Multi-sensor data fusion intelligent fault diagnosis system and method thereof

Abstract

The invention relates to the technical field of intelligent fault diagnosis systems, in particular to an intelligent fault diagnosis system and method for multi-sensor data fusion, and the system comprises an online diagnosis module and an offline training module, the online diagnosis module comprises a self-adaptive balance sub-module and a prototype network sub-module, and is used for processing sensor data in real time and carrying out fault diagnosis; the offline training module comprises a data preprocessing sub-module, a data fusion sub-module, a fault diagnosis sub-module, a data balance sub-module and a data interpretation sub-module, and is used for optimizing a diagnosis model; the data interpretation sub-module comprises an SHAP value extraction unit which is used for interpreting a diagnosis result; an online diagnosis module and an offline training module of the system form a closed-loop optimization mechanism, so that the diagnosis model can continuously learn and adapt to a new fault mode. The design ingeniously balances the requirement of real-time diagnosis and the requirement of continuous optimization of the model, and the long-term effectiveness of the system is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis systems, in particular to a multi-sensor data fusion intelligent fault diagnosis system and a method thereof. Background Art

[0002] With the advent of the Industrial 4.0 era, intelligent manufacturing and predictive maintenance have become important trends in the development of modern industry. In this context, the importance of equipment fault diagnosis technology has become increasingly prominent. Traditional fault diagnosis methods mainly rely on single sensor data and preset diagnostic rules. Although they can meet basic fault detection needs to a certain extent, their limitations are becoming increasingly obvious when faced with complex industrial environments and diverse fault types.

[0003] At present, the fault diagnosis methods commonly used in the industry are mostly based on a single data source such as vibration analysis or oil analysis. These methods perform well when dealing with simple fault types, but it is often difficult to accurately identify the root cause of the fault in complex systems. For example, in large rotating equipment, relying solely on vibration data may ignore important information such as temperature changes and pressure fluctuations, resulting in incomplete or inaccurate diagnostic results. In addition, these methods usually use fixed diagnostic thresholds, lack the ability to adapt to the dynamic working conditions of the equipment, and are prone to false positives or false negatives.

[0004] On the other hand, with the development of artificial intelligence technology, some researchers have begun to try to apply machine learning algorithms to the field of fault diagnosis. However, most of these methods are still limited to the analysis of a single data source and fail to fully utilize the complementarity of multi-source sensor data. At the same time, existing intelligent diagnosis systems often use "black box" models, and the diagnosis results lack interpretability, making it difficult to gain the trust and adoption of engineers.

[0005] In addition, the working environment and failure modes of industrial equipment are constantly changing, and most existing diagnostic systems lack adaptive learning capabilities and cannot effectively respond to new types of faults. This causes the system's diagnostic performance to gradually degrade over time, requiring frequent manual intervention and updates.

[0006] In the face of these challenges, there is an urgent need for an intelligent fault diagnosis system that can comprehensively utilize multi-source sensor data, has adaptive learning capabilities, and ensures the interpretability of diagnostic results. Summary of the invention

[0007] In the face of these challenges, the present invention proposes an intelligent fault diagnosis system and method based on multi-sensor data fusion. Through innovative system architecture design and advanced algorithm application, the present invention aims to solve key problems existing in the prior art, such as insufficient data utilization, low diagnostic accuracy, poor interpretability, and insufficient adaptability.

[0008] The present invention proposes an intelligent fault diagnosis system of multi-sensor data fusion, comprising:

[0009] Online diagnosis module and offline training module, including:

[0010] The online diagnosis module includes an adaptive balancing submodule and a prototype network submodule, which are used to process sensor data in real time and perform fault diagnosis;

[0011] The offline training module includes a data preprocessing submodule, a data fusion submodule, a fault diagnosis submodule, a data balancing submodule and a data interpretability submodule, which are used to optimize the diagnosis model;

[0012] The adaptive balancing submodule includes an attention unit and a data synthesis unit for dynamically adjusting data weights and generating balanced data;

[0013] The data preprocessing submodule includes a data acquisition unit and a data cleaning unit, which are used to acquire and preprocess multi-source sensor data;

[0014] The data fusion submodule includes a data screening unit and a feature extraction unit, which are used to screen valid data and extract features;

[0015] The fault diagnosis submodule includes an active learning strategy unit, which further includes a data annotation unit, a data prediction unit and an uncertainty assessment unit for continuously optimizing the diagnosis model;

[0016] The data interpretation submodule includes a SHAP value extraction unit for explaining the diagnosis results;

[0017] The online diagnosis module and the offline training module form a closed loop through data interaction to achieve real-time diagnosis and continuous optimization.

[0018] A multi-sensor data fusion fault diagnosis method comprises the following steps: (a) collecting and preprocessing multi-source sensor data through a data preprocessing submodule; (b) extracting and fusing features of the preprocessed data using a data fusion submodule; (c) performing diagnostic learning on the fused feature data using a fault diagnosis submodule; (d) interpreting the diagnosis result through a data interpretation submodule, and calculating the uncertainty using an uncertainty evaluation unit; (e) collecting and annotating new data based on the uncertainty evaluation result, and balancing the data set using a data balancing submodule; (f) updating the diagnosis model and adjusting the labels of the data set samples; (g) repeating steps (c) to (f) until a preset iteration stop condition is met; and (h) deploying the optimized model to an online diagnosis module to achieve real-time fault diagnosis.

[0019] Preferably, the feature extraction process in step (b) adopts an improved VGG-19 convolutional neural network, which obtains richer fault feature information by increasing the network depth. The specific steps are as follows: first, the preprocessed numerical data is input into the VGG-19 network; then, the deep features are extracted through the added convolutional layer and pooling layer; finally, the feature vector H with a dimension of 2048×1 is output, where H=[h1,h2,…,h 2048 ] T ,h i represents the i-th feature component.

[0020] Preferably, the data fusion process in step (b) adopts a weighted fusion algorithm assisted by an attention mechanism, and the specific steps are as follows: First, for the two feature vectors H1 and H2, calculate the fusion feature F: F = σ(W1H1+W2H2+b), where H i represents the i-th eigenvector, W i represents the weight matrix of the i-th feature vector, b represents the bias vector, and σ represents the activation function; then, the attention mechanism is introduced to calculate the attention weight: α i =softmax(v T tanh(WH i +b)),

[0021] Among them, v and W are learnable parameters, H i represents the i-th eigenvector; finally, calculate the fused eigenvector F′: F′=∑(α i H i ), where α i Represents the attention weight of the i-th feature vector; the fused feature vector F′ is input into the fully connected layer to obtain the final fused feature vector.

[0022] Preferably, the diagnostic learning process in step (c) uses a support vector machine combined with a prototype network for transfer learning, and the specific steps are as follows:

[0023] First, use multi-source datasets to train the prototype network to obtain prior knowledge;

[0024] Then, the source domain samples are assigned to their respective prototypes, and the target domain samples are assigned prior probabilities;

[0025] Next, the weighted prototype of the target domain is calculated to define the similarity between the source domain and the target domain:

[0026]

[0027] Among them, x t represents the tth sample in the target domain, x srepresents the sth sample in the source domain, σ represents the Gaussian kernel width parameter, then, set the similarity threshold θ, and perform similarity screening on the data set; calculate the similarity mean μ and standard deviation σ of the data set D:

[0028]

[0029] Among them, N represents the number of samples in the data set; finally, the data adaptive balance weight is calculated using the cosine similarity distance metric:

[0030]

[0031] Among them, cos(x t ,x s ) represents x t With x s The cosine similarity of .

[0032] Preferably, the uncertainty assessment process in step (d) adopts a method based on information entropy, and the specific steps are as follows:

[0033] If H(y|x)>θ1, the sample is marked as the uncertain set U1; if H(y|x)<θ2, the sample is marked as the uncertain set U2; where H(y|x) represents the conditional entropy of the output y given the input x, and θ1 and θ2 represent the preset thresholds; then, for the uncertain sets U1 and U2, their confidence is calculated Among them, K represents the number of categories; then, set the confidence threshold θ c , the confidence level is lower than θ c The data is marked as data to be labeled; finally, the confidence level is recalculated for the inaccurate diagnosis results. If the confidence level is lower than θ, c , it is marked as data to be labeled.

[0034] Preferably, the data balancing process in step (e) adopts an improved conditional generative adversarial network (CGAN), and the specific steps are as follows:

[0035] First, construct CGAN, including generator G and discriminator D, and its loss function is:

[0036] L CGAN =E x,y [logD(x,y)]+E x,z [log(1-D(G(z,y),y))]+λE x,z,y [∥xG(z,y)∥1]

[0037] Among them, x represents real data, y represents the label, z represents random noise, λ represents the L1 regularization coefficient, and G(z,y) represents the data generated by the generator;

[0038] Then, train the generator G and the discriminator D, and the optimization objective of the discriminator is:

[0039] maxE x,y [logD(x,y)]+E x,z [log(1-D(G(z,y),y))]

[0040] Finally, when the discriminator D cannot distinguish between the generated data and the original data, CGAN reaches equilibrium and completes the data balancing process.

[0041] Preferably, the data interpretation process in step (d) adopts an improved SHAP value calculation method, and the specific steps are as follows:

[0042] First, define the eigenvalue vector a and the sample vector x, and calculate the feature contribution:

[0043]

[0044] Where N represents the set of all features, S represents the subset that does not contain feature i, and v represents the feature value function;

[0045] Then, we calculate the mean and standard deviation of the eigenvalue vector contribution and the explanatory power of the eigenvector in the model:

[0046]

[0047] in, represents the expected SHAP value of feature i, represents its standard deviation;

[0048] Next, we introduce the feature importance weight w i , calculate the weighted SHAP value: WSHAP(a i )=w i ·SH AP (a i ), where w i It is determined by feature selection algorithms such as Lasso; finally, the contribution ratio of each feature is evaluated according to the WSHAP value.

[0049] Preferably, the active learning strategy in step (e) includes the following process:

[0050] First, calculate the data uncertainty U(x):

[0051] U(x)=-∑p(y|x)logp(y|x)

[0052] Among them, p(y|x) represents the conditional probability of output y given input x;

[0053] Then, calculate the sample representativeness R(x):

[0054]

[0055] Among them, sin(x,x i ) represents the difference between sample x and other samples x in the dataset i Similarity, N represents the size of the data set;

[0056] Next, considering uncertainty and representativeness, select the optimal sample x * :

[0057] x * =argmax(λU(x)+(1-λ)R(x))

[0058] Among them, λ represents the trade-off parameter;

[0059] Finally, for the selected sample x * Annotate them and add them to the training dataset.

[0060] Preferably, an unknown fault diagnosis function is also included, and its working process is as follows:

[0061] First, the new sensor data is fed into the diagnostic system.

[0062] Then, use One-Class SVM to perform anomaly detection to determine whether it is an unknown fault:

[0063] f(x)=sgn((w·φ(x))-ρ)

[0064] Among them, w represents the normal vector of the decision hyperplane, φ(x) represents the feature mapping function, and ρ represents the bias term;

[0065] Finally, if an unknown fault is detected, the incremental learning module is started and the model is updated online using a progressive network to adapt to the new fault type.

[0066] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0067] The multi-sensor data fusion intelligent fault diagnosis system of the present invention effectively solves the above technical problems through modular design and algorithm innovation. First, the system's online diagnosis module and offline training module form a closed-loop optimization mechanism, which enables the diagnosis model to continuously learn and adapt to new fault modes. This design cleverly balances the needs of real-time diagnosis and the need for continuous model optimization, greatly improving the long-term effectiveness of the system.

[0068] Secondly, the present invention uses an improved VGG-19 convolutional neural network for feature extraction and introduces a weighted fusion algorithm assisted by an attention mechanism to achieve effective integration of multi-source heterogeneous data. This method not only makes full use of the complementary information of various sensor data, but also dynamically adjusts the importance of various features according to different working conditions, significantly improving the accuracy and robustness of diagnosis.

[0069] Furthermore, the present invention innovatively combines the prototype network with the support vector machine to achieve knowledge transfer learning. This enables the system to quickly adapt to new equipment types or working conditions, greatly reducing the reliance on a large amount of labeled data and improving the generalization ability of the system. At the same time, by introducing SHAP value calculation, the system can provide a clear explanation for each diagnostic result, enhancing the credibility and operability of the diagnostic result.

[0070] In addition, the active learning strategy and data balancing module of the present invention work together to not only select the most informative samples for labeling in a targeted manner, but also effectively solve the common data imbalance problem in industrial sites. This design greatly improves the learning efficiency of the system and ensures the effective diagnosis of various faults, especially rare faults.

[0071] Finally, the unknown fault diagnosis function of the present invention combines One-Class SVM with incremental learning to provide the system with the ability to deal with new fault types. The addition of this function makes the system truly adaptive and can maintain efficient diagnostic performance in complex and changing industrial environments.

[0072] In summary, the present invention realizes the synergy between modules and the complementary enhancement between algorithms through the organic combination of multiple innovative points. This holistic optimization design not only greatly improves the accuracy, timeliness and reliability of fault diagnosis, but also enables the system to have the ability of continuous learning and self-improvement. This is of great significance for improving the reliability of industrial equipment, reducing maintenance costs, and improving production efficiency, and provides strong technical support for industrial intelligent upgrading. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is the overall system logic block diagram of the present invention.

[0074] Figure 2 It is the internal logic block diagram of the online diagnosis module of the present invention.

[0075] Figure 3 This is an internal logic block diagram of the offline training module of the present invention.

[0076] Figure 4 It is a logic block diagram of the method of the present invention.

[0077] Figure 5 It is a logic block diagram of the unknown fault diagnosis function of the present invention. DETAILED DESCRIPTION

[0078] In order to further explain the technical means and effects of the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments thereof. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics of one or more embodiments may be combined in any suitable form.

[0079] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0080] Example 1

[0081] See also Figure 1-5 The present invention relates to an intelligent fault diagnosis system of multi-sensor data fusion, which includes an online diagnosis module 1 and an offline training module 2. The online diagnosis module 1 includes an adaptive balance submodule 11 and a prototype network submodule 12, which are used to process sensor data in real time and perform fault diagnosis. The offline training module 2 includes a data preprocessing submodule 21, a data fusion submodule 22, a fault diagnosis submodule 23, a data balance submodule 24 and a data interpretability submodule 25, which are used to optimize the diagnosis model.

[0082] The adaptive balancing submodule 11 further includes an attention unit 111 and a data synthesis unit 112 for dynamically adjusting data weights and generating balanced data. The data preprocessing submodule 21 includes a data acquisition unit 211 and a data cleaning unit 212 for acquiring and preprocessing multi-source sensor data. The data fusion submodule 22 includes a data screening unit 221 and a feature extraction unit 222 for screening valid data and extracting features. The fault diagnosis submodule 23 includes an active learning strategy unit 231, which further includes a data annotation unit 2311, a data prediction unit 2312 and an uncertainty assessment unit 2313 for continuously optimizing the diagnosis model. The data interpretability submodule 25 includes a SHAP value extraction unit 251 for explaining the diagnosis results.

[0083] In one embodiment of the present invention, the system can be applied to fault diagnosis of large industrial equipment. For example, in a complex chemical production line, the system can simultaneously process data from reactor temperature sensors, pressure sensors, flow meters, and vibration sensors. Through real-time processing by the online diagnosis module 1, the system can quickly detect abnormal conditions of the equipment, such as abnormal temperature fluctuations in the reactor or excessive vibration of the pump body. At the same time, the offline training module 2 continuously optimizes the diagnosis model so that the system can adapt to changes in production processes and new types of faults.

[0084] The present invention also discloses intelligent fault diagnosis based on multi-sensor data fusion, which comprises the following steps:

[0085] (a) collecting and preprocessing multi-source sensor data through the data preprocessing submodule 21;

[0086] (b) extracting and fusing features of the preprocessed data using the data fusion submodule 22;

[0087] (c) using the fault diagnosis submodule 23 to perform diagnostic learning on the fused feature data;

[0088] (d) interpreting the diagnosis results through the data interpretation submodule 25 and calculating the uncertainty using the uncertainty assessment unit 2313;

[0089] (e) based on the uncertainty assessment results, new data is collected and annotated, and the data set is balanced using the data balancing submodule 24;

[0090] (f) Update the diagnostic model and adjust the labels of the dataset samples;

[0091] (g) Repeating steps (c) to (f) until a preset iteration stop condition is met;

[0092] (h) Deploy the optimized model to the online diagnosis module 1 to realize real-time fault diagnosis.

[0093] The advantage of this approach is that it forms a closed-loop optimization process. For example, in a wind farm application, the system first collects various sensor data from the wind turbine, such as wind speed, wind direction, generator speed, gearbox temperature, etc. Then, through data fusion and feature extraction, the system can comprehensively analyze this data and identify potential failure modes. If the system is uncertain about some emerging data patterns, it will actively ask the operation and maintenance personnel to mark them, thereby continuously improving its diagnostic capabilities.

[0094] In a specific embodiment of the present invention, the feature extraction process in step (b) adopts an improved VG G-19 convolutional neural network, which obtains richer fault feature information by increasing the network depth. The specific steps are as follows: first, the preprocessed numerical data is input into the VGG-19 network; then, the deep features are extracted through the added convolutional layer and pooling layer; finally, the feature vector H with a dimension of 2048×1 is output, where H=[h1,h2,…,h 2048 ] T ,h i represents the i-th feature component.

[0095] In practical applications, this step is crucial. For example, in a steel plant's blast furnace monitoring system, the raw data may contain a lot of noise and outliers. By setting reasonable thresholds (such as considering data exceeding 3 standard deviations of the historical average as outliers) and using appropriate filters (such as median filters), the system can effectively remove these interferences and provide a reliable data basis for subsequent analysis.

[0096] In a specific embodiment of the present invention, the data fusion process in step (b) adopts a weighted fusion algorithm assisted by an attention mechanism, and the specific steps are as follows: First, for two feature vectors H1 and H2, calculate the fusion feature F: F = σ(W1H1+W2H2+b), where H i represents the i-th eigenvector, W i represents the weight matrix of the i-th feature vector, b represents the bias vector, and σ represents the activation function; then, the attention mechanism is introduced to calculate the attention weight: α i =softmax(v T tanh(WH i +b)), where v and W are learnable parameters, H i represents the i-th eigenvector; finally, calculate the fused eigenvector F′: F′=∑(α i H i ), where α i Represents the attention weight of the i-th feature vector; the fused feature vector F′ is input into the fully connected layer to obtain the final fused feature vector.

[0097] In a specific embodiment of the present invention, the diagnostic learning process in step (c) uses a support vector machine combined with a prototype network for transfer learning, and the specific steps are as follows:

[0098] First, use multi-source datasets to train the prototype network to obtain prior knowledge;

[0099] Then, the source domain samples are assigned to their respective prototypes, and the target domain samples are assigned prior probabilities;

[0100] Next, the weighted prototype of the target domain is calculated to define the similarity between the source domain and the target domain:

[0101]

[0102] Among them, x t represents the tth sample in the target domain, x s represents the sth sample in the source domain, σ represents the Gaussian kernel width parameter, then, set the similarity threshold θ, and perform similarity screening on the data set; calculate the similarity mean μ and standard deviation σ of the data set D:

[0103]

[0104] Among them, N represents the number of samples in the data set; finally, the data adaptive balance weight is calculated using the cosine similarity distance metric:

[0105]

[0106] Among them, cos(x t ,x s ) represents x t With x s The cosine similarity of .

[0107] The present invention adopts an improved VGG-19 convolutional neural network for feature extraction, and introduces a weighted fusion algorithm assisted by an attention mechanism. In a practical application scenario, such as fault diagnosis of an oil drilling platform, the system may need to process acoustic signals, pressure signals, and vibration signals at the same time. Through a deep convolutional network, the system is able to extract high-level features from these heterogeneous data. For example, a specific frequency pattern is identified from an acoustic signal, which may be related to drill bit wear. The attention mechanism then dynamically adjusts the weights of different features based on the current diagnostic task. If the system detects an abnormal vibration pattern, it automatically increases the weight of the vibration-related features, thereby improving the accuracy of the diagnosis.

[0108] In a specific embodiment of the present invention, the uncertainty assessment process in step (d) adopts a method based on information entropy, and the specific steps are as follows:

[0109] If H(y|x)>θ1, the sample is marked as the uncertain set U1; if H(y|x)<θ2, the sample is marked as the uncertain set U2; where H(y|x) represents the conditional entropy of the output y given the input x, and θ1 and θ2 represent the preset thresholds; then, for the uncertain sets U1 and U2, their confidence is calculated Among them, K represents the number of categories; then, set the confidence threshold θ c , the confidence level is lower than θ cThe data is marked as data to be labeled; finally, the confidence level is recalculated for the inaccurate diagnosis results. If the confidence level is lower than θ, c , it is marked as data to be labeled.

[0110] In a specific embodiment of the present invention, the data balancing process in step (e) adopts an improved conditional generative adversarial network (CGAN), and the specific steps are as follows:

[0111] First, construct CGAN, including generator G and discriminator D, and its loss function is:

[0112] L CGAN =E x,y [logD(x,y)]+E x,z [log(1-D(G(z,y),y))]+λE x,z,y [∥xG(z,y)∥1]

[0113] Among them, x represents real data, y represents the label, z represents random noise, λ represents the L1 regularization coefficient, and G(z,y) represents the data generated by the generator;

[0114] Then, train the generator G and the discriminator D, and the optimization objective of the discriminator is:

[0115] maxE x,y [logD(x,y)]+E x,z [log(1-D(G(z,y),y))]

[0116] Finally, when the discriminator D cannot distinguish between the generated data and the original data, CGAN reaches equilibrium and completes the data balancing process.

[0117] In a specific embodiment of the present invention, the diagnostic learning process includes the application of support vector machines and prototype networks. An important advantage of this method is its transfer learning capability. For example, in the fault diagnosis of the cooling system of a large data center, the system may have learned certain types of refrigerator failure modes. When a new model of refrigerator needs to be diagnosed, even without a large amount of training data, the system can use the existing knowledge to quickly build an effective diagnostic model. Specifically, the system calculates the similarity between the new device data and the known failure mode. The similarity is calculated using a Gaussian kernel function, and the kernel width parameter σ is usually set to the square root of the feature space dimension. In this way, existing knowledge can be fully utilized while maintaining good generalization ability.

[0118] In a specific embodiment of the present invention, the data interpretation process in step (d) adopts an improved SH AP value calculation method, and the specific steps are as follows:

[0119] First, define the eigenvalue vector a and sample vector x, and calculate the feature contribution:

[0120]

[0121] Where N represents the set of all features, S represents the subset that does not contain feature i, and v represents the feature value function;

[0122] Then, we calculate the mean and standard deviation of the eigenvalue vector contribution and the explanatory power of the eigenvector in the model:

[0123]

[0124] in, represents the expected SHAP value of feature i, represents its standard deviation;

[0125] Next, we introduce the feature importance weight w i , calculate the weighted SHAP value: WSHAP(a i )=w i ·SH AP (a i ), where w i It is determined by feature selection algorithms such as Lasso; finally, the contribution ratio of each feature is evaluated according to the WSHAP value.

[0126] In a specific embodiment of the present invention, the active learning strategy in step (e) includes the following process:

[0127] First, calculate the data uncertainty U(x):

[0128] U(x)=-∑p(y|x)logp(y|x)

[0129] Where p(y|x) represents the conditional probability of output y given input x;

[0130] Then, calculate the sample representativeness R(x):

[0131]

[0132] Among them, sin(x,x i ) represents the difference between sample x and other samples x in the dataset i Similarity, N represents the size of the data set;

[0133] Next, considering uncertainty and representativeness, select the optimal sample x * :

[0134] x * =argmax(λU(x)+(1-λ)R(x))

[0135] Among them, λ represents the trade-off parameter;

[0136] Finally, for the selected sample x * Annotate them and add them to the training dataset.

[0137] In a specific embodiment of the present invention, uncertainty assessment and active learning strategies are extremely important in practical applications, especially when facing new equipment or rare failures. For example, in an intelligent manufacturing workshop, when a new production line is put into use, the system may encounter a large number of unseen data patterns. By calculating the information entropy and confidence of the data, the system can identify the samples that most need human intervention. Specifically, we can set the high uncertainty threshold θ1 to 0.8 and the low uncertainty threshold θ2 to 0.2. These thresholds are empirical values ​​based on a large number of experiments, which can achieve a good balance between maintaining system autonomy and the need for human intervention.

[0138] In a specific embodiment of the present invention, an unknown fault diagnosis function is also included, and its working process is as follows:

[0139] First, the new sensor data is input into the diagnostic system. Then, One-Class SVM is used to detect anomalies and determine whether it is an unknown fault:

[0140] f(x)=sgn((w·φ(x))-ρ)

[0141] Among them, w represents the normal vector of the decision hyperplane, φ(x) represents the feature mapping function, and ρ represents the bias term;

[0142] Finally, if an unknown fault is detected, the incremental learning module is started and the model is updated online using a progressive network to adapt to the new fault type.

[0143] This feature is particularly important in practical applications because new and unknown types of faults always appear in industrial environments. For example, in a large chemical plant, slight changes in the composition of raw materials may cause some unprecedented equipment anomalies. The system performs anomaly detection through One-Class SVM and can identify these deviations from normal working conditions. Once an unknown fault is detected, the system immediately starts the incremental learning module to quickly adapt to the new fault mode. This design allows the system to continue to evolve and always maintain efficient diagnostic capabilities.

[0144] In general, the multi-sensor data fusion intelligent fault diagnosis system and method proposed in the present invention realizes high-precision, high-reliability, explainable, and sustainable optimization fault diagnosis through innovative system architecture and advanced algorithm design. It can not only detect potential faults in a timely manner, but also continuously learn and adapt to new fault modes, providing a powerful tool for predictive maintenance of industrial equipment. This system has significant advantages in improving equipment reliability, reducing unexpected downtime, and reducing maintenance costs, and is of great significance to improving the overall operational efficiency of modern industry.

[0145] In order to verify the superiority of the multi-sensor data fusion intelligent fault diagnosis system of the present invention, we selected a typical industrial scenario for experiment. This experiment takes the catalytic cracking unit of a large petrochemical plant as the research object, which contains multiple key equipment, such as reactors, regenerators, air compressors, etc. We selected the main fan in the device as the specific research target because it is the key equipment of the entire system, and its failure will directly affect the operating efficiency and safety of the entire device.

[0146] Embodiment 1: Multi-sensor data fusion intelligent fault diagnosis system using the present invention

[0147] In this embodiment, we deployed the system of the present invention to simultaneously collect and analyze multiple sensor data of the main fan, including vibration, temperature, pressure, current, etc. The system adopts the innovative architecture and algorithm described above, including the improved VGG-19 network for feature extraction, the weighted fusion algorithm assisted by the attention mechanism, and the transfer learning method based on the prototype network.

[0148] Comparative Example 1: Traditional single sensor threshold-based fault diagnosis method

[0149] As a comparison, we chose the fault diagnosis method based on vibration monitoring that is commonly used in industry. This method mainly relies on the preset vibration threshold to determine whether the equipment is faulty, does not consider the data of other sensors, and does not have adaptive learning capabilities.

[0150] In order to comprehensively evaluate the performance of the two methods, we designed the following key indicators:

[0151] 1. Fault detection accuracy: the proportion of faults correctly identified.

[0152] 2. False alarm rate: the proportion of normal status being misjudged as a fault.

[0153] 3. Fault warning time: How long before the actual fault occurs can a warning be issued.

[0154] 4. Unknown fault recognition rate: the proportion of new faults that can be correctly identified.

[0155] 5. Diagnostic model update time: the time required for the model to adapt to new fault types.

[0156] We conducted a six-month follow-up test on these two methods, and the results are shown in the following table:

[0157] index Embodiment 1 (present invention) Comparative Example 1 (Traditional Method) Fault detection accuracy 97.8％ 85.3％ False Positive Rate 2.1％ 8.7％ Fault warning time Average 72 hours Average 24 hours Unknown fault recognition rate 89.5％ not applicable Diagnostic model update time Average 2.5 days not applicable

[0158] From the test results, it can be seen that the method of the present invention is significantly better than the traditional method in all indicators. Let us analyze these results in depth:

[0159] First, in terms of fault detection accuracy, the method of the present invention reached 97.8%, which is much higher than the 85.3% of the traditional method. This is mainly due to the application of multi-sensor data fusion and deep learning algorithms. For example, in the detection of a bearing fault, the vibration signal may not have reached the alarm threshold of the traditional method, but our system successfully identified the fault signs in advance by comprehensively analyzing multi-dimensional information such as temperature rise and current fluctuation.

[0160] Secondly, the method of the present invention significantly reduced the false alarm rate from 8.7% to 2.1%. This improvement is particularly important for industrial production, because frequent false alarms will lead to unnecessary downtime inspections, seriously affecting production efficiency. Our system effectively filters out possible misjudgments caused by a single sensor by cross-validating multiple sensor data. For example, in a false alarm caused by an external vibration source, our system correctly determined that there was no abnormality in the equipment itself by analyzing the current and pressure data.

[0161] In terms of fault warning time, the method of the present invention can issue a warning 72 hours in advance on average, which is three times that of traditional methods. This reserves sufficient preparation time for maintenance personnel and helps to transform passive maintenance into active prevention. In an actual case, the system predicted the possibility of bearing failure at an early stage that could not be detected by traditional methods by analyzing the slight but continuous upward trend of bearing temperature and combining it with changes in the vibration spectrum.

[0162] Particularly noteworthy is the unknown fault recognition rate. Traditional methods rely entirely on preset rules and are helpless against new faults. However, the method of the present invention successfully identified 89.5% of new faults through innovative anomaly detection algorithms and incremental learning capabilities. For example, during the test, an abnormal vibration pattern caused by a new lubricant appeared. Although this pattern had never been encountered before, our system still successfully captured this anomaly and quickly incorporated it into the diagnostic model.

[0163] Finally, the diagnostic model update time reflects the adaptive ability of the system. The method of the present invention takes an average of only 2.5 days to integrate newly discovered fault modes into the diagnostic model, while the traditional method requires manual analysis and rule updates, which usually takes weeks or even months. This rapid learning ability ensures that the system can continue to adapt to the dynamic changes of the equipment and the emergence of new fault types.

[0164] In summary, the multi-sensor data fusion intelligent fault diagnosis system of the present invention has shown significant advantages in terms of accuracy, timeliness and adaptability of fault detection. These improvements can not only greatly improve the reliability and production efficiency of equipment, but also reduce maintenance costs, providing strong technical support for the intelligent and refined management of industrial equipment. Especially in the face of complex working conditions and new faults, the strong adaptability and learning ability demonstrated by this system has laid a solid foundation for the realization of Industry 4.0 and intelligent manufacturing.

[0165] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. Multi-sensor data fusion fault diagnosis system, characterized in that: include: Online diagnosis module and offline training module, including: The online diagnosis module includes an adaptive balancing submodule and a prototype network submodule, which are used to process sensor data in real time and perform fault diagnosis; The offline training module includes a data preprocessing submodule, a data fusion submodule, a fault diagnosis submodule, a data balancing submodule and a data interpretability submodule, which are used to optimize the diagnosis model; The adaptive balancing submodule includes an attention unit and a data synthesis unit for dynamically adjusting data weights and generating balanced data; The data preprocessing submodule includes a data acquisition unit and a data cleaning unit, which are used to acquire and preprocess multi-source sensor data; The data fusion submodule includes a data screening unit and a feature extraction unit, which are used to screen valid data and extract features; The fault diagnosis submodule includes an active learning strategy unit, which further includes a data annotation unit, a data prediction unit and an uncertainty assessment unit for continuously optimizing the diagnosis model; The data interpretation submodule includes a SHAP value extraction unit for explaining the diagnosis results; The online diagnosis module and the offline training module form a closed loop through data interaction to achieve real-time diagnosis and continuous optimization.

2. The multi-sensor data fusion fault diagnosis method of the system according to claim 1 is characterized in that: The following steps are involved: (a) Collect and preprocess multi-source sensor data through the data preprocessing submodule; (b) Using the data fusion submodule to extract and fuse the preprocessed data; (c) Using the fault diagnosis submodule to perform diagnostic learning on the fused feature data; (d) Explain the diagnosis results through the data interpretation submodule and calculate the uncertainty using the uncertainty assessment unit; (e) Based on the uncertainty assessment results, collect new data and annotate them, and use the data balancing submodule to balance the data set; (f) Update the diagnosis model and adjust the labels of the data set samples; (g) Repeat steps (c) to (f) until the preset iteration stop condition is met; (h) Deploy the optimized model to the online diagnosis module to realize real-time fault diagnosis.

3. The method according to claim 2, characterized in that The feature extraction process in step (b) adopts an improved VGG-19 convolutional neural network, which obtains richer fault feature information by increasing the network depth. The specific steps are as follows: first, the preprocessed numerical data is input into the VGG-19 network; then, the deep features are extracted through the added convolutional layer and pooling layer; finally, the feature vector H with a dimension of 2048×1 is output, where H=[h1,h2,...,h 2048 ] T ,h i represents the i-th feature component.

4. The method according to claim 2, characterized in that: The data fusion process in step (b) adopts a weighted fusion algorithm assisted by an attention mechanism. The specific steps are as follows: First, for the two feature vectors H1 and H2, calculate the fusion feature F: F = σ(W1H1+W2H2+b), where H i represents the i-th eigenvector, W i represents the weight matrix of the i-th feature vector, b represents the bias vector, and σ represents the activation function; then, the attention mechanism is introduced to calculate the attention weight: α i =softmax(v T tanh(WH i +b)), where v and W are learnable parameters, H i represents the i-th eigenvector; finally, calculate the fused eigenvector F′: F′=∑(α i H i ), where α i Represents the attention weight of the i-th feature vector; the fused feature vector F′ is input into the fully connected layer to obtain the final fused feature vector.

5. The method according to claim 2, characterized in that: The diagnostic learning process in step (c) uses support vector machine combined with prototype network for transfer learning. The specific steps are as follows: First, use multi-source datasets to train the prototype network to obtain prior knowledge; Then, the source domain samples are assigned to their respective prototypes, and the target domain samples are assigned prior probabilities; Next, the weighted prototype of the target domain is calculated to define the similarity between the source domain and the target domain: Among them, x t represents the tth sample in the target domain, x s represents the sth sample in the source domain, σ represents the Gaussian kernel width parameter, then, set the similarity threshold θ to screen the data set for similarity; calculate the similarity mean μ and standard deviation σ of the data set D: Among them, N represents the number of samples in the data set; finally, the data adaptive balance weight is calculated using the cosine similarity distance metric: Among them, cos(x t , x s ) represents x t With x s The cosine similarity of .

6. The method according to claim 2, characterized in that The uncertainty assessment process in step (d) adopts a method based on information entropy. The specific steps are as follows: If H(y|x)>θ1, the sample is marked as the uncertain set U1; if H(y|x)<θ2, the sample is marked as the uncertain set U2; where H(y|x) represents the conditional entropy of the output y given the input x, and θ1 and θ2 represent the preset thresholds; then, for the uncertain sets U1 and U2, calculate their confidence C(x): Among them, K represents the number of categories; then, set the confidence threshold θ c , the confidence level is lower than θ c The data is marked as data to be labeled; finally, the confidence level is recalculated for the inaccurate diagnosis results. If the confidence level is lower than θ, c , it is marked as data to be labeled.

7. The method according to claim 2, characterized in that The data balancing process in step (e) adopts the improved conditional generative adversarial network CGAN, and the specific steps are as follows: First, construct CGAN, including generator G and discriminator D, and its loss function is: L CGAN =E x，y [logD(x,y)]+E x，z [log(1-D(G(z,y),y))]+λE x，z，y [||x-G(z,y)||1] Among them, x represents real data, y represents the label, z represents random noise, λ represents the L1 regularization coefficient, and G(z,y) represents the data generated by the generator; Then, train the generator G and the discriminator D, and the optimization objective of the discriminator is: maxE x，y [logD(x,y)]+E x，z [log(1-D(G(z,y),y))] Finally, when the discriminator D cannot distinguish between the generated data and the original data, CGAN reaches equilibrium and completes the data balancing process.

8. The method according to claim 2, characterized in that: The data interpretation process in step (d) adopts the improved SHAP value calculation method, and the specific steps are as follows: First, define the eigenvalue vector a and the sample vector x, and calculate the feature contribution: Where N represents the set of all features, S represents the subset that does not contain feature i, and v represents the feature value function; Then, we calculate the mean and standard deviation of the eigenvalue vector contribution and the explanatory power of the eigenvector in the model: in, represents the expected SHAP value of feature i, represents its standard deviation; Next, we introduce the feature importance weight w i , calculate the weighted SHAP value: WSHAP(a i )=w i SHAP(a i ), where w i It is determined by feature selection algorithms such as Lasso; finally, the contribution ratio of each feature is evaluated according to the WSHAP value.

9. The method according to claim 2, characterized in that: The active learning strategy in step (e) includes the following process: First, calculate the data uncertainty U(x): U(x)=-∑p(y|x)logp(y|x) Where p(y|x) represents the conditional probability of output y given input x; Then, calculate the sample representativeness R(x): Among them, sin(x,x i ) represents the difference between sample x and other samples x in the dataset i Similarity, N represents the size of the data set; Next, considering uncertainty and representativeness, select the optimal sample x * : x*=argmax(λU(x)+(1-λ)R(x)) Among them, λ represents the trade-off parameter; Finally, the selected samples x* are labeled and added to the training dataset.

10. The system according to claim 2, characterized in that It also includes unknown fault diagnosis function, and its working process is as follows: First, the new sensor data is fed into the diagnostic system. Then, use One-Class SVM to perform anomaly detection to determine whether it is an unknown fault: f(x)=sgn((w·φ(x))-ρ) Among them, w represents the normal vector of the decision hyperplane, φ(x) represents the feature mapping function, and ρ represents the bias term; Finally, if an unknown fault is detected, the incremental learning module is started and the model is updated online using a progressive network to adapt to the new fault type.

Images