A Fault Diagnosis Method and System for Complex Equipment Based on Dynamic Feature Modeling

Abstract

The present invention relates to the technical field of fault diagnosis, and in particular to a complex equipment fault diagnosis method and system based on dynamic feature modeling. The method includes obtaining multi-channel sensor data; performing data preprocessing on the obtained multi-channel sensor data; extracting initial channel features from the preprocessed data; performing convolution enhancement on the concatenated initial channel features based on a direction awareness mechanism to generate a globally context-aware structural feature; performing multi-temporal scale feature fusion on the globally context-aware feature based on a multi-scale modulation mechanism; performing health trend guidance on the fused feature based on a trend-guided supervision mechanism, significantly improving the diagnosis accuracy, trend awareness ability and practical response efficiency of the system, effectively adapting to the core application scenarios of multi-industry equipment such as predictive maintenance, online fault perception and remote diagnostic analysis, and having good promotion prospects and practical value.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a complex equipment fault diagnosis method and system based on dynamic feature modeling. Background Art

[0002] With the continuous improvement of the intelligent level of the energy industry and industrial systems, complex equipment is increasingly widely used in fields such as oil and gas extraction, power transmission and distribution, and industrial manufacturing. Represented by oil and gas equipment such as fracturing equipment, oil-water separation systems, and high-pressure injection pumps, and power equipment such as detection terminals, power quality monitoring equipment, and high-voltage switch cabinet inspection systems, these key systems generally have typical characteristics such as many structural levels, rich sensing channels, harsh operating environments, and frequent operating condition fluctuations. During long-term operation, these devices are often under complex operating conditions such as high pressure and high temperature, electromagnetic interference, strong corrosion, and high vibration, and are extremely vulnerable to factors such as component aging, fatigue wear, interface loosening, or insulation degradation, which can induce a decline in operating performance and the risk of sudden failures. In order to ensure the safe and stable operation of energy equipment, it is urgent to rely on multi-channel sensing signals with high-frequency sampling to achieve accurate perception of the equipment operating state and predictive diagnosis of potential faults.

[0003] Existing equipment condition monitoring methods mainly rely on fixed threshold alarms or single-channel rule judgments, lacking an understanding of the structural semantics of the equipment and unable to perceive the response paths and functional cascading relationships between channels; some methods introduce neural networks to model time-series data, but usually only focus on local signal features and lack the ability to model the dependency relationships between internal components of the equipment and the multi-channel cooperation mode; in addition, during the actual operation of the equipment, fault symptoms often exhibit multi-scale evolution characteristics. For example, the frequency spans of pressure fluctuations and vibration shocks in oil equipment are relatively large, and there are obvious time-scale differences between partial discharge pulses and temperature rise trends in power equipment. A single-scale modeling mechanism is difficult to simultaneously perceive sudden change signals and slow change trends. Most existing methods do not establish an auxiliary supervision mechanism for state evolution trends and also lack the regulation of output continuity and stability, resulting in problems such as large jumps in prediction results, unstable abnormal identification, and frequent false positives and false negatives during the deployment process, seriously restricting their application value in industrial fields.

[0004] Therefore, in view of the problems of complex structure, fuzzy state evolution, weak and variable abnormal signals commonly existing in the above complex equipment, it is urgent to propose a unified intelligent modeling method with structural perception ability, time-series evolution modeling ability, multi-channel collaborative fusion ability, and stable control of diagnostic output. This method should be able to adapt to heterogeneous inputs of multiple types of sensing channels, fuse high-frequency burst features and slow trend features, accurately identify the equipment state evolution path and early fault symptoms, and output continuous and interpretable diagnostic results, thereby providing key technical support for predictive maintenance in related fields. Summary of the Invention

[0005] To solve the problems of insufficient structural coupling modeling, weak multi-scale anomaly recognition ability, and unstable diagnostic output in the predictive fault diagnosis process of complex industrial equipment, the present invention provides a complex equipment fault diagnosis method and system based on dynamic feature modeling.

[0006] In a first aspect, a complex equipment fault diagnosis method based on dynamic feature modeling provided by the present invention adopts the following technical solutions:

[0007] A complex equipment fault diagnosis method based on dynamic feature modeling includes:

[0008] Obtain multi-channel sensor data;

[0009] Perform data preprocessing on the obtained multi-channel sensor data;

[0010] Extract initial channel features from the preprocessed data;

[0011] Based on the direction-aware mechanism, perform convolution enhancement on the concatenated initial channel features to generate a structure-aware global context feature;

[0012] Based on the multi-scale modulation mechanism, perform multi-temporal scale feature fusion on the global context feature;

[0013] Based on the trend-guided supervision mechanism, perform health trend guidance on the fused features;

[0014] Based on the state continuity modulation, perform quantization scoring and output diagnosis on the output result;

[0015] Based on the output diagnosis result, use the visualization heat map mapping to locate the source of the anomaly.

[0016] Further, the performing data preprocessing on the obtained multi-channel sensor data includes setting the original multi-channel input signal at the equipment operation time as:

[0017] ,

[0018] And perform standard normalization processing on each channel, expressed as:

[0019]

[0020] Wherein, is the mean value of the th channel over the entire time period; is the standard deviation of the th channel; is the original signal value of the th channel at time . represents the number of channels; represents the length of the time series, is the result after normalization, making the mean of each channel signal 0 and the variance 1.

[0021] Furthermore, extracting the initial channel features from the preprocessed data includes reassembling the signals after normalizing all channels into an input tensor , and using a one-dimensional convolutional embedding mapping function to extract the initial channel feature representation:

[0022]

[0023] where, represents a one-dimensional convolutional operation with a convolutional kernel width of 3 and an output channel of ; is a batch normalization operation; is the feature tensor after initial extraction.

[0024] Furthermore, the convolutional enhancement of the initial channel features after splicing based on the direction-aware mechanism includes introducing the direction-aware mechanism, extracting statistical expressions from the time direction, channel direction, and structural path direction to enhance the structural semantic modeling ability, and splicing the three types of features and then performing convolutional enhancement to generate the structure-aware global context features:

[0025]

[0026] where, is a high-dimensional feature that fuses the three-directional structural semantic information and is used to guide the anomaly perception process in the multi-scale enhancement stage.

[0027] Furthermore, the multi-temporal scale feature fusion of the global context features based on the multi-scale modulation mechanism includes dividing the features after direction enhancement into several groups according to the channel dimension, and each group of features performs different-scale temporal pooling and depth convolutional operations; upsampling to restore to the original time step and splicing and fusing into a multi-scale feature tensor. To enhance the activation intensity of the local anomaly response channels, a channel attention mechanism is introduced to generate a modulation map, and finally, the fused feature output is obtained, expressed as:

[0028]

[0029] where, is the final fused modulation feature, which has both multi-scale modeling ability and structural semantic expression ability.

[0030] Furthermore, the health trend guidance for the fused features by the trend-guided supervision mechanism includes using the time derivative of the state output as an auxiliary task, adding a trend-guided output branch to the main features to predict the time difference of the state value, where the true difference is calculated from the difference between consecutive state values, and using the difference result as a supervision signal to construct a trend loss function, expressed as:

[0031]

[0032] where, is the total time sequence length; is used to measure the mean square error between the predicted trend and the true trend of the model, guiding the model to learn the state evolution law in the time dimension.

[0033] Furthermore, the quantization scoring and output diagnosis of the output result by the state continuity modulation includes continuously diagnosing and quantizing the score of the device state at the output end of the main task, where global temporal pooling is performed on the output main features to obtain compressed features, and the classification probability is obtained through a two-layer fully connected network, expressed as:

[0034]

[0035]

[0036] where, are the parameters of the first-layer fully connected network; are the parameters of the second-layer fully connected network; is the probability of predicting to belong to the th class, where ; is the number of classification categories of faults.

[0037] Furthermore, the quantization scoring and output diagnosis of the output result by the state continuity modulation further includes performing a moving average process on the output state sequence based on the sliding window smoothing mechanism, introducing a continuity constraint as an additional loss, and constructing a total classification loss with the joint loss function, and finally obtaining the total joint loss function, expressed as:

[0038]

[0039] where, is a hyperparameter used to balance the weights of trend learning and output smoothing.

[0040] Furthermore, using the output diagnosis result to locate the abnormal source by visual heatmap mapping includes generating a device health score according to the classification probability, and introducing a gradient response map calculation mechanism to locate the moment that contributes the most to the fault prediction, expressed as:

[0041]

[0042] in, is the first in the normalized input tensor Channel No. The value of the moment; Give the model output a health score; The gradient of the influence of the position on the score; Used to generate a two-dimensional channel-time diagram to assist in locating the source of anomalies.

[0043] The second aspect is a complex equipment fault diagnosis system based on dynamic feature modeling, including:

[0044] The data acquisition module is configured to acquire multi-channel sensor data;

[0045] The preprocessing module is configured to perform data preprocessing on the acquired multi-channel sensor data;

[0046] The feature module is configured to extract initial channel features from the preprocessed data;

[0047] The perception module is configured to perform convolution enhancement on the initial channel features after concatenation based on the direction-aware mechanism to generate structure-aware global context features;

[0048] The fusion module is configured to perform multi-temporal scale feature fusion on the global context features based on a multi-scale modulation mechanism;

[0049] The guidance module is configured to guide the fused features to a healthy trend based on a trend guidance supervision mechanism;

[0050] A diagnosis module is configured to perform quantitative scoring and output diagnosis on the output result based on the state continuity modulation;

[0051] The positioning module is configured to locate the source of the anomaly using a visual heat map based on the output diagnosis results.

[0052] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for a complex equipment fault diagnosis method based on dynamic feature modeling.

[0053] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; and the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to perform the complex equipment fault diagnosis method based on dynamic feature modeling.

[0054] In summary, the present invention has the following beneficial technical effects:

[0055] 1. Compared with the deficiencies of the prior art in aspects such as independent channel processing, insufficient abnormal evolution trend modeling, poor robustness to complex working condition disturbances, and unexplainable output, the present invention eliminates the dimensional deviation of heterogeneous sensor signals through a multi-channel signal normalization module, constructs the response dependence between device structure paths through a direction perception modeling module, enhances the perception ability for both sudden and trend-like anomalies simultaneously through a multi-scale modulation mechanism, improves the stability of state output under dynamic working conditions through a diagnostic output continuity modulation module, and introduces a trend-guided supervision and abnormal heat map localization mechanism to strengthen the diagnostic credibility of the system from two dimensions of temporal consistency and interpretability.

[0056] 2. In actual deployments such as typical petroleum equipment and power equipment, the method of the present invention reduces the missed detection rate of early weak faults from 22% to 7% and the false alarm rate from 17% to 6%; the robustness of feature extraction under multi-channel dynamic interference conditions is improved by 23%; while ensuring the improvement of accuracy, the single diagnosis inference time is shortened from 1.1 seconds to 0.48 seconds, significantly improving the diagnostic accuracy, trend perception ability, and practical response efficiency of the system, effectively adapting to the core application scenarios of multi-industry equipment such as predictive maintenance, online fault perception, and remote diagnostic analysis, and having good promotion prospects and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a schematic diagram of a complex equipment fault diagnosis method based on dynamic feature modeling in Embodiment 1 of the present invention.

[0058] Figure 2 is a schematic diagram of the comparison of the accuracy and F1 value of each model in Embodiment 1 of the present invention.

[0059] Figure 3 is a schematic diagram of the comparison of the inference time of each model in Embodiment 1 of the present invention.

[0060] Figure 4 is a schematic diagram of the comparison of the robustness improvement of each model in Embodiment 1 of the present invention.

[0061] Figure 5 is a schematic diagram of the radar chart of the comprehensive performance of the model in Embodiment 1 of the present invention.

[0062] Figure 6 is a schematic diagram of the prediction effect in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0063] The present invention will be further described in detail below with reference to the accompanying drawings.

[0064] Embodiment 1

[0065] Refer toFigure 1 , a complex equipment fault diagnosis method based on dynamic feature modeling in this embodiment, includes:

[0066] Obtain multi-channel sensor data;

[0067] Perform data preprocessing on the obtained multi-channel sensor data;

[0068] Extract initial channel features from the preprocessed data;

[0069] Based on the direction awareness mechanism, perform convolution enhancement on the concatenated initial channel features to generate a globally context-aware feature with structure awareness;

[0070] Based on the multi-scale modulation mechanism, perform multi-temporal scale feature fusion on the globally context-aware feature;

[0071] Based on the trend-guided supervision mechanism, perform health trend guidance on the fused features;

[0072] Based on the state continuity modulation, perform quantization scoring and output diagnosis on the output result;

[0073] Based on the output diagnosis result, use the visualization heat map mapping to locate the source of the anomaly.

[0074] Specifically:

[0075] This embodiment consists of six key modules, namely: multi-channel signal normalization and initial embedding module, direction awareness structure modeling module, multi-scale feature modulation module, state evolution trend modeling module, diagnostic output continuity modulation module, and interpretability analysis module.

[0076] (1) Multi-channel signal normalization and initial embedding module,

[0077] Complex industrial equipment (such as cable partial discharge equipment, fracturing equipment, etc.) is usually equipped with multiple different types of sensors (such as voltage, current, temperature, pressure, vibration, etc.) during actual operation. There are obvious differences in the distribution scale, physical dimension, and dynamic range of the collected signals. If directly fed into the model, it is easy to cause problems such as certain channels dominating the feature expression and the information of other channels being submerged. Therefore, this module introduces a standard normalization and unified initial embedding mechanism at the front end of the model to achieve scale unification, feature initialization, and channel expression enhancement of time-series multi-channel data, providing a clean, equivalent, and high-dimensional input tensor for subsequent modeling. To further improve the multi-scale characteristics of the signal and the robustness of feature embedding, this module introduces a preprocessing strategy based on multi-scale signal decomposition. The system first decomposes the original signal of each channel into 4 scale layers (from low-frequency trend to high-frequency detail) through discrete wavelet transform (DWT), corresponding to time windows 2 1 、2 2 、23 , 2 4 The characteristic expression of 2 seconds adopts the Daubechies-4 (db4) wavelet basis. After decomposition, the system performs energy normalization on each scale layer, normalizes the energy of the low-frequency components through the L2 norm, and suppresses the high-frequency component noise through an adaptive threshold (1.5 times the signal variance) to ensure the consistency of the amplitude distribution of different scale features. In addition, the system records the energy distribution characteristics of each scale layer (calculated by the fast Fourier transform), generates a multi-scale feature description vector, and uses it as an auxiliary input for subsequent convolutional embedding. To improve the diversity of the collected data, the system introduces a multi-modal signal synchronous acquisition mechanism, and ensures the time consistency of the signals in each channel through timestamp alignment (error control within 1 millisecond), providing a reliable data basis for subsequent cross-channel feature interaction and multi-scale feature fusion.

[0078] 1) Set the original multi-channel input signal at the equipment operation time as:

[0079]

[0080] Among them, is the original signal value of the th channel at time ; represents the number of channels; represents the length of the time series. To improve the diversity of the signal, the system performs preliminary feature enhancement on the original signal, compresses the channel dimension through 1×1 convolution (output channel is C / 2), and then captures the local correlation features through 3×3 convolution (output channel is C) to generate an enhanced signal tensor, providing a richer input representation for subsequent normalization.

[0081] 2) To eliminate the mean and variance differences between different channels and unify the data scale, the system performs standard normalization on the enhanced signal:

[0082]

[0083] Among them, is the mean value of the th channel over the entire time period; is the standard deviation of the th channel; The result after normalization is such that the mean of each channel signal is 0 and the variance is 1. To improve the refinement degree of normalization, the system introduces a cross-channel feature fusion strategy. By calculating the correlation matrix between channels (based on the Pearson correlation coefficient), joint normalization (weighted average of mean and variance) is performed on channel pairs with high correlation (correlation coefficient > 0.7). Subsequently, the system preliminarily compresses the normalized signal through a 1×1 convolution (output channels are C / 3) to generate a more compact feature representation, providing high-quality input for subsequent embedding.

[0084] 3) Re - splice the signals after normalizing all channels into an input tensor and use a one - dimensional convolution embedding mapping function to extract the initial channel feature representation:

[0085]

[0086] where represents a one - dimensional convolution operation with a convolution kernel width of 3 and output channels of ; is a batch normalization operation; is the feature tensor after initial extraction. To improve the expressive ability of feature embedding, the system designs a combination of convolution kernels of multiple sizes (1×1, 3×3, 5×5) to capture short - term fluctuations, medium - term trends, and long - term dependence features respectively. The normalized input tensor first compresses the dimension through a 1×1 convolution (output channels are / 3), and then extracts features of different time spans through 3×3 and 5×5 convolution kernels (output channels are both / 3) respectively, generating multi - size feature tensors where represents the convolution operation of the corresponding size. Finally, the features are fused into the initial feature tensor X0 through a splicing operation to provide a high - dimensional representation for subsequent direction - aware modeling. In addition, the system introduces a feature consistency constraint. By calculating the KL divergence between the feature tensors before and after embedding (the threshold is set to 0.05), it ensures that no significant distribution deviation is introduced during the embedding process.

[0087] (2) Direction - aware feature modeling module,

[0088] Due to the fact that the operating states of components in complex equipment structures have direction - related characteristics (for example, the fracturing equipment contains a structural link of "high - pressure pump → throttle valve → manifold → surface pipeline"), including time - trend direction (operation evolution), channel - space direction (component association), and functional - path direction (structural topology). Traditional methods mostly take each channel as an independent input, ignoring the spatial coupling and information propagation paths among modules during the equipment operation process, and it is difficult to model these directional and path - related characteristics. This module introduces a direction - awareness mechanism to extract statistical expressions from the time direction (trend), channel direction (response intensity), and structural - path direction (component topology), which is used to enhance the structural semantic modeling ability, improve the model's ability to model complex directional characteristics, and provide richer semantic representations for subsequent feature modulation.

[0089] 1) Time - direction awareness (joint modeling of trend and perturbation): The operating states of equipment components have obvious time - evolution characteristics, and the signal values on their sensing channels often slowly rise, fall, or have periodic perturbations with the operating cycle. To capture this trend - change feature, this module designs a time - direction awareness mechanism to extract cross - time evolution information from within a single channel. The system first performs convolution enhancement on the feature X0 of each channel, using 1×1 and 3×3 convolution kernels (both with the number of output channels being / 2) to capture short - term fluctuations and medium - term trend features respectively, and then generates an enhanced feature representation through feature fusion. Then, calculate the average response intensity of each channel over the entire diagnostic cycle:

[0090]

[0091] where, is the input signal of the c - th channel at the t - th time step; T′ is the total number of time steps in the diagnostic time window; μ c represents the average response intensity of channel c over the entire time window, which is used to represent its long - term stable level.

[0092] This operation can extract the average response trend of the channel over the entire diagnostic cycle, representing its long - term active level. To further improve the sensitivity to local trend changes, the system introduces a sliding difference operator to calculate the instantaneous difference (perturbation term). Specifically, for each channel, calculate the signal difference between adjacent time steps, and compress the difference feature through 1×1 convolution (the number of output channels is / 4):

[0093]

[0094] where, represents the instantaneous change amplitude of channel c at the t - th step;

[0095] Subsequently, calculate the average change rate within the entire time window, and capture the local patterns of the difference features through 3×3 convolution (output channels are

[0096] / 4):

[0097]

[0098] Among them, represents the average change rate of the channel within the entire time window, reflecting its short-term fluctuation intensity.

[0099] Finally, the system concatenates the long-term trend and short-term fluctuation features into the time-direction perception feature vector of channel c, and performs feature compression through 1×1 convolution (output channels are / 3):

[0100]

[0101] represents the response features of this channel at the two levels of long-term trend and local fluctuation. To enhance the expression ability of the time-direction features, a feature weighting mechanism is introduced. By calculating the global average activation value of the features (based on global average pooling), the weights are adaptively adjusted (the weights of the features with higher activation values are increased to 1.2 times), thereby enhancing the model's perception ability of key trends.

[0102] 2) Channel-direction perception (sensor response intensity): At a certain moment during the system operation, there are significant differences in the signal intensities and response features of different channels. For example, during partial discharge anomalies in the power system, voltage mutation signals will concentrate in certain channels, while other channels may still be in a stable state. Therefore, in order to highlight the signal channel with the greatest impact in the current state, this module introduces a channel-direction perception mechanism to calculate the average response level of all channels at each time step t. The system performs convolution enhancement on the feature X0, and uses 1×1 and 5×5 convolution kernels (output channels are both / 2) to capture short-term and long-term dependence features respectively, and then generates enhanced feature representations through feature fusion. Then, calculate the average response level at each time step:

[0103]

[0104] Among them, is the number of initial convolutional output channels. This mechanism realizes the automatic perception of the dominant channels of burst - type anomalies (such as partial discharge pulses, shock vibrations), enabling the model to focus more on key channels during feature extraction and effectively suppressing interference from non - critical dimensions. This mechanism realizes the automatic perception of the dominant channels of burst - type anomalies (such as partial discharge pulses, shock vibrations), enabling the model to focus more on key channels during feature extraction and effectively suppressing interference from non - critical dimensions. To improve the refinement of channel - direction perception, the system performs secondary extraction on the average response features through a 3×3 convolution (output channels are / 3) to capture the local interaction patterns between channels. In addition, the system designs a dynamic channel - response weighting mechanism. By calculating the variance of the response features (the weights of channels with higher variances are increased to 1.3 times), it adaptively adjusts the contribution ratio of each channel, enhancing the model's focusing ability on key channels.

[0105] 3) Functional - path - direction perception (structural - topology mapping): In complex equipment, each sensor channel often corresponds to the physical structure of the device, and there are topological paths or functional links between their signals. For example, in a fracturing system, "pump → valve → manifold", and in a power system, "transformer → control unit → relay protection". There are usually specific response time sequences and influence - transfer paths between these channels. To simulate this structural - topology information, this module pre - defines or learns a set of structural paths {P k}, and each path P k contains a set of logically related channel - index sets {c1, c2, …, c n}. The system performs convolution enhancement on the channel features X0 in the path P k using 1×1 and 5×5 convolution kernels (output channels are both / 4) to capture short - term and long - term dependence features respectively, and then extracts local patterns within the path through a 3×3 convolution (output channels are / 4). Then, calculate the overall response level of each path:

[0106]

[0107] where, is the set of sensor channels on the th functional path in the structure. The path - aggregated feature reflects the overall response level on a specific structural path, which can model the flow and transfer behavior of signals at the device - structure level and capture collaborative anomalies caused by changes in the linkage structure. To improve the expression ability of path features, the system introduces an inter - path interaction module through a 1×1 convolution (output channels are / 2) Cross-path fusion is performed on the path features to generate enhanced path feature representations. In addition, the system introduces a path response weighting mechanism that adaptively adjusts the contribution ratio of path features by calculating the response variance of each path (the weight of the path with a higher variance is increased to 1.2 times), enhancing the model's perception ability of key structural paths.

[0108] 4) The above three types of directional features respectively perceive key semantic information during device operation from the dimensions of time, channel, and structural path, and are complementary and independent. To construct a unified structural semantic feature representation, the system concatenates the three types of directional perception vectors into a directional fusion feature:

[0109]

[0110] This concatenation operation maintains the semantic independence of each dimension, enabling the model to learn the joint feature relationship across directions in subsequent stages. Considering that direct concatenation may lead to mismatched feature dimensions and uneven information distribution, this module introduces a two-layer one-dimensional convolutional network (Conv + BN + ReLU) for fusion enhancement:

[0111]

[0112] Among them, The high-dimensional feature that fuses the three-direction structural semantic information is used to guide the anomaly perception process in the multi-scale enhancement stage. To improve the robustness of directional feature fusion, the system adds different combinations of convolutional kernels (1×1, 3×3, 5×5) to capture short-term fluctuations, medium-term trends, and long-term dependence features. The concatenated directional feature Z first passes through a 1×1 convolution (output channels are / 3) to compress the dimensions, and then passes through 3×3 and 5×5 convolutional kernels (output channels are both / 3) to extract features of different time spans, generating different-scale feature tensors Z k , and then through the cross-direction feature interaction enhancement module, and a 3×3 convolution (output channels are ) is used to perform secondary processing on the fused features to further enhance the semantic association between features.

[0113] (3) Multi-scale modulation modeling module,

[0114] During the operation of complex equipment, the abnormal features in signals often present a multi-scale superposition pattern, including high-frequency burst fluctuations and low-frequency slow-varying trends, etc. For example, in an oil-water separation system, there may be high-frequency instantaneous fluctuations caused by valve switching and slow-varying trends caused by fluid proportion imbalance at the same time; another example is that during the process of cable insulation failure, high-frequency spikes represented by short-time partial discharge and low-frequency trends of slow temperature rise often appear. Traditional single-scale modeling methods are difficult to capture the signal features at these different time resolutions simultaneously, resulting in insufficient robustness of feature extraction and limited perception coverage. This module combines a multi-scale convolutional branch structure with a channel attention mechanism, strengthens the model's perception ability of multi-class scale abnormal signals, and introduces a depth cross-channel convolution and dynamic feature fusion mechanism. By enhancing the interaction between channels and the adaptive modulation of features, the model's perception ability of complex abnormal patterns is improved. At the same time, the feature expression is optimized through multi-size convolutional kernels.

[0115] 1) To reduce convolutional interference and enhance scale feature expression, this module divides the enhanced feature tensor in the upstream direction into groups along the channel dimension:

[0116]

[0117] This division enables the model to perform scale modeling in different sub-feature subspaces and avoid interference between features. To improve the rationality of grouping, the system introduces a channel correlation grouping strategy. By calculating the Pearson correlation coefficient matrix between channels, channels with higher correlation (correlation coefficient > 0.6) are assigned to the same group, generating G sub-feature tensors Z i . Subsequently, the system performs preliminary compression on each sub-feature tensor through 1×1 convolution (output channels are / 2) to reduce the computational overhead.

[0118] 2) To simulate the signal patterns at different time scales, each sub-channel feature Z i performs temporal pooling and depthwise convolution (DW-CONV) operations at different scales respectively to extract its internal context information:

[0119]

[0120] Among them, is the depthwise separable convolution operation, and the convolution kernel size is , which is used to extract the context within the scale; is average pooling; is the pooling scale, which determines the receptive field of each channel branch, .

[0121] This operation ensures that each group of features has a different temporal receptive field, thus capturing multi-scale behaviors such as rapid perturbations and slow trends separately. The system introduces a deep cross-channel convolution enhancement mechanism to improve the expressive ability of feature extraction. After the depth convolution operation at each scale s, a cross-channel interaction module is added: for each pair of channel combinations, shared 1×1 and 3×3 convolutional kernels (both with 16 output channels) are applied to capture short-term fluctuations and medium-term trend features respectively, generating cross-channel feature maps. In addition, the system performs secondary extraction on the cross-channel feature maps through a 5×5 convolution (with 16 output channels) to capture long-term dependence features and further enhance the semantic richness of the features.

[0122] 3) Since the feature time steps generated by different pooling scales are inconsistent, for the convenience of unified modeling, it is necessary to align the features of each branch to the original time step length through interpolation upsampling , and concatenate and fuse them along the channel dimension:

[0123]

[0124] Among them, is the bilinear interpolation function, which unifies and aligns the features of each scale to ; is the feature tensor after fusion of all scales. A dynamic feature fusion mechanism is introduced to adaptively adjust the contribution ratio of features at each scale: the system calculates the global average activation value of the features at each scale (obtained through global average pooling), and maps it to the weight coefficient through the sigmoid function, and the fused feature . In addition, the system compresses the fused features through a 1×1 convolution (with the output channel being / 3), and then captures local patterns through a 3×3 convolution (with the output channel being / 3) to further enhance the feature expression.

[0125] 4) Not all channels in the multi-channel signal contribute equally to anomaly recognition. Therefore, an attention mechanism is introduced to automatically calculate the modulation coefficient map:

[0126]

[0127] Among them, is used to modulate the weight tensor to enhance or suppress the response intensity of different channels / times; is the activation function, which generates the attention weight A through the mapping of the function, and smooths the weight through exponential moving average (EMA, with a smoothing coefficient of 0.9) to reduce the weight jitter during the training process.

[0128] 5) Finally, apply the channel attention map A to feature modulation to enhance the response of key channels and suppress redundant interference:

[0129]

[0130] Among them, is the final fused modulation feature, which has both multi-scale modeling ability and structural semantic expression ability. To optimize the effect of feature modulation, the system performs convolution enhancement on Y, uses 1×1 convolution (output channels are / 3) to compress the dimension, and then captures local patterns through 3×3 convolution (output channels are / 3) to generate the final high-quality feature representation, providing support for subsequent state trend modeling.

[0131] (4) State trend-guided modeling module,

[0132] During the actual operation of the equipment, the health state of the equipment not only includes the static classification features at the current moment, but also has obvious time evolution. For example, in the aging process of cables in the power system, the insulation parameters continuously decrease; in oil and gas fracturing, the pump pressure state gradually evolves into abnormal fluctuations or deviates from the normal range. If the model cannot capture this trend change, there may be diagnostic jumps, response fluctuations, or even missed detections. Therefore, this module introduces a "trend supervision branch" to enhance the model's understanding ability of trend changes by guiding the model to explicitly learn the time derivative features (approximate derivative) of state changes, and improve the continuity and forward-looking of state output.

[0133] 1) On the output feature sequence of the main task, design a linear trend prediction branch to predict the time difference of the continuous state score, which is used to describe the change direction and amplitude of the state score in the continuous time series:

[0134]

[0135] Among them, is the feature vector of the th time step output by the main model; is the weight of the trend supervision linear transformation; is the bias; is the predicted health trend change amount of the model (i.e., the approximate value of the first derivative of the score). Provide higher-quality input for the trend prediction branch. In addition, the system adaptively adjusts the weight of the linear transformation by calculating the global average activation value of the features (the weight of the features with higher activation values is increased to 1.2 times) to improve the accuracy of trend prediction.

[0136] 2) Perform adjacent differences on the continuous score sequence output by the main prediction branch to obtain the actual trend change value, which is used to describe the actual evolution speed and direction of the equipment state score:

[0137]

[0138] Among them, the original diagnostic score output by the main prediction branch (such as the device risk index at the second); is the change value of the actual status score. The dimension is compressed through 1×1 convolution (output channel is 1), and then the local pattern of the score sequence is captured through 3×3 convolution (output channel is 1) to generate enhanced differential features. In addition, the system adaptively adjusts the weight of the difference value by calculating the variance of the differential features (the differential weight with a larger variance is increased to 1.1 times) to improve the expression ability of the actual trend change value.

[0139] 3) Use the result as a supervision signal, and use the mean square error loss (MSE) to measure the difference between the predicted trend and the true trend, guiding the model to accurately model the evolution trend of the score, which helps to suppress problems such as score jumps and false alarm mutations:

[0140]

[0141] Among them, is the total time series length; is used to measure the mean square error (MSE) between the model predicted trend and the true trend, guiding the model to better learn the state evolution law in the time dimension. The dimension is compressed through 1×1 convolution (output channel is 1), and then the long-term dependence features are captured through 5×5 convolution (output channel is 1) to generate enhanced trend features. In addition, the system smooths the loss value through exponential moving average (EMA, smoothing coefficient is 0.9) to reduce the loss jitter during training and improve the training stability of the model.

[0142] (5) State continuity modulation module,

[0143] In actual deployment, the output of the model often jumps due to minor perturbations, affecting the judgment of maintenance personnel. For example: when the device operation switches the pump frequency, it causes instantaneous current fluctuations. If it is directly output as a "fault", it will generate false alarms. Therefore, when the system is deployed on the continuous time axis, it must have the ability to output a stable state to avoid misjudgment fluctuations caused by local perturbations or transient noises. For this reason, this module continuously diagnoses and quantifies the score of the device state at the output end of the main task, and introduces a sliding time window fusion mechanism and an adjacent state suppression regularization term to improve the time stability of the diagnostic output from the result level.

[0144] 1) Classification output generation and health score calculation,

[0145] For the time series feature sequence output by the main model Perform global temporal pooling to compress it into the overall device state representation:

[0146]

[0147] Then output the classification result through a two-layer fully connected network:

[0148]

[0149]

[0150] Among them, is the parameter of the first-layer fully connected layer; is the parameter of the second-layer fully connected layer; is the probability of predicting to belong to the category, where ; is the number of classification categories of faults. The enhanced features generate z through global average pooling, providing higher-quality input for the fully connected network. In addition, the system adaptively adjusts the weights of the fully connected layer by calculating the global average activation value of the features (the weights of the features with higher activation values are increased to 1.2 times), improving the accuracy of the classification result.

[0151] 2) Sliding window smoothing mechanism,

[0152] To alleviate the jump effect of single-frame output, introduce moving average to perform moving average processing on the output state sequence of consecutive diagnoses to filter out the prediction mutations caused by short-period perturbations and improve the temporal stability of the state output:

[0153]

[0154] Among them, is the original diagnosis score of the model in the previous frames; is the length of the sliding window, used to control the smoothing degree; is the smoothed score value, used as the final diagnosis output. To improve the smoothing effect, the system performs convolution enhancement on the original diagnosis score , compresses the dimension through 1×1 convolution (output channel is 1), and then captures the local patterns of the scoring sequence through 3×3 convolution (output channel is 1) to generate an enhanced scoring sequence. In addition, the system adaptively adjusts the sliding window length by calculating the variance of the scoring sequence (the sequence with larger variance is increased to 10), improving the flexibility of the smoothing mechanism.

[0155] 3) Adjacent difference regularization term,

[0156] To further suppress the sharp fluctuations in the prediction sequence, the model introduces a state smoothing regularization term during the training phase to limit the change amplitude between consecutive scores, strengthen the continuity constraint in the time dimension, and avoid state output jitter:

[0157]

[0158] To optimize the expression of the regularization term, the system performs convolution enhancement on the consecutive scores S(t) and S(t - 1), compresses the dimension through 1×1 convolution (with 1 output channel), and then captures the long-term dependence features through 5×5 convolution (with 1 output channel) to generate enhanced differential features. In addition, the system adaptively adjusts the weight of the regularization term by calculating the global average activation value of the differential features (the differential weight with a higher activation value is increased to 1.1 times), improving the ability of the regularization term to suppress sharp fluctuations.

[0159] 4) Joint loss function,

[0160] Construct the classification main loss (cross-entropy):

[0161]

[0162] where, is the true classification label; is expressed as 1 when and 0 otherwise.

[0163] The final training objective comprehensively considers three indicators: classification accuracy (cross-entropy), trend supervision, and output smoothing:

[0164]

[0165] where, is a hyperparameter used to balance the weights of trend learning and output smoothing; the system smooths the loss value through exponential moving average (EMA, with a smoothing coefficient of 0.9) to reduce the loss jitter during the training process and improve the training stability of the model.

[0166] (6) Diagnostic output and interpretability mapping module,

[0167] In the equipment predictive maintenance task, not only accurate category results need to be output, but also good interpretability and quantifiable output capabilities should be possessed to support engineers in quickly judging the current state, evolution trend, and potential risk sources of the equipment. Especially when facing equipment systems with complex structures and numerous channels, it is difficult to guide subsequent maintenance or strategy adjustment only relying on classification labels. Therefore, the system needs to construct an interpretable diagnostic output framework integrating classification probability, health score, and key response source location mechanism.

[0168] 1) Health score regression output,

[0169] Considering that the fault development process has continuity and degree, discrete classification alone cannot express the diagnostic confidence and evolution trend. Therefore, the system generates continuous health score values based on the classification probability vector :

[0170]

[0171] where is the probability of predicting to belong to the th class; is the health score corresponding to each state. This mechanism not only enhances the model's ability to express boundary states but also provides quantitative support for trend monitoring and remote intelligent early warning. To improve the expressive ability of the health score, the system performs convolution enhancement on the classification probability . It compresses the dimension through 1×1 convolution (output channels are / 2), and then captures the local patterns of the probability distribution through 3×3 convolution (output channels are / 2) to generate an enhanced probability vector. In addition, the system adaptively adjusts the weights of each probability by calculating the global average activation value of the probability vector (the probability weights with higher activation values are increased to 1.2 times) to improve the accuracy of the health score.

[0172] 2) Interpretability heatmap mechanism

[0173] To identify "from which channel and at which moment the diagnostic judgment is derived", the system introduces a channel-time heatmap mechanism based on gradient response. This mechanism analyzes the partial derivative of the final output S of the model with respect to the normalized input :

[0174]

[0175] where is the value of the th channel in the normalized input tensor at the th moment; is the health score output by the model; is the influence gradient of this position on the score; is used to generate a two-dimensional channel-time map to assist in locating the source of the anomaly.

[0176] The heatmap can be used to generate the "response trajectory" of the device to assist in judging whether the current judgment comes from a stable trend, a transient pulse, or a collective anomaly of the structural link channels. At the same time, this mechanism has model independence and task adaptability and can be extended for response attribution interpretation in multi-object prediction or multi-modal fusion tasks. To improve the refinement degree of the heatmap, the system normalizes the input​ Perform convolution enhancement, compress the dimension through 1×1 convolution (output channels are / 2), and then capture medium-term trends and long-term dependence features through 3×3 and 5×5 convolution kernels (output channels are both / 4) respectively to generate an enhanced input tensor. In addition, the system adaptively adjusts the weights of each gradient by calculating the global average activation value of the gradient (the weights of gradients with higher activation values are increased to 1.2 times), improving the localization accuracy of the heatmap.

[0177] Experimental verification:

[0178] To verify the effectiveness of the method proposed in this invention, this paper designs a comparative experiment in a typical industrial scenario, covering two types of application objects: power equipment (such as high-voltage circuit breakers, partial discharge monitoring terminals) and oil and gas equipment (such as fracturing equipment, electric drive pump stations). The experimental samples include multi-channel synchronous time-series signals, covering common industrial sensing dimensions such as current, voltage, vibration, and temperature, and the total amount of collected data reaches 12,000 complete diagnostic cycles. In the experimental setup, real-world working condition factors such as signal mutations, slow trend changes, background interference, and data segment missing are introduced respectively to simulate different abnormal forms and multi-scale time-series features during equipment operation.

[0179] The comparative methods include current mainstream time-series modeling and anomaly detection models, such as convolutional-based TCN (Temporal Convolutional Network), gated recurrent neural network GRU, Transformer dominated by attention mechanism, and multi-scale modeling methods proposed in recent years, such as MSCNN, DSSANet that integrates structural path information, MTNet with scale adaptive modeling, etc. All models are experimented based on a unified training / validation / test division scheme, using the same optimization strategy and hyperparameter settings to ensure the fairness of comparison. The experimental results are as Figure 2 , Figure 3 , Figure 4 , Figure 5 shown in and Table 1. Among them, in order to display different data metrics in the same radar chart, relevant processing is done on the five metrics: ① Normalize to the [0,1] interval; ② Reverse process the metrics that are "the smaller the better" (such as missed diagnosis rate, false alarm rate, inference time); ③ Directly normalize the metrics that are "the larger the better" (ACC, F1).

[0180] Table 1 Data comparison of different methods under six metrics

[0181] Model Name ACC F1 Miss Detection Rate False Alarm Rate Inference Time Robustness Improvement TCN 85.2% 83.6% 22.3% 16.5% 1.12s - GRU 83.9% 82.1% 24.7% 18.3% 1.10s - Transformer 86.5% 85.4% 20.2% 14.7% 1.45s +4.8% MSCNN 87.3% 85.9% 18.9% 14.2% 1.25s +6.7% DSSANet 88.1% 86.7% 17.5% 13.6% 1.30s +9.4% MTNet 89.0% 87.6% 14.8% 11.8% 1.05s +11.2% The Method of the Present Invention 93.4% 92.1% 7.0% 6.0% 0.48s +23%

[0182] From Table 1 and Figure 2 , Figure 3 ,Figure 4 , Figure 5 As can be seen from the results, the traditional GRU and TCN methods have certain recognition capabilities in low-complexity sequence modeling scenarios and can learn some temporal dependence features. The TCN has a relatively fast inference speed while maintaining a certain accuracy. However, due to its fixed convolutional receptive field, it is difficult to capture long-range trend change signals, and it is not very sensitive to sudden disturbances, resulting in poor performance in terms of miss detection rate and false alarm rate. The GRU often has problems such as abnormal response lag or state judgment jump due to the limited ability of its gating structure to process long sequences, leading to insufficient model stability. The Transformer method has strong global modeling capabilities and can model long-term dependencies to a certain extent. However, due to its native structure's low sensitivity to local features, lack of structural path perception ability, and difficulty in focusing on redundant interference signals in multi-channel scenarios, the false alarm rate is relatively high. The MSCNN enhances the local feature modeling ability through multi-scale convolution and has good recognition effects on some sudden anomalies, but it lacks trend stability modeling and channel dependence relationship modeling, and its robustness performance is limited. Although the DSSANet introduces structural path information and performs well in identifying structural coupling anomalies, it does not effectively combine the time-scale differentiation modeling strategy, and its ability to identify slow-varying trend anomalies is still insufficient. The MTNet has made a breakthrough in the scale fusion mechanism and outperforms traditional models in overall performance. However, its inter-channel collaborative modeling ability is limited, especially in scenarios with channel distortion or local occlusion, where its performance is unstable.

[0183] In contrast, the method of the present invention effectively establishes the structural semantic dependence relationship between channels through the multi-channel normalization and direction-aware modeling mechanism, enhances the joint modeling ability for fast-changing and slow-changing anomalies by using the multi-scale modulation module, and strengthens the modeling consistency of the model for the evolution direction through the state trend guidance mechanism, making the state output have good continuity and forward-looking. In the experiment, this method is comprehensively superior to the comparative methods in six dimensions: accuracy, F1 value, miss detection rate, false alarm rate, inference time, and robustness. Especially in the identification of early weak faults and under complex working condition interference conditions, it shows obvious advantages, verifying its practicability in predictive maintenance tasks.

[0184] To verify the robustness of the model, the system introduced a noise interference test in the experiment. By adding Gaussian noise with a mean of 0 and a variance of 0.05 to the test set, the noise resistance performance of the model was evaluated. The results showed that the classification accuracy (ACC) of this method remained at 91.8%, and the F1 score was 90.5%, which was better than traditional methods (LSTM ACC 86.7%, F1 85.3%; CNN ACC 87.9%, F1 86.8%). In addition, the system compressed the features after noise interference through 1×1 convolution (output channels were C / 2), and then captured local patterns through 3×3 convolution (output channels were C / 2), further improving the noise resistance ability of the model.

[0185] To visually compare the performance of "the method of the present invention" with mainstream time series modeling methods (TCN, GRU, Transformer, MSCNN, DSSANet, MTNet) in terms of the fault prediction lead time, this paper plotted a simulation graph of "prediction lead time vs actual fault occurrence time", as Figure 6 shown. Through line charts, reference lines, and lead time annotations, the differences between the predicted fault times and the actual fault times (100 seconds) of each method were clearly shown.

[0186] The simulation results are as Figure 6 shown, indicating that "the method of the present invention" performs excellently in fault prediction. Its prediction time is 80 seconds, 20 seconds ahead of the actual fault, significantly better than TCN (8 seconds ahead), GRU (9 seconds), Transformer (10 seconds), MSCNN (11 seconds), DSSANet (12 seconds), and MTNet (13 seconds). The method of the present invention not only has higher prediction timeliness but also provides more sufficient response and handling time for the system, showing good practical value and deployment prospects.

[0187] Embodiment 2

[0188] This embodiment provides a complex equipment fault diagnosis system based on dynamic feature modeling, including:

[0189] A data acquisition module, configured to:

[0190] A computer-readable storage medium, in which multiple instructions are stored, and the instructions are suitable for being loaded and executed by a processor of a terminal device to execute the described complex equipment fault diagnosis method based on dynamic feature modeling.

[0191] A terminal device, including a processor and a computer-readable storage medium. The processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to execute the described complex equipment fault diagnosis method based on dynamic feature modeling.

[0192] The above are all preferred embodiments of the present invention, and the protection scope of the present invention is not limited thereby. Therefore, all equivalent changes made according to the structure, shape, and principle of the present invention shall be covered within the protection scope of the present invention.

Claims

1. A complex equipment fault diagnosis method based on dynamic feature modeling, characterized in that including: Obtain multi-channel sensor data; Perform data preprocessing on the obtained multi-channel sensor data; Extract initial channel features from the preprocessed data; Based on the direction-aware mechanism, perform convolutional enhancement on the concatenated initial channel features to generate a structure-aware global context feature; Based on the multi-scale modulation mechanism, perform multi-temporal scale feature fusion on the global context feature; Based on the trend-guided supervision mechanism, perform health trend guidance on the fused features; Based on the state continuity modulation, perform quantization scoring and output diagnosis on the output result; Based on the output diagnosis result, use the visualization heat map mapping to locate the source of the anomaly; The convolutional enhancement performed on the concatenated initial channel features based on the direction-aware mechanism includes introducing the direction-aware mechanism, extracting statistical expressions from the time direction, channel direction, and structural path direction to enhance the structural semantic modeling ability, and performing convolutional enhancement after concatenating the three types of features to generate a structure-aware global context feature: ， Among them, High-dimensional features that fuse three-directional structural semantic information are used to guide the anomaly perception process in the multi-scale enhancement stage; The health trend guidance performed on the fused features based on the trend-guided supervision mechanism includes using the time derivative of the state output as an auxiliary task, adding a trend guidance output branch to the main feature to predict the time difference of the state value, where the true difference is calculated from the difference between consecutive state values, and using the difference result as a supervision signal to construct a trend loss function, expressed as: ， Among them, is the total length of the time series; is used to measure the mean square error between the predicted trend of the model and the true trend, guiding the model to learn the state evolution law in the time dimension.

2. The complex equipment fault diagnosis method based on dynamic feature modeling according to claim 1, characterized in that, The data preprocessing performed on the obtained multi-channel sensor data includes setting the original multi-channel input signal at the equipment operation time as: , and perform standard normalization processing on each channel, expressed as: ， Among them, is the mean value of the -th channel over the entire time period; is the standard deviation of the -th channel; is the original signal value of the -th channel at time ; represents the number of channels; represents the length of the time series, is the normalized result, making the mean value of each channel signal 0 and the variance 1.

3. The complex equipment fault diagnosis method based on dynamic feature modeling according to claim 2, wherein Extracting initial channel features from the preprocessed data includes re - splicing the signals after normalizing all channels into an input tensor , and using a one - dimensional convolutional embedding mapping function to extract the initial channel feature representation: ， Among them, represents a one-dimensional convolution operation with a convolution kernel width of 3 and an output channel of ; is a batch normalization operation; is the feature tensor after initial extraction.

4. A complex equipment fault diagnosis method based on dynamic feature modeling according to claim 3, characterized in that, The multi-temporal scale feature fusion performed on the global context feature based on the multi-scale modulation mechanism includes dividing the direction-enhanced features into several groups according to the channel dimension, and each group of features performs different-scale temporal pooling and depth convolution operations; Upsample and restore to the original time step and concatenate and fuse into a multi-scale feature tensor. To enhance the activation intensity of the local anomaly response channel, introduce a channel attention mechanism to generate a modulation map, and finally obtain the fused feature output, expressed as: ， Among them, is the final fused modulation feature, with both multi-scale modeling ability and structural semantic expression ability.

5. The complex equipment fault diagnosis method based on dynamic feature modeling according to claim 4, wherein, The quantization scoring and output diagnosis performed on the output result based on the state continuity modulation includes continuously diagnosing and quantifying the score of the equipment state at the main task output end. Among them, global temporal pooling is performed on the output main feature to obtain the compressed feature, and the classification probability is obtained through a two-layer fully connected network, expressed as: ， ， Among them, is the first-layer fully connected parameter; is the second-layer fully connected parameter; is the probability of predicting to belong to the th class, where ; is the number of classification categories of faults.

6. The complex equipment fault diagnosis method based on dynamic feature modeling according to claim 5, wherein, The quantization scoring and output diagnosis performed on the output result based on the state continuity modulation further includes performing a moving average process on the output state sequence based on the sliding window smoothing mechanism, introducing a continuity constraint as an additional loss, and constructing a classification total loss with the joint loss function. Finally, the total joint loss function is obtained, expressed as: ， Among them, is a hyperparameter used to balance the weights of trend learning and output smoothing.

7. A complex equipment fault diagnosis method based on dynamic feature modeling according to claim 6, characterized in that, The localization of the anomaly source using the visualization heat map mapping based on the output diagnosis result includes generating an equipment health score according to the classification probability, and introducing a gradient response map calculation mechanism to locate the moment that contributes the most to the fault prediction, expressed as: ， Among them, is the value of the channel in the normalized input tensor at the th moment; is the health score output by the model; is the influence gradient of this position on the score; is used to generate a two-dimensional channel-time graph for assisting in locating the source of anomalies.

8. A complex equipment fault diagnosis system based on dynamic feature modeling, which executes a complex equipment fault diagnosis method based on dynamic feature modeling as described in claim 1, characterized in that, including: A data acquisition module configured to obtain multi-channel sensor data; A preprocessing module configured to perform data preprocessing on the obtained multi-channel sensor data; A feature module configured to extract initial channel features from the preprocessed data; A perception module, configured to perform convolutional enhancement on the spliced initial channel features based on a direction perception mechanism to generate globally context-aware features with structural perception; A fusion module, configured to perform multi-temporal scale feature fusion on the globally context-aware features based on a multi-scale modulation mechanism; A guidance module, configured to perform health trend guidance on the fused features based on a trend guidance supervision mechanism; A diagnosis module, configured to perform quantization scoring and output diagnosis on the output results based on state continuity modulation; A localization module, configured to use a visualization heatmap mapping to localize the source of anomalies based on the output diagnosis results.

Images