Intelligent diagnosis system for industrial equipment failure based on deep learning algorithm

Abstract

The application relates to the technical field of industrial equipment diagnosis, and discloses an intelligent industrial equipment fault diagnosis system based on a deep learning algorithm. A deep feature extraction module of the system extracts various types of features through parallel convolutional neural network branches and fuses the features to generate a high-order fault feature tensor; a time sequence evolution analysis module adopts a bidirectional long short-term memory network to capture the propagation law and change trend of the fault features, and outputs a fault development time sequence path; a multi-source feature correlation module matches equipment structure topology graph nodes, and establishes a mapping relationship between the fault features and the physical positions of the equipment; a fault mode judgment module utilizes a graph convolutional neural network in combination with a fault mode knowledge graph to output a fault type recognition result and a fault position coordinate set; and a diagnosis decision output module generates a decision instruction set containing a fault level, a component replacement priority and a maintenance strategy. The system improves the intelligent level and the precision level of industrial equipment fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment diagnosis, in particular to an intelligent industrial equipment fault diagnosis system based on a deep learning algorithm. BACKGROUND

[0002] In the field of industrial production, the stable operation of equipment is directly related to production efficiency and job safety, so timely diagnosis and accurate positioning of equipment faults have always been the focus of the industry. Traditional industrial equipment fault diagnosis methods rely on manual inspection and experience-based judgment, which is not only time-consuming and labor-intensive, but also difficult to meet the needs of multi-dimensional and high-dynamic operation state monitoring of equipment in complex industrial environments.

[0003] With the improvement of industrial automation level, the amount of sensor data generated by equipment operation grows exponentially, and multiple types of parameters such as vibration, temperature, and current are interrelated. Traditional single-dimensional feature analysis methods often cannot capture early signs and evolution laws of faults. For example, frequency spectrum analysis of vibration signals alone may misjudge the fault type and ignore the coupling relationship between temperature abnormalities and current fluctuations; and monitoring systems based on simple threshold values are difficult to adapt to the dynamic changes of equipment under different working conditions, resulting in high false positive and false negative rates.

[0004] Existing fault diagnosis systems have obvious limitations in feature extraction and fault pattern matching. On the one hand, manually designed feature extraction operators cannot adapt to the diversity of complex industrial scenarios, resulting in insufficient feature expression capability; on the other hand, fault pattern matching relies on static rule bases and lacks dynamic analysis of fault temporal evolution processes, making it difficult to accurately trace the propagation path and impact range of faults, thereby affecting the timeliness and effectiveness of maintenance decisions. The existence of these problems makes it difficult for industrial equipment fault diagnosis to achieve intelligent and accurate levels to meet the efficient operation and maintenance needs of modern industrial production. SUMMARY

[0005] The present application aims to provide an intelligent industrial equipment fault diagnosis system based on a deep learning algorithm to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides an intelligent industrial equipment fault diagnosis system based on a deep learning algorithm, which comprises:

[0007] A multi-dimensional operation state acquisition module acquires vibration sensor sequences, temperature distribution maps, and current waveform features of industrial equipment in real time, and constructs a multi-dimensional heterogeneous data stream;

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

[0009] The time evolution analysis module inputs the high-order fault feature tensor into a bidirectional long short-term memory network, captures the propagation law and intensity change trend of the fault feature in a continuous time window, and outputs a fault development time sequence path;

[0010] The multi-source feature correlation module matches the node positions in the device structure topology graph according to the fault development time sequence path, calculates the coupling degree of the vibration feature propagation path and the current abnormal area, and establishes a mapping relationship between the fault feature and the device physical position;

[0011] The fault mode determination module analyzes the feature correlation strength between the nodes of the device topology based on the mapping relationship using a graph convolutional neural network, combines the typical fault feature templates in the fault mode knowledge graph, and outputs a fault type recognition result and a fault position coordinate set;

[0012] The diagnostic decision output module generates a decision instruction set containing fault level quantitative indicators, component replacement priority list and maintenance strategy suggestions according to the fault type recognition result and the fault position coordinate set.

[0013] Preferably, the multi-dimensional running state acquisition module comprises:

[0014] The vibration sensing unit acquires three-axis vibration signals of key bearings of the device through a distributed acceleration sensor array, and generates vibration sensing sequences after wavelet denoising processing;

[0015] The infrared scanning unit periodically acquires the temperature distribution of the device shell using a thermal imager, and generates a temperature distribution map through pixel-level temperature calibration;

[0016] The current monitoring unit captures the fundamental wave and harmonic components of the motor driving current using a high-precision current transformer, and extracts current waveform features through Fourier transform.

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

[0018] The vibration feature extraction submodule processes the vibration sensing sequence using a one-dimensional hollow convolutional neural network, extracts impact pulse features in the time domain, and extracts resonance frequency band energy proportion in the frequency domain;

[0019] The temperature feature extraction submodule analyzes the spatial gradient change of the temperature distribution map through a two-dimensional convolutional neural network, identifies the boundary of the local overheating area and the heat conduction direction;

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

[0021] The feature fusion sub-module concatenates the impact pulse feature, resonance frequency band energy proportion, local overheating region boundary, heat conduction direction, fundamental wave distortion component and specific harmonic amplitude to generate a high-order fault feature tensor through a full connection layer.

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

[0023] The high-order fault feature tensor is input into a bidirectional long short-term memory network through equal-time length slicing;

[0024] The feature evolution mode of the historical time slice is extracted through a forward hidden layer;

[0025] The feature attenuation trend of the future time slice is captured through a backward hidden layer;

[0026] The output states of the forward hidden layer and the backward hidden layer are fused to generate a gradient vector representing the rate of change of the fault feature intensity;

[0027] According to the alignment result of the peak position of the gradient vector and the equipment operation cycle, a fault development time sequence path is output.

[0028] Preferably, the multi-source feature correlation module comprises:

[0029] A topology loading sub-module calls the three-dimensional coordinates and connection relationships of each node in the device structure topology graph;

[0030] A path mapping sub-module maps the feature intensity peak in the fault development time sequence path to the corresponding node in the topology graph;

[0031] A coupling analysis sub-module calculates the spatial overlap degree of the vibration feature propagation path and the current abnormal region, and generates a mapping relationship between the fault feature and the physical position of the equipment in combination with the temperature distribution abnormal region coordinates.

[0032] Preferably, the fault mode judgment module comprises:

[0033] A knowledge graph loading sub-module reads the typical fault feature templates of bearing wear, gear tooth breakage and winding short circuit in the fault mode knowledge graph;

[0034] A node correlation analysis sub-module traverses the device structure topology graph using a graph convolutional neural network to calculate the feature correlation strength of each node in the mapping relationship;

[0035] A mode matching sub-module performs similarity calculation on the feature correlation strength and the typical fault feature templates to output a fault type recognition result and a fault position coordinate set.

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

[0037] The level quantization submodule divides the three-level quantization indexes of slight anomaly, moderate deterioration and serious fault according to the feature intensity change rate in the fault development time sequence path;

[0038] The component analysis submodule generates a component replacement priority list according to the component replacement cost and downtime influence coefficient in the equipment spare parts library matched with the fault position coordinate set;

[0039] The strategy generation submodule formulates three maintenance strategy suggestions of immediate shutdown maintenance, online monitoring operation and preventive maintenance plan combined with the fault level quantization index and the component replacement priority list, and forms a decision instruction set.

[0040] Preferably, the system further comprises:

[0041] The real-time feedback module receives sensor verification data in the equipment maintenance process, and updates the acquisition parameters of the multi-dimensional running state acquisition module;

[0042] The real-time feedback module inputs the equipment running state data after maintenance into the deep feature extraction module to generate a verified high-order fault feature tensor;

[0043] The time sequence evolution analysis module corrects the fault development time sequence path according to the verified high-order fault feature tensor.

[0044] Preferably, the system further comprises:

[0045] The model optimization module adjusts the convolution kernel size of the convolutional neural network layer in the deep feature extraction module according to the comparison between the actual fault type in the maintenance record and the prediction result;

[0046] The model optimization module updates the weight parameters of the bidirectional long short-term memory network in the time sequence evolution analysis module by using the contrast loss function;

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

[0048] Preferably, the system further comprises:

[0049] The visualization interface module receives the fault level quantization index and the component replacement priority list in the decision instruction set, and generates a fault positioning annotation diagram of the equipment three-dimensional model;

[0050] The visualization interface module converts the maintenance strategy suggestion into an operation guidance flowchart and synchronously pushes it to the mobile terminal and the central control console;

[0051] The visualization interface module dynamically displays the projection track of the fault development time sequence path output by the time sequence evolution analysis module on the equipment running time axis.

[0052] Compared with the prior art, the application has the beneficial effects that:

[0053] The industrial equipment fault intelligent diagnosis system based on the deep learning algorithm can fully capture the vibration, temperature, current and other multi-type parameter characteristics of the industrial equipment in the running process by constructing a multi-dimensional heterogeneous data stream through the multi-dimensional running state acquisition module, breaks the limitation of traditional single-dimensional data acquisition, and makes the perception of the system to the equipment running state more comprehensive and stereoscopic.

[0054] The deep feature extraction module adopts a parallel convolutional neural network branch to extract the features of different types of data and fuse to generate a high-order fault feature tensor, compared with the traditional manually designed feature operator mode, can automatically mine the complex correlation features hidden in the data, improve the expression ability of the features, and make the system more suitable for diversified data features in complex industrial scenes.

[0055] The time series evolution analysis module introduces a bidirectional long short-term memory network, effectively captures the propagation law and intensity change trend of the fault features in the continuous time window, changes the mode of the traditional system for static analysis of the fault, can dynamically track the evolution process of the fault, and provides strong support for accurately judging the development trend of the fault.

[0056] The multi-source feature correlation module combines the fault development time series path and the device structure topology graph, calculates the coupling degree between different features and establishes the mapping relationship with the physical position of the device, overcomes the problem that the features and the device position are disconnected in the traditional system, makes the fault features accurately correspond to the specific components of the device, and improves the accuracy of fault positioning.

[0057] The fault mode judgment module analyzes the feature correlation strength between the device topology nodes by using a graph convolutional neural network, and matches the fault mode knowledge graph, not only considers the correlation between the components of the device, but also learns from the typical fault feature template, improves the accuracy of fault type identification, and reduces the occurrence of misjudgment.

[0058] The diagnosis decision output module generates a decision instruction set containing fault level quantitative indicators, component replacement priority list and maintenance strategy suggestions, directly converts the fault diagnosis result into specific operation and maintenance guidance, makes the maintenance work more targeted and planned, helps to optimize the operation and maintenance process, and reduces the operation and maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The working principle diagram of the industrial equipment fault intelligent diagnosis system based on the deep learning algorithm described in the application;

[0060] Figure 2 The working flowchart of the multi-dimensional running state acquisition module;

[0061] Figure 3 Workflow diagram of the deep feature extraction module;

[0062] Figure 4 Workflow diagram of the fault mode determination module. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0064] Referring to Figure 1 The present application provides an industrial equipment fault intelligent diagnosis system based on a deep learning algorithm, which comprises:

[0065] The multi-dimensional running state acquisition module is used to acquire vibration sensing sequences, temperature distribution maps and current waveform features of the industrial equipment in real time, and to construct a multi-dimensional heterogeneous data stream.

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

[0067] The multi-source feature correlation module matches the node positions in the equipment structure topology graph according to the time sequence path, calculates the coupling degree of the vibration feature propagation path and the current abnormal area, and establishes a mapping relationship between the fault features and the physical positions of the equipment.

[0068] The fault mode determination module uses a graph convolutional neural network to analyze the feature correlation strength between the nodes of the equipment topology based on the mapping relationship, combines a typical fault feature template in a fault mode knowledge graph, and outputs a fault type recognition result and a fault position coordinate set.

[0069] The diagnosis decision output module generates a decision instruction set comprising a fault level quantitative index, a component replacement priority list and a maintenance strategy suggestion.

[0070] Embodiment 1: Referring to Figure 2The multi-dimensional running state acquisition module is composed of three units that work cooperatively to complete the all-around monitoring of the running state of the industrial equipment. The vibration sensing unit deploys a distributed acceleration sensor array, which covers the axial, radial and tangential positions of the key bearings of the equipment. Each sensor captures three-axis vibration original signals at a sampling rate of no less than 20 kHz. The original signals are transmitted to the preprocessing unit, which uses a wavelet transform algorithm for noise reduction processing, selects a suitable wavelet basis function to decompose the signal level, eliminates environmental noise interference through threshold filtering, and retains the impact components reflecting the mechanical state of the equipment. The noise-reduced signals are reconstructed into time-domain continuous vibration sensing sequences, which contain time stamps, three-axis 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 surface of the equipment shell at a preset scanning period. During the scanning process, the thermal imager captures the infrared radiation intensity of each pixel point in real time, and dynamically compensates for the environmental temperature, humidity and emissivity parameters. Through a pixel-level temperature calibration algorithm, environmental radiation interference is eliminated. The calibration process introduces a blackbody radiation reference source as a reference to ensure that the temperature error of each pixel point is controlled within a reasonable range. Finally, a temperature distribution map is generated, which is stored in a matrix structure, with each matrix element corresponding to the temperature value of a specific coordinate point on the equipment surface, with a spatial resolution of no less than 0.5 degrees Celsius. The scanning time and regional coordinate boundaries are also labeled.

[0072] The current monitoring unit uses a closed-loop high-precision current transformer, which is directly coupled to the motor drive cable to capture three-phase current waveforms in real time. The original current signals are subjected to amplitude adaptation and filtering processing by a signal conditioning circuit, and then converted into digital signals by a high-speed ADC module. Fast Fourier transform algorithm is used to analyze the frequency spectrum of the digital signals, calculate the fundamental frequency amplitude, total harmonic distortion rate 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 features. The generated current waveform features include fundamental effective value, 3-25 harmonic amplitude ratio, harmonic distortion rate and characteristic frequency energy aggregation coefficient.

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

[0074] The layout of the acceleration sensor array in the vibration sensing unit is designed according to the vibration transmission characteristics of the device. Three groups of three-axis sensors are arranged at the positions of the input end bearing, the output end bearing and the coupling side bearing of the gear box, and each group contains three orthogonally distributed sensors. The sensor signals are connected to an anti-interference acquisition card through a shielded cable, and the acquisition card is built-in with a programmable amplifier and a 24-bit analog-to-digital converter. The wavelet denoising process uses a multi-level decomposition and reconstruction strategy, which retains the approximate coefficients of low-frequency vibration components and sets the zero detail coefficients of high-frequency noise components. The reconstructed signal retains the effective frequency band below 5 kHz.

[0075] The gimbal control system of the infrared scanning unit moves the lens of the thermal imager according to the preset path, and the path planning covers all the key heat dissipation areas of the device shell. In each scanning period, the thermal imager acquires not less than 100 frames of infrared images, and the image registration algorithm eliminates the pixel shift caused by device vibration. In the temperature calibration stage, the reference blackbody source is placed in the same scene as the device, and by comparing the theoretical radiation value of the blackbody with the actual imaging value, a pixel gray scale-temperature mapping lookup table is established. The final output temperature distribution map is accompanied by a spatial coordinate system, and the origin of the coordinate system is aligned with the reference point of the device base.

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

[0077] The construction process of the multi-dimensional data stream includes time and space alignment operations. The time alignment module detects the timestamp difference of the three types of data, and when the difference exceeds 10 milliseconds, it starts the interpolation compensation algorithm. The space alignment module converts the pixel coordinates of the temperature distribution map into three-dimensional coordinates in the global coordinate system of the device, and the conversion process is based on the pose parameters of the thermal imager gimbal and the lens distortion correction parameters. The final generated data stream is packaged and transmitted at fixed time intervals, and each data packet contains the vibration waveform, temperature matrix and current spectrum vector within a complete time slice, and the data packet header marks the device number and the time interval.

[0078] Example 2: see Figure 3, the deep feature extraction module is composed of four sub-modules that work collaboratively to process raw information from multi-dimensional heterogeneous data streams. The vibration feature extraction sub-module receives a vibration sensor sequence, which is a stream of three-axis acceleration values in the time domain. A one-dimensional dilated convolutional neural network processes this sequence, with a network structure containing two main processing stages. The first processing stage sets the dilation factor to two, and the convolution kernel skips some data points when sliding along the time axis, expanding the receptive field range. This stage focuses on identifying transient impact components in the vibration signal, extracting features such as pulse peak amplitude, rise time, and repetition period. The second processing stage connects a frequency spectrum analysis layer, converting the convolution output to the frequency domain and calculating the proportion of energy in a specific frequency band range to the total energy, focusing on characteristic frequency bands that reflect the state of bearings or gears. The frequency spectrum analysis uses a sliding window strategy, with the window length synchronized with the device rotation period.

[0079] The temperature feature extraction sub-module processes temperature distribution maps, which are represented as two-dimensional matrices of device surface temperature fields. A two-dimensional convolutional neural network analyzes the matrix, with a convolution kernel size of three by three, moving along the spatial dimensions of the device surface. The first convolution operation detects temperature differences between adjacent pixels, outputting a feature map identifying temperature mutation regions. Subsequent convolution layers gradually expand the receptive field, identifying temperature gradient change directions and heat flow transmission paths. The deep network output includes a set of local high-temperature region boundary coordinates and a vector direction data from the high-temperature center to the low-temperature region, with the vector length representing heat conduction strength.

[0080] The current feature extraction sub-module processes current waveform feature vectors, including fundamental wave and harmonic component parameters. The convolutional attention mechanism network first decomposes the input features through a one-dimensional convolution layer, separating the fundamental wave distortion component and quantifying its deviation from the standard sinusoidal wave. The attention mechanism acts on the harmonic components, calculating feature weights for different harmonics. Certain harmonic components are given higher weights, and their amplitude features are extracted. The network output includes the fundamental wave distortion rate, weighted specific harmonic amplitudes, and harmonic energy aggregation coefficients.

[0081] The feature fusion sub-module receives the output data from the above three sub-modules. The vibration sub-module provides an impact pulse feature vector and a frequency band energy proportion scalar. The temperature sub-module outputs a set of boundary coordinates and a heat conduction direction vector. The current sub-module submits a fundamental wave distortion rate and a harmonic amplitude sequence. The fusion process first tensor concatenates the six types of heterogeneous features to form a six-channel feature map. The concatenated tensor is input into a fully connected layer network, with the first fully connected layer compressing the feature dimension to sixty percent of the original size, and the activation function introducing a non-linear transformation. The second fully connected layer further reduces the dimension, finally outputting a fixed two hundred and fifty-six-dimensional high-order fault feature tensor. Each dimension of this tensor represents a composite fault feature that fuses vibration, temperature, and current information.

[0082] The time series evolution analysis module processes the high-order fault feature tensor sequence. The tensor sequence is divided by fixed time length slicing, and each slice covers ten seconds of device running time. The slice data is input into a bidirectional long short-term memory network, which contains two processing chains, forward and backward. The forward processing chain reads the historical eight time slices in order, and the hidden layer state records the pattern of feature enhancement over time. The backward processing chain reads the future four time slices in reverse order, and the hidden layer state captures the feature attenuation trend. The bidirectional hidden layer state is spliced and fused at the output end, and the fused state is converted into a gradient vector through a linear layer. The gradient vector represents the rate of change of the fault feature intensity per unit time, and its local peak point corresponds to the mutation time of the fault feature. The network output is a time-feature intensity sequence, which labels the feature intensity value and change rate of each time point, forming a fault development time series path. This path describes the whole process of the fault feature from generation, enhancement to attenuation.

[0083] In the vibration feature extraction submodule, the residual connection structure is adopted in the cavity convolutional neural network. The output of each convolutional layer is added to the input to avoid gradient disappearance in the training process of the deep network. The adaptive frequency band selection mechanism is set in the spectrum analysis layer to dynamically adjust the feature frequency band center frequency according to the device speed. The attention network of the current feature extraction submodule contains two mechanisms of channel attention and spatial attention. The channel attention recalibrates the importance weight of each harmonic component, and the spatial attention locates the key distortion event along the time axis. The batch normalization processing is used in the fully connected layer of the feature fusion submodule to accelerate the network convergence speed. The bidirectional long short-term memory network of the time series analysis module sets a gating mechanism, the input gate controls the proportion of new information flow, the forgetting gate adjusts the retention degree of historical memory, and the output gate determines the exposure amount of the current state. The peak detection algorithm is used in the gradient vector generation stage to identify the turning point of the feature intensity change rate, and the turning point is aligned with the device running cycle marker to determine the fault development stage division.

[0084] Embodiment 3: see Figure 4The multi-source feature correlation module processes the fault development time sequence path from the time sequence evolution analysis module, which contains time-feature intensity sequence information. The module operation starts with the topology loading submodule accessing the device structure topology graph database, which stores the three-dimensional coordinates and mechanical connection relationships of each component of the device in a graph structure. The nodes in the topology graph represent physical components of the device, such as bearings, gears, or motor windings, and each node is assigned a unique identifier and three-dimensional coordinate value; the edges represent the mechanical coupling relationships between components, such as shaft connections or gear meshing, and the edge attributes include connection type and distance coefficient. The topology graph data format uses an adjacency list representation for fast query of node reachability. The path mapping submodule receives the fault development time sequence path and parses the time points and peak amplitudes corresponding to the feature intensity peaks in the path. The peak detection algorithm identifies the mutation time based on the gradient change rate threshold, and maps each peak time point to the topology graph node. The mapping process aligns the time points with the device state snapshots based on the device operation log, for example, the time point t = 125 seconds corresponds to bearing node B7, and the peak amplitude is recorded as 0.75 normalized units. The mapping result generates a node-peak correspondence table, which includes node ID, coordinates, peak time, and intensity value.

[0085] The coupling analysis submodule uses the node-peak correspondence table to analyze the spatial correlation of multi-source features. The submodule first extracts the vibration feature propagation path, which is derived from the path mapping result, representing the transmission direction of the fault feature in the device, such as from node B7 to node G12 to node M3, forming a directed sequence. At the same time, the current abnormal region is identified from the current monitoring data, which is defined as the node set where the harmonic distortion rate in the current waveform feature exceeds the limit, such as motor winding nodes M3 to M5. The spatial overlap degree calculation focuses on the intersection of the vibration path node set and the current abnormal node set. The calculation process introduces the node coincidence rate index to quantify the matching degree of the two sets. When the coincidence rate exceeds the preset threshold, and combined with the temperature distribution map to verify that there is a local temperature rise anomaly in the region, the mapping relationship between the fault feature and the physical location is generated. The mapping relationship table structure includes 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 abnormal 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 adds an entry: node M3 is associated with vibration impact and current harmonic features, coordinates (x = 1.2m, y = 0.8m, z = 0.5m), and coupling strength is set to 0.8.

[0086] The fault mode determination module works based on the above mapping relationship table. The knowledge graph loading submodule calls the pre-constructed fault mode knowledge graph, which is stored in the form of a graph database. The nodes represent fault types, such as bearing wear, gear tooth breakage, or winding short circuit. The edges represent the causal relationship or concurrent probability between faults. Each fault type node is attached to a typical feature template, which is derived from a historical fault case library. For example, the bearing wear template contains the envelope shape parameters of the vibration impact pulse, the gear tooth breakage template records the frequency offset value of the harmonic sideband, and the winding short circuit template defines the threshold range of the amplitude of odd harmonics. The template is encoded in the form of a feature vector, with the dimension matching the node embedding space. The node association analysis submodule uses a graph convolutional neural network to process the device topology graph, taking the nodes in the mapping relationship table as input vertices and the mechanical connection edges as graph edges. The network structure contains three layers of graph convolution layers, each of which performs a neighborhood feature aggregation operation: the first layer aggregates the features of directly adjacent nodes, and the subsequent layers extend to the second-order neighborhood. The aggregation process outputs a node embedding vector, with a dimension of 128, representing the association strength of the node in fault propagation. For example, for node M3, the embedding vector is calculated by integrating its own features and the influence of neighbor nodes G12 and M4.

[0087] The pattern matching submodule compares the similarity between the node embedding vector and the fault feature template, and outputs the fault type recognition result. The similarity calculation uses cosine similarity measurement, with the formula as follows:

[0088]

[0089] wherein: represents the node embedding vector, derived 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 the preset matching threshold, for example, the bearing wear template threshold is set to 0.85. If the embedding vector of node M3 has a similarity greater than 0.85 with the bearing wear template, the fault type is determined to be "bearing wear". At the same time, the associated fault position coordinate set is extracted, with the three-dimensional coordinates of the corresponding nodes extracted from the mapping relationship table, such as {(x=1.2m, y=0.8m, z=0.5m)}. The output result includes the fault type label, the coordinate list, and the similarity confidence score. The entire determination process ensures the consistency of multi-source features, for example, when the vibration and current features point to the same node and the temperature anomaly is confirmed, the fault positioning accuracy is improved.

[0091] In the topology loading submodule, the device structure topology graph is imported through a computer-aided design model, the model format is STEP or IGES, and the model is stored in a graph database after conversion. The node coordinates are in meters, with an accuracy of up to millimeters; the edge attributes include the connection stiffness coefficient, which is used for weighted graph calculation. The peak mapping of the path mapping submodule uses the nearest neighbor algorithm, and when there is no direct log matching at the peak time point, the node is located based on the device operation cycle interpolation. The spatial overlap calculation of 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 in the overlapping node area exceeds the 15-degree Celsius threshold, which is set according to the heat capacity of the device material. When calculating the overlap rate, the size of the node set is dynamically adjusted to avoid distortion of small sets. The construction of the fault mode knowledge graph relies on the annotation of domain experts, and the template vector is reduced in dimension through principal component analysis, retaining the key feature dimensions. The graph convolution network of the node association analysis submodule is set to skip connection to prevent overfitting, and random dropout regularization is used during training. The cosine similarity calculation of the pattern matching submodule is implemented on an embedded system to optimize vector operation speed, and the similarity result is post-processed to filter low-confidence matches. The final output interface encapsulates the fault type and coordinate set as a structured message and transmits it to the diagnostic decision module. The entire implementation emphasizes data-driven, without human intervention, and the processing delay is controlled within milliseconds to meet real-time diagnostic needs.

[0092] In the diagnostic decision output module, the fault type recognition result and the fault location coordinate set are received, and a bearing wear fault is taken as an example for specific implementation. The level quantization submodule analyzes the feature intensity changes in the fault development time sequence path, assuming that the feature intensity F = 0.25 at time point t = 120 seconds, F = 0.72 at time point t = 180 seconds, and the time interval Δt = 60 seconds, the calculation unit time change amount is

[0093]

[0094] According to the preset threshold rule: as a slight abnormality, as moderate degradation, as a serious fault, 0.0078 belongs to a slight abnormality in this example, and the fault level quantization indicator "Level1" is output.

[0095] The component analysis submodule queries the device spare parts library to match 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 three-dimensional coordinates of the component, the name, the replacement cost (thousands of yuan), and the downtime impact coefficient (1-10 level). The downtime impact coefficient is determined by the criticality of the device function: the coefficient of the main transmission component is greater than 8, and the coefficient of the auxiliary component is less than 4. In this example, the coordinates match the "input shaft bearing", and its attributes are shown in the following table:

[0096]

[0097] Priority weight calculation rule: weight = replacement cost x downtime impact factor. Input shaft bearing weight = 2.0 x 8.0 = 16.0, generate component replacement priority list in descending order of weight:

[0098] Input shaft bearing (weight 16.0)

[0099] Cooling fan (weight 2.4)

[0100] Motor carbon brush (weight 2.0)

[0101] Strategy generation submodule combines fault level "Level 1" and priority list to develop maintenance strategy:

[0102] Immediate shutdown for repair: weight > 10.0 and fault level ≥ moderate deterioration.

[0103] Online monitoring operation: weight ≤ 5.0 and fault level = slight anomaly.

[0104] 48-hour preventive maintenance: other combined conditions.

[0105] In this example, the highest weight 16.0 > 10.0 but the fault level is slight anomaly, triggering "online monitoring operation + 48-hour preventive maintenance" instruction. Instruction details:

[0106] Maintain equipment running state, collect vibration data every 10 minutes.

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

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

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

[0110] Real-time feedback module intervenes during maintenance. When the technician uses the portable infrared calibrator to detect the bearing seat temperature, it is found that the original temperature distribution map has a deviation of ±3℃. After uploading the calibration data, the module updates the pixel calibration parameters of the infrared scanning unit: adjust the original environmental radiation compensation coefficient 0.92 to 0.89. After maintenance, the equipment restarts operation, and the multi-dimensional running state acquisition module obtains new data:

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

[0112] Temperature distribution map: the highest temperature in the bearing seat area decreases from 68℃ to 52℃.

[0113] Current waveform feature: 3rd harmonic amplitude ratio decreased from 5.2% to 1.8%.

[0114] The new data input deep feature extraction module generates the verified high-order fault feature tensor. The time evolution analysis module reprocesses the tensor sequence: the feature intensity of the original fault development time sequence path 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 that the fault development rate slows down, and the system automatically adjusts the fault level from "Level 1" to "monitoring state".

[0115] The entire implementation process embodies dynamic closed-loop control: the initial diagnosis generates decisions based on sensor raw data, feedback field verification data during maintenance action execution, real-time update of acquisition module parameters (such as infrared calibration coefficients), re-calculation of data-driven feature tensors after maintenance, and time sequence path correction triggering iterative diagnosis conclusions.

[0116] For example, after replacing the input shaft bearing, the fault position coordinate set is emptied, the component replacement priority list removes this item, and the strategy generation submodule terminates the preventive maintenance instruction and outputs the "device running normally" status code. Real-time feedback data is also stored in the historical case library for optimizing subsequent diagnosis thresholds.

[0117] Example 5: The model optimization module is activated after the completion of equipment maintenance operations. The module first reads the maintenance record document, which records the comparison information between the actual fault type and the system prediction result. When the related misjudgment of vibration features exceeds the set threshold, the module starts the convolution kernel adjustment program of the deep feature extraction module. The adjustment process analyzes the spectral distribution characteristics of the misjudgment cases, and if the lack of low-frequency feature recognition is the main error cause, the size of the first convolution kernel of the one-dimensional convolutional neural network is expanded by two units from the original specification, and the expanded convolution kernel covers a wider time window. The convolution kernel weights are reinitialized, and new parameters are generated using truncated normal distribution. The attention network of the current feature extraction submodule is also optimized, and the number of attention heads is dynamically increased according to the complexity of the harmonic components, and new attention heads focus on specific harmonic combination patterns.

[0118] The optimization of the time 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 timestamp of the fault feature mutation point, and the module calculates the absolute difference between this timestamp and the predicted peak time point. When the difference continuously exceeds the allowed range, the bidirectional long short-term memory network's weight update is driven by the comparison loss function. The training process fixes the number of forward hidden layer units and adjusts the forget gate bias term of the backward hidden layer. The bias correction amount is directly proportional to the average time deviation. The gradient vector generation layer at the network output end introduces a time alignment loss term, forcing the gradient peak position to approach the actual mutation time. The updated network parameters are stored in a difference file, which is loaded and run the next time the system starts.

[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 the adjustment records of the model optimization module are stored in the version database. Each optimization generates an independent version number, and the adjustment item type and parameter snapshots before and after modification are recorded. The visualization interface module provides version comparison function, and the three-dimensional model label style can switch the effect display of different optimization versions. The operation guidance flowchart retains historical version backtracking, and maintenance personnel can view the iterative process of strategy suggestions. The time series analysis curve supports superimposed display of the prediction paths of multiple optimization versions, and the actual fault development trajectory is displayed as a thick solid line as a reference benchmark.

[0123] The sensor calibration data during maintenance triggers a real-time feedback loop. After the real temperature value of the equipment surface collected by the infrared calibrator is input into the system, the temperature distribution map generation algorithm updates the pixel interpolation parameters. The calibrated temperature matrix participates in the construction of the high-order fault feature tensor again, and the change of the tensor triggers the recalculation of the time series path. The recalculation result is displayed in the visualization interface, the original prediction path is converted into a light gray background track, and the new path is displayed in color solid line. The mobile terminal synchronously receives the path update notification, and the maintenance personnel can compare the feature intensity difference before and after calibration in real time. The whole optimization and visualization process forms a closed loop, and the system version continuously evolves with the accumulation of maintenance experience, without interrupting the equipment operation.

[0124] It should be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0125] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. An intelligent diagnosis system for industrial equipment failure based on a deep learning algorithm, characterized in that, The system comprises: The multi-dimensional running state acquisition module acquires vibration sensing sequences, temperature distribution maps and current waveform features of the industrial equipment in real time, and constructs multi-dimensional heterogeneous data streams; The deep feature extraction module receives the multi-dimensional heterogeneous data streams, extracts time-frequency domain features of vibration frequency spectrum, spatial gradient features of temperature distribution and energy aggregation features of current harmonics through parallel convolutional neural network branches respectively, and fuses to generate high-order fault feature tensors; The time sequence evolution analysis module inputs the high-order fault feature tensors into a bidirectional long short-term memory network, captures the propagation law and intensity change trend of fault features in a continuous time window, and outputs a fault development time sequence path; The multi-source feature correlation module matches node positions in a device structure topology graph according to the fault development time sequence path, calculates the coupling degree of vibration feature propagation paths and current abnormal regions, and establishes a mapping relationship between fault features and device physical positions; The fault mode determination module analyzes the feature correlation strength between device topology nodes based on the mapping relationship, combines typical fault feature templates in a fault mode knowledge graph, and outputs a fault type recognition result and a fault position coordinate set; The diagnosis decision output module generates a decision instruction set including fault level quantitative indicators, component replacement priority lists and maintenance strategy suggestions according to the fault type recognition result and the fault position coordinate set; The multi-source feature correlation module comprises: The topology loading submodule calls the three-dimensional coordinates and connection relationships of each node in the device structure topology graph; The path mapping submodule maps the feature intensity peaks in the fault development time sequence path to the corresponding nodes of the topology graph; The coupling analysis submodule calculates the spatial overlap degree of vibration feature propagation paths and current abnormal regions, combines temperature distribution abnormal region coordinates, and generates a mapping relationship between fault features and device physical positions; The fault mode determination module comprises: The knowledge graph loading submodule reads typical fault feature templates of bearing wear, gear tooth breakage and winding short circuit in the fault mode knowledge graph; The node correlation analysis submodule traverses the device structure topology graph using a graph convolutional neural network, and calculates the feature correlation strength of each node in the mapping relationship; The mode matching submodule performs similarity calculation on the feature correlation strength and the typical fault feature templates, and outputs a fault type recognition result and a fault position coordinate set. 2.The deep learning algorithm-based intelligent diagnosis system for industrial equipment faults according to claim 1, characterized in that, The multi-dimensional running state acquisition module comprises: The vibration sensing unit acquires three-axis vibration signals of key bearings of the equipment through a distributed acceleration sensor array, generates vibration sensing sequences after wavelet denoising processing; The infrared scanning unit periodically acquires temperature distribution of the equipment shell using a thermal imager, and generates a temperature distribution map through pixel-level temperature calibration; The current monitoring unit captures fundamental wave and harmonic components of motor driving current using a high-precision current transformer, and extracts current waveform features through Fourier transform. 3.The industrial equipment fault intelligent diagnosis system based on deep learning algorithm according to claim 1, characterized in that, The deep feature extraction module comprises: The vibration feature extraction submodule processes the vibration sensing sequences using a one-dimensional hollow convolutional neural network, extracts impact pulse features in the time domain, and extracts resonance frequency band energy proportion in the frequency domain; The temperature feature extraction submodule analyzes the spatial gradient change of the temperature distribution map through a two-dimensional convolutional neural network, identifies the boundary of the local overheating area and the heat conduction direction; The current feature extraction submodule separates the fundamental wave distortion component and the specific harmonic amplitude in the current waveform feature using a convolutional attention mechanism network, and quantifies the energy aggregation characteristics of each harmonic; The feature fusion submodule tensor splices the impact pulse feature, the resonance frequency band energy proportion, the local overheating area boundary, the heat conduction direction, the fundamental wave distortion component and the specific harmonic amplitude, and generates a high-order fault feature tensor through a fully connected layer. 4.The industrial equipment fault intelligent diagnosis system based on deep learning algorithm according to claim 3, characterized in that, The time evolution analysis module performs the following operations: The high-order fault feature tensor is input into a bidirectional long short-term memory network by equal-time length slicing; The feature evolution pattern of the historical time slice is extracted through the forward hidden layer; The feature attenuation trend of the future time slice is captured through the backward hidden layer; The gradient vector representing the change rate of the fault feature intensity is generated by fusing the output states of the forward hidden layer and the backward hidden layer; According to the alignment result of the gradient vector peak value position and the device operation cycle, the fault development time sequence path is output. 5.The industrial equipment fault intelligent diagnosis system based on deep learning algorithm according to claim 1, characterized in that, The diagnostic decision output module includes: The level quantization submodule divides the three-level quantization indicators of slight anomaly, moderate degradation and serious fault according to the feature intensity change rate in the fault development time sequence path; The component analysis submodule generates a component replacement priority list by matching the component replacement cost and the downtime influence coefficient in the equipment spare parts library according to the fault position coordinate set; The strategy generation submodule formulates three maintenance strategy suggestions of immediate shutdown for repair, online monitoring operation and preventive maintenance plan, and forms a decision instruction set in combination with the fault level quantization indicator and the component replacement priority list. 6.The industrial equipment fault intelligent diagnosis system based on deep learning algorithm according to claim 5, characterized in that, The system further includes: The real-time feedback module receives sensor verification data during equipment maintenance, and updates the acquisition parameters of the multi-dimensional running state acquisition module; The real-time feedback module inputs the equipment running state data after maintenance into the deep feature extraction module to generate the verified high-order fault feature tensor; The time evolution analysis module corrects the fault development time sequence path according to the verified high-order fault feature tensor. 7.The industrial equipment fault intelligent diagnosis system based on deep learning algorithm according to claim 6, characterized in that, The system further includes: The model optimization module adjusts the convolution kernel size of the convolutional neural network layer in the deep feature extraction module according to the comparison of the actual fault type and the prediction result in the maintenance record; The model optimization module updates the weight parameters of the bidirectional long short-term memory network in the time evolution analysis module using the contrast loss function; The model optimization module optimizes the adjacency matrix calculation method of the graph convolutional neural network based on the fault mode judgment error rate. 8.The industrial equipment fault intelligent diagnosis system based on deep learning algorithm of claim 1, wherein, The system further includes: The visualization interface module receives the fault level quantization indicator and the component replacement priority list in the decision instruction set, and generates a fault positioning annotation map of the equipment three-dimensional model; The visualization interface module converts the maintenance strategy suggestion into an operation guidance flowchart and synchronously pushes it to the mobile terminal and the central control console; The visualization interface module dynamically displays the projection trajectory of the fault development time sequence path output by the time evolution analysis module on the equipment operation time axis.

Images