Industrial equipment fault intelligent diagnosis system based on deep learning algorithm

Abstract

The invention relates to the technical field of industrial equipment diagnosis, and discloses an industrial equipment fault intelligent diagnosis system based on a deep learning algorithm. A deep feature extraction module of the system extracts various features through parallel convolutional neural network branches and fuses the features to generate a high-order fault feature tensor; the time sequence evolution analysis module captures a propagation rule and a change trend of fault features by adopting a bidirectional long-short-term memory network, and outputs a fault development time sequence path; the multi-source feature association module is matched with equipment structure topological graph nodes, and a mapping relation between fault features and equipment physical positions is established; the fault mode judgment module outputs a fault type recognition result and a fault position coordinate set by using a graph convolutional neural network in combination with the fault mode knowledge graph; and the diagnosis decision output module generates a decision instruction set containing a fault level, a component replacement priority and a maintenance strategy. According to the system, the intelligence and precision level of industrial equipment fault diagnosis is improved.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment diagnostic technology, specifically to an intelligent fault diagnosis system for industrial equipment based on deep learning algorithms. Background Technology

[0002] In the industrial production sector, the stable operation of equipment is directly related to production efficiency and operational safety. Therefore, timely diagnosis and accurate location of equipment faults have always been a key focus of the industry. Traditional methods of diagnosing industrial equipment faults mostly rely on manual inspections and experience-based judgment. This approach is not only time-consuming and labor-intensive, but also struggles to meet the multi-dimensional and highly dynamic operational status monitoring needs of equipment in complex industrial environments.

[0003] With the improvement of industrial automation, the amount of sensor data generated by equipment operation is growing exponentially. Various parameters such as vibration, temperature, and current are interconnected, and traditional single-dimensional feature analysis methods often fail to capture early signs and evolution patterns of faults. For example, relying solely on the spectrum analysis of vibration signals may misjudge the fault type and ignore the coupling relationship between temperature anomalies and current fluctuations. On the other hand, monitoring systems based on simple thresholds are difficult to adapt to the dynamic changes of equipment under different operating conditions, resulting in high false alarm and false negative rates.

[0004] Existing fault diagnosis systems have significant limitations in feature extraction and fault mode matching. On the one hand, manually designed feature extraction operators struggle to adapt to the diversity of complex industrial scenarios, resulting in insufficient feature representation capabilities. On the other hand, fault mode matching largely relies on static rule bases, lacking dynamic analysis of the fault's temporal evolution and failing to accurately trace the fault's propagation path and impact range, thus affecting the timeliness and effectiveness of maintenance decisions. These problems make it difficult for the level of intelligence and precision in industrial equipment fault diagnosis to meet the efficient operation and maintenance needs of modern industrial production. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent fault diagnosis system for industrial equipment based on deep learning algorithms, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an intelligent fault diagnosis system for industrial equipment based on deep learning algorithms, the system comprising:

[0007] The multi-dimensional operation status acquisition module acquires the vibration sensing sequence, temperature distribution spectrum and current waveform characteristics of industrial equipment in real time, and constructs a multi-dimensional heterogeneous data stream.

[0008] The deep feature extraction module receives the multidimensional heterogeneous data stream and extracts the time-frequency domain features of the vibration spectrum, the spatial gradient features of the temperature distribution, and the energy accumulation features of the current harmonics through parallel convolutional neural network branches, and fuses them to generate a high-order fault feature tensor.

[0009] The temporal evolution analysis module inputs the high-order fault feature tensor into a bidirectional long short-term memory network to capture the propagation pattern and intensity change trend of fault features within a continuous time window, and outputs the temporal path of fault development.

[0010] The multi-source feature association module matches the node positions in the equipment structure topology diagram according to the fault development time sequence path, calculates the coupling degree between the vibration feature propagation path and the current abnormal area, and establishes the mapping relationship between fault features and equipment physical location.

[0011] Based on the mapping relationship, the fault mode determination module uses a graph convolutional neural network to analyze the feature correlation strength between the topology nodes of the equipment, and combines the typical fault feature templates in the fault mode knowledge graph to output the fault type identification result and the fault location coordinate set.

[0012] The diagnostic decision output module generates a set of decision instructions based on the fault type identification results and the fault location coordinate set, which includes fault level quantification indicators, component replacement priority list and maintenance strategy suggestions.

[0013] Preferably, the multi-dimensional operation status acquisition module includes:

[0014] The vibration sensing unit collects triaxial vibration signals of the key bearings of the equipment through a distributed accelerometer array, and generates a vibration sensing sequence after wavelet noise reduction processing.

[0015] The infrared scanning unit uses a thermal imager to periodically acquire the temperature distribution of the device casing and generates a temperature distribution map through pixel-level temperature calibration.

[0016] The current monitoring unit uses a high-precision current transformer to capture the fundamental and harmonic components of the motor drive current, and extracts the current waveform characteristics through Fourier transform.

[0017] Preferably, the deep feature extraction module includes:

[0018] The vibration feature extraction submodule uses a one-dimensional dilated convolutional neural network to process the vibration sensing sequence, extracting impact pulse features in the time domain and extracting the energy ratio of the resonant frequency band in the frequency domain.

[0019] The temperature feature extraction submodule analyzes the spatial gradient changes of the temperature distribution map through a two-dimensional convolutional neural network to identify the boundaries of local overheated regions and the direction of heat conduction.

[0020] The current feature extraction submodule uses a convolutional attention mechanism network to separate the fundamental distortion component and the amplitude of specific harmonics in the current waveform features, and quantifies the energy accumulation characteristics of each harmonic.

[0021] The feature fusion submodule splices the impact pulse features, resonant frequency band energy ratio, local overheating region boundary, heat conduction direction, fundamental distortion component and specific subharmonic amplitude into tensors, and generates a high-order fault feature tensor through dimensionality reduction by a fully connected layer.

[0022] Preferably, the time-series evolution analysis module performs the following operations:

[0023] The higher-order fault feature tensor is sliced ​​into equal-length segments and input into a bidirectional long short-term memory network.

[0024] Extract feature evolution patterns of historical time slices through forward hidden layers;

[0025] Capture the feature decay trend of future time slices through backward hidden layers;

[0026] By fusing the output states of the forward hidden layer and the backward hidden layer, a gradient vector representing the rate of change of fault feature intensity is generated.

[0027] Based on the alignment of the gradient vector peak position with the equipment operating cycle, the fault development timeline path is output.

[0028] Preferably, the multi-source feature association module includes:

[0029] The topology loading submodule calls the three-dimensional coordinates and connection relationships of each node in the device structure topology diagram;

[0030] The path mapping submodule maps the feature intensity peaks in the fault development time-series path to the corresponding nodes in the topology graph;

[0031] The coupling analysis submodule calculates the spatial overlap between the vibration characteristic propagation path and the current anomaly region, and combines the coordinates of the temperature distribution anomaly region to generate a mapping relationship between fault characteristics and the physical location of the equipment.

[0032] Preferably, the fault mode determination module includes:

[0033] The knowledge graph loading submodule reads typical fault feature templates such as bearing wear, gear tooth breakage, and winding short circuit from the fault mode knowledge graph;

[0034] The node association analysis submodule uses a graph convolutional neural network to traverse the device structure topology graph and calculate the feature association strength of each node in the mapping relationship.

[0035] The pattern matching submodule calculates the similarity between the feature association strength and the typical fault feature template, and outputs the fault type identification result and the fault location coordinate set.

[0036] Preferably, the diagnostic decision output module includes:

[0037] The grading and quantification submodule classifies three levels of quantitative indicators—minor anomalies, moderate degradation, and severe faults—based on the rate of change of characteristic intensity in the fault development timeline.

[0038] The component analysis submodule matches the replacement cost and downtime impact coefficient of components in the equipment spare parts library with the set of fault location coordinates to generate a priority list of component replacements.

[0039] The strategy generation submodule combines fault level quantification indicators with component replacement priority lists to formulate three maintenance strategy recommendations: immediate shutdown for repair, online monitoring operation, and preventive maintenance plan, forming a decision instruction set.

[0040] Preferably, the system further includes:

[0041] The real-time feedback module receives sensor verification data during equipment maintenance and updates the acquisition parameters of the multi-dimensional operating status acquisition module.

[0042] The real-time feedback module inputs the maintained equipment operating status data into the deep feature extraction module to generate a verified high-order fault feature tensor.

[0043] The time-series evolution analysis module corrects the fault development time-series path based on the verified higher-order fault feature tensor.

[0044] Preferably, the system further includes:

[0045] The model optimization module adjusts the kernel size of the convolutional neural network layer in the deep feature extraction module by comparing the actual fault types in the maintenance records with the prediction results.

[0046] The model optimization module uses the contrastive loss function to update the weight parameters of the bidirectional long short-term memory network in the temporal evolution analysis module;

[0047] The model optimization module is based on the adjacency matrix calculation method of the fault mode determination error rate optimization graph convolutional neural network.

[0048] Preferably, the system further includes:

[0049] The visualization interface module receives the fault level quantification index and component replacement priority list from the decision instruction set and generates a fault location annotation map of the equipment's three-dimensional model.

[0050] The visualization interface module transforms maintenance strategy suggestions into operation guidance flowcharts, which are then simultaneously pushed to mobile terminals and the central control console.

[0051] The visualization interface module dynamically displays the projection trajectory of the fault development timeline output by the timeline evolution analysis module onto the device's operating timeline.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] This intelligent fault diagnosis system for industrial equipment based on deep learning algorithms constructs a multi-dimensional heterogeneous data stream through a multi-dimensional operation status acquisition module. It can comprehensively capture the characteristics of various parameters such as vibration, temperature, and current of industrial equipment during operation, breaking the limitations of traditional single-dimensional data acquisition and making the system's perception of the equipment's operating status more comprehensive and three-dimensional.

[0054] The deep feature extraction module uses parallel convolutional neural network branches to extract features from different types of data and fuse them to generate a high-order fault feature tensor. Compared with the traditional method of manually designing feature operators, it can automatically mine the complex correlation features hidden in the data, improve the expressive power of the features, and make the system more adaptable to the diverse data features in complex industrial scenarios.

[0055] The temporal evolution analysis module introduces a bidirectional long short-term memory network, which effectively captures the propagation pattern and intensity change trend of fault characteristics within a continuous time window. This changes the traditional model of static fault analysis and enables dynamic tracking of the fault evolution process, providing strong support for accurately judging the fault development trend.

[0056] The multi-source feature association module combines the fault development time sequence path with the equipment structure topology map, calculates the coupling degree between different features and establishes a mapping relationship with the physical location of the equipment. This overcomes the problem of feature decoupling from equipment location in traditional systems, enabling fault features to be accurately mapped to specific components of the equipment and improving the accuracy of fault location.

[0057] The fault mode determination module uses graph convolutional neural networks to analyze the feature correlation strength between topological nodes of the equipment and combines it with the fault mode knowledge graph for matching. It not only considers the correlation between various components of the equipment, but also draws on typical fault feature templates, which improves the accuracy of fault type identification and reduces the occurrence of misjudgments.

[0058] The diagnostic decision output module generates a set of decision instructions that includes quantitative indicators of fault level, a list of component replacement priorities, and maintenance strategy recommendations. It directly transforms fault diagnosis results into specific operation and maintenance guidance, making maintenance work more targeted and planned, which helps to optimize operation and maintenance processes and reduce operation and maintenance costs. Attached Figure Description

[0059] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent fault diagnosis system for industrial equipment based on deep learning algorithms described in this invention.

[0060] Figure 2 A flowchart of the multi-dimensional operational status acquisition module;

[0061] Figure 3 A flowchart illustrating the workflow of the deep feature extraction module;

[0062] Figure 4 This is a flowchart of the fault mode determination module. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Please see Figure 1 This invention provides an intelligent fault diagnosis system for industrial equipment based on deep learning algorithms, the system comprising:

[0065] The multi-dimensional operation status acquisition module acquires the vibration sensing sequence, temperature distribution spectrum and current waveform characteristics of industrial equipment in real time, and constructs a multi-dimensional heterogeneous data stream.

[0066] The deep feature extraction module receives the data stream and uses parallel convolutional neural network branches to extract the time-frequency domain features of the vibration spectrum, the spatial gradient features of the temperature distribution, and the energy accumulation features of the current harmonics, respectively, and fuses them to generate a high-order fault feature tensor. The temporal evolution analysis module inputs this tensor into a bidirectional long short-term memory network to capture the propagation law and intensity change trend of the fault features within a continuous time window, and outputs the fault development time-series path.

[0067] The multi-source feature association module matches the node positions in the equipment structure topology diagram based on the time sequence path, calculates the coupling degree between the vibration feature propagation path and the current abnormal area, and establishes the mapping relationship between fault features and the physical location of the equipment.

[0068] Based on this mapping relationship, the fault mode determination module uses a graph convolutional neural network to analyze the feature correlation strength between the topology nodes of the equipment, and combines the typical fault feature templates in the fault mode knowledge graph to output the fault type identification results and the fault location coordinate set.

[0069] Based on this, the diagnostic decision output module generates a set of decision instructions that includes quantitative indicators of fault level, a list of component replacement priorities, and maintenance strategy recommendations.

[0070] Example 1: See Figure 2The multi-dimensional operational status acquisition module consists of three collaborative units that work together to comprehensively monitor the operational status of industrial equipment. The vibration sensing unit deploys a distributed array of accelerometers, covering the axial, radial, and tangential positions of key bearings. Each sensor captures raw triaxial vibration signals at a sampling rate of at least 20kHz. The raw signals are transmitted to the preprocessing unit, where wavelet transform algorithms are used for noise reduction. Appropriate wavelet basis functions are selected to decompose the signal levels, and threshold filtering eliminates environmental noise interference while retaining the impact components reflecting the mechanical state of the equipment. The denoised signals are then reconstructed into a continuous time-domain vibration sensing sequence, including timestamps, triaxial acceleration values, and spectral feature identifiers.

[0071] The infrared scanning unit is equipped with a high-resolution thermal imager, which is fixed on a rotatable gimbal to perform full-coverage temperature measurement on the device's outer surface at preset scanning cycles. During the scanning process, the thermal imager captures the infrared radiation intensity of each pixel in real time and dynamically compensates for it by combining ambient temperature, humidity, and emissivity parameters. A pixel-level temperature calibration algorithm eliminates environmental radiation interference, and a blackbody radiation reference source is used as a benchmark during the calibration process to ensure that the temperature error of each pixel is controlled within a reasonable range. Finally, a temperature distribution map is generated, stored in a matrix structure, where each matrix element corresponds to the temperature value at a specific coordinate point on the device surface, with a spatial resolution of no less than 0.5 degrees Celsius. The scan time and area coordinate boundaries are also marked.

[0072] The current monitoring unit employs a closed-loop high-precision current transformer, directly coupled to the motor drive cable, to capture the three-phase current waveform in real time. The raw current signal undergoes amplitude adaptation and filtering processing by a signal conditioning circuit, and is then converted into a digital signal by a high-speed ADC module. A Fast Fourier Transform (FFT) algorithm is used to perform spectral analysis on the digital signal, calculating the fundamental frequency amplitude, total harmonic distortion (THD), and amplitude-phase difference of each harmonic component. For specific high-frequency harmonic components, a sliding time window is set to extract transient harmonic energy characteristics. The final generated current waveform features include parameters such as the fundamental effective value, the percentage of amplitudes of the 3rd to 25th harmonics, the harmonic distortion rate, and the energy concentration coefficient of the characteristic frequencies.

[0073] The output data from the three units are integrated using a unified timing framework. The data synchronization engine uses the clock signal of the equipment control system as a reference, adding millisecond-level precision timestamps to the vibration sensing sequence, temperature distribution map, and current waveform characteristics. When the three types of data at the same timestamp are detected, a structured multidimensional data stream is constructed: vibration data is stored in a time-amplitude sequence format, temperature data in a spatial coordinate-temperature value matrix format, and current data in a spectral component-energy value vector format. This data stream is transmitted to the deep feature extraction module via a high-speed industrial bus, with a data packet verification mechanism maintaining data integrity during transmission.

[0074] In the vibration sensing unit, the layout of the accelerometer array is designed based on the vibration transmission characteristics of the equipment. Triaxial sensor groups are arranged at the input bearing, output bearing, and coupling-side bearing positions of the gearbox, with each group containing three orthogonally distributed sensors. Sensor signals are connected to an anti-interference acquisition card via shielded cables. The acquisition card has a built-in programmable amplifier and a 24-bit analog-to-digital converter. The wavelet denoising process employs a multi-level decomposition and reconstruction strategy, retaining approximate coefficients for low-frequency vibration components and setting detail coefficients to zero for high-frequency noise components. The reconstructed signal retains the effective frequency band below 5kHz.

[0075] The pan-tilt control system of the infrared scanning unit moves the thermal imager lens along a preset path, covering all critical heat dissipation areas of the equipment casing. In each scanning cycle, the thermal imager acquires no fewer than 100 infrared images, and an image registration algorithm eliminates pixel shifts caused by equipment vibration. During the temperature calibration phase, a reference blackbody source is placed in the same scene as the equipment. By comparing the theoretical radiation value of the blackbody with the actual image value, a pixel grayscale-temperature mapping lookup table is established. The final output temperature distribution map includes a spatial coordinate system, with the origin aligned with the reference point of the equipment base.

[0076] The current monitoring unit's transformer bandwidth covers DC to 10kHz, with a linearity error not exceeding 0.2%. The signal conditioning circuit includes a second-order low-pass filter and a programmable gain amplifier; the filter cutoff frequency is set to 45% of the sampling rate to prevent aliasing. The Fast Fourier Transform uses a Hanning window function to reduce spectral leakage, and the calculation window length is matched to an integer multiple of the device's power cycle. The harmonic analysis module automatically identifies harmonic components with amplitudes exceeding 1% of the fundamental frequency, recording their frequency, amplitude, and phase offset angle relative to the fundamental frequency.

[0077] The construction process of the multidimensional data stream includes spatiotemporal alignment operations. The temporal alignment module detects the timestamp differences among the three types of data, and initiates an interpolation compensation algorithm when the difference exceeds 10 milliseconds. The spatial alignment module converts the pixel coordinates of the temperature distribution map into three-dimensional coordinates in the device's global coordinate system. The conversion process is based on the pose parameters of the thermal imager's gimbal and the lens distortion correction parameters. The final generated data stream is encapsulated and transmitted at fixed time intervals. Each data packet contains the vibration waveform, temperature matrix, and current spectrum vector within the complete time slice. The data packet header indicates the device number and the acquisition time interval.

[0078] Example 2: See Figure 3The deep feature extraction module consists of four collaborative sub-modules that process raw information from multidimensional heterogeneous data streams. The vibration feature extraction sub-module receives a vibration sensing sequence, which is a three-axis acceleration numerical stream in the time domain. A one-dimensional dilated convolutional neural network processes this sequence, and the network structure includes two main processing stages. The first processing stage sets the dilation factor to two, and the convolutional kernel skips some data points as it slides along the time axis, expanding the receptive field. This stage focuses on identifying transient impact components in the vibration signal, extracting pulse peak amplitude, rise time, and repetition period features. The second processing stage connects to the spectrum analysis layer, converting the convolutional output to the frequency domain, calculating the proportion of energy in a specific frequency band to the total energy, and focusing on characteristic frequency bands reflecting the state of the bearing or gear. The spectrum analysis employs a sliding window strategy, with the window length synchronized with the equipment's rotation period.

[0079] The temperature feature extraction submodule processes the temperature distribution map, which represents the temperature field on the device surface as a two-dimensional matrix. A two-dimensional convolutional neural network analyzes this matrix, with a 3x3 kernel size that moves along the spatial dimensions of the device surface. The first convolutional layer detects the temperature difference between adjacent pixels, outputting a feature map to identify regions of abrupt temperature changes. Subsequent convolutional layers progressively expand the receptive field, identifying the direction of temperature gradient changes and heat transfer paths. The deep network output includes a set of boundary coordinates for local high-temperature regions, as well as vector direction data pointing from the high-temperature center to the low-temperature region; the vector magnitude represents the heat conduction intensity.

[0080] The current feature extraction submodule processes the current waveform feature vector, including parameters of the fundamental wave and each harmonic component. The convolutional attention mechanism network first decomposes the input features through a one-dimensional convolutional layer, separating the fundamental wave distortion component and quantifying its deviation from a standard sine wave. The attention mechanism is applied to the harmonic components, calculating the feature weights for different harmonics. Specific harmonic components are assigned higher weights, and their amplitude features are extracted more effectively. The network output includes the fundamental wave distortion rate, the weighted amplitude of the specific harmonic, and the harmonic energy concentration coefficient.

[0081] The feature fusion submodule receives the output data from the three submodules mentioned above. The vibration submodule provides the impact pulse feature vector and the frequency band energy ratio scalar. The temperature submodule outputs the boundary coordinate set and the heat conduction direction vector. The current submodule submits the fundamental distortion rate and harmonic amplitude sequence. The fusion process first concatenates the six heterogeneous features into a six-channel feature map. The concatenated tensor is input into a fully connected layer network. The first fully connected layer compresses the feature dimension to 60% of the original size, and the activation function introduces a nonlinear transformation. The second fully connected layer further reduces the dimensionality, finally outputting a high-order fault feature tensor with a fixed 256 dimensions. Each dimension of this tensor represents a composite fault feature that integrates vibration, temperature, and current information.

[0082] The temporal evolution analysis module processes high-order fault feature tensor sequences. The tensor sequences are divided into fixed-length slices, each covering ten seconds of equipment operation time. Slice data is input into a bidirectional long short-term memory (LSTM) network, which contains forward and backward processing chains. The forward processing chain sequentially reads eight historical time slices, with hidden layer states recording the feature enhancement patterns over time. The backward processing chain reads the next four time slices in reverse order, with hidden layer states capturing feature decay trends. The bidirectional hidden layer states are concatenated and fused at the output, and the fused state is converted into a gradient vector through a linear layer. The gradient vector represents the rate of change of fault feature intensity per unit time, with local peaks corresponding to abrupt changes in fault feature intensity. The network output is a time-feature intensity sequence, annotating the feature intensity value and rate of change at each time point, forming a fault development temporal path. This path describes the entire time trajectory of fault features from generation, enhancement, to decay.

[0083] In the vibration feature extraction submodule, the dilated convolutional neural network employs a residual connection structure. The output of each convolutional layer is added to the input to avoid gradient vanishing during deep network training. The spectrum analysis layer uses an adaptive frequency band selection mechanism to dynamically adjust the center frequency of the feature band based on the device's rotational speed. The current feature extraction submodule's attention network incorporates both channel attention and spatial attention mechanisms. Channel attention recalibrates the importance weights of each harmonic component, while spatial attention locates key distortion events along the time axis. The fully connected layers in the feature fusion submodule use batch normalization to accelerate network convergence. The bidirectional long short-term memory network in the time-series analysis module uses a gating mechanism: the input gate controls the proportion of new information inflow, the forget gate adjusts the degree of historical memory retention, and the output gate determines the current state exposure. The gradient vector generation stage uses a peak detection algorithm to identify the inflection point of the feature intensity change rate. After aligning the inflection point with the device's operating cycle marker, the fault development stage is determined.

[0084] Example 3: See Figure 4The multi-source feature association module processes the fault development time-series path from the time-series evolution analysis module. This path contains time-feature intensity sequence information. The module operation begins with the topology loading submodule accessing the equipment structure topology database. The database stores the three-dimensional coordinates and mechanical connections of each component of the equipment in a graphical structure. Nodes in the topology graph represent physical components of the equipment, such as bearings, gears, or motor windings. Each node is assigned a unique identifier and three-dimensional coordinate values. Edges represent mechanical coupling relationships between components, such as shaft connections or gear meshing. Edge attributes include connection type and distance coefficient. The topology graph data format uses an adjacency list for easy and quick querying of node reachability. The path mapping submodule receives the fault development time-series path and parses the time point and peak amplitude corresponding to the feature intensity peaks in the path. The peak detection algorithm identifies abrupt change moments based on the gradient change rate threshold and maps each peak time point to a topology graph node. The mapping process aligns the time points with the equipment state snapshots based on the equipment operation log. For example, time point t = 125 seconds corresponds to bearing node B7, and the peak amplitude is recorded as 0.75 normalized units. The mapping results generate a node-peak correspondence table, in which each entry includes the node ID, coordinates, peak time, and intensity value.

[0085] The coupling analysis submodule utilizes a node-peak correspondence table to analyze the spatial correlation of multi-source features. The submodule first extracts the vibration feature propagation path, derived from the path mapping result, representing the direction of fault feature transmission within the equipment. For example, it connects node B7 via an edge to node G12 and then to node M3, forming a directed sequence. Simultaneously, it identifies abnormal current regions from current monitoring data. This region is defined as the set of nodes where the harmonic distortion rate in the current waveform exceeds the limit, such as nodes M3 to M5 in the motor windings. Spatial overlap calculation focuses on the intersection of the vibration path node set and the abnormal current node set. The calculation process introduces a node overlap rate index to quantify the matching degree between the two sets. When the overlap rate exceeds a preset threshold, and combined with a temperature distribution map, it verifies the existence of a local temperature rise anomaly in the region, generating a mapping relationship between fault features and physical locations. The mapping relationship table structure includes a node list, feature type, spatial coordinates, and coupling strength coefficient. For example, if the vibration path node set {V} is {B7,G12,M3} and the current anomaly node set {C} is {M3,M4,M5}, the intersection calculation shows that node M3 coincides, and the temperature data shows a temperature rise of 18 degrees Celsius at this coordinate, then the mapping relationship is updated with the following entry: Node M3 is associated with vibration impact and current harmonic characteristics, coordinates (x=1.2m,y=0.8m,z=0.5m), and the coupling strength is set to 0.8.

[0086] The fault mode determination module operates based on the aforementioned mapping table. The knowledge graph loading submodule calls a pre-built fault mode knowledge graph, stored in a graph database format. Nodes represent fault types, such as bearing wear, gear tooth breakage, or winding short circuit; edges represent causal relationships or concurrency probabilities between faults. Each fault type node is appended with a typical feature template. Template data originates from a historical fault case database. For example, the bearing wear template includes the envelope shape parameters of vibration and impact pulses, the gear tooth breakage template records the frequency offset values ​​of harmonic sidebands, and the winding short circuit template defines the threshold range of odd-order harmonic amplitudes. Templates are encoded as feature vectors, with dimensions matching the node embedding space. The node association analysis submodule uses a graph convolutional neural network to process the device topology graph, using nodes in the mapping table as input vertices and mechanical connection edges as graph edges. The network structure contains three graph convolutional layers, each performing a neighborhood feature aggregation operation: the first layer aggregates features of directly adjacent nodes, and subsequent layers expand to second-order neighborhoods. The aggregation process outputs node embedding vectors with a dimension of 128, representing the association strength of nodes in fault propagation. For example, for node M3, the embedding vector calculation integrates its own features and the influence of its neighboring nodes G12 and M4.

[0087] The pattern matching submodule compares the similarity between node embedding vectors and fault feature templates, outputting the fault type identification result. The similarity calculation uses the cosine similarity metric, as shown in the formula below:

[0088]

[0089] in: This represents the node embedding vector, which originates from the output of the graph convolutional neural network. represents the fault feature template vector, loaded from the knowledge graph; · represents the vector dot product operation; ‖·‖ represents the Euclidean norm of the vector; sim represents the similarity score, ranging from -1 to 1.

[0090] After calculation, the similarity score is compared with a preset matching threshold, for example, a threshold of 0.85 for the bearing wear template. If the similarity between the embedding vector of node M3 and the bearing wear template is greater than 0.85, the fault type is determined to be "bearing wear". Simultaneously, a set of fault location coordinates is associated, which extracts the three-dimensional coordinates of the corresponding node from a mapping table, such as {(x=1.2m, y=0.8m, z=0.5m)}. The output includes a fault type label, a list of coordinates, and a similarity confidence score. The entire determination process ensures consistency of multi-source features; for example, when vibration and current features both point to the same node and an abnormal temperature is confirmed, the fault location accuracy is improved.

[0091] In the topology loading submodule, the equipment structure topology diagram is imported through a computer-aided design model in STEP or IGES format, and then stored in a graph database after conversion. Node coordinates are in meters with millimeter-level precision; edge attributes include connection stiffness coefficients used for weighted graph calculations. The peak mapping in the path mapping submodule uses a nearest neighbor algorithm; when there is no direct log match for the peak time point, nodes are located by interpolation based on the equipment's operating cycle. The spatial overlap calculation in the coupling analysis submodule also involves a temperature verification step: the temperature distribution map data of the infrared scanning unit is queried to check whether the temperature rise value of the overlapping node region exceeds a 15-degree Celsius threshold, which is set according to the heat capacity of the equipment materials. When calculating the overlap rate, the node set size is dynamically adjusted to avoid distortion from small sets. The construction of the fault mode knowledge graph relies on domain expert annotation; the template vectors are dimensionality-reduced through principal component analysis to retain key feature dimensions. The graph convolutional network in the node association analysis submodule is configured to skip connections to prevent overfitting, and random deactivation regularization is used during training. The cosine similarity calculation in the pattern matching submodule is implemented in an embedded system, optimizing vector operation speed. Low-confidence matches are filtered out through post-processing of the similarity results. The final output interface encapsulates the fault type and coordinate set into a structured message, which is then transmitted to the diagnostic decision module. The entire implementation emphasizes data-driven approaches, requires no manual intervention, and keeps processing latency within milliseconds to meet real-time diagnostic needs.

[0092] Example 4: The diagnostic decision output module receives the fault type identification result and the fault location coordinate set, taking bearing wear fault as an example. The level quantification submodule analyzes the feature intensity changes in the fault development time path. Assuming the feature intensity F = 0.25 at the path data recording time point t = 120 seconds, F = 0.72 at t = 180 seconds, and the time interval Δt = 60 seconds, the change per unit time is calculated.

[0093]

[0094] According to the preset threshold rules: It is a minor abnormality. Moderate degradation. This is a serious fault. In this example, 0.0078 is a minor anomaly, and the fault level quantification index "Level 1" is output.

[0095] The component analysis submodule queries the equipment spare parts library and matches the component attributes corresponding to the fault location coordinate set {(x=1.2m, y=0.8m, z=0.5m)}. The spare parts library relationship table contains the component's three-dimensional coordinates, name, replacement cost (thousand yuan), and downtime impact coefficient (level 1-10). The downtime impact coefficient is determined by the criticality of the equipment function: coefficient > 8 for main transmission components and < 4 for auxiliary components. In this example, the coordinates match to "input shaft bearing", whose attributes are shown in the table below:

[0096]

[0097] Priority weight calculation rule: Weight = Replacement cost × Downtime impact coefficient. Input shaft bearing weight = 2.0 × 8.0 = 16.0, generating a list of component replacement priorities sorted in descending order of weight:

[0098] Input shaft bearing (weight 16.0)

[0099] Cooling fan (weight 2.4)

[0100] Motor carbon brushes (weight 2.0)

[0101] The strategy generation submodule combines the fault level "Level 1" with the priority list to formulate maintenance strategies:

[0102] Immediate shutdown and maintenance: weight > 10.0 and fault level ≥ moderate degradation.

[0103] Online monitoring operation: weight ≤ 5.0 and fault level = minor abnormality.

[0104] 48-hour preventative maintenance: Other combinations.

[0105] In this example, the highest weight (16.0) is greater than 10.0, but the fault level is minor, triggering the "Online Monitoring Operation + 48-Hour Preventive Maintenance" command. Command details:

[0106] Maintain equipment operation and collect vibration data every 10 minutes.

[0107] Complete the preparation for input shaft bearing replacement within 48 hours.

[0108] Spare parts list: Input shaft bearing (model B-207C).

[0109] Tools required: hydraulic puller, torque wrench (range 50-200Nm).

[0110] The real-time feedback module intervened during maintenance. When technicians used a portable infrared calibrator to check the bearing housing temperature, they found a ±3℃ deviation in the original temperature distribution map. After the calibration data was uploaded, the module updated the pixel calibration parameters of the infrared scanning unit: adjusting the original environmental radiation compensation coefficient from 0.92 to 0.89. After maintenance, the equipment restarted, and the multi-dimensional operating status acquisition module acquired the new data.

[0111] Vibration sensing sequence: peak amplitude decreased from 1.5g to 0.8g.

[0112] Temperature distribution map: The highest temperature in the bearing housing area drops from 68℃ to 52℃.

[0113] Current waveform characteristics: The proportion of the third harmonic amplitude decreased from 5.2% to 1.8%.

[0114] The new data input deep feature extraction module generates a validated high-order fault feature tensor. The temporal evolution analysis module reprocesses the tensor sequence: the feature strength of the original fault development timeline at t=180 seconds is corrected from 0.72 to 0.35, and the gradient vector slope is reduced from 0.0078 to 0.0033. The corrected path indicates a slower fault development rate, and the system automatically adjusts the fault level from "Level 1" to "Monitoring Status".

[0115] The entire implementation process embodies dynamic closed-loop control: the initial diagnosis generates decisions based on the raw sensor data, the maintenance action provides feedback on the on-site verification data, the acquisition module parameters are updated in real time (such as infrared calibration coefficients), the data after maintenance drives the recalculation of feature tensors, and the timing path correction triggers the iteration of diagnostic conclusions.

[0116] For example, after the input shaft bearing is replaced, the fault location coordinate set is cleared, the item is removed from the component replacement priority list, the strategy generation submodule terminates the preventive maintenance instruction, and instead outputs a "equipment operating normally" status code. Real-time feedback data is also stored in the historical case library for optimizing subsequent diagnostic thresholds.

[0117] Example 5: The model optimization module is activated after equipment maintenance is completed. The module first reads the maintenance record document, which contains comparison information between the actual fault type and the system's prediction results. When the misjudgment related to vibration characteristics exceeds a set threshold, the module initiates the convolution kernel adjustment program of the deep feature extraction module. The adjustment process analyzes the spectral distribution characteristics of the misjudged cases. If the lack of low-frequency feature recognition is the main cause of the error, the size of the first-layer convolution kernel of the one-dimensional convolutional neural network is expanded by two units from the original specification, resulting in a wider time window coverage. The convolution kernel weights are reinitialized, using a truncated normal distribution to generate new parameters. The attention network of the current feature extraction submodule is simultaneously optimized, with the number of attention heads dynamically increasing according to the complexity of harmonic components, and new attention heads focusing on specific harmonic combinations.

[0118] The optimization of the temporal evolution analysis module is based on the deviation data between the actual fault development timeline and the predicted path. The maintenance record contains the exact timestamps of the fault feature abrupt changes, and the module calculates the absolute difference between this timestamp and the predicted peak time. When the difference consistently exceeds the allowable range, a contrastive loss function drives the weight update of the bidirectional long short-term memory network. During training, the number of units in the forward hidden layer is fixed, and the forget gate bias term in the backward hidden layer is adjusted, with the bias term correction proportional to the mean of the time deviation. A time alignment loss term is introduced into the gradient vector generation layer at the network output, forcing the gradient peak position to approximate the actual abrupt change time. The updated network parameters are stored in a difference file and loaded for execution upon the next system startup.

[0119] The optimization of the fault mode determination module focuses on the reconstruction of the adjacency matrix using a graph convolutional neural network. The original adjacency matrix is ​​a binary matrix, where elements represent the existence of physical connections between nodes. The optimization module analyzes historical error cases and finds that mechanical connections cannot fully reflect the intensity of fault propagation. The reconstruction process introduces a vibration propagation attenuation coefficient as edge weights, calculated using the coherence function of the sensor array. Specifically, it iterates through all connected node pairs in the equipment topology graph, querying the integral value of the vibration signal transfer function amplitude spectrum of that node pair during maintenance. This integral value is normalized and used as the weight of the adjacency matrix. When the new weighted adjacency matrix is ​​input into the graph convolutional layer, the feature aggregation process strengthens the information transmission of high-weight edges. The template vector of the knowledge graph is simultaneously fine-tuned; the feature patterns of newly added maintenance cases are updated using a moving average method after principal component analysis.

[0120] The visualization interface module receives structured data from the diagnostic decision output module. Fault level quantification indicators are converted into color-coding rules for the 3D model. The color mapping is as follows: minor anomalies are marked in yellow, moderate degradation in orange, and severe faults are marked with flashing red. A component replacement priority list drives the generation of spatial markers in the 3D model. The first component on the list is highlighted with a border at its corresponding coordinates on the equipment model, with the component number and fault type abbreviation superimposed within the border. Maintenance strategy suggestion text is processed by a natural language processing engine to extract key operational verbs and object nouns, generating operation guidance flowchart nodes. Flowchart nodes are arranged chronologically, with the starting node containing equipment shutdown instructions, intermediate nodes describing disassembly steps and inspection actions, and the ending node indicating component replacement or parameter adjustment operations. When the flowchart is pushed to mobile terminals, it automatically adapts to the screen size, and touch interaction supports step-by-step preview and single-step zoom-in.

[0121] The visualization of the fault development timeline uses a dual-axis curve. The horizontal axis represents the continuous operating time of the equipment, with the time unit dynamically scaled according to the total duration. The vertical axis displays the normalized feature intensity values; the main curve trajectory reflects changes in feature intensity, while the auxiliary trajectory displays the rate of change gradient as a semi-transparent band. Inverted triangle markers are embedded at the curve peaks, with the predicted fault time point displayed within each marker. Equipment operating cycle markers are overlaid in the background to help observe the phase relationship between feature mutations and equipment operating conditions. The entire display interface supports multi-touch zooming, and the timeline can be expanded to view minute-level evolution details.

[0122] All adjustment records in the model optimization module are stored in the version database. Each optimization generates a unique version number, recording the type of adjustment and a snapshot of the parameters before and after the modification. The visualization interface module provides version comparison functionality, and the 3D model annotation style can switch between different optimization versions to display their effects. The operation guidance flowchart retains historical version backtracking, allowing maintenance personnel to view the iteration process of strategy suggestions. The time series analysis curves support the overlay display of predicted paths from multiple optimization versions, and the actual fault development trajectory is displayed as a reference benchmark by a bold solid line running through it.

[0123] During maintenance, sensor calibration data triggers a real-time feedback loop. After the actual surface temperature values ​​of the equipment collected by the infrared calibrator are input into the system, the temperature distribution map generation algorithm updates the pixel interpolation parameters. The calibrated temperature matrix is ​​then re-involved in the construction of the higher-order fault feature tensor, and changes in the tensor trigger recalculation of the time-series path. The recalculation results are displayed in the visualization interface; the original predicted path is converted to a light gray background trajectory, and the new path is displayed as a colored solid line. Mobile terminals simultaneously receive path update notifications, allowing maintenance personnel to compare the differences in feature intensity before and after calibration in real time. The entire optimization and visualization process forms a closed loop, and the system version continues to evolve with accumulated maintenance experience without interrupting equipment operation.

[0124] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0125] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent fault diagnosis system for industrial equipment based on deep learning algorithms, characterized in that, The system includes: The multi-dimensional operation status acquisition module acquires the vibration sensing sequence, temperature distribution spectrum and current waveform characteristics of industrial equipment in real time, and constructs a multi-dimensional heterogeneous data stream. The deep feature extraction module receives the multidimensional heterogeneous data stream and extracts the time-frequency domain features of the vibration spectrum, the spatial gradient features of the temperature distribution, and the energy accumulation features of the current harmonics through parallel convolutional neural network branches, and fuses them to generate a high-order fault feature tensor. The temporal evolution analysis module inputs the high-order fault feature tensor into a bidirectional long short-term memory network to capture the propagation pattern and intensity change trend of fault features within a continuous time window, and outputs the temporal path of fault development. The multi-source feature association module matches the node positions in the equipment structure topology diagram according to the fault development time sequence path, calculates the coupling degree between the vibration feature propagation path and the current abnormal area, and establishes the mapping relationship between fault features and equipment physical location. Based on the mapping relationship, the fault mode determination module uses a graph convolutional neural network to analyze the feature correlation strength between the topology nodes of the equipment, and combines the typical fault feature templates in the fault mode knowledge graph to output the fault type identification result and the fault location coordinate set. The diagnostic decision output module generates a set of decision instructions based on the fault type identification results and the fault location coordinate set, which includes fault level quantification indicators, component replacement priority list and maintenance strategy suggestions.

2. The intelligent fault diagnosis system for industrial equipment based on deep learning algorithm according to claim 1, characterized in that, The multi-dimensional operation status acquisition module includes: The vibration sensing unit collects triaxial vibration signals of the key bearings of the equipment through a distributed accelerometer array, and generates a vibration sensing sequence after wavelet noise reduction processing. The infrared scanning unit uses a thermal imager to periodically acquire the temperature distribution of the device casing and generates a temperature distribution map through pixel-level temperature calibration. The current monitoring unit uses a high-precision current transformer to capture the fundamental and harmonic components of the motor drive current, and extracts the current waveform characteristics through Fourier transform.

3. The intelligent fault diagnosis system for industrial equipment based on deep learning algorithm according to claim 1, characterized in that, The deep feature extraction module includes: The vibration feature extraction submodule uses a one-dimensional dilated convolutional neural network to process the vibration sensing sequence, extracting impact pulse features in the time domain and extracting the energy ratio of the resonant frequency band in the frequency domain. The temperature feature extraction submodule analyzes the spatial gradient changes of the temperature distribution map through a two-dimensional convolutional neural network to identify the boundaries of local overheated regions and the direction of heat conduction. The current feature extraction submodule uses a convolutional attention mechanism network to separate the fundamental distortion component and the amplitude of specific harmonics in the current waveform features, and quantifies the energy accumulation characteristics of each harmonic. The feature fusion submodule splices the impact pulse features, resonant frequency band energy ratio, local overheating region boundary, heat conduction direction, fundamental distortion component and specific subharmonic amplitude into tensors, and generates a high-order fault feature tensor through dimensionality reduction by a fully connected layer.

4. The intelligent fault diagnosis system for industrial equipment based on deep learning algorithm according to claim 3, characterized in that, The time-series evolution analysis module performs the following operations: The higher-order fault feature tensor is sliced ​​into equal-length segments and input into a bidirectional long short-term memory network. Extract feature evolution patterns of historical time slices through forward hidden layers; Capture the feature decay trend of future time slices through backward hidden layers; By fusing the output states of the forward hidden layer and the backward hidden layer, a gradient vector representing the rate of change of fault feature intensity is generated. Based on the alignment of the gradient vector peak position with the equipment operating cycle, the fault development timeline path is output.

5. The intelligent fault diagnosis system for industrial equipment based on deep learning algorithm according to claim 1, characterized in that, The multi-source feature association module includes: The topology loading submodule calls the three-dimensional coordinates and connection relationships of each node in the device structure topology diagram; The path mapping submodule maps the feature intensity peaks in the fault development time-series path to the corresponding nodes in the topology graph; The coupling analysis submodule calculates the spatial overlap between the vibration characteristic propagation path and the current anomaly region, and combines the coordinates of the temperature distribution anomaly region to generate a mapping relationship between fault characteristics and the physical location of the equipment.

6. The intelligent fault diagnosis system for industrial equipment based on deep learning algorithm according to claim 5, characterized in that, The fault mode determination module includes: The knowledge graph loading submodule reads typical fault feature templates such as bearing wear, gear tooth breakage, and winding short circuit from the fault mode knowledge graph; The node association analysis submodule uses a graph convolutional neural network to traverse the device structure topology graph and calculate the feature association strength of each node in the mapping relationship. The pattern matching submodule calculates the similarity between the feature association strength and the typical fault feature template, and outputs the fault type identification result and the fault location coordinate set.

7. The intelligent fault diagnosis system for industrial equipment based on deep learning algorithm according to claim 1, characterized in that, The diagnostic decision output module includes: The grading and quantification submodule classifies three levels of quantitative indicators—minor anomalies, moderate degradation, and severe faults—based on the rate of change of characteristic intensity in the fault development timeline. The component analysis submodule matches the replacement cost and downtime impact coefficient of components in the equipment spare parts library with the set of fault location coordinates to generate a priority list of component replacements. The strategy generation submodule combines fault level quantification indicators with component replacement priority lists to formulate three maintenance strategy recommendations: immediate shutdown for repair, online monitoring operation, and preventive maintenance plan, forming a decision instruction set.

8. The intelligent fault diagnosis system for industrial equipment based on deep learning algorithm according to claim 7, characterized in that, The system also includes: The real-time feedback module receives sensor verification data during equipment maintenance and updates the acquisition parameters of the multi-dimensional operating status acquisition module. The real-time feedback module inputs the maintained equipment operating status data into the deep feature extraction module to generate a verified high-order fault feature tensor. The time-series evolution analysis module corrects the fault development time-series path based on the verified higher-order fault feature tensor.

9. The intelligent fault diagnosis system for industrial equipment based on deep learning algorithm according to claim 8, characterized in that, The system also includes: The model optimization module adjusts the kernel size of the convolutional neural network layer in the deep feature extraction module by comparing the actual fault types in the maintenance records with the prediction results. The model optimization module uses the contrastive loss function to update the weight parameters of the bidirectional long short-term memory network in the temporal evolution analysis module; The model optimization module is based on the adjacency matrix calculation method of the fault mode determination error rate optimization graph convolutional neural network.

10. The intelligent fault diagnosis system for industrial equipment based on deep learning algorithm according to claim 1, characterized in that, The system also includes: The visualization interface module receives the fault level quantification index and component replacement priority list from the decision instruction set and generates a fault location annotation map of the equipment's three-dimensional model. The visualization interface module transforms maintenance strategy suggestions into operation guidance flowcharts, which are then simultaneously pushed to mobile terminals and the central control console. The visualization interface module dynamically displays the projection trajectory of the fault development timeline output by the timeline evolution analysis module onto the device's operating timeline.

Images