Clutch state diagnosis and fault early warning method and system based on multi-source information fusion

Abstract

The embodiment of the application provides a clutch state diagnosis fault early warning method and system based on multi-source information fusion, belongs to the clutch fault early warning technical field, solves the problem that when the existing method evaluates the meshing state of the clutch, the data information source is single, and early and weak degradation signs of the clutch performance are difficult to capture, the method comprises preprocessing of multi-source state information, fusion processing of the preprocessed information set by a state evaluation model, determination of a clutch state evaluation coefficient based on a multi-modal fusion feature and a control chart algorithm, and time sequence dynamic analysis of the clutch based on a trend discrimination method; the application acquires multi-source state information of the clutch in real time based on a distributed sensing terminal, combines the state evaluation model to perform fusion processing on the preprocessed information set, can comprehensively evaluate the clutch state from multiple dimensions, realizes accurate identification and positioning of different fault modes, and greatly reduces the misdiagnosis and missed diagnosis rates.

Description

Technical Field

[0001] This invention relates to the field of clutch fault early warning technology, specifically to a clutch state diagnosis fault early warning method and system based on multi-source information fusion. Background Technology

[0002] In modern industrial production and power transmission systems, the SSS clutch, as a key mechanical component, is widely used in wind power generation, ship propulsion, industrial transmission, and other fields. It enables the connection and disengagement between two rotating shafts, thereby controlling power transmission and playing a crucial role in ensuring the normal operation of the entire system. However, during long-term operation, the SSS clutch is prone to various malfunctions due to factors such as mechanical wear, poor lubrication, overload impact, and environmental corrosion. If these malfunctions are not detected and addressed in a timely manner, they may lead to a decline in clutch performance, or even cause serious equipment damage and production accidents, resulting in significant economic losses and safety hazards.

[0003] Chinese patent application CN115289154A discloses a method, apparatus, and medium for monitoring clutch engagement state. In this method, a displacement sensor is installed on the clutch to acquire the displacement changes of each shaft. After receiving the displacement changes detected by the sensor, the processor determines the clutch engagement state based on the displacement changes and then adjusts the concentricity of the shafts accordingly. When the two gears of the clutch are engaged, if the clutch engagement state is ideal, the displacement change of the shafts should be close to zero; if the clutch engagement state is non-ideal, the clutch will be subjected to shear force and bending moment, resulting in a large displacement change of the shafts. However, existing methods for evaluating clutch engagement state rely on a single data source, and a single displacement signal is insufficient to distinguish different fault modes, easily leading to misjudgments or omissions, thus making it difficult to detect early and subtle signs of clutch performance degradation.

[0004] To address the aforementioned problems, this invention proposes a clutch condition diagnosis and fault early warning method and system based on multi-source information fusion. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems existing in the prior art, and provides a clutch state diagnosis fault early warning method and system based on multi-source information fusion.

[0006] In a first aspect, embodiments of the present invention provide a clutch condition diagnosis fault early warning method based on multi-source information fusion, the method comprising: Based on the real-time acquisition of multi-source state information of the clutch by distributed sensing terminals, the multi-source state information is preprocessed and a preprocessed information set is output. A pre-built state evaluation model based on deep learning is introduced into the state evaluation model. A multi-module supervision strategy is introduced into the state evaluation model. The pre-built state evaluation model is deployed based on the multi-module supervision strategy, historical state information is captured, and noise perturbation is added to the historical state information to obtain noise perturbation samples. The state evaluation model is iteratively trained based on the noise perturbation samples to output a converged state evaluation model. A preprocessed information set is acquired, and the state evaluation model performs fusion processing on the preprocessed information set to obtain multimodal fusion features characterizing the real-time state of the clutch. Based on the multimodal fusion features and the control chart algorithm, the clutch state evaluation coefficients are determined, and a set of state evaluation coefficients is output. A set of loading state evaluation coefficients is used. The state evaluation model performs time-series dynamic analysis of the clutch based on the trend discrimination method and outputs a comprehensive evaluation result.

[0007] Optionally, the preprocessing of multi-source state information includes: With the system master clock as a constraint, a unified time base is allocated to the multi-source state information to obtain multi-source state information after time and space stamp alignment. It is determined whether the information sampling frequency is high-frequency sampling information, and the equalization fusion period is preset based on the type of multi-source state information. If the information sampling frequency is high-frequency sampling information, the multi-source state information is interpolated and resampled to obtain the multi-source state information after being resampled to a set fusion period. If the information sampling frequency is not high-frequency sampling information, the multi-source state information is filled in forward to obtain the multi-source state information after being filled to the set fusion period. The synchronized multi-source state information is loaded, and adaptive outlier detection is performed on the multi-source state information based on the threshold detection method to obtain the outlier-cleaned multi-source state information. After outlier cleaning, multi-source state information is obtained, and a hybrid filtering strategy is used to filter and denoise the multi-source state information, outputting a pre-processed information set after filtering and denoising.

[0008] Optionally, the preset equalization fusion period based on multi-source state information types includes: Acquire multi-source state information sources, classify multi-source state information into high-frequency dynamic type, medium-frequency slowly changing type and low-frequency command type based on the information source, and record the typical sampling period and maximum allowable delay of the information type; Traverse the typical sampling period and maximum allowable delay of information types, and determine the local processing period of multi-source state information based on the information type; The equalization fusion cycle constraint principle is preset. The equalization fusion cycle constraint principle is the minimum common cycle in which all information types participating in the fusion can complete a synchronous update within this cycle. The equalization fusion cycle is determined based on the equalization fusion cycle calculation formula.

[0009] Optionally, the adaptive outlier detection of multi-source state information based on the threshold detection method includes: Obtain the information type classification results of multi-source state information, determine the initial static threshold of information type based on the information type classification results, and set the upper and lower limits of the standard deviation of the initial static threshold of information type. Within the local processing cycle of information type, the mean and standard deviation of multi-source state information are statistically analyzed, and an adaptive threshold for information type is determined based on the initial static threshold, the mean and standard deviation of multi-source state information; Load multi-source status information, determine whether the multi-source status information exceeds the adaptive threshold. If the multi-source status information exceeds the adaptive threshold, determine the multi-source status information as abnormal and mark it. If the multi-source status information does not exceed the adaptive threshold, retain the multi-source status information.

[0010] Optionally, the step of using a hybrid filtering strategy to filter and reduce noise from multi-source state information includes: The outlier-cleaned multi-source state information is loaded, and a low-pass Butterworth filter is used to perform sliding filtering on the outlier-cleaned multi-source state information to output the multi-source state information after filtering out and smoothing high-frequency noise. The process involves acquiring multi-source state information after filtering out and smoothing high-frequency noise, performing frequency domain analysis on the multi-source state information, identifying transient anomalies in the multi-source state information, and removing transient anomalies from the multi-source state information. Based on the equalization fusion cycle, multi-source state information is segmented and extracted. Then, a label field is added to the segmented multi-source state information features. The segmented multi-source state information with added label fields is then standardized by Z-score to unify the dimensionality, and a preprocessed information set is output.

[0011] Optionally, the state evaluation model is based on an LSTM network model and includes an input layer, an LSTM network model, and an output layer. A multimodal fusion layer is set between the input layer and the LSTM network model. The multimodal fusion layer is a graph neural network architecture, and a cross-modal attention mechanism is introduced in the graph neural network architecture. A dynamic analysis layer is introduced between the LSTM network model and the output layer.

[0012] Optionally, the determination of clutch state evaluation coefficients based on multimodal fusion features combined with control chart algorithms includes: Obtain the preprocessed information set, extract the information feature variables corresponding to the preprocessed information set based on the Transformer time encoder, and output the time-encoded feature variable set. The feature variable set is used as the node of the multimodal graph structure, and the relationship between the information feature variables in the feature variable set is used as the connection edge of the multimodal graph structure. The heterogeneous multimodal graph structure is updated based on the nodes and connection edges. Modal embedding encoding is performed on the nodes in the multimodal graph structure. The nodes in the multimodal graph structure are fused based on the multi-module supervision strategy, and the weighted multimodal fusion feature representing the real-time state of the clutch is output. Based on the LSTM network, the fusion feature sequence that represents the dynamic evolution trend of multimodal fusion features is captured. The standard dimension that is strongly correlated with the health of the clutch is selected as the monitoring variable, and the modulus of the fusion feature sequence is used as the comprehensive health index. Principal component dimensionality reduction is performed on the comprehensive health indicators to obtain the dimensionality reduction representation of the comprehensive monitoring indicators. The dimensionality reduction vector of the indicators is output. The dimensionality reduction vector of the indicators is then weighted and fused based on the standard dimension to output the multi-indicator fusion vector. Based on Shewhart control charts, multi-indicator fusion vectors are used to monitor the trends of mean and standard deviation, and a state evaluation function is constructed to determine the clutch state evaluation coefficients.

[0013] Optionally, the state assessment model performs time-series dynamic analysis of the clutch based on the trend discriminant method, including: Load the state evaluation coefficient set, and the dynamic analysis layer establishes a sliding window polynomial fitting trend model based on the state evaluation coefficients and multi-index fusion vectors. The sliding window polynomial fitting trend model captures the trend slope, trend significance, and trend direction of the multi-indicator fusion vector, and performs weighted voting fusion on the trend slope, trend significance, and trend direction of the multi-indicator fusion vector. The comprehensive confidence level of the multi-indicator fusion vector is calculated based on the trend confidence function, and the comprehensive confidence level is used as the comprehensive evaluation value. The health level of the clutch is determined based on the comprehensive evaluation value. Determine if the clutch health level exceeds the preset health threshold. If the clutch health level exceeds the preset health threshold, trigger a clutch warning command.

[0014] Secondly, embodiments of the present invention provide a clutch condition diagnosis fault early warning system based on multi-source information fusion, used to implement the clutch condition diagnosis fault early warning method based on multi-source information fusion described above, the system comprising: The preprocessing module collects multi-source state information of the clutch in real time based on distributed sensing terminals, preprocesses the multi-source state information, and outputs a preprocessed information set. The model building module is used to pre-build a deep learning-based state evaluation model. A multi-module supervision strategy is introduced into the state evaluation model. The pre-built state evaluation model is deployed based on the multi-module supervision strategy, historical state information is captured, and noise perturbation is added to the historical state information to obtain noise perturbation samples. The state evaluation model is iteratively trained based on the noise perturbation samples to output a converged state evaluation model. The comprehensive evaluation module is used to acquire a preprocessed information set. The state evaluation model performs fusion processing on the preprocessed information set to obtain multimodal fusion features that characterize the real-time state of the clutch. Based on the multimodal fusion features and the control chart algorithm, the clutch state evaluation coefficients are determined and the state evaluation coefficient set is output. The state evaluation model performs time-series dynamic analysis of the clutch based on the trend discrimination method and outputs the comprehensive evaluation results.

[0015] Optionally, the comprehensive evaluation module includes: The fusion coding unit is used to acquire the preprocessed information set. The state evaluation model performs fusion processing on the preprocessed information set to obtain multimodal fusion features that characterize the real-time state of the clutch. The evaluation coefficient calculation unit determines the clutch state evaluation coefficients based on multimodal fusion features and control chart algorithms, and outputs a set of state evaluation coefficients. The timing analysis unit performs timing dynamic analysis on the clutch based on the trend discrimination method and outputs a comprehensive evaluation result.

[0016] Compared with the prior art, the embodiments of the present invention have the following main advantages: In this embodiment of the invention, multi-source state information of the clutch is collected in real time based on a distributed sensing terminal. The pre-processed information set is fused and processed in conjunction with a state assessment model. The clutch is then subjected to time-series dynamic analysis based on a trend discrimination method. Multi-source heterogeneous data of various physical quantities, including but not limited to vibration, temperature, pressure, sound, and current, are collected, providing a data foundation for comprehensive evaluation. Through multi-source information fusion and deep feature extraction, the limitations of a single sensor are overcome. The clutch state can be comprehensively evaluated from multiple dimensions, enabling accurate identification and location of different fault modes. This greatly reduces the rate of misdiagnosis and missed diagnosis, and can keenly capture early and subtle signs of clutch performance degradation, issuing early warnings before faults occur. This provides maintenance personnel with sufficient reaction time, transforming passive maintenance into proactive prevention, thereby avoiding major equipment failures, reducing downtime losses, and extending equipment lifespan.

[0017] In this embodiment of the invention, when preprocessing multi-source state information, allocating a unified time reference for the multi-source state information ensures that all subsequent analyses and fusions are based on real state snapshots at the same time or within the same time period, guaranteeing the accuracy and effectiveness of data fusion. When using adaptive outlier detection, an adaptive threshold can be dynamically determined based on the real-time statistical characteristics and operating conditions of each information type to identify and remove outliers. This effectively removes real outliers while preserving data fluctuations caused by changes in normal operating conditions to the maximum extent. Furthermore, after hybrid filtering and standardization, the real state characteristics in the data are completely preserved and highlighted, while various irrelevant noises and interferences can be suppressed to the lowest level.

[0018] In this embodiment of the invention, when the equalization fusion period is preset based on the multi-source state information type, the multi-source state information is divided into high-frequency dynamic type, medium-frequency slowly changing type, and low-frequency command type. This allows for targeted processing based on the characteristics of different data types. This classification method makes subsequent processing strategies more precise, avoiding a "one-size-fits-all" approach and better adapting to the characteristics of different data types. Furthermore, based on the equalization fusion period constraint principle, it ensures that all information types participating in the fusion can complete a synchronous update within this period. This minimum common period constraint method guarantees that data with different sampling frequencies can be updated synchronously during fusion, avoiding data misalignment or update delays caused by differences in sampling frequencies. By determining the equalization fusion period, the update frequencies of different data types can be reasonably arranged, avoiding unnecessary data processing and transmission, and improving the efficiency of data fusion.

[0019] In this embodiment of the invention, when using a hybrid filtering strategy to filter and denoise multi-source state information, a low-pass Butterworth filter is used to perform sliding filtering on the outlier-cleaned multi-source state information. This effectively filters out high-frequency noise. The Butterworth filter has excellent frequency characteristics, enabling it to retain important low-frequency trend information in the data while filtering out high-frequency noise. Furthermore, through frequency domain analysis, transient anomalies in the multi-source state information can be identified. After identifying transient anomalies, they are removed from the data, further improving data quality. After removing transient anomalies, the data is smoother and more representative, reducing misjudgments caused by sudden anomalies and improving the reliability of subsequent analysis.

[0020] In this embodiment of the invention, when determining the clutch state evaluation coefficients based on multimodal fusion features combined with control chart algorithms, a Transformer time-series encoder extracts temporal feature variables from the preprocessed information set. These feature variables are then used as nodes in the multimodal graph structure, with the relationships between the feature variables serving as connecting edges to construct a heterogeneous multimodal graph structure. This method fully utilizes the temporal characteristics and inherent relationships of multi-source information to extract more comprehensive and richer feature representations. Furthermore, by capturing the dynamic evolution trend of the multimodal fusion features using an LSTM network, the temporal changes in clutch state can be effectively identified. This capture of dynamic evolution trends allows the state evaluation to reflect changes in clutch health status in real time, rather than being based on static data. This helps to promptly detect potential clutch faults, improving the real-time and dynamic nature of the state evaluation. Detecting the mean and standard deviation trends of the multi-index fusion vector using a Shewhart control chart enables real-time monitoring of clutch state changes and timely detection of anomalies. Control chart algorithms are a classic statistical process control method that can effectively identify abnormal changes in clutch status and provide early warnings. By constructing a status evaluation function, the mean and standard deviation trends of a multi-index fusion vector can be transformed into specific clutch status evaluation coefficients. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 A schematic diagram of the implementation process of a clutch condition diagnosis and fault early warning method based on multi-source information fusion is shown. Figure 2 This diagram illustrates the implementation flow of a preprocessing method for multi-source state information. Figure 3 This diagram illustrates the implementation process of a method based on a preset equalization fusion cycle for multi-source state information types. Figure 4 This diagram illustrates the implementation process of an adaptive outlier detection method for multi-source state information based on threshold detection. Figure 5 This diagram illustrates the implementation process of a method for filtering and denoising multi-source state information using a hybrid filtering strategy. Figure 6 This diagram illustrates the implementation process of a method for determining clutch state evaluation coefficients based on multimodal fusion features combined with a control chart algorithm. Figure 7A schematic diagram of the implementation process of the time-series dynamic analysis method for clutch based on the trend discriminant method is shown in the state assessment model. Figure 8 A schematic diagram of a clutch condition diagnosis and fault early warning system based on multi-source information fusion is shown. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0025] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0026] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0027] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0028] Existing methods for assessing clutch engagement status rely on a single data source, and a single displacement signal is insufficient to distinguish different fault modes, leading to misjudgments or omissions, thus failing to capture early and subtle signs of clutch performance degradation. To address these issues, this invention proposes a clutch status diagnosis and fault early warning method and system based on multi-source information fusion. In short, the method first collects multi-source clutch status information in real time using distributed sensing terminals. This information is preprocessed, and a status assessment model is iteratively trained based on noise disturbance samples, outputting a converged status assessment model. The status assessment model then fuses the preprocessed information set to obtain multimodal fusion features characterizing the real-time clutch status. Based on these multimodal fusion features and a control chart algorithm, clutch status assessment coefficients are determined. Finally, the status assessment model performs time-series dynamic analysis of the clutch using a trend discriminant method and outputs a comprehensive assessment. The invention utilizes distributed sensing terminals to collect multi-source state information of the clutch in real time. This information is then fused using a state assessment model, and a trend-based dynamic analysis of the clutch is performed. The system collects multi-source heterogeneous data, including but not limited to vibration, temperature, pressure, sound, and current, providing a data foundation for comprehensive evaluation. Through multi-source information fusion and deep feature extraction, the limitations of a single sensor are overcome, enabling a comprehensive assessment of the clutch state from multiple dimensions. This allows for accurate identification and location of different fault modes, significantly reducing misdiagnosis and missed diagnosis rates. Furthermore, it can keenly detect early, subtle signs of clutch performance degradation, issuing warnings before faults occur, providing maintenance personnel with sufficient reaction time, transforming passive maintenance into proactive prevention, thereby avoiding major equipment failures, reducing downtime losses, and extending equipment lifespan.

[0029] This invention provides a clutch condition diagnosis and fault early warning method based on multi-source information fusion. Figure 1 A schematic diagram of the implementation process of a clutch condition diagnosis fault early warning method based on multi-source information fusion is shown. The clutch condition diagnosis fault early warning method based on multi-source information fusion specifically includes: S10: Based on the distributed sensing terminal, the multi-source status information of the clutch is collected in real time, the multi-source status information is preprocessed, and the preprocessed information set is output. S20: A pre-built state assessment model based on deep learning is constructed. A multi-module supervision strategy is introduced into the state assessment model. The pre-built state assessment model is deployed based on the multi-module supervision strategy, historical state information is captured, and noise perturbation is added to the historical state information to obtain noise perturbation samples. The state assessment model is iteratively trained based on the noise perturbation samples to output a converged state assessment model. The multi-module supervision strategy and noise perturbation sample training enable the model to learn to accurately assess the state even when there is noise and perturbation in the data, with extremely strong generalization ability and robustness. S30, acquire the preprocessed information set, the state evaluation model performs fusion processing on the preprocessed information set to obtain multimodal fusion features characterizing the real-time state of the clutch, and determines the clutch state evaluation coefficients based on the multimodal fusion features combined with the control chart algorithm, and outputs the state evaluation coefficient set. S40, load the set of state evaluation coefficients. The state evaluation model performs time-series dynamic analysis of the clutch based on the trend discrimination method and outputs a comprehensive evaluation result.

[0030] In this embodiment of the invention, multi-source state information of the clutch is collected in real time based on a distributed sensing terminal. The pre-processed information set is fused and processed in conjunction with a state assessment model. The clutch is then subjected to time-series dynamic analysis based on a trend discrimination method. Multi-source heterogeneous data of various physical quantities, including but not limited to vibration, temperature, pressure, sound, and current, are collected, providing a data foundation for comprehensive evaluation. Through multi-source information fusion and deep feature extraction, the limitations of a single sensor are overcome. The clutch state can be comprehensively evaluated from multiple dimensions, enabling accurate identification and location of different fault modes. This greatly reduces the rate of misdiagnosis and missed diagnosis, and can keenly capture early and subtle signs of clutch performance degradation, issuing early warnings before faults occur. This provides maintenance personnel with sufficient reaction time, transforming passive maintenance into proactive prevention, thereby avoiding major equipment failures, reducing downtime losses, and extending equipment lifespan.

[0031] This invention provides a method for preprocessing multi-source state information. Figure 2 This diagram illustrates the implementation flow of a multi-source state information preprocessing method, which specifically includes: S101, constrained by the system master clock, allocates a unified time base to the multi-source state information to obtain the multi-source state information after time and space stamp alignment. The allocation of a unified time base to the multi-source state information ensures that all subsequent analysis and fusion are based on the real state snapshot at the same time or within the same time period, thus guaranteeing the accuracy and effectiveness of data fusion. S102, determine whether the information sampling frequency is high-frequency sampling information, and preset the equalization fusion period based on the multi-source state information type. The preset equalization fusion period avoids the processing bottleneck or resource waste caused by waiting for asynchronous data, so that the system can perform collaborative analysis and state assessment of multi-source information with the highest efficiency. S103, if the information sampling frequency is high-frequency sampling information, perform interpolation and resampling processing on the multi-source state information to obtain the multi-source state information after resampling to the set fusion period; S104, If the information sampling frequency is not high-frequency sampling information, perform forward padding and completion processing on the multi-source state information to obtain the multi-source state information after being completed to the set fusion period. S105, Load the synchronized multi-source state information, perform adaptive outlier detection on the multi-source state information based on the threshold detection method, and obtain the outlier-cleaned multi-source state information; S106: Obtain multi-source state information after outlier cleaning, use a hybrid filtering strategy to filter and denoise the multi-source state information, and output the preprocessed information set after filtering and denoising.

[0032] In this embodiment of the invention, when preprocessing multi-source state information, allocating a unified time reference for the multi-source state information ensures that all subsequent analyses and fusions are based on real state snapshots at the same time or within the same time period, guaranteeing the accuracy and effectiveness of data fusion. When using adaptive outlier detection, an adaptive threshold can be dynamically determined based on the real-time statistical characteristics and operating conditions of each information type to identify and remove outliers. This effectively removes real outliers while preserving data fluctuations caused by changes in normal operating conditions to the maximum extent. Furthermore, after hybrid filtering and standardization, the real state characteristics in the data are completely preserved and highlighted, while various irrelevant noises and interferences can be suppressed to the lowest level.

[0033] This invention provides a method for pre-setting equalization fusion period based on multi-source state information types. Figure 3 This diagram illustrates the implementation process of a method for pre-setting an equalization fusion period based on multi-source state information types. The method specifically includes: S1021, acquire multi-source state information sources, classify multi-source state information into high-frequency dynamic type, medium-frequency slow-change type and low-frequency command type based on information sources, and record the typical sampling period and maximum allowable delay of information type; S1022, Traverse the typical sampling period and maximum allowable delay of information types, and determine the local processing period of multi-source state information based on information type; The formula for determining the local processing cycle is expressed as follows:

[0034] in, Indicates information type Local processing cycle, Information type The processing adjustment coefficient and delay margin coefficient, Information type Typical sampling period and maximum allowable delay, Number of information types;

[0035] in, Indicates the adjustment constant. These are the local maximum processing cycle and the local minimum processing cycle, respectively. S1023, Preset equalization fusion cycle constraint principle. The equalization fusion cycle constraint principle is the minimum common cycle in which all information types participating in the fusion can complete a synchronous update within this cycle. The equalization fusion cycle is determined based on the equalization fusion cycle calculation formula. The formula for calculating the equalization fusion cycle is as follows:

[0036] in, Indicates the equalization and integration cycle. It is the minimum common multiple of the local processing cycle.

[0037] In this embodiment of the invention, when the equalization fusion period is preset based on the multi-source state information type, the multi-source state information is divided into high-frequency dynamic type, medium-frequency slowly changing type, and low-frequency command type. This allows for targeted processing based on the characteristics of different data types. This classification method makes subsequent processing strategies more precise, avoiding a "one-size-fits-all" approach and better adapting to the characteristics of different data types. Furthermore, based on the equalization fusion period constraint principle, it ensures that all information types participating in the fusion can complete a synchronous update within this period. This minimum common period constraint method guarantees that data with different sampling frequencies can be updated synchronously during fusion, avoiding data misalignment or update delays caused by differences in sampling frequencies. By determining the equalization fusion period, the update frequencies of different data types can be reasonably arranged, avoiding unnecessary data processing and transmission, and improving the efficiency of data fusion.

[0038] This invention provides an adaptive outlier detection method for multi-source state information based on threshold detection. Figure 4 This diagram illustrates the implementation flow of an adaptive outlier detection method for multi-source state information based on threshold detection. The method specifically includes: S1051, obtain the information type classification result of multi-source state information, determine the initial static threshold of information type based on the information type classification result, and set the upper and lower limits of the standard deviation of the initial static threshold of information type. The introduction of the initial static threshold and standard deviation boundary can set a basic and safe initial threshold range for each information type and set the upper and lower limits of its standard deviation. The information operating condition factor can be adjusted according to different operating modes (such as idling, constant speed cruising, rapid acceleration, and hill climbing). S1052 calculates the mean and standard deviation of multi-source state information within the local processing cycle of the information type. Based on the initial static threshold and the mean and standard deviation of the multi-source state information, an adaptive threshold for the information type is determined. The adaptive threshold is set within a dynamic range determined by empirical knowledge, real-time data, and operating conditions. This approach balances expert macro-level control with micro-level data verification, making the anomaly detection logic chain more complete and reliable. The adaptive threshold is expressed as:

[0039]

[0040] in, An adaptive threshold representing the information type. These represent the mean and standard deviation of the information condition factor and the multi-source state information, respectively. This is the initial static threshold; S1053, load multi-source state information, determine whether the multi-source state information exceeds the adaptive threshold. If the multi-source state information exceeds the adaptive threshold, determine that the multi-source state information is abnormal and mark it. If the multi-source state information does not exceed the adaptive threshold, retain the multi-source state information. In this embodiment of the invention, by setting an adaptive threshold and using a detection method based on statistical characteristics, it is possible to dynamically adapt to changes in data characteristics, reduce false positives and false negatives, and improve the accuracy and reliability of outlier detection.

[0041] This invention provides a method for filtering and denoising multi-source state information using a hybrid filtering strategy. Figure 5 This diagram illustrates the implementation flow of a method for filtering and denoising multi-source state information using a hybrid filtering strategy. The method specifically includes: S1061 loads the outlier-cleaned multi-source state information and uses a low-pass Butterworth filter to perform sliding filtering on the outlier-cleaned multi-source state information, outputting multi-source state information after filtering out smooth high-frequency noise. The use of a low-pass Butterworth filter for sliding averaging aims to filter out the most common and random smooth high-frequency noise in the data. The Butterworth filter is designed to achieve a maximally flat passband response, meaning that while filtering out high-frequency noise, it can retain the phase and amplitude information of the signal to the greatest extent, avoiding signal distortion. Frequency domain elimination precisely targets the identified "abnormal frequency components," with minimal impact on normal signals. This refined processing method ensures the preservation of key characteristics strongly correlated with the clutch's health status. S1062, acquire the multi-source state information after filtering out and smoothing high-frequency noise, perform frequency domain analysis on the multi-source state information, identify transient abnormal information in the multi-source state information, and remove transient abnormal information in the multi-source state information. S1063 extracts multi-source state information in segments based on the equalization fusion cycle, adds a label field to the features of the segmented multi-source state information, performs Z-score standardization and unified dimensional processing on the segmented multi-source state information after adding the label field, and outputs a preprocessed information set.

[0042] In this embodiment of the invention, when using a hybrid filtering strategy to filter and denoise multi-source state information, a low-pass Butterworth filter is used to perform sliding filtering on the outlier-cleaned multi-source state information. This effectively filters out high-frequency noise. The Butterworth filter has excellent frequency characteristics, enabling it to retain important low-frequency trend information in the data while filtering out high-frequency noise. Furthermore, through frequency domain analysis, transient anomalies in the multi-source state information can be identified. After identifying transient anomalies, they are removed from the data, further improving data quality. After removing transient anomalies, the data is smoother and more representative, reducing misjudgments caused by sudden anomalies and improving the reliability of subsequent analysis.

[0043] In this embodiment of the invention, the state assessment model is based on an LSTM network model, including an input layer, an LSTM network model, and an output layer. A multimodal fusion layer is set between the input layer and the LSTM network model. The multimodal fusion layer is a graph neural network architecture, and a cross-modal attention mechanism is introduced into the graph neural network architecture. The multimodal fusion layer is used to acquire a preprocessed information set, extract the information feature variables corresponding to the preprocessed information set based on the Transformer temporal encoder, output the temporally encoded feature variable set, and use the feature variable set as nodes of the multimodal graph structure. The information feature variables in the feature variable set are used as the connection edges of the multimodal graph structure. The heterogeneous multimodal graph structure is updated based on the nodes and connection edges. Modal embedding encoding is performed on the nodes in the multimodal graph structure. The nodes in the multimodal graph structure are fused based on a multi-module supervision strategy, and a weighted multimodal fusion feature representing the real-time state of the clutch is output. The LSTM network model is used to acquire the multimodal fusion feature, capture the fusion feature sequence representing the dynamic evolution trend of the multimodal fusion feature based on the LSTM network, and select a standard dimension strongly correlated with the clutch health as a monitoring variable. Using the modulus of the fused feature sequence as a comprehensive health indicator, principal component dimensionality reduction is performed on the comprehensive health indicator to obtain a dimensionality-reduced representation of the comprehensive monitoring indicator, and the dimensionality-reduced vector of the indicator is output. The dimensionality-reduced vector of the indicator is weighted and fused based on the standard dimension to output a multi-indicator fused vector. The mean and standard deviation trends of the multi-indicator fused vector are monitored based on Shewhart control charts, and a state evaluation function is constructed. The clutch state evaluation coefficient is determined based on the state evaluation function. A dynamic analysis layer is introduced between the LSTM network model and the output layer. The dynamic analysis layer establishes a sliding window polynomial fitting trend model based on the state evaluation coefficient and the multi-indicator fused vector. The sliding window polynomial fitting trend model captures the trend slope, trend significance, and trend direction of the multi-indicator fused vector. The trend slope, trend significance, and trend direction of the multi-indicator fused vector are weighted and fused by voting. The comprehensive confidence of the multi-indicator fused vector is calculated based on the trend confidence function. The comprehensive confidence is used as the comprehensive evaluation value. The health level of the clutch is determined based on the comprehensive evaluation value, and it is judged whether the health level of the clutch exceeds the preset health threshold. If the health level of the clutch exceeds the preset health threshold, a clutch warning command is triggered.

[0044] The multimodal fusion feature output is represented as follows:

[0045] in, These represent the updated output representation of the multimodal fusion feature m, the feature representation of the multimodal fusion feature m at layer l, and the feature representation of the associated feature u of the multimodal fusion feature m, respectively. For node update functions, This represents the connection weights between the multimodal fusion feature m and the associated feature u; The state evaluation function is expressed as:

[0046]

[0047] in, Represents the state evaluation coefficient. These are the multi-indicator fusion vector and the mean of the multi-indicator fusion vector, respectively. These are the continuous trend index and the standard deviation of the multi-indicator fusion vector, respectively. These are the Shewhart control chart control limits, deviation from trend constant, and number of indicators.

[0048] This invention provides a method for determining clutch state evaluation coefficients based on multimodal fusion features combined with a control chart algorithm. Figure 6 This diagram illustrates the implementation flow of a method for determining clutch state evaluation coefficients based on multimodal fusion feature analysis combined with a control chart algorithm. The method specifically includes: S201, Obtain the preprocessed information set, extract the information feature variables corresponding to the preprocessed information set based on the Transformer time encoder, and output the time-encoded feature variable set. Using the Transformer encoder, not only can the deep features of each sensor signal be extracted, but also the long-range dependencies of the features on the time axis can be captured. S202, the feature variable set is used as the nodes of the multimodal graph structure, and the relationship between the information feature variables in the feature variable set is used as the connection edge of the multimodal graph structure. The heterogeneous multimodal graph structure is updated based on the nodes and connection edges. Modal embedding encoding is performed on the nodes in the multimodal graph structure. The nodes in the multimodal graph structure are fused based on the multi-module supervision strategy, and the weighted multimodal fusion feature representing the real-time state of the clutch is output. When the graph neural network (GNN) is updated, the message passing mechanism of the GNN allows the nodes (features) to obtain information from the neighboring nodes (related features), thereby dynamically learning and updating their own representations. This allows a vibration feature to not only represent itself, but also to be associated with related information such as temperature and pressure. The attention mechanism makes the fusion process context-aware. In different health states, the dominant factors affecting the clutch state are different. This embodiment can automatically adjust the weights based on the attention mechanism, always grasping the most important contradiction at the moment, so that the fused feature can most accurately represent the real comprehensive state of the clutch at any time. S203, based on LSTM network to capture fusion feature sequence that represents the dynamic evolution trend of multimodal fusion features, selects standard dimensions that are strongly correlated with clutch health as monitoring variables, and uses the modulus of fusion feature sequence as comprehensive health index. S204 performs principal component dimensionality reduction on the comprehensive health indicators to obtain the dimensionality reduction representation of the comprehensive monitoring indicators, outputs the indicator dimensionality reduction vector, and performs weighted fusion on the indicator dimensionality reduction vector based on the standard dimension to output the multi-indicator fusion vector. S205 uses Shewhart control charts to detect the mean and standard deviation trends of a multi-indicator fusion vector, and constructs a state evaluation function. Based on this function, it determines the clutch state evaluation coefficients. The Shewhart control chart-based detection of the mean and standard deviation trends of the multi-indicator fusion vector enables real-time monitoring of clutch state changes and timely detection of anomalies. Control chart algorithms are a classic statistical process control method that effectively identifies abnormal changes in clutch state and provides early warnings.

[0049] In this embodiment of the invention, when determining the clutch state evaluation coefficients based on multimodal fusion features combined with control chart algorithms, a Transformer time-series encoder extracts temporal feature variables from the preprocessed information set. These feature variables are then used as nodes in the multimodal graph structure, with the relationships between the feature variables serving as connecting edges to construct a heterogeneous multimodal graph structure. This method fully utilizes the temporal characteristics and inherent relationships of multi-source information to extract more comprehensive and richer feature representations. Furthermore, by capturing the dynamic evolution trend of the multimodal fusion features using an LSTM network, the temporal changes in clutch state can be effectively identified. This capture of dynamic evolution trends allows the state evaluation to reflect changes in clutch health status in real time, rather than being based on static data. This helps to promptly detect potential clutch faults, improving the real-time and dynamic nature of the state evaluation. Detecting the mean and standard deviation trends of the multi-index fusion vector using a Shewhart control chart enables real-time monitoring of clutch state changes and timely detection of anomalies. Control chart algorithms are a classic statistical process control method that can effectively identify abnormal changes in clutch status and provide early warnings. By constructing a status evaluation function, the mean and standard deviation trends of a multi-index fusion vector can be transformed into specific clutch status evaluation coefficients.

[0050] This invention provides a method for time-series dynamic analysis of clutches based on a trend discriminant analysis model. Figure 7 The diagram illustrates the implementation flow of a clutch time-series dynamic analysis method based on trend discriminant analysis for state assessment modeling. This method specifically includes: S301 loads the state evaluation coefficient set. The dynamic analysis layer establishes a sliding window polynomial fitting trend model based on the state evaluation coefficients and multi-index fusion vectors. By loading the state evaluation coefficient set and using the sliding window polynomial fitting trend model, it captures the trend slope, trend significance, and trend direction of the multi-index fusion vectors, effectively capturing the dynamic changes in clutch state. The sliding window polynomial fitting can adapt to local changes in data, providing more flexible trend analysis. The sliding window mechanism allows the model to be updated in real time. As new data is added, the model can dynamically adjust the trend analysis results, ensuring real-time monitoring and dynamic evaluation of clutch state. S302, based on a sliding window polynomial fitting trend model, captures the trend slope, trend significance, and trend direction of a multi-indicator fusion vector. It then performs a weighted voting fusion of these factors, comprehensively considering the impact of multiple trend features on the clutch state. The weighted voting fusion method assigns different weights to each trend feature based on its importance, thus highlighting the influence of key features in the comprehensive evaluation. This method enhances the robustness of decision-making, comprehensively considers multiple trend features, more accurately reflects the changing trends of the clutch state, improves the accuracy of trend analysis, and provides a more reliable basis for assessing the clutch's health status. S303 calculates the comprehensive confidence level of the multi-indicator fusion vector based on the trend confidence function. This comprehensive confidence level is used as the overall evaluation value to determine the clutch's health level. Calculating the comprehensive confidence level of the multi-indicator fusion vector using the trend confidence function quantifies the reliability of the trend analysis results. The comprehensive confidence level provides a quantitative reliability indicator for assessing the clutch's condition, helping to more scientifically determine the clutch's health status. S304 determines whether the health level of the clutch exceeds the preset health threshold. If the health level of the clutch exceeds the preset health threshold, a clutch warning command is triggered. By determining whether the health level of the clutch exceeds the preset health threshold, potential clutch faults can be detected in time and a warning command can be triggered, thereby helping maintenance personnel to take measures in advance to avoid sudden equipment failures and reduce equipment downtime and maintenance costs.

[0051] In this embodiment of the invention, when the state assessment model performs time-series dynamic analysis of the clutch based on the trend discriminant method, it can more accurately capture the dynamic changes in the clutch state by using a sliding window polynomial fitting trend model, weighted voting fusion, and a trend credibility function, thereby improving the accuracy and reliability of the state assessment. Furthermore, by quantifying trend credibility and establishing an early warning mechanism based on health levels, it provides clear quantitative basis for equipment maintenance and management, supports scientific decision-making, and optimizes equipment maintenance strategies.

[0052] On the other hand, embodiments of the present invention also provide a clutch condition diagnosis fault early warning system based on multi-source information fusion. Figure 8 A schematic diagram of a clutch condition diagnosis fault early warning system based on multi-source information fusion is shown. The clutch condition diagnosis fault early warning system based on multi-source information fusion specifically includes: The preprocessing module 100 collects multi-source state information of the clutch in real time based on the distributed sensing terminal, preprocesses the multi-source state information, and outputs a preprocessed information set. The model building module 200 is used to pre-build a deep learning-based state evaluation model. A multi-module supervision strategy is introduced into the state evaluation model. The pre-built state evaluation model is deployed based on the multi-module supervision strategy, historical state information is captured, and noise perturbation is added to the historical state information to obtain noise perturbation samples. The state evaluation model is iteratively trained based on the noise perturbation samples to output a converged state evaluation model. The comprehensive evaluation module 300 is used to acquire a preprocessed information set. The state evaluation model performs fusion processing on the preprocessed information set to obtain multimodal fusion features that characterize the real-time state of the clutch. Based on the multimodal fusion features and the control chart algorithm, the clutch state evaluation coefficients are determined and the state evaluation coefficient set is output. The state evaluation model performs time-series dynamic analysis on the clutch based on the trend discrimination method and outputs the comprehensive evaluation results.

[0053] In this embodiment of the invention, the comprehensive evaluation module 300 includes: The fusion coding unit 310 is used to acquire a preprocessed information set. The state evaluation model performs fusion processing on the preprocessed information set to obtain multimodal fusion features that characterize the real-time state of the clutch. The evaluation coefficient calculation unit 320 determines the clutch state evaluation coefficients based on multimodal fusion features and control chart algorithm, and outputs a set of state evaluation coefficients. The timing analysis unit 330 performs timing dynamic analysis on the clutch based on the trend discrimination method and outputs a comprehensive evaluation result.

[0054] It should be noted that each module in the clutch state diagnosis fault early warning system based on multi-source information fusion in the embodiments of the present invention corresponds to the clutch state diagnosis fault early warning method based on multi-source information fusion, and will not be described again here.

[0055] In summary, this invention provides a clutch condition diagnosis and fault early warning method and system based on multi-source information fusion. In the embodiments of this invention, multi-source condition information of the clutch is collected in real time by a distributed sensing terminal. The pre-processed information set is fused and processed in combination with a condition assessment model. The clutch is then subjected to time-series dynamic analysis based on trend discrimination. Multi-source heterogeneous data of various physical quantities, including but not limited to vibration, temperature, pressure, sound, and current, are collected, providing a data foundation for comprehensive evaluation. Through multi-source information fusion and deep feature extraction, the limitations of a single sensor are overcome. The clutch condition can be comprehensively evaluated from multiple dimensions, achieving accurate identification and location of different fault modes. This greatly reduces the false diagnosis and false negative rates, and can keenly capture early and subtle signs of clutch performance degradation, issuing early warnings before faults occur. This provides maintenance personnel with sufficient reaction time, transforming passive maintenance into proactive prevention, thereby avoiding major equipment failures, reducing downtime losses, and extending equipment lifespan.

[0056] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0057] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A clutch condition diagnosis and fault early warning method based on multi-source information fusion, characterized in that, The method includes: Based on the real-time acquisition of multi-source state information of the clutch by distributed sensing terminals, the multi-source state information is preprocessed and a preprocessed information set is output. A pre-built state evaluation model based on deep learning is introduced into the state evaluation model. A multi-module supervision strategy is introduced into the state evaluation model. The pre-built state evaluation model is deployed based on the multi-module supervision strategy, historical state information is captured, and noise perturbation is added to the historical state information to obtain noise perturbation samples. The state evaluation model is iteratively trained based on the noise perturbation samples to output a converged state evaluation model. A preprocessed information set is acquired, and the state evaluation model performs fusion processing on the preprocessed information set to obtain multimodal fusion features characterizing the real-time state of the clutch. Based on the multimodal fusion features and the control chart algorithm, the clutch state evaluation coefficients are determined, and a set of state evaluation coefficients is output. A set of loading state evaluation coefficients is used. The state evaluation model performs time-series dynamic analysis of the clutch based on the trend discrimination method and outputs a comprehensive evaluation result.

2. The clutch condition diagnosis and fault early warning method based on multi-source information fusion according to claim 1, characterized in that, The preprocessing of multi-source state information includes: With the system master clock as a constraint, a unified time base is allocated to the multi-source state information to obtain multi-source state information after time and space stamp alignment. It is determined whether the information sampling frequency is high-frequency sampling information, and the equalization fusion period is preset based on the type of multi-source state information. If the information sampling frequency is high-frequency sampling information, the multi-source state information is interpolated and resampled to obtain the multi-source state information after being resampled to a set fusion period. If the information sampling frequency is not high-frequency sampling information, the multi-source state information is filled in forward to obtain the multi-source state information after being filled to the set fusion period. The synchronized multi-source state information is loaded, and adaptive outlier detection is performed on the multi-source state information based on the threshold detection method to obtain the outlier-cleaned multi-source state information. After outlier cleaning, multi-source state information is obtained, and a hybrid filtering strategy is used to filter and denoise the multi-source state information, outputting a pre-processed information set after filtering and denoising.

3. The clutch condition diagnosis and fault early warning method based on multi-source information fusion according to claim 2, characterized in that, The preset equalization fusion period based on multi-source state information types includes: Acquire multi-source state information sources, classify multi-source state information into high-frequency dynamic type, medium-frequency slowly changing type and low-frequency command type based on the information source, and record the typical sampling period and maximum allowable delay of the information type; Traverse the typical sampling period and maximum allowable delay of information types, and determine the local processing period of multi-source state information based on the information type; The equalization fusion cycle constraint principle is preset. The equalization fusion cycle constraint principle is the minimum common cycle in which all information types participating in the fusion can complete a synchronous update within this cycle. The equalization fusion cycle is determined based on the equalization fusion cycle calculation formula.

4. The clutch condition diagnosis and fault early warning method based on multi-source information fusion according to claim 3, characterized in that, The adaptive outlier detection based on threshold detection method for multi-source state information includes: Obtain the information type classification results of multi-source state information, determine the initial static threshold of information type based on the information type classification results, and set the upper and lower limits of the standard deviation of the initial static threshold of information type. Within the local processing cycle of information type, the mean and standard deviation of multi-source state information are statistically analyzed, and an adaptive threshold for information type is determined based on the initial static threshold, the mean and standard deviation of multi-source state information; Load multi-source status information, determine whether the multi-source status information exceeds the adaptive threshold. If the multi-source status information exceeds the adaptive threshold, determine the multi-source status information as abnormal and mark it. If the multi-source status information does not exceed the adaptive threshold, retain the multi-source status information.

5. The clutch condition diagnosis and fault early warning method based on multi-source information fusion according to claim 2, characterized in that, The method of using a hybrid filtering strategy to filter and reduce noise from multi-source state information includes: The outlier-cleaned multi-source state information is loaded, and a low-pass Butterworth filter is used to perform sliding filtering on the outlier-cleaned multi-source state information to output the multi-source state information after filtering out and smoothing high-frequency noise. The process involves acquiring multi-source state information after filtering out and smoothing high-frequency noise, performing frequency domain analysis on the multi-source state information, identifying transient anomalies in the multi-source state information, and removing transient anomalies from the multi-source state information. Based on the equalization fusion cycle, multi-source state information is segmented and extracted. Then, a label field is added to the segmented multi-source state information features. The segmented multi-source state information with added label fields is then standardized by Z-score to unify the dimensionality, and a preprocessed information set is output.

6. The clutch condition diagnosis and fault early warning method based on multi-source information fusion according to claim 1, characterized in that, The state evaluation model is based on an LSTM network model and includes an input layer, an LSTM network model, and an output layer. A multimodal fusion layer is set between the input layer and the LSTM network model. The multimodal fusion layer is a graph neural network architecture, and a cross-modal attention mechanism is introduced into the graph neural network architecture. A dynamic analysis layer is introduced between the LSTM network model and the output layer.

7. The clutch condition diagnosis and fault early warning method based on multi-source information fusion according to claim 6, characterized in that, The determination of clutch state evaluation coefficients based on multimodal fusion features combined with control chart algorithm includes: Obtain the preprocessed information set, extract the information feature variables corresponding to the preprocessed information set based on the Transformer time encoder, and output the time-encoded feature variable set. The feature variable set is used as the node of the multimodal graph structure, and the relationship between the information feature variables in the feature variable set is used as the connection edge of the multimodal graph structure. The heterogeneous multimodal graph structure is updated based on the nodes and connection edges. Modal embedding encoding is performed on the nodes in the multimodal graph structure. The nodes in the multimodal graph structure are fused based on the multi-module supervision strategy, and the weighted multimodal fusion feature representing the real-time state of the clutch is output. Based on the LSTM network, the fusion feature sequence that represents the dynamic evolution trend of multimodal fusion features is captured. The standard dimension that is strongly correlated with the health of the clutch is selected as the monitoring variable, and the modulus of the fusion feature sequence is used as the comprehensive health index. Principal component dimensionality reduction is performed on the comprehensive health indicators to obtain the dimensionality reduction representation of the comprehensive monitoring indicators. The dimensionality reduction vector of the indicators is output. The dimensionality reduction vector of the indicators is then weighted and fused based on the standard dimension to output the multi-indicator fusion vector. Based on Shewhart control charts, multi-indicator fusion vectors are used to monitor the trends of mean and standard deviation, and a state evaluation function is constructed to determine the clutch state evaluation coefficients.

8. The clutch condition diagnosis and fault early warning method based on multi-source information fusion according to claim 7, characterized in that, The state assessment model performs time-series dynamic analysis of the clutch based on the trend discriminant method, including: Load the state evaluation coefficient set, and the dynamic analysis layer establishes a sliding window polynomial fitting trend model based on the state evaluation coefficients and multi-index fusion vectors. The sliding window polynomial fitting trend model captures the trend slope, trend significance, and trend direction of the multi-indicator fusion vector, and performs weighted voting fusion on the trend slope, trend significance, and trend direction of the multi-indicator fusion vector. The comprehensive confidence level of the multi-indicator fusion vector is calculated based on the trend confidence function, and the comprehensive confidence level is used as the comprehensive evaluation value. The health level of the clutch is determined based on the comprehensive evaluation value. Determine if the clutch health level exceeds the preset health threshold. If the clutch health level exceeds the preset health threshold, trigger a clutch warning command.

9. A clutch condition diagnosis fault early warning system based on multi-source information fusion, used to implement the clutch condition diagnosis fault early warning method based on multi-source information fusion according to any one of claims 1 to 8, characterized in that, The system includes: The preprocessing module collects multi-source state information of the clutch in real time based on distributed sensing terminals, preprocesses the multi-source state information, and outputs a preprocessed information set. The model building module is used to pre-build a deep learning-based state evaluation model. A multi-module supervision strategy is introduced into the state evaluation model. The pre-built state evaluation model is deployed based on the multi-module supervision strategy, historical state information is captured, and noise perturbation is added to the historical state information to obtain noise perturbation samples. The state evaluation model is iteratively trained based on the noise perturbation samples to output a converged state evaluation model. The comprehensive evaluation module is used to acquire a preprocessed information set. The state evaluation model performs fusion processing on the preprocessed information set to obtain multimodal fusion features that characterize the real-time state of the clutch. Based on the multimodal fusion features and the control chart algorithm, the clutch state evaluation coefficients are determined and the state evaluation coefficient set is output. The state evaluation model performs time-series dynamic analysis of the clutch based on the trend discrimination method and outputs the comprehensive evaluation results.

10. The clutch condition diagnosis and fault early warning system based on multi-source information fusion according to claim 9, characterized in that, The comprehensive evaluation module includes: The fusion coding unit is used to acquire the preprocessed information set. The state evaluation model performs fusion processing on the preprocessed information set to obtain multimodal fusion features that characterize the real-time state of the clutch. The evaluation coefficient calculation unit determines the clutch state evaluation coefficients based on multimodal fusion features and control chart algorithms, and outputs a set of state evaluation coefficients. The timing analysis unit performs timing dynamic analysis on the clutch based on the trend discrimination method and outputs a comprehensive evaluation result.

Images