Industrial equipment fault diagnosis and tracing method based on multi-source heterogeneous parameter fusion and deep learning network

Abstract

This invention relates to the field of industrial equipment diagnostic technology and discloses a method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning networks. The method includes: acquiring a multi-dimensional heterogeneous signal sequence representing the operating state of the controlled object; mapping the signal sequence into a two-dimensional spatial image tensor using the Gram angle field algorithm; constructing a dual-stream feature extraction network containing graph convolution branches and cyclic unit branches to extract spatial topological features and temporal evolution features respectively; determining the logical offset matrix based on the spatial flow feature gradient divergence, and resampling the temporal flow features accordingly to achieve spatiotemporal causal phase alignment, and outputting diagnostic analysis results. This invention compensates for the physical system response hysteresis through asynchronous resampling operators in the computational domain, eliminates feature phase contamination caused by conventional rigid alignment, enhances the model's ability to capture weak abnormal signals, and achieves causal consistency in fault root cause localization.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment diagnostic technology, and in particular relates to a method for industrial equipment fault diagnosis and tracing based on multi-source heterogeneous parameter fusion and deep learning network. Background Technology

[0002] Currently, sensor networks are used to collect heterogeneous parameters such as temperature, pressure, vibration signals, heater power, rotation speed, and flow rate. Deep learning networks are used to extract multidimensional features, and tensor fusion is performed based on the absolute timestamp of the sampling time. The industry is trying to introduce algorithmic models to improve the logical reasoning ability at the diagnostic level. For example, Chinese invention patent application with publication number CN118313417A discloses a multi-expert LoRA fine-tuning method for industrial equipment fault diagnosis. By constructing a multi-expert dataset and fine-tuning the base model with instructions, modular diagnosis for different types of equipment can be achieved. This kind of fine-tuning scheme belongs to the application-level logical mapping optimization. The technical concept relies on textual fault logs or structured data and fails to penetrate to the underlying dynamic characteristics of physical signals.

[0003] Energy transfer and matter flow in physical systems inherently involve transmission delays, leading to cross-domain spatiotemporal lag effects in the occurrence mechanisms of state parameters across different dimensions. Conventional computational logic employs rigid alignment, forcibly splicing feature tensors at different causal evolution phases. This alignment introduces cross-phase redundant noise when processing heterogeneous parameters, masking the weak causal relationships of minor faults in their early stages. Although increasing the sampling frequency or expanding the network depth can alleviate these contradictions, such linear improvement paths cannot correct the phase distortion of tensors in the feature space and cause excessive consumption of computational resources. Existing technologies mainly suffer from the following shortcomings: 1. Rigid alignment mechanisms cause phase misalignment of heterogeneous tensors; 2. Cross-phase splicing introduces redundant noise and masks weak fault characteristics; 3. The computational model lacks the ability to dynamically compensate for physical lag laws.

[0004] Therefore, the technical problem to be solved by this invention is how to compensate for the state transmission lag of physical entities in situ in the tensor dimension and eliminate the feature distortion caused by forced cross-phase splicing without adding physical sensors or changing the mechanical structure, by optimizing the fusion mechanism of the computational model. Summary of the Invention

[0005] This invention provides a method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning networks, comprising the following steps: Step S1: Obtain a multidimensional heterogeneous signal sequence characterizing the operating state of the controlled object in the physical information system. The multidimensional heterogeneous signal sequence contains heterogeneous time-series variables characterizing the internal energy exchange state and mechanical conduction state of the controlled object. Step S2: The Gram angle field algorithm is used to perform nonlinear mapping and polar coordinate transformation on the multidimensional heterogeneous signal sequence to generate a two-dimensional spatial image tensor containing the implicit spatial coupling characteristics between physical quantities of different dimensions, so as to construct the global topological constraint relationship between different physical parameters inside the controlled object. Step S3: Construct a dual-stream feature extraction network that includes graph convolution branches and recurrent unit branches. Use the graph convolution branches to perform spatial domain feature aggregation on the two-dimensional spatial image tensor to generate spatial flow features. At the same time, use the recurrent unit branches to extract the temporal evolution features of multidimensional heterogeneous signal sequences to generate temporal flow features. Step S4: Calculate the feature gradient divergence generated during the transmission of spatial flow features in the multi-layer network, and use the feature gradient divergence to measure the causal evolution differences between different physical fields inside the controlled object, thereby generating a logical offset matrix that characterizes the response hysteresis inside the controlled object. Step S5: Using cross-domain asynchronous alignment gating, asynchronous resampling processing based on sampling index bias is performed on the temporal flow features according to the logical offset matrix to generate an alignment feature tensor. Then, the alignment feature tensor and spatial flow features are dynamically weighted and fused using an attention mechanism to output the diagnosis and tracing results of the controlled object.

[0006] Preferably, step S1 includes: acquiring a parameterized signal stream with a sampling frequency of not less than 0.5Hz through a signal acquisition node, representing the thermal field equilibrium, vacuum pressure, vibration mode, and excitation power state of the controlled object; filtering out high-frequency disturbance noise in the parameterized signal stream using a mean filter operator; and mapping the filtered signal to a preset [0,1] numerical range using a Z-score normalization algorithm to generate a standardized multidimensional heterogeneous signal sequence with dimension normalization properties.

[0007] Preferably, based on the two-dimensional spatial image tensor generated in step S2, the method further includes: constructing a generative adversarial model composed of a generative network and a discriminative network; using the generative network to learn the probability distribution of scarce abnormal categories in the sample space to generate virtual feature samples that conform to the statistical characteristics of real working conditions; and using an adaptive sliding window sampling algorithm to resample and mix the original sample set and the virtual feature samples to correct the imbalance of class weights in the training dataset and enhance the model's ability to identify small abnormal boundaries.

[0008] Preferably, the operation of mapping using the Gram angle field algorithm in step S2 includes: calculating the polar coordinate mapping components of the standardized signal sequence in each dimension within the unit circle; generating the Gram summation field matrix and the Gram difference field matrix using trigonometric function difference logic; and nonlinearly superimposing the Gram summation field matrix and the Gram difference field matrix using the channel dimension splicing operator to construct a two-dimensional spatial image tensor that represents the transient spatial coupling correlation between multidimensional physical parameters.

[0009] Preferably, the operation of extracting features using graph convolution branches in step S3 includes: establishing an initial adjacency matrix based on the prior physical causal strength between various physical quantities within the controlled object; using the graph feature aggregation operator in the graph convolution branch to dynamically update the edge weights of the initial adjacency matrix during network training, so as to extract the cross-parameter coupling fluctuation features in the two-dimensional spatial image tensor that represent the early stage of abnormal state evolution, thereby achieving accurate representation of local topological changes.

[0010] Preferably, the operation of generating the logical offset matrix in step S4 includes: calculating the tensor gradient operator of the spatial flow characteristics at time t. The dynamic offset is determined using a preset nonlinear mapping function. Its logical calculation rules are as follows: ,in, Let be the dynamic offset at time t. Let be the tensor gradient operator of the spatial flow characteristics at time t, α be the preset spatiotemporal correlation strength weight, and β be the inherent minimum response delay constant of the system. represents the L2 norm operation, and round represents the integer operation.

[0011] Preferably, the asynchronous resampling operation in step S5 includes: based on the dynamic offset Nonlinear reverse backtracking sampling is performed on the feature sequences in the cyclic unit branches; the time-domain features are rearranged using logical shift operators to compensate for the phase deviation of the observed signal caused by the physical time delay of internal thermal equilibrium or mechanical conduction of the controlled object, so that the feature representation in the computational domain returns to the physical causal evolution phase reference of the controlled object.

[0012] Preferably, the dynamic weighted fusion operation using the attention mechanism in step S5 includes: introducing a cross-domain alignment fusion operator to map the spatial flow features and the aligned feature tensors to the same latent feature space; using the Softmax function to calculate the mutual information weights between the spatial domain features and the aligned temporal domain features, and performing point-to-point nonlinear weighted summation on the heterogeneous features based on the weights to generate a fusion feature vector containing spatiotemporal causal consistency.

[0013] Preferably, the diagnostic and tracing results are obtained through the following tracing analysis operations: calculating the sensitivity Jacobian matrix of the global diagnostic classification loss function relative to the feature level of the input signal; identifying the key signal component dimensions in the sensitivity Jacobian matrix whose values ​​exceed the preset tracing criterion threshold through the global average pooling operator; determining the key signal component dimensions as the original physical causes that trigger the abnormal evolution of the controlled object's state, and outputting an analysis report containing fault root cause location information.

[0014] Preferably, the controlled object is a single-crystal silicon growth furnace system with nonlinear time-varying thermal field distribution characteristics. Its multi-source heterogeneous parameters include cooling water inlet and outlet temperatures, furnace vacuum pressure, heater power, and argon flow rate, and the corresponding signal sampling interval is set to 2s. The diagnostic results are used to determine the abnormal state distribution inside the single-crystal silicon growth furnace system and provide logical driving instructions for the closed-loop process adjustment of the controlled object.

[0015] Compared with existing technologies, the industrial equipment fault diagnosis and tracing method based on multi-source heterogeneous parameter fusion and deep learning networks of this invention has the following advantages: 1. In industrial equipment fault diagnosis, a fault sample expansion mechanism based on generative adversarial networks is constructed. By utilizing generative networks to learn the distribution patterns of scarce fault categories, virtual fault samples consistent with the distribution of real working conditions are generated. Combined with the adaptive sliding window method, the balance and completeness of the training dataset are improved without changing the sensor deployment. This ensures that the deep learning network can obtain sufficient feature boundary information when processing data streams with extremely scarce fault samples, such as those from industrial equipment. This avoids the risk of missed detections caused by class weight imbalance and improves the system's sensitivity to capturing minor abnormal states.

[0016] 2. A collaborative processing mechanism of Gram corner field and dual-stream network architecture is adopted to realize the transformation of multi-dimensional temporal parameters into two-dimensional spatial images. The graph convolution branch is used to extract local and global coupling features between pixels. This mapping method of converting one-dimensional temporal correlation into two-dimensional spatial topological relationship can uncover the nonlinear spatial coupling features hidden in the original temporal sequence. Combined with the temporal evolution law extracted by the cyclic unit branch, a complete representation of heterogeneous parameters is formed. Compared with single-dimensional feature modeling, this dual-stream parallel architecture, through deep aggregation of space and time, enables the weak cross-parameter fluctuations exhibited by minor faults in the early stage of evolution to be amplified in multiple dimensions in the feature space.

[0017] 3. Construct a temporal offset measurement mechanism based on spatial flow feature gradient divergence to achieve dynamic phase alignment of multi-source heterogeneous parameters in the computational domain. This mechanism uses the spatial state transition intensity generated by graph convolution path to determine the logical offset of heterogeneous temporal tensors, guiding the branches of the loop unit to complete nonlinear resampling, thereby compensating for the inherent energy transfer hysteresis of the physical entity system in situ in the computational model. This processing method eliminates cross-phase feature pollution caused by conventional rigid timestamp alignment, so that the weak causal correlation signal of minor faults in the bud stage can be completely preserved in the feature space. This mechanism reconstructs the physical causal chain of fault evolution at the computational logic level, enhancing the model's causal identification capability under strong noise conditions. Attached Figure Description

[0018] Figure 1This is the main flowchart of the fault diagnosis and tracing method under the multi-source heterogeneous parameter fusion of the present invention; Figure 2 This is a breakdown diagram of the multi-dimensional feature processing logic architecture of the fault diagnosis and tracing system of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0021] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0022] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0023] A method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning networks includes the following steps: Step S1: Obtain a multidimensional heterogeneous signal sequence characterizing the operating state of the controlled object in the physical information system. The multidimensional heterogeneous signal sequence contains heterogeneous time-series variables characterizing the internal energy exchange state and mechanical conduction state of the controlled object. Step S2: The Gram angle field algorithm is used to perform nonlinear mapping and polar coordinate transformation on the multidimensional heterogeneous signal sequence to generate a two-dimensional spatial image tensor containing the implicit spatial coupling characteristics between physical quantities of different dimensions, so as to construct the global topological constraint relationship between different physical parameters inside the controlled object. Step S3: Construct a dual-stream feature extraction network that includes graph convolution branches and recurrent unit branches. Use the graph convolution branches to perform spatial domain feature aggregation on the two-dimensional spatial image tensor to generate spatial flow features. At the same time, use the recurrent unit branches to extract the temporal evolution features of multidimensional heterogeneous signal sequences to generate temporal flow features. Step S4: Calculate the feature gradient divergence generated during the transmission of spatial flow features in the multi-layer network, and use the feature gradient divergence to measure the causal evolution differences between different physical fields inside the controlled object, thereby generating a logical offset matrix that characterizes the response hysteresis inside the controlled object. Step S5: Using cross-domain asynchronous alignment gating, asynchronous resampling processing based on sampling index bias is performed on the temporal flow features according to the logical offset matrix to generate an alignment feature tensor. Then, the alignment feature tensor and spatial flow features are dynamically weighted and fused using an attention mechanism to output the diagnosis and tracing results of the controlled object.

[0024] Preferably, step S1 includes: acquiring a parameterized signal stream with a sampling frequency of not less than 0.5Hz through a signal acquisition node, representing the thermal field equilibrium, vacuum pressure, vibration mode, and excitation power state of the controlled object; filtering out high-frequency disturbance noise in the parameterized signal stream using a mean filter operator; and mapping the filtered signal to a preset [0,1] numerical range using a Z-score normalization algorithm to generate a standardized multidimensional heterogeneous signal sequence with dimension normalization properties.

[0025] Preferably, based on the two-dimensional spatial image tensor generated in step S2, the method further includes: constructing a generative adversarial model composed of a generative network and a discriminative network; using the generative network to learn the probability distribution of scarce abnormal categories in the sample space to generate virtual feature samples that conform to the statistical characteristics of real working conditions; and using an adaptive sliding window sampling algorithm to resample and mix the original sample set and the virtual feature samples to correct the imbalance of class weights in the training dataset and enhance the model's ability to identify small abnormal boundaries.

[0026] Preferably, the operation of mapping using the Gram angle field algorithm in step S2 includes: calculating the polar coordinate mapping components of the standardized signal sequence in each dimension within the unit circle; generating the Gram summation field matrix and the Gram difference field matrix using trigonometric function difference logic; and nonlinearly superimposing the Gram summation field matrix and the Gram difference field matrix using the channel dimension splicing operator to construct a two-dimensional spatial image tensor that represents the transient spatial coupling correlation between multidimensional physical parameters.

[0027] Preferably, the operation of extracting features using graph convolution branches in step S3 includes: establishing an initial adjacency matrix based on the prior physical causal strength between various physical quantities within the controlled object; using the graph feature aggregation operator in the graph convolution branch to dynamically update the edge weights of the initial adjacency matrix during network training, so as to extract the cross-parameter coupling fluctuation features in the two-dimensional spatial image tensor that represent the early stage of abnormal state evolution, thereby achieving accurate representation of local topological changes.

[0028] Preferably, the operation of generating the logical offset matrix in step S4 includes: calculating the tensor gradient operator of the spatial flow characteristics at time t. The dynamic offset is determined using a preset nonlinear mapping function. Its logical calculation rules are as follows: ,in, Let be the dynamic offset at time t. Let be the tensor gradient operator of the spatial flow characteristics at time t, α be the preset spatiotemporal correlation strength weight, and β be the inherent minimum response delay constant of the system. represents the L2 norm operation, and round represents the integer operation.

[0029] Preferably, the asynchronous resampling operation in step S5 includes: based on the dynamic offset Nonlinear reverse backtracking sampling is performed on the feature sequences in the cyclic unit branches; the time-domain features are rearranged using logical shift operators to compensate for the phase deviation of the observed signal caused by the physical time delay of internal thermal equilibrium or mechanical conduction of the controlled object, so that the feature representation in the computational domain returns to the physical causal evolution phase reference of the controlled object.

[0030] Preferably, the dynamic weighted fusion operation using the attention mechanism in step S5 includes: introducing a cross-domain alignment fusion operator to map the spatial flow features and the aligned feature tensors to the same latent feature space; using the Softmax function to calculate the mutual information weights between the spatial domain features and the aligned temporal domain features, and performing point-to-point nonlinear weighted summation on the heterogeneous features based on the weights to generate a fusion feature vector containing spatiotemporal causal consistency.

[0031] Preferably, the diagnostic and tracing results are obtained through the following tracing analysis operations: calculating the sensitivity Jacobian matrix of the global diagnostic classification loss function relative to the feature level of the input signal; identifying the key signal component dimensions in the sensitivity Jacobian matrix whose values ​​exceed the preset tracing criterion threshold through the global average pooling operator; determining the key signal component dimensions as the original physical causes that trigger the abnormal evolution of the controlled object's state, and outputting an analysis report containing fault root cause location information.

[0032] Preferably, the controlled object is a single-crystal silicon growth furnace system with nonlinear time-varying thermal field distribution characteristics. Its multi-source heterogeneous parameters include cooling water inlet and outlet temperatures, furnace vacuum pressure, heater power, and argon flow rate, and the corresponding signal sampling interval is set to 2s. The diagnostic results are used to determine the abnormal state distribution inside the single-crystal silicon growth furnace system and provide logical driving instructions for the closed-loop process adjustment of the controlled object.

[0033] Example 1: In the operating conditions of a single-crystal silicon growth furnace system with nonlinear time-varying thermal field distribution characteristics, the multi-source heterogeneous parameter fusion and deep learning network fault diagnosis and tracing method of the present invention works according to the preset computing architecture. The single-crystal silicon growth furnace system has cross-domain spatiotemporal hysteresis physical characteristics of micro-fault evolution. Multi-dimensional heterogeneous signals, including cooling water inlet and outlet temperatures, furnace vacuum pressure, heater power, and argon flow rate, have energy transfer delay phenomena in the underlying fault occurrence mechanism. When conventional rigid time synchronization alignment computing logic processes the above-mentioned multi-source heterogeneous parameters, it forcibly splices physical state parameters that are in different causal evolution phases, causing the deep learning network to introduce cross-phase redundant tensor noise in the feature aggregation stage, which masks the weak causal correlation signals exhibited by the equipment micro-fault in the bud stage.

[0034] To address the contradiction in computational model reconstruction between multidimensional parametric spatial coupling and temporal evolution hysteresis, the signal acquisition node collects multidimensional heterogeneous signal sequences with a sampling frequency greater than or equal to 0.5Hz. The system synchronously acquires standard current signals of 4mA to 20mA output from cooling water inlet and outlet temperature transmitters through a 12-bit analog-to-digital converter chip, and uses a 5th-order Butterworth low-pass filter to filter out power frequency interference above 10Hz. The difference between the sampling mean of each sensor channel within 10s and the preset process reference value is defined as the thermal field equilibrium state quantity. This feature extraction method based on physical quantity deviation... This approach eliminates the interference of sensor zero-point drift on the deep learning network. The system uses the Gram angle field algorithm to process the multidimensional heterogeneous signal sequence, completing nonlinear mapping and polar coordinate transformation. The system maintains a sliding buffer with a length of 60 sampling points, storing historical running data from the past 120 seconds in real time. Whenever a signal acquisition node generates 5 new sampling points (i.e., every 10 seconds), a Gram angle field transformation operator is triggered to generate a two-dimensional spatial image tensor with a pixel dimension of 60 x 60. This high overlap sampling strategy ensures that adjacent feature images have 91.6% temporal overlap. Overlapping is used to ensure the model's capture of minute thermal field perturbations has temporal continuity. A two-dimensional spatial image tensor containing implicit spatial coupling features between physical quantities of different dimensions is generated. Global topological constraints between different physical parameters within the controlled object are established. Simultaneously, the system constructs a dual-stream feature extraction network including graph convolution branches and recurrent unit branches. The graph convolution branches aggregate the two-dimensional spatial image tensor to generate spatial flow features, and the recurrent unit branches extract the temporal evolution features of multidimensional heterogeneous signal sequences to generate temporal flow features. The spatial flow features extracted by the graph convolution branches provide local and global topological data benchmarks for quantifying the state transition of the physical field. This data benchmark serves as the phase correction input for the temporal evolution features of the recurrent unit branches. The system calculates the feature gradient divergence generated by the spatial flow features during the transmission of the multi-layer network. This feature gradient divergence is used to quantify the causal evolution differences between different physical fields within the controlled object, generating a logical offset matrix characterizing the system's response hysteresis. This logical offset matrix enables the temporal flow features to obtain physical phase correction weights based on the spatial state gradient, establishing a collaborative amplification computation link where the underlying parameters of the spatial flow features and the temporal flow features are interdependent.

[0035] The system transforms the state propagation hysteresis law of the physical world into an asynchronous resampling operator of feature tensors within the computational domain, redefines the fusion alignment boundary of heterogeneous parametric tensors, and calculates the tensor gradient operator of the spatial flow features at time t. The dynamic offset is determined based on the nonlinear mapping function. The logical calculation rule is as follows: multiply the calculated spatial flow feature gradient norm by a dimensionless spatiotemporal correlation strength weight of 1.5, add a minimum response delay offset constant of 1.0, and finally output the dimensionless sampling index offset by rounding to zero. Its physical mapping logic is that for every 0.1 unit change in the spatial flow feature gradient, the corresponding physical field hysteresis evolution step increases by 1 sampling period (2 seconds). This allows for index backtracking of the discrete step count, canceling out the time phase delay caused by heat conduction within the computational domain. The logical calculation rule is as follows: ,in Let be the dimensionless dynamic offset at time t. For spatial flow characteristics in the first The dimensionless tensor gradient operator at time t, where α is the weight of the spatiotemporal correlation strength constant, and β is the dimensionless bias constant characterizing the inherent minimum response delay of the system. For L2 norm operation, round is the rounding operation. The dimensionless bias constant β is determined by the sum of the physical transmission link delay from the signal acquisition node to the processing terminal and the inherent hardware response time of the sensor. It is calibrated in the range of 0.8 to 1.2. The transmission time of the background signal under constant temperature conditions is collected before the equipment enters the online diagnostic process. The spatiotemporal correlation strength weight is dynamically mapped based on the correlation coefficient between the heater excitation power and the vacuum pressure and temperature gradient in the furnace. The grid search program is used to traverse power disturbances of different intensities in the simulation environment.

[0036] In the experiment, the correlation coefficient corresponding to the minimum value of the classification loss function was extracted as a benchmark. The weight was locked in the range of 1.5 to 3.0 so that the logical offset matrix could represent the causal phase lag caused by thermal inertia. Using cross-domain asynchronous alignment gating, the system nonlinearly backtracks the feature sequence in the loop unit branch based on the dynamic offset in the logical offset matrix to generate an aligned feature tensor. When the cross-domain asynchronous alignment gating performs asynchronous resampling, a nonlinear backtracking algorithm based on linear interpolation is used to perform temporal rearrangement of the time stream features output by the loop unit branch. The processor calculates the dynamic offset based on time t. The system determines the corresponding historical feature index position. When the index position is not an integer, it reads the feature tensors of two adjacent sampling times and performs a distance weight linear compensation operation to generate the aligned feature tensor of the current time. The aligned feature tensor and the spatial flow feature are input into the same latent feature space for mutual information calculation. The attention weights are calculated using the Softmax function to perform point-to-point weighted summation on heterogeneous features, eliminating phase contamination of the feature tensor caused by energy transmission delay and achieving causal evolution phase alignment within the computational domain. The system uses an attention mechanism to dynamically weight and fuse the aligned feature tensor with the spatial flow feature. The mutual information weights of heterogeneous features are calculated using the Softmax function, and the features are nonlinearly weighted and summed to output a deep fused feature vector containing spatiotemporal causal consistency. The fused feature vector is then... After being input into the classification network, the system calculates the sensitivity Jacobian matrix of the global diagnostic classification loss function relative to the feature level of the input signal. By using the global average pooling operator, it identifies the key signal component dimension in the sensitivity Jacobian matrix whose element value is greater than the preset traceability criterion threshold. The system determines that the key signal component dimension is the original physical cause that triggers the abnormal evolution of the controlled object's state and outputs an analysis report containing specific abnormal parameters and fault root cause location information. The cross-domain asynchronous resampling fusion operation reconstructs the state transmission law of the industrial physical entity in the computational network model and eliminates cross-phase pseudo-correlation state noise, so that the output diagnostic traceability results fit the physical mechanism evolution chain of the single crystal silicon growth furnace system, and establishes the underlying diagnostic traceability mechanism of computational domain deterministic causal inference.

[0037] Example 2: This example utilizes a single-crystal silicon physical simulation platform to verify the fault diagnosis and tracing capabilities of multi-source heterogeneous parameter fusion and deep learning networks. The single-crystal silicon physical simulation platform is equipped with a temperature sensor array with a measurement resolution of 0.1℃ and a high-frequency vacuum gauge covering a measurement range of 0.1Pa to 1000Pa to generate physical evolution data sequences. To test the computational model's data parsing capability against interference, the test unit superimposes Gaussian white noise with a signal-to-noise ratio of 20.5dB and injects 50Hz power frequency harmonic disturbances into the raw temperature and pressure signal sequences collected by the sensors. The system is set to control the spatiotemporal correlation strength of the dynamic offset. The weighting of the system faces a technical trade-off between balancing the sensitivity of capturing transient anomalies and the computational load of tensor resampling. Its decision rule is determined based on the product of the variance of the argon flow fluctuation and the divergence of the temperature gradient within the controlled object. When the product of the variance of the flow fluctuation and the divergence of the temperature gradient is greater than the preset steady-state benchmark value, it indicates that the causal phase hysteresis caused by thermal convection in the furnace is aggravated. At this time, the weight of the system driving the spatiotemporal correlation strength constant tends to the upper limit of the set interval to increase the compensation depth of the characteristic phase. Based on this decision logic and under the current operating conditions, the weight of the spatiotemporal correlation strength constant α is selected as 2.5, and the bias constant β, which represents the inherent minimum response delay, is set to 1.

[0038] The test unit constructs a multi-dimensional comparison system to quantify the underlying computational mechanism of the heterogeneous parameter fusion model. Samples using the conventional rigid time-synchronous alignment fusion operator are designated as control group one; samples with a spatiotemporal correlation strength constant weight α of 0.5 are designated as control group two; samples with a spatiotemporal correlation strength constant weight α of 4.5 are designated as control group three; and samples using the asynchronous resampling alignment operator of this invention with α of 2.5 are designated as the experimental group. The system injects heater degradation faults with power deviations of 1.0%, 3.0%, and 5.0% into the single-crystal silicon physical simulation platform to construct the problem intensity gradient. For a power deviation of 1.0%, the signal-to-noise ratio of the tensor fusion feature output by control group one is 1.2 dB. The branching of the loop unit in the experimental group is based on the tensor gradient operator of the spatial flow feature at time t. Calculated dynamic offset Time-division reverse backtracking sampling of internal feature sequences generates aligned feature tensors with a signal-to-noise ratio of 5.8dB and filters out 50Hz power frequency harmonic disturbances. When dealing with a 3.0% power deviation condition, the deep fusion feature vector output by the experimental group drives the classification network to achieve a fault tracing accuracy of 97.8%. The control group, due to the lack of an asynchronous phase compensation operator based on spatial topology, causes tensor misalignment and collision between temporal and spatial features, resulting in a fault tracing accuracy of 68.5%. The numerical difference between the aforementioned intermediate feature signal-to-noise ratio and the final tracing accuracy confirms the synergistic effect of the graph convolution spatial features guiding the dynamic resampling of the cyclic unit temporal features. As the intensity of the degraded fault increases, the diagnostic performance data output by the system exhibits nonlinear inflection point characteristics constrained by the laws of physical evolution.

[0039] Under a 5.0% power deviation, the fault tracing accuracy of control group 2 dropped to 72.4% because the value of the spatiotemporal correlation strength constant weight α was too low, resulting in the phase compensation of the time flow feature being unable to cover the actual thermal inertia hysteresis difference of the physical thermal field. In control group 3, the value of the spatiotemporal correlation strength constant weight α exceeded the upper limit of physical constraints, causing overcompensation of dynamic offset, breaking the inherent causal evolution time window of physical parameters, and leading to over-correction of features and a surge in pseudo-correlation, resulting in a drop in fault tracing accuracy to 81.2%. The fault tracing accuracy of the experimental group remained at 97.5%. The inflection point phenomenon in the test data sequence confirmed that the value range of the spatiotemporal correlation strength constant weight α from 1.5 to 3.0 constitutes the optimal calculation window that conforms to the energy transfer law of the controlled physical system. The system calculates the sensitivity Jacobian matrix of the global diagnostic classification loss function relative to the feature level of the input signal based on the deep fusion feature vector within the optimal window, identifies the specific causes of thermal field boundary instability that cause abnormalities, and outputs the tracing results including the location number of the root heating area.

[0040] Example 3: Addressing the technical obstacles of opaque computational paths and ambiguous network parameter settings in single-crystal silicon growth furnace systems when processing cross-dimensional parameters, the multi-source heterogeneous parameter fusion and deep learning network fault diagnosis and tracing method of this invention operates according to a defined internal algorithm architecture. The system acquires a multi-dimensional heterogeneous signal sequence with a sampling frequency of 0.5Hz, selects the maximum and minimum values ​​of this multi-dimensional heterogeneous signal sequence within a preset time sliding window, calculates the difference between the maximum and minimum values ​​as the proportional denominator, and uses this proportional denominator to linearly map the physical feature values ​​in the multi-dimensional heterogeneous signal sequence to negative one and positive... Within a closed interval, a standardized sequence is generated. The inverse cosine of the standardized sequence values ​​is calculated to generate a polar angle tensor. Simultaneously, the discrete timestamps corresponding to each sampling point are directly set as polar radius tensors. The system calculates the cosine of the sum of elements of different polar angle tensors under the same polar radius tensor to form a Gram summation field matrix, and calculates the sine of the subtraction of elements of different polar angle tensors to form a Gram difference field matrix. The system concatenates the Gram summation field matrix and the Gram difference field matrix along the channel dimension and outputs a two-dimensional spatial image tensor. In the feature extraction stage, the system constructs a graph convolution branch, which internally contains three... The graph attention network (GAN) generates an initial adjacency matrix based on the reciprocal of the three-dimensional geometric distances between the physical sensor nodes within the monocrystalline silicon growth furnace system. The edge weights of the initial adjacency matrix are determined by the reciprocal of the three-dimensional geometric distances between the physical sensor nodes within the monocrystalline silicon growth furnace and the gain coefficient of the preset heat transfer model. The network reads the three-dimensional coordinate data of each sensor in the vacuum chamber, calculates the Euclidean distance between any two nodes, and sets the reciprocal of the distance as the initial physical connectivity strength. The GAN dynamically adjusts the weights based on the cross-correlation matrix of the input signals. When the time delay between the change in heater power and the change in graph crucible temperature exceeds a preset threshold, the connectivity weight of the corresponding path in the adjacency matrix is ​​reduced, so that the extracted spatial flow features reflect the nonlinear time-varying characteristics of the internal thermal field distribution. The GAN reads the two-dimensional spatial image tensor and the initial adjacency matrix, calculates the inner product of the node feature vectors during the forward propagation of the network, updates the connectivity weights of the initial adjacency matrix based on the inner product, and outputs the spatial flow features. The system simultaneously constructs a loop unit branch containing two layers of gated loop units to extract the temporal flow features of multidimensional heterogeneous signal sequences. The above surface analysis process demonstrates the specific hardware and software operation logic of feature mapping and network topology architecture.

[0041] To address the data distribution imbalance caused by the scarcity of fault categories in the training sample set, the system constructs a generative adversarial model (GAP) consisting of a generator network and a discriminator network to expand the fault feature space. The generator network receives a random Gaussian noise vector of dimension 100 as input. Through a nonlinear transformation chain consisting of three fully connected layers and exponential linear units, this random Gaussian noise vector is mapped into virtual feature samples consistent with the statistical distribution of the original multidimensional heterogeneous signal sequence. The discriminator network adopts a topology corresponding to the graph convolution branch to calculate the probability score of the input sample belonging to the real physical evolution data sequence. To establish the convergence criterion for the GAP model, the system defines a total loss function. Total loss function The calculation formula is as follows: ,in The total loss score for the generative adversarial model is given by E, where E represents the expected value, x represents the real signal sample from the sensor of the controlled object, and G(z) represents the virtual feature sample generated by the generative network based on the noise vector z. To determine the classification probability value output by the network, the system uses the backpropagation algorithm to alternately update the weight parameters of the generator network and the discriminator network until the total loss function is reached. Once the preset equilibrium steady-state threshold is reached, the virtual feature samples generated by the network can simulate the energy transfer law of physical parameters under real working conditions. The virtual feature samples are then injected into the original sample set to perform adaptive sliding window resampling and mixing, which corrects the perception weights of the deep learning network for small abnormal boundaries.

[0042] Tensor gradient operator for system computation of spatial flow characteristics Determining dynamic offset using a nonlinear mapping function The system uses this dynamic offset. The system performs reverse backtracking resampling on the feature sequences in the loop unit branches to generate aligned feature tensors. After fusing the aligned feature tensors with spatial flow features, the system inputs them into the classification network. Based on the classification network output, the system calculates the sensitivity Jacobian matrix of the global diagnostic classification loss function relative to the feature level of the input signal. To establish a statistical benchmark for feature determination thresholds, the system retrieves the historical sensitivity Jacobian matrix set generated by the single-crystal silicon growth furnace system under 100 hours of continuous, fault-free, and stable operation. The system calculates the mean and mean of elements corresponding to each physical parameter dimension in this historical sensitivity Jacobian matrix set. The standard deviation is defined as the sum of the element mean and three times the element standard deviation of the preset traceability criterion threshold. The system compares the element values ​​of each dimension of the currently calculated sensitivity Jacobian matrix with the preset traceability criterion threshold, extracts the signal dimensions with values ​​greater than the preset traceability criterion threshold as the physical causes of abnormal equipment status, and outputs a diagnostic report containing the name of specific abnormal parameters and the identification of the underlying components of the equipment. This mechanism, by introducing a dynamic mean square error statistical law to set the threshold, suppresses the index drift phenomenon caused by empirical parameter settings and establishes the data consistency of fault traceability judgment.

[0043] Example 4: When the system faces the initial connection of the physical equipment of the single crystal silicon growth furnace, this method performs on-site deployment pre-calibration and baseline data filling procedures before starting the dual-stream feature extraction network. The system control equipment enters the no-load constant temperature maintenance state and continuously collects the static background signal sequence of each sensor. The arithmetic mean of the static background signal sequence is calculated as the zero-point drift compensation amount of the corresponding physical channel. The zero-point drift compensation amount is subtracted from the subsequently collected multidimensional heterogeneous signals to generate a zero-alignment signal sequence. The equipment is driven to execute the standard production process and the zero-alignment signal sequence under continuous operation is recorded. The network forward propagation calculation is performed on the zero-alignment signal sequence to generate the corresponding historical sensitivity Jacobian matrix set. The system calculates the discrete coefficient of the feature dimension in the historical sensitivity Jacobian matrix set and removes the abnormal tensor components with excessive deviation. The filtered matrix set is written into the database unit to construct the initial knowledge base representing the current physical equipment health steady-state boundary.

[0044] Based on the constructed initial knowledge base, the system extracts the maximum time offset from the historical causal relationships of multidimensional physical parameters as the specific calibration value of the dimensionless offset constant. The system determines the initial dynamic offset based on this dimensionless offset constant and the calculation operator to initiate cross-domain asynchronous resampling gating operation on the input signal sequence. After the equipment enters a continuous production cycle, the system reads the latest generated sensitivity Jacobian matrix of the model with a sliding time window, calculates the real-time data distribution characteristics within the current window, and covers the oldest timestamp data in the initial knowledge base according to the first-in-first-out principle. The system periodically recalculates the mean and standard deviation based on the updated knowledge base and adjusts the preset source tracing criterion threshold used for fault classification. Under the condition of maintaining the unchanged network topology architecture, the underlying data monitoring system completes the quantitative convergence of the diagnostic judgment boundary according to the aging law of physical entities.

[0045] Example 5: Addressing the uncertainty in underlying topology connection parameters caused by batch deployment of a single-crystal silicon growth furnace system, the system executes an offline generation of the benchmark matrix and bias constant calibration procedure during the initial construction phase of the dual-stream feature extraction network. This acquires ten sets of historical steady-state multidimensional heterogeneous signal sequences containing specific heater substrate materials and specific vacuum pump mechanical wear states. The Gram angle field algorithm is used to map each set of multidimensional heterogeneous signal sequences to generate corresponding two-dimensional spatial image tensors. The root mean square error (RMSE) of each set of two-dimensional spatial image tensors across all channels is calculated. The arithmetic mean of the ten RMSEs is set as the benchmark attenuation constant for the connectivity strength of the initial adjacency matrix nodes. This is then used to analyze the data from each physical sensor node. The reciprocal of the three-dimensional spatial geometric distance is used to perform a weighted scaling operation on the benchmark attenuation constant to generate a modified adjacency matrix adapted to the current physical structure. For argon flow rate and heater power parameters that do not have geometric coordinate attributes, the initial physical connectivity between their nodes is equivalently replaced by the thermal response time constant. That is, during the equipment commissioning phase, a 10% excitation power step signal is injected into the heater end, and the time interval required for the argon flow rate to reach 90% steady-state output is measured. The reciprocal of this time interval is set as the initial weight and calibrated between 0.15 and 0.45 to replace the Euclidean distance in the edge weight calculation of the adjacency matrix, ensuring that the nonlinear thermal field coupling logic has physical support.

[0046] The system inputs the simulation data sequence with pre-injected power deviation variables into the graph convolution branch equipped with the corrected adjacency matrix, calculates and extracts the timestamp of the first abrupt change point corresponding to the divergence value jump in the feature gradient divergence evolution curve of the spatial flow features, calculates the absolute time difference between the timestamp of the abrupt change point and the physical timestamp of the actual injected power deviation variables in the simulation data sequence, sets the absolute time difference as the specific operating value of the dimensionless bias constant β for feature resampling in the loop unit branch, processes the test data sequence using the corrected adjacency matrix and the calibrated dimensionless bias constant β, and performs cross-domain asynchronous resampling operation to generate a test deep fusion feature vector, calculates the spatial distribution cluster boundary coordinates of the test deep fusion feature vector after dimensionality reduction by principal component analysis, and locks the corrected adjacency matrix and the dimensionless bias constant β as fixed pre-call parameters for the device to enter the online fault diagnosis process when the spatial distribution cluster boundary coordinates are confirmed to be within the preset physical system response hysteresis steady-state reference range.

[0047] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning networks, characterized in that, Includes the following steps: Step S1: Obtain a multidimensional heterogeneous signal sequence characterizing the operating state of the controlled object in the physical information system. The multidimensional heterogeneous signal sequence contains heterogeneous time-series variables characterizing the internal energy exchange state and mechanical conduction state of the controlled object. Step S2: The Gram angle field algorithm is used to perform nonlinear mapping and polar coordinate transformation on the multidimensional heterogeneous signal sequence to generate a two-dimensional spatial image tensor containing the implicit spatial coupling characteristics between physical quantities of different dimensions, so as to construct the global topological constraint relationship between different physical parameters inside the controlled object. Step S3: Construct a dual-stream feature extraction network that includes graph convolution branches and recurrent unit branches. Use the graph convolution branches to perform spatial domain feature aggregation on the two-dimensional spatial image tensor to generate spatial flow features. At the same time, use the recurrent unit branches to extract the temporal evolution features of multidimensional heterogeneous signal sequences to generate temporal flow features. Step S4: Calculate the feature gradient divergence generated during the transmission of spatial flow features in the multi-layer network, and use the feature gradient divergence to measure the causal evolution differences between different physical fields inside the controlled object, thereby generating a logical offset matrix that characterizes the response hysteresis inside the controlled object. Step S5: Using cross-domain asynchronous alignment gating, asynchronous resampling processing based on sampling index bias is performed on the temporal flow features according to the logical offset matrix to generate an alignment feature tensor. Then, the alignment feature tensor and spatial flow features are dynamically weighted and fused using an attention mechanism to output the diagnosis and tracing results of the controlled object.

2. The method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning network according to claim 1, characterized in that, Step S1 includes: acquiring a parameterized signal stream with a sampling frequency of not less than 0.5Hz through a signal acquisition node, representing the thermal field equilibrium, vacuum pressure, vibration mode, and excitation power state of the controlled object; filtering out high-frequency disturbance noise in the parameterized signal stream using a mean filter operator; and mapping the filtered signal to a preset [0,1] numerical range using a Z-score normalization algorithm to generate a standardized multidimensional heterogeneous signal sequence with dimension normalization properties.

3. The method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning network according to claim 1, characterized in that, Based on the two-dimensional spatial image tensor generated in step S2, the method further includes: constructing a generative adversarial model composed of a generative network and a discriminative network; using the generative network to learn the probability distribution of scarce abnormal categories in the sample space to generate virtual feature samples that conform to the statistical characteristics of real working conditions; and using an adaptive sliding window sampling algorithm to resample and mix the original sample set and the virtual feature samples to correct the class weight imbalance of the training dataset.

4. The method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning network according to claim 1, characterized in that, The operations in step S2 using the Gram angle field algorithm for mapping include: calculating the polar coordinate mapping components of the standardized signal sequence in each dimension within the unit circle; generating the Gram summation field matrix and the Gram difference field matrix using trigonometric function difference logic; and nonlinearly superimposing the Gram summation field matrix and the Gram difference field matrix using the channel dimension splicing operator to construct a two-dimensional spatial image tensor representing the transient spatial coupling correlation between multidimensional physical parameters.

5. The method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning network according to claim 1, characterized in that, The feature extraction operation in step S3 using graph convolution branches includes: establishing an initial adjacency matrix based on the prior physical causal strength between various physical quantities within the controlled object; and dynamically updating the edge weights of the initial adjacency matrix during network training using graph feature aggregation operators in the graph convolution branches, in order to extract cross-parameter coupling fluctuation features representing the early stage of abnormal state evolution in the two-dimensional spatial image tensor.

6. The method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning network according to claim 1, characterized in that, The operation of generating the logical offset matrix in step S4 includes: calculating the tensor gradient operator of the spatial flow characteristics at time t. The dynamic offset is determined using a preset nonlinear mapping function. Its logical calculation rules are as follows: ,in, Let be the dynamic offset at time t. Let be the tensor gradient operator of the spatial flow characteristics at time t, α be the preset spatiotemporal correlation strength weight, and β be the inherent minimum response delay constant of the system. represents the L2 norm operation, and round represents the integer operation.

7. The method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning network according to claim 6, characterized in that, The asynchronous resampling operation in step S5 includes: based on the dynamic offset Nonlinear reverse backtracking sampling is performed on the feature sequences in the cyclic unit branches; the time-domain features are rearranged using logical shift operators to compensate for the phase deviation of the observed signal caused by the physical time delay of internal thermal equilibrium or mechanical conduction of the controlled object, so that the feature representation in the computational domain returns to the physical causal evolution phase reference of the controlled object.

8. The method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning network according to claim 1, characterized in that, The dynamic weighted fusion operation using the attention mechanism in step S5 includes: introducing a cross-domain alignment fusion operator to map the spatial flow features and the aligned feature tensors to the same latent feature space; using the Softmax function to calculate the mutual information weights between the spatial domain features and the aligned temporal domain features, and performing point-to-point nonlinear weighted summation on the heterogeneous features based on the weights to generate a fusion feature vector containing spatiotemporal causal consistency.

9. The method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning network according to claim 1, characterized in that, The diagnostic and source tracing results are obtained through the following source tracing analysis operations: calculating the sensitivity Jacobian matrix of the global diagnostic classification loss function relative to the feature level of the input signal; identifying the key signal component dimensions in the sensitivity Jacobian matrix whose values ​​exceed the preset source tracing criterion threshold through the global average pooling operator; determining the key signal component dimensions as the original physical causes that trigger the abnormal evolution of the controlled object's state, and outputting an analysis report containing fault root cause location information.

10. The method for fault diagnosis and tracing of industrial equipment based on multi-source heterogeneous parameter fusion and deep learning network according to claim 1, characterized in that, The controlled object is a single-crystal silicon growth furnace system with nonlinear time-varying thermal field distribution characteristics. Its multi-source heterogeneous parameters include cooling water inlet and outlet temperatures, furnace vacuum pressure, heater power, and argon flow rate. The corresponding signal sampling interval is set to 2s. The diagnostic results are used to determine the abnormal state distribution inside the single-crystal silicon growth furnace system and provide logical driving instructions for the closed-loop process adjustment of the controlled object.

Images