A data and knowledge double-driven non-intrusive equipment group intelligent diagnosis system

Abstract

This invention discloses a non-intrusive intelligent diagnostic system for equipment groups driven by both data and knowledge, relating to the field of industrial equipment condition monitoring technology. The system includes: acquiring non-intrusive multimodal monitoring data of the equipment group during its current operating cycle; generating a spatiotemporal state tensor in this spatiotemporal dimension; performing feature dimensionality reduction extraction through multi-scale convolutional layers in the data feature extraction network to output a data-driven representation vector; outputting a knowledge-driven representation vector by matching entity relationships in a preset graph database; inputting the data-driven representation vector and the knowledge-driven representation vector into a preset cross-attention fusion module for weighted fusion calculation; and determining the target fault classification result of the equipment group based on the calculated fusion feature vector. This invention achieves integrated fault diagnosis and location, avoiding the damage to equipment caused by invasive installation and the efficiency loss of downtime monitoring.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring technology, specifically to a non-intrusive intelligent diagnostic system for equipment clusters driven by both data and knowledge. Background Technology

[0002] In modern industrial manufacturing systems, equipment groups consisting of multiple CNC machine tools, wind turbines, and other heavy rotating machinery are the core carriers of production. In order not to damage the original structure of the equipment, non-invasive monitoring methods are usually used in industrial sites, that is, acoustic emission, vibration, or magnetic flux sensors are deployed on the outside of the equipment casing to collect the operating status data of the equipment.

[0003] Currently, the processing of such non-intrusive monitoring data heavily relies on purely data-driven deep learning models. However, in actual industrial edge computing scenarios, such purely data-driven solutions face a surge in false alarm rates due to a lack of cross-condition feature generalization capabilities. In industrial settings, equipment often experiences unknown variable loads or speeds. When extracting features, pure data models essentially perform statistical data distribution fitting, completely failing to understand the mechanical logic behind equipment operation, material fatigue evolution, or physical constraints of the transmission chain. When encountering variable operating condition fluctuations not seen during the training phase, the data feature distribution extracted by the model undergoes severe boundary shifts. This shift causes the model to incorrectly identify normal signal bandwidth broadening or energy surges caused by variable loads as faults such as bearing spalling or gear tooth breakage, resulting in extremely high false positive rates. This data processing mechanism, lacking prior knowledge constraints, significantly reduces the usability of non-intrusive monitoring systems in complex industrial settings. Summary of the Invention

[0004] To address the aforementioned technical problems, a non-intrusive intelligent diagnostic system for equipment clusters, driven by both data and knowledge, is provided. This technical solution resolves the problems described above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A non-intrusive intelligent diagnostic system for equipment swarms, driven by both data and knowledge, includes: S101, acquire non-intrusive multimodal monitoring data of the equipment group during the current operating cycle; S102, the non-invasive multimodal monitoring data is mapped to a preset device space topology matrix according to the spatial coordinates of the monitoring points to generate a spatiotemporal state tensor in this spatiotemporal dimension; S103, the spatiotemporal state tensor is input into a preset data feature extraction network, and feature dimensionality reduction is performed through the multi-scale convolutional layer in the data feature extraction network to output a data-driven representation vector; S104, input the node attribute features in the spatiotemporal state tensor into the preset knowledge graph reasoning engine, and output the knowledge-driven representation vector by matching the entity relationships in the preset graph database; S105, the data-driven representation vector and the knowledge-driven representation vector are input into a preset cross-attention fusion module for weighted fusion calculation, and the target fault classification result of the equipment group is determined based on the fusion feature vector output by the calculation.

[0006] Preferably, the step of acquiring non-invasive multimodal monitoring data of the equipment group during the current operating cycle specifically includes: Based on acoustic emission sensors and stray magnetic flux sensors deployed outside the equipment group, raw monitoring signals are collected at a preset sampling frequency. The original monitoring signal is segmented and truncated using a sliding time window of preset length to obtain multimodal monitoring data containing multiple one-dimensional time series arrays.

[0007] Preferably, the step of mapping the non-invasive multimodal monitoring data to a preset device spatial topology matrix according to the spatial coordinates of the monitoring points specifically includes: Based on the physical distances between the devices in the equipment group and the impedance of the vibration transmission medium, a spatial topology matrix of the equipment, consisting of nodes and edges, is constructed. The non-invasive multimodal monitoring data is used as node attribute features to fill the corresponding nodes of the device spatial topology matrix, generating the spatiotemporal state tensor containing spatial adjacency relationships and time series features.

[0008] Preferably, the step of inputting the spatiotemporal state tensor into a preset data feature extraction network specifically includes: The first-scale one-dimensional convolutional kernel and the second-scale one-dimensional convolutional kernel in the data feature extraction network are used to perform sliding inner product operations on the spatiotemporal state tensor, respectively. The feature maps obtained from the computation are processed by a linear rectified activation function and then concatenated along the feature channel dimension to generate the data-driven representation vector.

[0009] Preferably, the step of inputting the node attribute features in the spatiotemporal state tensor into a preset knowledge graph inference engine specifically includes: Extract the frequency band energy peak value from the node attribute features as the query entity; The system traverses the triplet relationships containing the query entity in the preset graph database, extracts the fault representation entities associated with the query entity, and transforms the fault representation entities into multi-dimensional floating-point vectors as the knowledge-driven representation vector.

[0010] Preferably, the step of inputting the data-driven representation vector and the knowledge-driven representation vector into a preset cross-attention fusion module for weighted fusion calculation specifically includes: The data-driven representation vector is linearly transformed to generate a query matrix; The knowledge-driven representation vector is linearly transformed to generate the key matrix and the value matrix; Calculate the dot product of the query matrix and the key matrix, and obtain the attention weight matrix using the normalized exponential function; Multiplying the attention weight matrix by the value matrix yields a data fusion feature vector containing knowledge constraints.

[0011] Preferably, the step of determining the target fault classification result of the equipment group based on the calculated fusion feature vector specifically includes: The data fusion feature vector containing knowledge constraints is input into a preset fully connected classification layer; The fully connected classification layer outputs the probability distribution values ​​of the equipment group under each preset fault category. The fault category corresponding to the highest probability distribution value is determined as the target fault classification result.

[0012] Preferably, after the step of processing the obtained feature map using a linear rectified activation function, the method further includes: The activated feature map is downsampled using max pooling to retain the maximum feature activation value within the local receptive field of the feature map, thereby reducing the spatial resolution dimension of the spatiotemporal state tensor.

[0013] Preferably, the step of converting the fault representation entity into a multi-dimensional floating-point vector specifically includes: The preset graph embedding transformation matrix is ​​invoked to perform index matching between the fault representation entities in discrete character form and the preset entity dictionary; Based on the index value obtained from the matching, the corresponding one-dimensional floating-point vector is extracted from the graph embedding transformation matrix as the knowledge-driven representation vector.

[0014] Preferably, the step of calculating the dot product of the query matrix and the key matrix further includes: The dot product result is scaled by dividing it by the square root of the column dimension of the key matrix to prevent the subsequent normalized exponential function from entering the gradient saturation region during computation.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention introduces a preset knowledge graph of equipment failure mechanisms as a predetermined prior constraint, anchoring the amorphous fluctuation data collected by non-invasive sensors within a feature space that conforms to the mechanical operation logic.

[0016] 2. In the diagnostic logic of this invention, the implicit representation vector of real-time data and the explicit inference vector of prior knowledge are extracted respectively. The cross-attention mechanism is used to make the two interact by dot product in the vector space. The weight of the knowledge vector is used to correct the feature shift of the data vector under varying working conditions.

[0017] 3. In view of the spatial distribution characteristics of equipment groups, this invention constructs a graph-structured topological mapping matrix to map discrete external monitoring point signals into spatiotemporal state tensors with spatial correlation, thereby eliminating signal source aliasing interference caused by non-intrusive monitoring. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the execution flow of a non-intrusive intelligent diagnostic system for equipment clusters driven by both data and knowledge, provided as an embodiment of this disclosure. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the protection scope of this application.

[0020] This application provides a non-intrusive intelligent diagnostic system for equipment clusters driven by both data and knowledge. Please refer to [link / reference]. Figure 1 The implementation method of a non-intrusive intelligent diagnostic system for equipment clusters driven by both data and knowledge, provided in this application embodiment, includes the following steps: Step 101: Collect the raw multimodal monitoring signals of the equipment group through a non-invasive sensor network, and obtain standardized multimodal monitoring data by segmenting and preprocessing through a sliding time window; Step 102: Map the multimodal monitoring data to the preset equipment spatial topology matrix according to the spatial coordinates of the monitoring points, and fuse the time series features to generate a spatiotemporal state tensor; Step 103: Input the spatiotemporal state tensor into the data feature extraction network, and after multi-scale convolution and pooling processing, extract signal features and convert them into data-driven representation vectors; Step 104: Input the spatiotemporal state tensor into the knowledge graph reasoning engine, and transform the fault mechanism knowledge into a knowledge-driven representation vector through entity query and graph embedding transformation. Step 105: Input the data-driven representation vector and the knowledge-driven representation vector into the cross-attention fusion module, and obtain the data fusion feature vector through weighted fusion; Step 106: Input the data fusion feature vector into the fully connected classification layer, output the fault probability distribution, determine the target fault classification result, and output the fault location information.

[0021] The above scheme employs non-invasive sensors to achieve non-stop, non-destructive monitoring of equipment groups, avoiding the impact of invasive installations on equipment structure and operation. It constructs a spatial topology matrix for the equipment and generates a spatiotemporal state tensor, explicitly preserving the spatial distribution and temporal series characteristics of signals while suppressing signal aliasing and crosstalk from multiple devices. A dual-branch parallel extraction structure, combining data and knowledge, is used to simultaneously mine implicit signal features and prior knowledge of fault mechanisms. A cross-attention mechanism enables deep constraint fusion of data and knowledge features, improving diagnostic stability and generalization ability under varying operating conditions. Finally, standardized fault classification results and location information are output, achieving accurate and robust intelligent diagnosis of the equipment group's health status. This scheme effectively addresses the technical pain points of traditional single-data-driven diagnostic models, such as poor generalization ability and cross-condition failure, as well as the difficulty of single-knowledge-driven models adapting to complex real-time monitoring data. It achieves the organic integration of data observation and mechanistic knowledge, improving diagnostic accuracy and reliability.

[0022] In some embodiments, raw multimodal monitoring signals from a group of equipment are collected via a non-invasive sensor network, and standardized multimodal monitoring data is obtained through sliding time window segmentation and preprocessing, including: External adsorption-type acoustic emission sensors and stray magnetic flux sensors are deployed at key locations such as the main shaft shell and motor base of the equipment group to construct a non-invasive sensor network. Acoustic emission signals and magnetic flux signals are collected according to a preset sampling frequency to obtain continuous non-stationary raw monitoring signals. The raw monitoring signals are overlapped and segmented using a sliding time window to obtain multiple sets of fixed-dimensional one-dimensional time series arrays. The segmented time series are preprocessed by detrending and normalization to remove interference from environmental baseline drift and amplitude differences, generating standardized multimodal monitoring data.

[0023] Specifically, the non-invasive sensor network uses an adsorption-type installation, which eliminates the need to disassemble the equipment's main shaft sealing structure and stop the production line, enabling online monitoring of equipment status. This effectively avoids the damage to equipment sealing performance and operational accuracy caused by invasive installation, while also preventing production efficiency losses due to downtime monitoring. The acoustic emission sensor is used to collect high-frequency elastic wave signals, sensitively capturing transient fault characteristics such as impact, friction, and cracks in rotating components such as gears and bearings. Its acquisition frequency band covers 100kHz-1MHz, enabling accurate capture of instantaneous elastic wave radiation signals when a fault occurs. The stray magnetic flux sensor is used to collect low-frequency electromagnetic fluctuation signals, with an acquisition frequency band covering 10Hz-1kHz. It can effectively reflect electrical and mechanical coupling anomalies in the motor stator, rotor, and conduction path, and is particularly suitable for early fault monitoring of motor-type equipment.

[0024] The sliding time window employs an overlapping truncation method with an overlap rate of 50%, transforming an infinitely long continuous industrial signal stream into locally stationary, dimensionally uniform discrete time series segments. This preserves the temporal correlation of the signal, preventing the loss of fault features due to segmentation, and provides a standardized input format for subsequent tensor operations, avoiding system memory overflow and computational anomalies. The segmented signals are sequentially detrended, using a linear fitting detrending algorithm to eliminate temperature drift and electromagnetic baseline interference, ensuring consistent signal baselines. Normalization maps the signal amplitude to a unified range of [0,1], eliminating interference caused by differences in signal amplitude under different sensors and operating conditions, resulting in standardized multimodal monitoring data. This provides high-quality input for subsequent spatial topology mapping, ensuring the accuracy of subsequent feature extraction and fusion.

[0025] In some embodiments, multimodal monitoring data is mapped to a preset device spatial topology matrix according to the spatial coordinates of monitoring points, and time series features are fused to generate a spatiotemporal state tensor, including: Based on the physical layout of the equipment group, the sensor installation positions, and the relationship between the vibration transmission medium and impedance, a spatial topology matrix of the equipment is constructed, consisting of nodes and edges. Each node corresponds to a single sensor point of a single device, and the edge weights characterize the spatial distance, signal crosstalk, and coupling strength between sensors. The closer the distance and the stronger the coupling, the larger the edge weight. The preprocessed multimodal monitoring data is used as the node attribute features and filled into the corresponding nodes of the spatial topology matrix according to the point coordinates. The time window sequence and spatial topology are fused, and the multi-time topology matrix is ​​stacked in the channel dimension to generate a spatiotemporal state tensor containing spatial adjacency relationships, time series, and modal features.

[0026] Specifically, the equipment spatial topology matrix is ​​a structured data representation based on graph theory, used to quantitatively describe the spatial interference and signal transmission relationship between multiple sensors within an equipment group. Its matrix dimension is consistent with the total number of sensors, and the matrix elements correspond to the edge weights between sensors. For heavy CNC machine tool groups that are densely arranged in workshops and have mutually interfering working conditions, traditional monitoring methods are prone to misjudgment of faults due to signal crosstalk. By explicitly modeling the spatial topology, the spatial relationship and coupling strength between sensors are quantified into edge weights, which can effectively remove the background crosstalk noise transmitted by adjacent equipment and solve the problem of signal aliasing and crosstalk that is difficult to distinguish in non-intrusive monitoring.

[0027] Each sensor is treated as an independent node, with node attributes containing acoustic emission and magnetic flux modal data collected by that sensor. Physical priors such as device spacing, vibration transmission attenuation, and electromagnetic shielding effectiveness are transformed into edge weight values. The closer the distance, the smaller the vibration transmission attenuation, and the worse the electromagnetic shielding effect, the larger the edge weight value, thus achieving a quantitative expression of spatial coupling relationships. Acoustic emission and magnetic flux data within the same time window are respectively filled into a matrix as node attributes of different modes, and then stacked in the depth dimension according to the time window order to form a three-dimensional spatiotemporal state tensor. The dimensions of this tensor are: number of sensors × number of sensors × number of time windows × number of modes. This allows the model to simultaneously perceive the temporal evolution and spatial distribution characteristics of signals, improving the ability to analyze fault propagation across devices and locations, and providing comprehensive spatiotemporal modal information for subsequent dual-branch feature extraction.

[0028] In some embodiments, the spatiotemporal state tensor input data feature extraction network is subjected to multi-scale convolution and pooling processing to extract signal features and convert them into data-driven representation vectors, including: A multi-scale one-dimensional convolutional feature extraction network is used to drive branches of the spatiotemporal state tensor input data. Parallel sliding convolution operations are performed on the tensor using one-dimensional convolution kernels of two different scales to extract high-frequency transient impact features and low-frequency modulation envelope features, respectively. The convolution output feature map is nonlinearly activated by a linear rectified activation function. A max pooling layer with a preset window is used to downsample the activated feature map to retain local salient features and reduce dimensional redundancy. Multi-scale pooling features are concatenated in the channel dimension, unfolded, and mapped into a dense vector of fixed dimensions to obtain the data-driven representation vector.

[0029] Specifically, two different scales of one-dimensional convolutional kernels are adapted to capture different types of fault features. The large-size convolutional kernel is used to capture long-term, low-frequency modulation and envelope features of the signal, and is adapted to gradual faults such as wear and loosening. These fault signals change slowly and have a low frequency, so a large-size convolutional kernel is needed to expand the receptive field in order to fully capture their change patterns. The small-size convolutional kernel is used to capture short-term, high-frequency transient impact features, and is adapted to sudden faults such as cracks, pitting, and peeling. These fault signals change rapidly and have a high frequency, and the small-size convolutional kernel can accurately capture their instantaneous impact peak. Parallel multi-scale convolution can simultaneously cover the full-band features of the fault signal, avoid feature omissions caused by single-scale convolution, and improve the sensitivity to perceive different types and degrees of defects.

[0030] The linear rectified activation function introduces a nonlinear mapping, effectively solving the gradient vanishing problem and enhancing the network's ability to fit complex, non-stationary industrial signals. It can accurately capture nonlinear fault characteristics in the signal. Max pooling downsampling uses a fixed-size window to retain the maximum activation value within the local receptive field. While compressing feature dimensions and reducing computational load, it highlights the abnormal mutation characteristics caused by faults, suppresses random noise interference, and ensures that the extracted features have strong discriminative power. The feature maps after multi-scale pooling are concatenated along the channel dimension, and high-frequency and low-frequency features are statistically analyzed. A fixed-dimensional dense vector is obtained through mapping via a fully connected layer, forming a data-driven representation vector. This achieves purification from highly redundant spatiotemporal tensors to compact, highly discriminative implicit features, providing high-quality data representation for subsequent knowledge fusion and ensuring that the data features accurately reflect the equipment's operating status.

[0031] In some embodiments, the spatiotemporal state tensor is input into the knowledge graph reasoning engine, and through entity querying and graph embedding transformation, the fault mechanism knowledge is transformed into a knowledge-driven representation vector, including: Fault-sensitive statistics such as frequency band energy, peak value, kurtosis, and margin of each sensor node are extracted from the spatiotemporal state tensor and used as query entities in the knowledge graph. The triple relationships containing the query entities are traversed in the pre-constructed equipment fault knowledge graph to match the corresponding fault mode, fault location, and fault mechanism entities. The matched fault representation entities are input into the graph embedding and transformation module. Through the mapping of the preset entity dictionary index and the graph embedding matrix, the discrete character-type fault entities are transformed into continuous floating-point vectors of the same dimension as the data-driven vector to obtain the knowledge-driven representation vector.

[0032] Specifically, the equipment fault knowledge graph is a pre-built relational knowledge base, constructed based on equipment fault mechanisms, expert diagnostic experience, and historical fault data. It stores typical fault mechanisms, frequency domain characteristics, vibration performance, electromagnetic characteristics, and expert diagnostic experience of gears and bearings in the form of entities, relations, and entity triples. The knowledge graph also includes related information such as fault level, fault development trend, and maintenance suggestions to ensure that the fault information obtained through reasoning is comprehensive and complete.

[0033] The knowledge graph reasoning engine uses signal statistical features as the query entry point and combines forward reasoning with reverse verification to achieve structured reasoning from observed features to fault mechanisms and fault modes, avoiding reasoning bias caused by single features. The graph embedding transformation module maps symbolic fault entities into continuous numerical vectors in a vector space through a pre-trained graph embedding matrix. This graph embedding matrix is ​​trained with a large number of fault samples and can accurately represent the correlation between fault entities, transforming expert knowledge and physical mechanisms into computable and fusionable mathematical expressions. The knowledge-driven representation vector and the data-driven representation vector maintain the same dimension, providing a unified spatial foundation for subsequent cross-attention fusion. This aligns the black-box data model and the white-box knowledge model at the mathematical level, ensuring that the two can be effectively integrated and leverage their respective advantages.

[0034] In some embodiments, the data-driven representation vector and the knowledge-driven representation vector are input into the cross-attention fusion module, and a data fusion feature vector is obtained through weighted fusion, including: A linear transformation is performed on the data-driven representation vector to generate a query matrix; the knowledge-driven representation vector is then linearly transformed to generate a key matrix and a value matrix; the dot product of the query matrix and the key matrix is ​​calculated and scaled by dividing by the square root of the column dimension of the key matrix to maintain numerical stability; the attention weight matrix is ​​calculated using a normalized exponential function on the scaling result; the attention weight matrix and the value matrix are then multiplied by weight to obtain a fusion feature map with knowledge constraints; the fusion feature map is then dimensionally normalized to obtain a data fusion feature vector.

[0035] Specifically, cross-attention fusion uses knowledge features as constraints and data features as queries to achieve deep entanglement between data observation and prior knowledge. Its core advantage lies in its ability to dynamically adjust the weight allocation of data features based on the correlation of knowledge features. The query matrix comes from actual monitoring signals, representing current data features and reflecting the real-time operating status of the equipment; the key matrix and value matrix come from the knowledge graph, representing fault mechanisms and expert experience, providing prior constraints for the interpretation of data features and avoiding misjudgments due to noise interference.

[0036] Scaling the dot product operation avoids excessively large dot product values ​​due to increased vector dimension, prevents the normalized exponential function from entering the gradient saturation region, ensures training and inference stability, and guarantees the accuracy of attention weight calculation. The attention weight matrix automatically learns the correlation strength between data features and knowledge entities. When data features are highly correlated with a certain type of fault knowledge, higher weights are assigned. When data features deviate from the normal range due to noise interference, correction and anchoring are performed through the constraints of knowledge features. Compared with simple concatenation or weighted summation, cross-attention achieves dynamic alignment of data and knowledge at the feature semantic level, significantly improving the model's diagnostic robustness and generalization ability under varying speeds, varying loads, and multi-device interference. It effectively solves the problem of single data-driven models failing across operating conditions, ensuring the accuracy and reliability of diagnostic results.

[0037] In some embodiments, the data fusion feature vector is input into a fully connected classification layer, the fault probability distribution is output, the target fault classification result is determined, and fault location information is output, including: The data fusion feature vector is input into a preset fully connected classification layer; through the fully connected layer and the classification activation function, the probability distribution of the equipment group under each preset fault category is output; the preset fault categories include normal state, light wear, moderate spalling, severe fracture, bearing failure, and gear failure; the fault category corresponding to the maximum value in the probability distribution is selected as the target fault classification result of the current equipment; the node positions in the spatiotemporal state tensor are traced back to locate the equipment to which the fault belongs, the sensor points, and the approximate component positions, and the fault location information and diagnostic conclusions are output.

[0038] Specifically, the fully connected classification layer consists of multiple perceptrons. Each perceptron achieves nonlinear mapping through an activation function, mapping high-dimensional fused features to a preset fault label space, thus completing the dimensionality reduction output from abstract features to industrially understandable fault types. The classification activation function adopts a normalized exponential function, outputting the confidence probability of each fault category. The probability value ranges from [0,1]. The higher the probability value, the greater the possibility that the equipment belongs to that fault category, making the diagnostic results interpretable and reliable, and facilitating on-site maintenance personnel to judge the severity of the fault.

[0039] The system not only outputs fault classification results, but also accurately locates the faulty equipment, sensor points, and approximate component locations by tracing back the node positions in the spatiotemporal state tensor and combining them with the node coordinate information of the equipment spatial topology matrix. This achieves integrated output of fault diagnosis and fault location, solving the pain point of traditional diagnostic methods that can only determine the fault type but cannot locate the fault location. The final results can be directly uploaded to the edge platform or workshop MES system, forming a closed loop of monitoring-analysis-diagnosis-location. This provides on-site maintenance personnel with executable equipment maintenance and alarm commands, facilitating timely maintenance work and reducing equipment downtime and maintenance costs.

[0040] In some embodiments, the method further includes performing structural sensitivity verification and coordinate compensation on the fault region: The system retrieves the node spatial coordinates and topological adjacency relationships of the fault region corresponding to the target fault classification results; it retrieves the propagation path, typical amplitude, and frequency band threshold of the corresponding fault from the knowledge graph; if the fault characteristic amplitude and spatial distribution exceed the allowable range of the knowledge graph, it is determined to be environmental crosstalk or structural interference, and false faults are eliminated; if the fault is located in a high vibration transmission path or a strong electromagnetic coupling area, interference compensation is performed based on the topological edge weights to correct the fault probability and location coordinates.

[0041] Specifically, in non-intrusive monitoring, environmental noise, vibration crosstalk between equipment, and electromagnetic interference can easily lead to false fault signals. Without verification, these factors significantly increase the false alarm rate and affect system reliability. This step performs secondary verification of the data-driven diagnostic results using knowledge constraints. It retrieves the propagation path, typical amplitude, and frequency threshold of the corresponding fault from the knowledge graph and compares the current fault characteristics with the standard features in the knowledge graph. If the fault characteristic amplitude and spatial distribution exceed the allowable range of the knowledge graph and there is no reasonable fault propagation path to support it, it is determined to be environmental crosstalk or structural interference, eliminating false faults and reducing the false alarm rate. If the fault is located in a high vibration transmission path or a strong electromagnetic coupling area, it indicates that the fault characteristics may be deviated due to interference from adjacent equipment. The interference coefficient is calculated based on the topological edge weights to correct the fault probability and location coordinates, ensuring the accuracy of fault location. This step, through secondary verification and coordinate compensation, distinguishes between real faults and operational interference or signal crosstalk, while correcting location deviations caused by interference, reducing the false alarm and missed alarm rates of non-intrusive monitoring, and further improving the diagnostic reliability of the system in complex equipment cluster scenarios.

[0042] This invention introduces a pre-defined knowledge graph reasoning engine and a cross-attention fusion module to form a feature constraint mechanism based on mechanical prior knowledge in the system processing logic. After obtaining the spatiotemporal state tensor with spatial correlation, the system no longer relies solely on blindly fitting data branches, but instead triggers knowledge reasoning branches in parallel. The cross-attention module uses knowledge-driven representation vectors as keys and values ​​to perform weighted correction on the query features of the data-driven representation vectors. This enables the system to forcibly pull the offset data features back into a reasonable fault mechanism expression space when facing unseen variable load signals, relying on the pre-defined mechanical operation law knowledge baseline. This fundamentally eliminates the catastrophic forgetting and boundary failure risks brought about by pure statistical fitting, significantly reduces the false alarm rate of the system under complex working conditions, and significantly improves the accuracy of diagnosis and classification.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A non-intrusive intelligent diagnostic system for equipment clusters driven by both data and knowledge, characterized in that, include: Acquire non-intrusive multimodal monitoring data of the equipment group during the current operating cycle; The non-invasive multimodal monitoring data is mapped to a preset device space topology matrix according to the spatial coordinates of the monitoring points to generate a spatiotemporal state tensor in this spatiotemporal dimension. The spatiotemporal state tensor is input into a preset data feature extraction network, and feature dimensionality reduction is performed through the multi-scale convolutional layer in the data feature extraction network to output a data-driven representation vector. The node attribute features in the spatiotemporal state tensor are input into a preset knowledge graph inference engine, and a knowledge-driven representation vector is output by matching entity relationships in a preset graph database. The data-driven representation vector and the knowledge-driven representation vector are input into a preset cross-attention fusion module for weighted fusion calculation. The target fault classification result of the equipment group is determined based on the fusion feature vector output by the calculation.

2. The system according to claim 1, characterized in that, The steps for acquiring non-invasive multimodal monitoring data of the equipment group during the current operating cycle specifically include: Based on acoustic emission sensors and stray magnetic flux sensors deployed outside the equipment group, raw monitoring signals are collected at a preset sampling frequency. The original monitoring signal is segmented and truncated using a sliding time window of preset length to obtain multimodal monitoring data containing multiple one-dimensional time series arrays.

3. The system according to claim 1, characterized in that, The step of mapping the non-invasive multimodal monitoring data to a preset device spatial topology matrix according to the spatial coordinates of the monitoring points specifically includes: Based on the physical distances between the devices in the equipment group and the impedance of the vibration transmission medium, a spatial topology matrix of the equipment, consisting of nodes and edges, is constructed. The non-invasive multimodal monitoring data is used as node attribute features to fill the corresponding nodes of the device spatial topology matrix, generating the spatiotemporal state tensor containing spatial adjacency relationships and time series features.

4. The system according to claim 1, characterized in that, The step of inputting the spatiotemporal state tensor into a preset data feature extraction network specifically includes: The first-scale one-dimensional convolutional kernel and the second-scale one-dimensional convolutional kernel in the data feature extraction network are used to perform sliding inner product operations on the spatiotemporal state tensor, respectively. The feature maps obtained from the computation are processed by a linear rectified activation function and then concatenated along the feature channel dimension to generate the data-driven representation vector.

5. The system according to claim 1, characterized in that, The step of inputting the node attribute features in the spatiotemporal state tensor into a preset knowledge graph inference engine specifically includes: Extract the frequency band energy peak value from the node attribute features as the query entity; The system traverses the triplet relationships containing the query entity in the preset graph database, extracts the fault representation entities associated with the query entity, and transforms the fault representation entities into multi-dimensional floating-point vectors as the knowledge-driven representation vector.

6. The system according to claim 1, characterized in that, The step of inputting the data-driven representation vector and the knowledge-driven representation vector into a preset cross-attention fusion module for weighted fusion calculation specifically includes: The data-driven representation vector is linearly transformed to generate a query matrix; The knowledge-driven representation vector is linearly transformed to generate the key matrix and the value matrix; Calculate the dot product of the query matrix and the key matrix, and obtain the attention weight matrix using the normalized exponential function; Multiplying the attention weight matrix by the value matrix yields a data fusion feature vector containing knowledge constraints.

7. The system according to claim 6, characterized in that, The step of determining the target fault classification result of the equipment group based on the calculated fusion feature vector specifically includes: The data fusion feature vector containing knowledge constraints is input into a preset fully connected classification layer; The fully connected classification layer outputs the probability distribution values ​​of the equipment group under each preset fault category. The fault category corresponding to the highest probability distribution value is determined as the target fault classification result.

8. The system according to claim 4, characterized in that, After the step of processing the obtained feature map using a linear rectified activation function, the method further includes: The activated feature map is downsampled using max pooling to retain the maximum feature activation value within the local receptive field of the feature map, thereby reducing the spatial resolution dimension of the spatiotemporal state tensor.

9. The system according to claim 5, characterized in that, The step of converting the fault representation entity into a multi-dimensional floating-point vector specifically includes: The preset graph embedding transformation matrix is ​​invoked to perform index matching between the fault representation entities in discrete character form and the preset entity dictionary; Based on the index value obtained from the matching, the corresponding one-dimensional floating-point vector is extracted from the graph embedding transformation matrix as the knowledge-driven representation vector.

10. The system according to claim 6, characterized in that, The step of calculating the dot product of the query matrix and the key matrix further includes: The dot product result is scaled by dividing it by the square root of the column dimension of the key matrix to prevent the subsequent normalized exponential function from entering the gradient saturation region during computation.

Images