Transmission system cross-sensor collaborative diagnosis method based on dynamic competition force balance

Abstract

The application discloses a drive system cross-sensor collaborative diagnosis method based on dynamic competition force balance, which comprises the following steps: obtaining and preprocessing monitoring data of sensors in a drive system; inputting the preprocessed monitoring data as a sample set, constructing a cross-sensor collaborative distillation network, and the cross-sensor collaborative distillation network comprising a teacher model and a student model; the teacher model and the student model respectively performing up-to-standard discrimination, and constructing a heterogeneous sensor dynamic competition force balance mechanism after reaching the standard; iteratively training the student model through optimization of distillation loss and dynamic competition force balance mechanism regularization loss; inputting monitoring data to be diagnosed into an optimal model obtained based on iterative training to perform diagnosis and output corresponding diagnosis results; relying on the cross-sensor collaborative distillation network, establishing the association between multi-sensor fusion fault features and single-sensor local features, and enabling missing data to effectively participate in model training without generating additional pseudo data in the whole process.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for electromechanical composite transmission systems, and more specifically to a cross-sensor collaborative diagnosis method for transmission systems based on dynamic competitive balance. Background Technology

[0002] Currently, as the core power transmission unit in aerospace, military equipment, high-end industrial equipment and rail transportation, the stable operation of the transmission system is directly related to the reliability of the equipment, mission safety and energy efficiency.

[0003] However, in complex working conditions and harsh environments, problems such as sensor hardware aging, signal transmission link interruption, and acquisition timing mismatch frequently occur, leading to an imbalance in the missing data rate of multi-sensor monitoring data. Some sensors, due to their harsh working environment, have a data missing rate of over 35%, while the missing rate of auxiliary sensors is only 5%. Moreover, samples of core fault types such as "bearing wear" and "rotor imbalance" often have additional missing data, creating a complex situation of "coexistence of high / low missing rate sensors, with even worse data quality for key faults." Traditional diagnostic methods either directly remove missing data or use single interpolation reconstruction. Furthermore, because they do not address the problem of "low missing rate sensors dominating training and high missing rate sensor features being marginalized," the diagnostic accuracy of the model for faults associated with high missing rates is significantly degraded, making it difficult to meet the needs of complex scenarios.

[0004] With the deep integration of artificial intelligence and big data technologies, deep learning provides an innovative path for the intelligent processing of multi-source heterogeneous missing data. In current research, generative adversarial networks can generate pseudo-data to compensate for missing data through game theory, but they rely on a large amount of auxiliary complete data and are prone to deviating from physical laws; physical information diffusion methods can embed constraints, but the models are complex and the pre-training cost is high; shared representation learning methods such as cross-domain feature transfer are highly adaptable, but they have not yet broken through the bottlenecks of "cross-sensor dynamic interaction missing data" and "unquantified relative competitiveness of sensors"—they cannot achieve balanced extraction of multi-source features, and it is difficult to identify the marginalization risk of sensors with high missing rates, resulting in insufficient robustness of the model in scenarios with unbalanced missing rates.

[0005] Therefore, how to solve the problem of unbalanced multi-source feature extraction and difficulty in identifying high missing rate sensors, which leads to insufficient robustness of the model in scenarios with unbalanced missing rate, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of the above problems, the present invention is proposed to provide a cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance that overcomes or at least partially solves the above problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a cross-sensor collaborative diagnostic method for a transmission system based on dynamic competitive balance; including the following steps: S1, acquiring monitoring data from sensors in the transmission system and performing preprocessing;

[0009] S2. Using the preprocessed monitoring data as the input sample set, a cross-sensor collaborative distillation network is constructed. The cross-sensor collaborative distillation network includes a teacher model and a student model. The teacher model and the student model respectively perform the target discrimination. After the target is achieved, a dynamic competitiveness balancing mechanism for heterogeneous sensors is constructed.

[0010] S3. Optimize the loss function of the cross-sensor collaborative distillation network and the regularization loss function of the heterogeneous sensor dynamic competitiveness balancing mechanism, and iteratively train the student model.

[0011] S4. Based on the optimal model obtained through iterative training, input the monitoring data to be diagnosed into the model for diagnosis and output the corresponding diagnostic results.

[0012] Furthermore, the preprocessing in step S1 specifically includes:

[0013] Missing data is marked on sensor monitoring data to distinguish between incomplete and complete missing data.

[0014] Improve the quality of monitoring data through noise reduction and standardized operations;

[0015] A non-overlapping sliding window is used to construct the sample set, which is then divided into a support set and a query set.

[0016] Simultaneously, the sample set is split into a training set and a test set according to a preset ratio, and category information and input shape are configured.

[0017] Furthermore, step S2 specifically includes:

[0018] S21. The teacher model performs signal layer timing alignment and point-by-point logit alignment, semantic layer prototype calculation and probability distribution transfer on the support set, generates fault category prototypes by weighting sensor contribution, and optimizes fault category prototypes.

[0019] S22. Determine whether the prototype optimization of the teacher model meets the standard. If it does not meet the standard, readjust the prototype optimization process of the teacher model. If it meets the standard, proceed to the student model training stage.

[0020] S23. The student model receives the query set, and based on the fault category prototype and fusion knowledge of the teacher model, it realizes the signal layer Logits collaboration and feature layer semantic transfer through the temperature softening strategy, and completes the prediction by using the distillation loss composed of KL divergence and feature transfer loss.

[0021] S24. The student model is used to determine whether the distillation loss meets the standard. If it does not meet the standard, the student model's distillation loss calculation process is readjusted. Once the standard is met, the heterogeneous sensor dynamic competitiveness balancing stage is entered.

[0022] S25. Construct a dynamic competitiveness balancing mechanism for heterogeneous sensors.

[0023] Furthermore, the specific implementation of signal layer timing alignment and point-by-point logit alignment in step S21 includes: converting multi-sensor time-domain signals into time-frequency diagrams through short-time Fourier transform to achieve timing feature alignment; for any fault type c Sensors The data defines the point-by-point cooperative distillation loss, and the specific formula is as follows:

[0024]

[0025] in, L point，a For single sensor The pointwise loss, KL(·||·) is the Kullback-Leibler divergence, and T is the temperature hyperparameter. softmax (·) represents the softmax function. and These represent the fault types. c exist b Multi-sensor m-fusion features of layer and single sensor Its characteristics.

[0026] Furthermore, the specific implementation of semantic layer prototype calculation in step S21 is as follows: calculate the fault category prototype of the multi-sensor fusion model. With a single sensor Model Fault Category Prototype The formula is:

[0027]

[0028] in, and These represent the fault types when the initial layer b=0. c Multi-sensor fusion features and single-sensor The characteristic output, y n Indicates sample c The n The true label of each data point This is an indicator function.

[0029] Furthermore, the semantic similarity calculation for the semantic layer probability distribution transfer in step S21 is specifically implemented as follows: cosine similarity is used to calculate the semantic similarity between the single sensor features and the prototype. S a (n) Semantic similarity of multi-sensor fusion S m (n), the formula is:

[0030]

[0031] in, Cos (·) represents the cosine similarity function. For a single sensor exist b Layer n The feature values ​​of each feature point For multi-sensor fusion models in b Layer n The feature values ​​of each feature point For a single sensor Corresponding fault type c The category prototype, Fault types corresponding to multi-sensor fusion models c The category prototype.

[0032] Furthermore, the specific implementation of defining the fault semantic knowledge gap and calculating the transfer loss in step S21 for the semantic layer probability distribution transfer is as follows: defining the fault semantic knowledge gap between single sensors and multiple sensors. The formula is:

[0033]

[0034] in, for 2-norm, n For feature point index, c Index for fault type;

[0035] Semantic knowledge transfer is achieved by minimizing the knowledge transfer loss, and the loss formula is as follows:

[0036]

[0037] in, To address the gap in semantic knowledge regarding faults, c For fault type index, Index for sensor categories.

[0038] Furthermore, the dynamic competitiveness balancing mechanism for heterogeneous sensors includes:

[0039] Define fault type c and sensors Relevant weight coefficient w c，a Calculate the weighted average distance of each sensor. D c，a Based on the weighted average distance D c，a Define the relative competitiveness of sensors RP a Combining task regularization and gradient regularization, using task masks... λ a Trigger regularized training of the slow-learning sensor by using gradient weighting coefficients. θ a The learning rate of each sensor is dynamically adjusted.

[0040] Furthermore, in step S2, the specific implementation of the weighted average distance calculation is as follows: Define the fault type. c and sensors Relevant weight coefficient w c，a The weighted average distance of each sensor is calculated based on the weighting coefficients. D c，a The specific formula is as follows:

[0041]

[0042] in, and These represent the fault types. c exist b Multi-sensor fusion features of layers and single sensor Features for 2-norm, For sensors The sum of the weighting coefficients corresponding to all fault types.

[0043] The specific implementation of the relative competitiveness of sensors is based on weighted average distance. D c，a Define the relative competitiveness of sensors RP a The specific formula is as follows:

[0044]

[0045] in, sign (·) represents the sign function, and A represents the total number of sensor categories. To remove the sensor Other sensor indexes, To remove the sensor The weighted average distance mean of all other sensors.

[0046] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance, which has the following beneficial effects:

[0047] 1. By leveraging a cross-sensor collaborative distillation network, a correlation is established between multi-sensor fusion fault features and single-sensor local features. Even if a sensor with a high missing rate only retains short-term valid data fragments, its implicit fault information can still be transferred to the diagnostic model through knowledge distillation. The entire process allows missing data to effectively participate in model training without generating additional pseudo-data.

[0048] 2. A dynamic competitiveness balancing mechanism for heterogeneous sensors was constructed. For weak sensors, task regularization was used to force their features to align with the fused features, and gradient weights were adjusted to accelerate the learning of effective features. For strong sensors, gradient weights were appropriately reduced to avoid overfitting. This dynamic compensation adjustment balances the contribution of sensors with different missing rates at the training mechanism level, ensuring that the model's diagnostic capabilities for various faults are more balanced. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0050] Figure 1 This is a flowchart of the cross-sensor collaborative diagnostic method provided in this embodiment of the invention;

[0051] Figure 2 This is a flowchart of the model training process provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the dynamic competitiveness balancing mechanism for heterogeneous sensors provided in this embodiment of the invention.

[0053] Figure 4 The figure shows the ablation experiment verification results of the key model provided in the embodiments of the present invention;

[0054] Figure 5 This is a comparison chart showing the diagnostic accuracy and F1 score of the method of the present invention and existing data missing reconstruction and data imbalance optimization methods provided in this embodiment of the invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] This invention discloses a cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balancing, such as... Figure 1 As shown, the specific steps include:

[0057] S1. Acquire monitoring data from sensors in the transmission system and perform preprocessing;

[0058] S2. Using the preprocessed monitoring data as the input sample set, a cross-sensor collaborative distillation network is constructed. The cross-sensor collaborative distillation network includes a teacher model and a student model. The teacher model and the student model respectively perform the target discrimination. After the target is met, a dynamic competitiveness balancing mechanism for heterogeneous sensors is constructed.

[0059] S3. Optimize the regularization loss through distillation loss and dynamic competitiveness equilibrium mechanism, and iteratively train the student model;

[0060] S4. Based on the optimal model obtained through iterative training, input the monitoring data to be diagnosed into the model for diagnosis and output the corresponding diagnostic results.

[0061] This invention focuses on the problem of unbalanced missing data rates in multi-sensor transmission systems. Relying on big data-driven multi-source feature mining and the adaptive adjustment capabilities of artificial intelligence algorithms, it constructs an intelligent diagnostic framework of "cross-sensor collaborative distillation + dynamic competitiveness balancing". It forms a closed-loop process of "data preprocessing - feature alignment - competitiveness assessment - model optimization - diagnostic reasoning". By mining the value of missing data through big data and adaptively balancing training biases through artificial intelligence algorithms, it breaks through the performance bottleneck of traditional methods under imperfect information, significantly improves the accuracy and robustness of transmission system fault diagnosis under complex working conditions, and provides a technical paradigm of deep collaboration between "data-algorithm-model" for intelligent operation and maintenance of equipment.

[0062] The following is a detailed discussion of each of the above steps:

[0063] The specific process of model training involved in this invention is as follows: Figure 2 As shown, the specific process is as follows:

[0064] First, in step S1, the monitoring data from the sensors in the transmission system is acquired and preprocessed.

[0065] By acquiring multi-source heterogeneous data such as acceleration and current signals from the drive end and fan end of the transmission system, covering various fault types, the data is marked with missing information to distinguish between incomplete and complete missing data. Then, the data quality is improved through noise reduction, standardization and other operations. Finally, a sample set is constructed using a non-overlapping sliding window of length 1024, which is divided into a support set and a query set. At the same time, the training set and the test set are split in an 8:2 ratio, and the category information and input shape are configured to provide a data foundation for subsequent model training.

[0066] The data collection and preprocessing process ensures the diversity and representativeness of the data, providing comprehensive data support for fault identification; at the same time, the data is standardized to unify the data scale, improve the overall data quality, and lay the foundation for the subsequent model training process.

[0067] Secondly, in step S2, a cross-sensor collaborative distillation network is constructed, and a dynamic competitiveness balancing mechanism for heterogeneous sensors is built by performing a target-based discrimination. The specific process is as follows:

[0068] The cross-sensor collaborative distillation network comprises a teacher model and a student model. The teacher model integrates a 1D convolutional residual network and a sensor attention mechanism. It first aligns the signal layer temporally using a short-time Fourier transform on the support set, extracts single-sensor and multi-sensor fusion features, and then calculates pointwise loss to constrain the consistency between the output of fusion features and single-sensor features. Subsequently, it performs semantic layer prototype calculation and probability distribution transfer to generate fault category prototypes and fused soft labels, and optimizes the prototypes. Then, it checks whether the prototype optimization meets the criteria. If not, the teacher model's prototype optimization process is readjusted; if it meets the criteria, it proceeds to the student model training phase. The student model adopts the same architecture as the teacher model but removes the sensor attention mechanism. After receiving the query set, it achieves feature semantic alignment through cosine similarity constraints, generates fault classification probabilities using a temperature softening strategy, and calculates the distillation loss, which consists of Kullback-Leibler divergence and feature transfer loss. Then, it checks whether the distillation loss meets the criteria. If not, the student model's distillation loss calculation process is readjusted; if it meets the criteria, it proceeds to the heterogeneous sensor dynamic competitiveness balancing stage.

[0069] Then, a dynamic competitiveness balancing mechanism for heterogeneous sensors is implemented: To address the problem of "strong sensors dominating and weak sensors marginalizing" caused by the imbalance of missing data rates among multiple sensors, weights are first set inversely proportional to the sensor missing rate, and the weighted average distance between single-sensor features and multi-sensor fused features is calculated; then, the average distance between the current sensor and other sensors is compared to determine the relative competitiveness of the sensor; subsequently, for sensors with weak competitiveness, task regularization is triggered by task masking, and an additional L2 loss is added to align the student probability with the fusion probability to force feature alignment; at the same time, the gradient weighting coefficients of the sensors are initialized, the learning rate of each sensor is dynamically adjusted according to the iterative average competitiveness, and the adjusted gradient weights are embedded into the gradient descent process of the student model.

[0070] By leveraging a cross-sensor collaborative distillation network, this approach first aligns the temporal sequence of multi-sensor signals through short-time Fourier transform, ensuring that the locally effective signals from sensors with high and low missing rates remain synchronized over time. Then, utilizing semantic layer prototype transfer and pointwise logit constraints from the teacher model, it establishes a correlation between multi-sensor fusion fault features and single-sensor local features. Even if a high-missing-rate sensor retains only short-term effective data fragments, its implicit fault information can still be transferred to the diagnostic model through knowledge distillation. The entire process requires no additional pseudo-data generation, allowing missing data to effectively participate in model training. Simultaneously, a dynamic competitiveness balancing mechanism for heterogeneous sensors quantifies the difference between single-sensor features and multi-sensor fusion features through weighted average distance, and identifies the "strong / weak" attributes of each sensor in real time using relative competitiveness assessment. For weak sensors, task regularization forces their features to align with the fusion features, and gradient weights are adjusted to accelerate the learning of effective features. For strong sensors, gradient weights are appropriately reduced to avoid overfitting. This dynamic compensation adjustment balances the contribution of sensors with different missing rates at the training mechanism level, ensuring a more balanced diagnostic capability for various faults.

[0071] In step S3, as Figure 3 As shown, the student model is iteratively trained by optimizing the distillation loss and the regularization loss of the dynamic competitiveness equilibrium mechanism. The specific process is as follows:

[0072] Figure 3 In this context, Epoch represents the number of times the training dataset has been fully traversed; S1-S5 in the Epoch part represent heterogeneous sensors.

[0073] , , To determine the threshold for model compliance, , , Let be the sample feature vector under the c-th type of fault.

[0074] This is the formula for relative competitiveness;

[0075] in, For sensors For the relative competitiveness of fault type c, G ca G is the generalized inverse gain factor. ca It is a sensor Weights associated with fault type c For sensors The weighted average distance, Let sgn() be the mean of the weighted average distances from the sensors, and let sgn() be the sign function.

[0076] when When the value is greater than 0, the sensor is highly competitive.

[0077] when When =0, the sensor competitiveness is balanced;

[0078] when When the value is less than 0, the sensor's competitiveness is weak.

[0079] When the sensor is less competitive, i.e. When <0, =λ, at which point the gradient weighting coefficient is positive, which will increase the learning rate of the sensor, accelerate its training process, and prevent it from being marginalized;

[0080] When sensor competitiveness is strong or balanced, that is When ≥0, =0, at this time the gradient weighting coefficient is 0, no additional acceleration is performed, and the learning rate is appropriately reduced according to other rules.

[0081] As a loss term in the overall optimization objective, it is used to combine distillation loss and regularization loss of dynamic competitiveness balancing mechanism during training to achieve balanced optimization of multi-sensor collaborative diagnosis.

[0082] in, γ 1. γ 2 represents the equilibrium hyperparameter. For sensors The weighting coefficient for fault type c, This includes distillation losses, such as point-by-point co-distillation losses and fusion soft-label losses. For regularization loss, These are the gradient weighting coefficients in gradient regularization;

[0083] The overall optimization objective is "distillation loss + regularization loss from dynamic competitiveness equilibrium mechanism".

[0084]

[0085] in, , To balance hyperparameters, For distillation loss, To mitigate the loss, the Adam optimizer is used to iteratively train the student model. The initial learning rate is set to 0.001, with a total of 500 epochs. The learning rate is decayed by a factor of 0.1 every 100 epochs. The model is considered converged when the accuracy fluctuation is less than 0.5% for 50 consecutive training epochs. After convergence, the multi-sensor data to be diagnosed is input into the model, and the fault diagnosis results are output. The effectiveness of the diagnosis results is evaluated under various scenarios, including no missing data, uniform missing data, and typical non-uniform missing data, based on accuracy, F1 score, and feature transfer alignment. If the result is invalid, the process returns to the multi-sensor monitoring data acquisition and preprocessing stage and is re-executed. If the result is valid, the process ends.

[0086] The training of the model is completed based on the above process; this architecture enables the model to stably adapt to the complex operating conditions of the transmission system, and maintain the stability of diagnostic performance even under extreme conditions.

[0087] In the final step S4, based on the optimal model obtained after iterative training, the monitoring data to be diagnosed is input into the model for diagnosis and the corresponding diagnostic results are output.

[0088] In summary, the technical solution of this invention forms a complete technology chain of "data preprocessing - cross-sensor feature alignment - dynamic competitiveness evaluation - model adaptive optimization - fault diagnosis reasoning". Compared with existing methods, it has significant advantages in terms of missing data utilization, sensor contribution balance, and robustness under complex operating conditions. Moreover, it can be directly adapted without additional modifications to existing monitoring hardware, effectively improving the operational reliability of transmission systems, reducing the probability of unplanned downtime and maintenance costs, and providing efficient and reliable technical support for the intelligent operation and maintenance of transmission systems in aerospace, high-end industry, and rail transportation.

[0089] The following is a further explanation of the above steps:

[0090] 1. Multi-source heterogeneous data acquisition: Taking the electromechanical composite transmission system as the monitoring object, dedicated sensors are deployed in key parts: three triaxial acceleration sensors are installed at both the drive end and the fan end, and four Hall current sensors are deployed in the power distribution box. The acquisition process covers eight typical fault states of the transmission system, and records data under different operating conditions to ensure that the data can reflect the complete characteristics of the equipment under normal operation, various faults, and different operating conditions, avoiding insufficient model generalization ability due to incomplete data coverage.

[0091] 2. Missing Data Labeling and Classification: Based on the causes of missing data in actual industrial scenarios, the definition criteria for two types of missing data are clarified: "Incompletely Missing" data refers to data in which the sensor loses signals only for part of the time period, but retains the key effective segments when the fault occurs. This type of data needs to be labeled with the missing time period and the location of the effective segments to retain core information for subsequent feature extraction; "Completely Missing" data refers to data in which the sensor has no effective output for a certain period of time. The start / end time of the missing data and the corresponding operating conditions need to be labeled to facilitate cross-sensor inference based on the correlation features of other sensors and avoid the data chain being broken due to the complete loss of a single sensor.

[0092] 3. Preprocessing including noise reduction and data standardization: The noise reduction stage focuses on typical interferences in the transmission system monitoring data, employing wavelet noise reduction technology to optimize signal purity. For vibration signals collected by the accelerometers at the drive / fan ends, a "db4" wavelet basis is selected to decompose the signal into 6 scales—low-scale components correspond to low-frequency characteristics of equipment faults, while high-scale components are mostly environmental interference; heuristic threshold filtering is used for high-scale noise components to retain effective fluctuations related to faults, eliminate meaningless interference, and reconstruct the signal, ensuring that key features such as bearing wear and cage faults are clearly preserved. For current sensor data, electromagnetic interference is removed using the same wavelet decomposition logic to prevent noise from masking the current fluctuation characteristics of faults such as stator inter-turn short circuits and voltage imbalances.

[0093] The standardization process eliminates the impact of sensor differences on the model. Each sensor data is processed independently, and the data from the accelerometer and the current sensor are adjusted according to their own data distribution. All types of signals are mapped to a unified numerical scale. This avoids the imbalance of feature weights caused by the magnitude of the acceleration signal being much smaller than that of the current signal, and also eliminates the interference caused by differences in the units and precision of different sensors. This ensures that vibration and current signals have equal participation in subsequent feature extraction.

[0094] 4. Sample Set Construction and Partitioning: A non-overlapping sliding window of length 1024 is used to slice the preprocessed time-domain signal, ensuring that each sample can completely contain at least one fault feature cycle. The generated samples are divided into a support set and a query set according to "functional requirements"—the support set, accounting for 70% of the total sample size, is used for fault prototype learning of the teacher model in the subsequent cross-sensor collaborative distillation network and needs to cover typical samples of various fault-condition combinations; the query set, accounting for 30%, is used for knowledge transfer training of the student model, focusing on including missing data samples to verify the model's adaptability to incomplete information. At the same time, the total sample is split into a training set and a test set in an "8:2" ratio. The training set is used for iterative optimization of model parameters, and the test set is used for independent performance verification. When dividing, it is ensured that the fault types and operating condition distributions of the two datasets are consistent to avoid model evaluation distortion due to data distribution bias.

[0095] 5. Construction and training of teacher models:

[0096] The teacher model, serving as the "core of knowledge supply," provides the student model with dual knowledge benchmarks—multi-sensor and single-sensor—through fine-grained feature constraints and semantic prototype shaping. The training process revolves around "architecture design - signal alignment - point-by-point constraints - semantic prototype generation - iterative optimization."

[0097] Architecture Design: Building a 1D Convolutional Residual Network: Adopting a "convolutional block + residual connection" structure. The convolutional block contains multiple layers of 1D convolutions, combined with batch normalization and ReLU activation, to extract local time-varying features of sensor time-series signals; the residual connection alleviates gradient vanishing through "identity mapping", allowing the network to learn long-range dependency features across time steps.

[0098] Design a sensor attention mechanism: Based on the "gated sigmoid mechanism", for each sensor feature, first compress the dimension through a fully connected layer, and then calculate the attention weights.

[0099]

[0100] in, For sensors Features , These are learnable parameters. Weights It is used for weighted fusion of features from multiple sensors, enabling the model to adaptively focus on sensors with high information content, thereby enhancing the effectiveness of multi-sensor fusion.

[0101] Signal layer alignment:

[0102] Perform short-time Fourier transform: Divide the continuous time-domain signal from multiple sensors into short-time windows, perform Fourier transform on each window to generate a time-frequency graph, achieve time-series alignment and extract frequency domain features.

[0103] Extracting single / multi-sensor features: Multi-sensor fusion features are extracted through "multi-channel convolution + global pooling"; single-sensor features are extracted by "local convolution + pooling" on the time-frequency map of a single sensor, preserving the individual time-frequency characteristics of a single sensor.

[0104] Point-by-point co-distillation losses:

[0105] Introduce a temperature hyperparameter T: "soften" the probability distribution of the softmax function. This reduces the discriminative power of feature classes to expose inter-class details.

[0106] KL divergence calculation: for fault type c, network layer b, and sensor Calculate multi-sensor fusion features Features of a single sensor through The subsequent distribution differences force single-sensor features to align with the multi-sensor fusion feature distribution point-by-point and layer-by-layer, achieving fine-grained knowledge transfer. The formula is:

[0107]

[0108] in, For single sensor The pointwise loss, KL(·||·) is the Kullback-Leibler divergence, and T is the temperature hyperparameter. For the softmax function, and These represent the multi-sensor fusion characteristics and single-sensor characteristics of fault type c at layer b, respectively. Its characteristics.

[0109] Semantic layer fault prototype generation:

[0110] Calculate the fault category prototype: Select the initial layer of the network Filter samples by fault type c (using indicator function) (Take 1 if the condition is true, otherwise take 0), and then apply the filtered multi-sensor fusion features. Single sensor characteristics Take the average of each to obtain the prototype. , The formula is:

[0111]

[0112] in, and These represent the multi-sensor fusion features and single-sensor features of fault type c when the initial layer b=0, respectively. The characteristic output, This represents the true label of the nth data point in sample c. This is an indicator function.

[0113] Fault category prototype optimization:

[0114] Incorporate sensor effectiveness constraints: Introduce a "data quality index" for the sensor to weight its features. For example, if the sensor... The missing rate is Then the weight is set to During prototype updates, features The weight is This ensures that the characteristics of high-reliability sensors have a greater impact on the prototype, and avoids low-quality sensors misleading the prototype's learning.

[0115] Iterative optimization logic: Combining "point-by-point loss" and "sensor effectiveness weights", continuously update... and If the "feature distance" or "prototype-based classification accuracy" between the prototype and the real sample does not reach the preset threshold, the teacher model parameters will continue to be adjusted; if they reach the threshold, the "knowledge condensation" of the teacher model is completed, and the student model training begins.

[0116] 6. Construction and training of student models:

[0117] The student model aims to "inherit teacher knowledge in a lightweight manner." Through semantic alignment, temperature softening, and multi-loss fusion, it retains fault identification capabilities while reducing complexity. The training process revolves around "lightweight architecture - semantic alignment - temperature softening - distillation loss optimization," and the workflow is as follows:

[0118] Lightweight architecture design: It adopts a network with the same structure as the teacher model, but removes the sensor attention mechanism, reduces the number of parameters, and forces students to learn single-to-multi-sensor knowledge associations from the "features themselves".

[0119] Feature semantic alignment: Constraining the semantic consistency between student features and teacher prototypes through cosine similarity.

[0120] Calculate the semantic similarity between single-sensor features and the teacher's single-sensor prototype. :

[0121]

[0122] Calculate the semantic similarity between multi-sensor fusion features and the teacher's multi-sensor prototype. :

[0123]

[0124] in, The cosine similarity function is used. For a single sensor The feature value of the nth feature point in layer b This represents the feature value of the nth feature point in layer b of the multi-sensor fusion model. For a single sensor The corresponding category prototype for fault type c, This is the category prototype for fault type c corresponding to the multi-sensor fusion model.

[0125] Temperature softening prediction:

[0126] The student model's output needs to be softened at a temperature of 1.5 to generate fault classification probabilities. The softened softmax function is...

[0127]

[0128] when When this is the case, the output probability distribution is smoother (the difference in probability between classes is reduced). Setting This is to ensure that the student model's output not only includes "class classification" but also preserves subtle probability differences between classes. If using... The probability tends to concentrate in a few categories, losing the inter-class relationship; if T is too large, the distribution is too flat, losing class discrimination. Therefore, 1.5 is a compromise between "smoothness and discrimination".

[0129] Distillation loss fusion:

[0130] Distillation loss is the core optimization objective for student model training. It is a weighted combination of three parts: pointwise loss, soft label fusion loss, and semantic knowledge transfer loss, ensuring the transfer of multi-dimensional knowledge.

[0131] Pointwise loss (weight 0.6): compared to the teacher model The calculation methods are consistent, and the point-by-point feature distributions are aligned.

[0132] Fusion soft label loss (weight 0.4): Calculate the L2 norm of "teacher multi-sensor fusion soft label (probability distribution)" and "student model output probability" to constrain the consistency of the global probability distribution.

[0133] Semantic knowledge transfer loss: Defines the gap in fault semantic knowledge between single-sensor and multi-sensor systems. The formula is:

[0134]

[0135] in, for 2-norm, where n is the feature point index and c is the fault type index;

[0136] Semantic knowledge transfer is achieved by minimizing the knowledge transfer loss, and the loss formula is as follows:

[0137]

[0138] in, To represent the gap in semantic knowledge about faults, c is the fault type index. Index for sensor categories.

[0139] This narrows the gap between the "single-sensor semantics" and "multi-sensor semantics" of student models, ensuring the transfer of semantic layer knowledge.

[0140] Loss optimization logic: Weighted summation of the three parts of the loss.

[0141]

[0142] Ldistill This represents the global distillation loss;

[0143] L point To compensate for point-by-point coordinated distillation losses;

[0144] L softlabel To incorporate soft label loss;

[0145] L KDE Estimate the loss for kernel density;

[0146] The student model parameters are updated through backpropagation. If the distillation loss does not reach the preset threshold, training continues; if it does, the student model training is complete, and the dynamic competitiveness equilibrium mechanism is activated.

[0147] 7. Weighted average distance calculation: quantifying the differences in features between single and multiple sensors.

[0148] First, define the fault type c and the sensor. Relevant weight coefficient The weighted average distance of each sensor is calculated based on the weighting coefficients. This is used to quantify the difference between "single-sensor features" and "multi-sensor fusion features": The formula is:

[0149]

[0150] in, and These represent the multi-sensor fusion characteristics and single-sensor characteristics of fault type c at layer b, respectively. Features for 2-norm, For sensors The sum of the weighting coefficients corresponding to all fault types.

[0151] 8. Sensor relative competitiveness calculation: Determining the strength of competitiveness.

[0152] Based on the above weighted average distance The relative competitiveness of sensors is defined by a formula. Clarify the competitiveness level of each sensor in the dimension of "single-to-multi-sensor feature differences":

[0153]

[0154] in, Let A be the sign function, and A be the total number of sensor categories. To remove the sensor Other sensor indexes, To remove the sensor The weighted average distance mean of all other sensors.

[0155] when At that time, (sensor (Strong competitiveness)

[0156] when At that time, (sensor (weak competitiveness)

[0157] when At that time, (sensor (Competitive balance or unavailable).

[0158] 9. Task Regularization: Constraining the Probability Distribution of Weakly Competitive Sensors

[0159] against The "slow learning sensor" sets a task mask. ( For indicator functions, Output 1 if the condition is met, otherwise output 0), and force the feature layer probability distribution of the weakly competitive sensor to align with that of the multi-sensor system by minimizing the "sum of squared differences loss":

[0160]

[0161] Where T is the temperature hyperparameter. and These represent the multi-sensor / m-fused characteristics and single-sensor characteristics of fault type c at layer b, respectively. Features For the softmax function, Sum the losses for all fault types c and feature layers b.

[0162] 10. Gradient Regularization: Dynamically Adjusting the Learning Rate

[0163] To enable sensors with varying levels of competitiveness to "adjust their learning efficiency as needed," gradient updates need to be regularized.

[0164] Gradient weighting coefficient initialization: Let the sensor The initial gradient weighting coefficients are

[0165]

[0166] Where A is the total number of sensor types, used to balance the impact of sensor quantity on the gradient; calculate the iterative average relative competitiveness.

[0167] Iterative average competitiveness calculation: During the i-th training epoch, the sensor performance is calculated. "Iterative average relative competitiveness"

[0168]

[0169] Where i represents the current training epoch. For the sensor at the k-th epoch The relative competitiveness; the learning speed of each epoch is dynamically adjusted through gradient descent, as shown in the formula:

[0170]

[0171] in, The update rate is represented by i, which indicates the epoch; when (sensor When learning is too fast, reduce To slow down the learning speed; when (sensor When learning is slow, increase To improve learning speed.

[0172] Through the process of "quantifying feature differences → determining competitiveness → task + gradient dual regularization", this mechanism enables heterogeneous sensors with different performance to achieve dynamic balance of competitiveness, and ultimately efficiently and collaboratively complete the fault identification task.

[0173] 11. The total loss function is used to construct the total loss of the student model. Combining the two parts of loss:

[0174] Distillation loss It includes point-by-point collaborative distillation loss, fusion soft label loss, and semantic knowledge transfer loss, and is the core constraint of knowledge transfer from teacher to student.

[0175] Regularization loss in dynamic competitive equilibrium Loss due to task regularization This loss is a combination of the loss from gradient regularization and the loss from gradient regularization. The total loss formula is:

[0176]

[0177] in, The weighting coefficients for the regularization loss.

[0178] 12. Optimizer and Training Strategy:

[0179] The Adam optimizer is used for parameter updates, leveraging its adaptive momentum adjustment characteristics to balance the gradient update efficiency across different layers and sensor features. The training strategy is designed as follows:

[0180] Initial learning rate: set to 0.001 to allow the model to quickly explore the parameter space in the early stages of training;

[0181] Stepped learning rate decay: The learning rate decays to 0.1 times its original value every 100 epochs, ensuring that parameters can be finely adjusted later.

[0182] Total epochs: set to 500, covering the entire stage of "fast convergence + fine-tuning".

[0183] 13. Convergence determination mechanism:

[0184] Set "the accuracy fluctuation of the validation set is less than 0.5% in 50 consecutive training rounds" as the convergence condition: record the fault classification accuracy of the validation set in each round. If the absolute value of the difference between the accuracy of adjacent rounds does not exceed 0.5% in 50 consecutive rounds, the model parameters are considered to be stable and the training process has converged.

[0185] 14. Fault Diagnosis and Multi-Dimensional Assessment:

[0186] After the model converges, perform fault diagnosis and effect evaluation:

[0187] Fault diagnosis execution: Input the multi-sensor time-series data to be diagnosed into the model. The model outputs the probability of the fault category through "lightweight network + transferred multi-sensor knowledge" and finally determines the fault type.

[0188] Multi-indicator evaluation:

[0189] Classification performance: Accuracy and F1 score were used.

[0190] Knowledge transfer effect: Calculate feature transfer alignment.

[0191] 15. Process Cycle and Validity Determination:

[0192] If the evaluation results show that the diagnostic effectiveness is insufficient, return to the "Acquisition and Preprocessing of Multi-Sensor Monitoring Data" stage and re-execute the entire process of "Data Acquisition → Preprocessing → Model Training → Competitiveness Balancing → Diagnostic Evaluation"; if the evaluation indicators of each scenario meet the preset requirements, the process is deemed effective and training and diagnosis are terminated.

[0193] To verify the practicality of the method of this invention, an empirical experiment was conducted using an electromechanical composite transmission system of a certain device as the research object. The experiment relied on a simulation platform for the electromechanical composite transmission system to complete multi-dimensional data acquisition and performance verification. Verification schemes were designed from three dimensions: experimental data acquisition, ablation study of core modules, and comparative experiments with existing methods. The system systematically verified the practical application effectiveness of the method of this invention under complex working conditions and various data loss scenarios.

[0194] The experimental data acquisition and processing are as follows: The experiment relied on a simulation test bench for electromechanical composite transmission systems to complete multi-sensor signal acquisition. The data design fully covered the complexity and information imperfections of real working conditions, including 9 sensor channels with a sampling frequency of 25.6kHz. The original signal was divided into segments using a non-overlapping sliding window of length 1024. 1200 samples were extracted for each motor operating speed and divided into training and test sets in an 8:2 ratio. Eight motor health states were simulated. At the same time, scenarios with no missing data, uniform missing data, and three typical non-uniform missing data were constructed to provide real support for verifying the robustness of the method.

[0195] To verify the superiority of the cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balancing, an ablation study was designed to compare the performance of three schemes under scenarios of "no missing data, uniform missing data, and typical non-uniform missing data," specifically: 1) missing cross-sensor collaborative distillation network (Method A); 2) missing heterogeneous sensor dynamic competitiveness balancing mechanism (Method B); 3) the method of this study. The three methods were validated under the same network architecture and training parameters. Quantitative results are as follows: Figure 4 As shown.

[0196] Experimental results show that Method A retains only the basic lightweight network, removing knowledge transfer modules such as signal layer time-frequency alignment and point-by-point collaborative distillation loss. Because it cannot transfer multi-sensor fusion knowledge to the student model, it can only simply mask or inefficiently utilize missing data. The diagnostic accuracy is below 75% in all three missing data scenarios, confirming the core role of cross-sensor collaborative distillation networks in "single-to-multi-sensor knowledge alignment".

[0197] Method B retains the cross-sensor collaborative distillation network but removes the weighted average distance calculation, task regularization, and gradient regularization modules. Because it cannot quantify sensor contributions or adjust learning weights, sensors with high missing rates are marginalized, leading to a bias in training where "strongly competitive sensors overfit and weakly competitive sensors underlearn." The diagnostic accuracy is improved by approximately 10% compared to Method A, but still lower than the method used in this study.

[0198] This study integrates a "cross-sensor collaborative distillation network" with a "heterogeneous sensor dynamic competitiveness balancing mechanism." Fine-grained knowledge transfer is achieved through the collaborative distillation network, and the sensor weights are then balanced via the dynamic competitiveness balancing mechanism. This effectively overcomes the dual interference of data loss and competitiveness differences, achieving a diagnostic accuracy exceeding 92% in three scenarios with data loss, thus validating the necessity of dual-module collaboration.

[0199] The experiment used a cross-sensor collaborative distillation network framework as its basic architecture, and its basic structure is shown in Table 1.

[0200] Table 1. Summary of Basic Network Structure

[0201]

[0202] To further verify the performance of the method of this invention, experiments were conducted to compare it with three existing fault diagnosis methods addressing data imbalance. These three methods include: Method C, which utilizes adaptive bandwidth local density calculation and global density optimization to enhance classifier robustness; employing a misclassification rate feedback mechanism and a hybrid density weight normalization strategy to optimize sample weight distribution and ensure model learning balance; Method D, which converts one-dimensional rolling bearing vibration data into two-dimensional time-frequency images using short-time Fourier transform; improving the diffusion model backbone based on Unet to generate high-quality minority class time-frequency samples; and employing a horizontal slicing strategy combined with Vision Transformer to extract horizontal features from the time-frequency images to alleviate data imbalance; and Method E, which uses a pre-trained Vision Transformer to extract global data information, calculates cosine similarity between this information and the features extracted by the diffusion model, and uses this cosine similarity as a regularization term to guide the diffusion model to generate fault samples with more comprehensive representations.

[0203] The comparison results show that... Figure 5 As shown:

[0204] 1) No missing data scenario: The proposed method achieves a diagnostic accuracy of over 98.5% and an F1-score of 0.981; Method C, due to its reliance on a single sensor, has an accuracy that is more than 9 percentage points lower than the proposed method; Methods D and E, limited by single-sensor optimization or lack of multi-sensor collaboration, have accuracy that is 7 and 5 percentage points lower, respectively.

[0205] 2) 20% uniform missing scenario: The accuracy of the proposed method only slightly decreased to 97.2%, and the F1-score remained at 0.975; Method C's accuracy decreased to 86.7% due to the lack of cross-sensor information compensation; Methods D and E's accuracy dropped to 87.2% and 89.2% respectively due to the lack of missing rate adaptation design.

[0206] 3) Non-uniform missing scenarios: The proposed method maintains a high accuracy of 95.8% and an F1-score of 0.970 thanks to the dual advantages of "cross-sensor collaborative distillation" and "dynamic competitive balance". In contrast, methods C, D, and E, due to their inability to cope with the imbalance of missing rates, have accuracy drops to 83.4%, 85.6%, and 87.6%, respectively, showing a significant difference.

[0207] Under three experimental conditions, the method of this invention outperforms three existing methods in both diagnostic accuracy and F1-score. This indicates that the method of this invention exhibits higher robustness and diagnostic efficiency under conditions of missing data and imbalanced missing data rates.

[0208] Through the above experiments and comparative analysis, the practicality and superiority of the method of this invention in dealing with complex fault diagnosis problems can be seen. Combining a cross-sensor collaborative distillation network with a dynamic competitiveness balancing mechanism, the method of this invention can effectively solve the common problem of data missing rate imbalance in practical applications, and has high application value and technical advantages.

[0209] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0210] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance, characterized in that, Includes the following steps: S1. Acquire monitoring data from sensors in the transmission system and perform preprocessing; S2. Using the preprocessed monitoring data as input samples, a cross-sensor collaborative distillation network is constructed. The cross-sensor collaborative distillation network includes a teacher model and a student model. The teacher model and the student model respectively perform a pass / fail judgment. After passing the judgment, a heterogeneous sensor dynamic competitiveness balancing mechanism is constructed. The difference between single sensor features and multi-sensor fusion features is quantified by weighted average distance. Combined with relative competitiveness evaluation, the strong / weak attributes of each sensor are identified in real time. The heterogeneous sensor dynamic competitiveness balancing mechanism is as follows: for weak sensors, task regularization is used to force their features to align with the fusion features, and the gradient weights are adjusted to accelerate the learning of effective features; for strong sensors, the gradient weights are appropriately reduced to avoid overfitting. S3. Regularize the loss using the distillation loss function and the dynamic competitiveness balancing mechanism of heterogeneous sensors, and iteratively train the student model. S4. Based on the optimal model obtained through iterative training, input the monitoring data to be diagnosed into the model for diagnosis and output the corresponding diagnostic results.

2. The cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance as described in claim 1, characterized in that, The preprocessing process in step S1 specifically includes: Missing data is marked on sensor monitoring data to distinguish between incomplete and complete missing data. Improve the quality of monitoring data through noise reduction and standardized operations; A non-overlapping sliding window is used to construct the sample set, which is then split into a training set and a test set according to a preset ratio. At the same time, the training set is split into a support set and a query set, which are used as inputs for the model training phase; The test set is split into a support set and a query set, which serve as inputs for the model testing phase. Configure category information and input shape.

3. The cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance as described in claim 1, characterized in that, Step S2 specifically includes: S21. Using the preprocessed monitoring data as input samples, a cross-sensor collaborative distillation network is constructed. The cross-sensor collaborative distillation network includes a teacher model and a student model. The teacher model performs signal layer temporal alignment and point-by-point logit alignment, semantic layer prototype calculation and probability distribution transfer on the support set. It generates fault category prototypes by weighting the sensor contribution and optimizes the fault category prototypes. S22. Determine whether the prototype optimization of the teacher model meets the standard. If it does not meet the standard, readjust the prototype optimization process of the teacher model. If it meets the standard, proceed to the student model training stage. S23. The student model receives the query set, and based on the fault category prototype and fusion knowledge of the teacher model, it realizes the signal layer Logits collaboration and feature layer semantic transfer through the temperature softening strategy, and completes the prediction by using the distillation loss composed of KL divergence and feature transfer loss. S24. The student model is used to determine whether the distillation loss meets the standard. If it does not meet the standard, the student model's distillation loss calculation process is readjusted. Once the standard is met, the heterogeneous sensor dynamic competitiveness balancing stage is entered. S25. Construct a dynamic competitiveness balancing mechanism for heterogeneous sensors.

4. The cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance as described in claim 3, characterized in that, The specific implementation of signal layer timing alignment and point-by-point logit alignment in step S21 includes: converting the multi-sensor time-domain signals into a time-frequency diagram through short-time Fourier transform to achieve timing feature alignment; for sensors with any fault type c... The data defines the point-by-point cooperative distillation loss, and the specific formula is as follows: in, For single sensor The pointwise loss, KL(·||·) is the Kullback-Leibler divergence, and T is the temperature hyperparameter. For the softmax function, and These represent the multi-sensor / m-fused features and single-sensor features of fault type c at layer b, respectively. Its characteristics.

5. The cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance as described in claim 4, characterized in that, The specific implementation of semantic layer prototype calculation in step S21 is as follows: calculate the fault category prototype of the multi-sensor fusion model. With a single sensor Model Fault Category Prototype The formula is: in, and The multi-sensor m-fusion features and single-sensor features of fault type c when the initial layer b=0 are respectively represented. The characteristic output, Indicates sample c The n The true label of each data point This is an indicator function.

6. The cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance as described in claim 5, characterized in that, The semantic similarity calculation for semantic layer probability distribution transfer in step S21 is specifically implemented as follows: cosine similarity is used to calculate the semantic similarity between single sensor features and the prototype. Semantic similarity of multi-sensor fusion The formula is: in, The cosine similarity function is used. For a single sensor exist b Layer n The feature values ​​of each feature point This represents the feature value of the nth feature point in layer b of the multi-sensor fusion model. For a single sensor Corresponding fault type c The category prototype, This is the prototype of the category corresponding to fault type c in the multi-sensor fusion model.

7. The cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance as described in claim 6, characterized in that, The specific implementation of defining the fault semantic knowledge gap and calculating the transfer loss in step S21, which involves semantic layer probability distribution transfer, is as follows: defining the fault semantic knowledge gap between single sensors and multiple sensors. The formula is: in, for 2-norm, n For feature point index, c Index for fault type; Semantic knowledge transfer is achieved by minimizing the knowledge transfer loss, and the loss formula is as follows: in, To address the gap in semantic knowledge regarding faults, c For fault type index, Index for sensor categories.

8. The cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance as described in claim 1, characterized in that, The dynamic competitiveness balancing mechanism for heterogeneous sensors includes: Define fault type c and sensors Relevant weight coefficient Calculate the weighted average distance of each sensor. Based on the weighted average distance Define the relative competitiveness of sensors Combining task regularization and gradient regularization, using task masks... Trigger regularized training of the slow-learning sensor by using gradient weighting coefficients. The learning rate of each sensor is dynamically adjusted.

9. The cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance as described in claim 8, characterized in that, In step S2, the specific implementation of the weighted average distance calculation is as follows: defining the fault type. c and sensors Relevant weight coefficient The weighted average distance of each sensor is calculated based on the weighting coefficients. The specific formula is as follows: in, and These represent the fault types. c exist b Multi-sensor m-fusion features of layer and single sensor Features for 2-norm, For sensors The sum of the weighting coefficients corresponding to all fault types.

10. The cross-sensor collaborative diagnostic method for transmission systems based on dynamic competitiveness balance as described in claim 9, characterized in that, The specific implementation of the sensor's relative competitiveness is as follows: based on weighted average distance. Define the relative competitiveness of sensors The specific formula is as follows: in, Let A be the sign function, and A be the total number of sensor categories. To remove the sensor Other sensor indexes, To remove the sensor The weighted average distance mean of all other sensors.

Images