slow varying signals; IEEE 1588 time synchronization: use path delay calculation through IEEE 1588 master-slave clock architecture and hardware timestamp technology; Working condition label generation: read device running parameters from PLC, encode as one-hot working condition label after fuzzy logic classification; Structured data output: perform short-time Fourier transform on vibration signals to generate time-frequency spectrogram, and combine time domain statistical features, frequency domain features and working condition labels to form structured data package, providing input for feature extraction layer and causal reasoning layer.

3. The system for diagnosing and detecting faults of a production automation device according to claim 1, wherein The recursive block convolution module is constructed, L1 layer gated convolution unit is used to capture transient impact features within 4 time steps, L2 layer sparse attention mechanism is used to associate 16 time step cross-block periodic degradation mode, L3 layer global context node is used to aggregate multi-sensor spatio-temporal features, and through dynamic sparse mask generator to generate redundant connections based on device working condition, for multi-scale feature extraction, including the following steps: L1 layer transient feature capture: use gated convolution unit to extract local features of 4 time steps of sensor data; L2 layer cross-block periodic association: 16 time steps are divided into 4 data blocks, and cross-block association is established through sparse self-attention mechanism, represented as wherein Q / K / V are query / key / value matrices, d k is the key vector dimension, Mask(·) is a dynamically generated sparse mask, and the mask generation formula is Mask = I(σ(MLP(op_status))>τ), wherein I(·) is an indicator function, τ is a working condition perception threshold, and op_status is a device operating parameter; L3 layer global context aggregation: build global context nodes, aggregate multi-sensor spatio-temporal features through graph attention network, represented as where, is the i-th node feature output by the L2 layer, W a is a shared weight matrix, and the attention coefficient a ij is generated by the constraints of the physical topology of the device (such as the bearing-gear transmission chain); Dynamic feature re-labeling: the final feature is re-labeled by a dynamic sparse mask: where Conv 1×1 is a 1x1 convolution to realize channel compression, Mask is dynamically generated by the working condition perception generator, and the calculation formula is Mask = σ(W m ·op_status + b m )⊙base_mask, where base_mask is a static topology mask, W m and b m are learnable parameters.

4. The system for diagnosing and detecting faults of a production automation device according to claim 1, wherein The establishing physical constraint driven causal graph engine learns the directed acyclic graph topology of the root cause variables, observed variables and disturbance variables through structural equation modeling, including the following steps: Variable definition and hypothesis space construction: define the set of root cause variables, the set of observed variables and the set of environmental disturbance variables based on the physical mechanism of the equipment, and assume causal sufficiency to exclude unobserved confounding variables; Structural equation modeling: For each observed variable o j The structural equation is established, and the observed variable is linearly combined by the root cause variable, the environmental disturbance variable and the error term, which is expressed as Wherein, and are the root cause parent node and the environmental disturbance parent node of o j , β ji and γ jk are path coefficients, representing the strength of causal effect, ε j ~ Laplace(0, b j ) is a non-Gaussian error term, and the error term independence constraint is imposed Causal discovery based on PC algorithm: perform causal discovery using the PC algorithm, perform conditional independence test through chi-square goodness-of-fit test, gradually delete redundant edges and determine the undirected skeleton, and then generate a directed acyclic graph using the V-structure rule and Meek propagation orientation rule; Causal graph output and verification: evaluate the model goodness-of-fit through the AIC criterion, select the optimal causal structure, and form an interpretable causal graph model.

5. The production automation equipment failure diagnostic detection system of claim 4, wherein, The specific steps of the counterfactual reasoning module for calculating the intervention effect are as follows: Counterfactual world construction: perform forced assignment on the target variable using hard intervention operation, and cut off the original causal chain by reconstructing the structural equation model; Counterfactual sample generation: generate counterfactual samples that meet the physical constraints using a generative adversarial network, and force the counterfactual samples to meet the data distribution and causal mechanism constraints through a counterfactual consistency loss function; Intervention effect quantification: calculate the state change before and after the intervention, and decompose the natural direct effect and total indirect effect to distinguish direct and indirect influence paths; Fault propagation path probability update: update the posterior distribution of path coefficients based on Bayesian inference, dynamically adjust the weighted causal propagation probability, and thus construct a causal evidence chain from virtual intervention to real decision-making.

6. The production automation equipment failure diagnostic detection system of claim 5, wherein, The specific steps of the output of the fault propagation path with probability weight are as follows: Path decomposition: for any observed variable o j , the probability distribution is decomposed as where n j is the number of parent nodes of o j . Path strength quantification: compute the contribution of each causal path c i → o j to the final outcome. Belief propagation: Probabilistic inference on a causal graph, denoted as Posterior probabilities are computed by a message passing algorithm; Output weighted causal graph: for each directed edge c i → o j , compute its propagation probability, generate a set of weighted fault propagation paths Π = {(c i → o j →... → i k ), w i→k} where the weight w i→k is the product of all edge probabilities on the path.

7. The production automation equipment failure diagnostic detection system of claim 1, wherein, The specific steps of the state modeling layer for device full life cycle health state prediction and remaining useful life estimation are as follows: Continuous health state modeling: a continuous dynamics model is constructed, which is centered on the physical degradation law of the device, and is expressed as where h(t)∈R d is the device health state vector, x(t) is the multi-modal observation data, f θ (·) is a learnable neural network parameterized function, which uses a parameterized neural network to describe the nonlinear evolution process of the health state vector over time, and realizes high-precision state advance through the Dormand-Prince numerical solver, and at the same time, the Paris crack propagation law is transformed into a regularization term embedded in the loss function; Three-level memory fusion mechanism: establish a three-level memory fusion architecture, the short-term memory unit uses an LSTM network to capture transient degradation features in minute-level time series data, the medium-term memory stores event-level memory through a differentiable neural dictionary, and the long-term memory associates device full life cycle data through a knowledge graph; Memory fusion and state update: feature fusion is performed through cross-time scale attention mechanism, continuous evolution state and discrete memory information are integrated through feedforward network, health index is generated by integrating physical law and data-driven, and Weibull proportional risk model is used to map health index to probability distribution of remaining useful life.

8. The production automation equipment failure diagnostic detection system of claim 1, wherein, The specific steps of the decision support layer for generating personalized maintenance work order are as follows: State space modeling: model the device health index, working condition parameters and maintenance cost through multi-dimensional state vector, and use Min-Max normalization to eliminate dimension differences; Action space design: design a discrete-continuous hybrid action space, parameterize the continuous parameter action, and set a discrete action set; Reward function construction: construct a composite reward function, and use a multi-objective optimization mechanism of device availability reward item, maintenance cost penalty item (combined with the budget proportion of actual cost and expected downtime loss) and operation compliance reward item to guide the agent to learn the optimal strategy; Reinforcement Learning: Deep Q-network architecture is adopted to deal with high-dimensional state input, priority experience replay mechanism is used to store key state transition samples, and the balance between exploration and utilization is achieved through the soft update of the target network and the decay of the ε-greedy exploration strategy. Maintenance Order Generation: Select the action combination that maximizes the Q value based on the current state, output a structured order containing specific operation instructions, priority labels, and safety notes, and display the decision basis through a visual interface while setting up a manual confirmation link.

Images