A TBM electric drive system fault diagnosis method of an adaptive sparse fusion network

Abstract

The application discloses a TBM electric drive system fault diagnosis method based on an adaptive sparse fusion network, and belongs to the technical field of TBM electric drive control. Multi-source operation data of the electric drive system are denoised and time-frequency domain features are extracted to construct a multi-modal fusion data set; a multi-task diagnosis framework is established, and an optimal modal combination is dynamically allocated through a learnable modal weight vector; a double support vector quantile regression algorithm is proposed to fit upper and lower threshold values of temperature variation, and real-time temperature anomaly early warning is realized; a task conflict graph is constructed based on task gradient similarity, and the activation range of neurons of the conflict task is controlled by using a dynamic sparse path; a fusion modulator is used to weight and compensate the confidence of each task result, and a fault type and diagnosis basis are output. The application solves the problem that weak fault signals are covered under strong coupling of multiple motors, and the misdiagnosis and missed diagnosis rate is high, improves the robustness and accuracy of fault diagnosis, and guarantees the operation stability and safety of the TBM electric drive system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of TBM electric drive control, and particularly relates to a TBM electric drive system fault diagnosis method based on an adaptive sparse fusion network. BACKGROUND

[0002] A TBM (tunnel boring machine) is major equipment for hard rock tunnel construction, has the advantages of high tunneling efficiency, good safety, environmental friendliness, etc., and has been widely used in tunneling construction. As the power core of the whole machine, the TBM electric drive system undertakes multiple key tasks such as driving the cutter head to rotate, moving the propulsion system, adjusting the guide mechanism, etc., and its stable operation is the basic condition for ensuring the normal tunneling of the whole machine. In the long-distance, high-intensity and continuous operation of the tunnel construction environment, if the electric drive system fails, it is easy to cause problems such as cutter head jamming, propulsion failure, and cooling system abnormalities, which will lead to the interruption, lag or even stoppage of the tunneling operation, and thus affect the project cycle and cost. At the same time, the electric drive system failure may be accompanied by phenomena such as electric arc, overcurrent and overheating, and in severe cases, it can cause damage to electrical equipment and even cause safety accidents such as fire, explosion and electric shock, which directly threatens the operators and equipment. Therefore, the running state of the TBM electric drive system not only determines the running efficiency and economy of the tunneling equipment, but also directly relates to the safety of the construction site.

[0003] In recent years, research on TBM electric drive system fault modeling and health state monitoring has gradually attracted attention, and some theoretical methods have been successfully applied to key component anomaly detection, abnormal data detection, motor temperature rise monitoring and control strategy optimization, effectively improving the fault response capability and system stability of the tunneling process, and promoting the development of TBM intelligent operation and maintenance.

[0004] However, the TBM electric drive system usually consists of 9-14 motors working together, with various fault types, large data volume and strong signal coupling. Traditional fault diagnosis methods usually focus on a single fault type and can only monitor short circuit or open circuit faults, and may misjudge or miss the diagnosis when multiple types of faults occur; the TBM working conditions are harsh, often accompanied by strong vibration, high temperature, high humidity and strong dust, resulting in signals containing strong disturbances and noise, and multi-motor cooperation also brings signal coupling problems, making it difficult for traditional methods to extract effective fault features, especially the weak features in the early stage of failure, which are easily "overwhelmed"; although there are some multi-task fault diagnosis models, these models are prone to task conflicts, unreasonable resource allocation, delayed diagnosis and low diagnosis accuracy when facing the complex faults of the TBM electric drive system. SUMMARY

[0005] In view of the above technical problems, the application provides a TBM electric drive system fault diagnosis method based on an adaptive sparse fusion network.

[0006] The application adopts the following technical scheme: a TBM electric drive system fault diagnosis method of an adaptive sparse fusion network, comprising the following steps:

[0007] The original data is denoised and effective signals are obtained, time domain features and frequency domain features are extracted based on signal characteristics, and multi-modal fusion data sets are fused;

[0008] A multi-task diagnosis model including a temperature anomaly detection auxiliary module and multiple task feature recognition modules is built, and each module shares a core diagnosis module while retaining an independent feature recognition branch;

[0009] The multi-task diagnosis model is trained, a learnable modal selection weight vector is introduced for each module, and the modal selection weight vector is used to dynamically allocate the optimal modal data combination to the corresponding module;

[0010] The temperature anomaly detection auxiliary module uses a double support vector quantile regression algorithm to fit the temperature change threshold and establish a threshold dynamic updating mechanism;

[0011] The task feature recognition module calculates the conflict degree between tasks based on task gradient similarity, and constructs a task conflict graph; the task conflict graph generates a sparse mask parameter guiding the generation of tasks , and regulates the neuron activation range to obtain a preliminary decoding result;

[0012] A temperature monitoring and fault correlation mechanism is established, and the preliminary decoding result is subjected to confidence weighting and decision compensation through a fusion modulator to generate a final fusion decision.

[0013] In further embodiments, the original data includes current data, torque data and temperature data;

[0014] The time domain features include peak value, skewness and kurtosis;

[0015] The frequency domain features include spectral energy, spectral bandwidth and peak frequency.

[0016] In further embodiments, for task , the allocation process of the optimal modal data combination is as follows:

[0017] The selection weight vector is composed of the selection weights of the modal , and the weighted fusion of multi-modal features is realized by the following formula:

[0018] ;

[0019] In the formula,​​ For the task A weighted fusion mechanism for multimodal features. For the number of modes, , For modality The output of the feature extractor For modality The input to the feature extractor;

[0020] The selected weight vector is optimized using gradient descent. Combined with the loss function for each task, the total loss is minimized using backpropagation and the Adam optimizer, and the selected weights are then updated. ;

[0021] Updated selection weights The modality corresponding to the weight threshold is determined as the task. The optimal combination of modal data.

[0022] In a further embodiment, the fitting process for the temperature change threshold is as follows:

[0023] A TSVQR algorithm supporting parallel computation of biquantiles is constructed, and a biobjective boundary function for the TSVQR algorithm is established, wherein the biobjective boundary function is a nonlinear fitting of kernel function mapping; the lower and upper bound quantiles are constrained according to the temporal characteristics of temperature data.

[0024] Define the boundary constraints of the TSVQR algorithm, use the boundary constraints to simultaneously optimize the parameters of the dual-objective boundary function, and record the change data of the dual-objective boundary function; save the upper and lower bounds of the change data as temperature change thresholds;

[0025] Correspondingly, the update process of the threshold dynamic update mechanism is as follows:

[0026] If the number of temperature data exceeds the specified number, the input temperature data is updated and the TSVQR algorithm is run again to update the temperature change threshold.

[0027] In a further embodiment, the process for constructing the task conflict graph is as follows:

[0028] For any task and tasks Calculate the gradient vector of the backbone network and the gradient vector of the backbone network Cosine similarity between And calculate the task and tasks The degree of conflict between , Main network gradient vector and the gradient vector of the backbone network The angle between them;

[0029] Based on the conflict level Construct a task conflict matrix , R is the real number field, and N is the total number of tasks;

[0030] Set conflict threshold Traverse the task conflict matrix Each element in, if Then, in representing the task and tasks Add a conflict edge between nodes and the degree of conflict Assigning weights to the edges creates a task conflict graph. : , A collection of task nodes. It is a set of weighted conflicting edges.

[0031] In a further embodiment, the sparse mask parameters Express it in the following form:

[0032] ,in, The elements representing the mask parameters are either 0 or 1, where 0 represents masking irrelevant features and 1 represents preserving relevant features. The dimension of the mask.

[0033] In a further embodiment, the process of obtaining the preliminary decoding result is as follows:

[0034] ;

[0035] In the formula, Represents element-wise product. The basic feature extraction function, For the task The optimal combination of input modal data, These are the model parameters for the multi-task diagnostic model. For the task Preliminary decoding results.

[0036] In a further embodiment, the temperature monitoring and fault association mechanism is as follows:

[0037] If the current temperature If the temperature change threshold is exceeded, the temperature-fault correlation judgment process is triggered. Based on the preliminary decoding results of each task feature recognition module, the following correlation verification is performed:

[0038] Viewing the diagnostic confidence of each task, if the diagnostic confidence of a certain task is higher than the corresponding confidence threshold, it is determined that the temperature anomaly is caused by the task, and the association result containing the fault type and the temperature anomaly is output;

[0039] If the diagnostic confidence of all tasks is lower than the corresponding confidence threshold, a warning is issued and manual intervention is prompted.

[0040] In further embodiments, the confidence weighting and decision compensation process of the preliminary decoding result is as follows: the diagnostic confidence of the task result corresponding to the task of acquiring the task result in the preliminary decoding result , , is generated using the following formula to generate the final fusion decision :

[0041] wherein, is the normalized weight of the task , , and N is the total number of tasks.

[0042] The beneficial effects of the present application: the present application fuses the multi-source operation data of current, torque, temperature, etc. of the TBM electric drive system with wavelet denoising and time-frequency domain analysis, and skillfully constructs a multi-source information fusion data set of the TBM electric drive system. At the same time, a high-dimensional, multi-modal fusion feature vector is constructed, which fundamentally improves the distinguishability of fault features and effectively enhances the detectability of fault signals.

[0043] The present application proposes a learnable modal selection weight mechanism for short circuit, open circuit and temperature anomaly, realizes the autonomous selection of the optimal modal of each task, realizes the precise construction of the task-specific modal, effectively improves the accuracy of fault feature extraction, and effectively alleviates the problem of fault feature "submersion" caused by signal coupling of TBM multi-motor system.

[0044] For the first time, temperature anomaly detection is combined with a fault diagnosis model as an auxiliary judgment algorithm for fault diagnosis tasks. The TSVQR algorithm used for temperature anomaly detection has good human boundary fitting ability and data sensitivity, and can accurately fit the change rule of complex data and identify outliers. In addition, temperature anomaly, as a representation of TBM electric drive system faults, can effectively assist other fault diagnosis tasks and avoid misjudgment and omission caused by complex and severely coupled data.

[0045] ​The application fuses gradient calculation similarities of different tasks, constructs a task conflict matrix, and constructs a task conflict graph using the task conflict matrix. The task conflict graph can display the conflict degree between different tasks, enhance the model interpretability, effectively alleviate the defect that the traditional deep learning model has poor interpretability, and is beneficial to the later maintenance of engineering personnel.

[0046] The application constructs a sparse mask control neuron activation area. According to the task conflict degree indicated by the task conflict graph, the sparse mask is controlled, tasks with high similarity share more neurons, and tasks with low similarity share fewer neurons. Through sparse mask control, the gradient interference between conflicting tasks is reduced, and the stability and task generalization ability of the feature extractor are improved. The introduction of the sparse mask effectively alleviates the problem of complex fault information caused by the cooperation of multiple motors in the TBM electric drive system, helps to improve the efficiency of the utilization of computing resources, and makes the allocation of computing resources more reasonable. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a flowchart of a TBM electric drive system fault diagnosis method.

[0048] Figure 2 (a) in is the original waveform diagram of the current signal.

[0049] Figure 2 (b) in is the waveform diagram after filtering the current signal.

[0050] Figure 3 is a TSVQR algorithm temperature change upper and lower bound fitting result diagram.

[0051] Figure 4 is a training flowchart of the adaptive sparse fusion network and the temperature anomaly diagnosis auxiliary module. DETAILED DESCRIPTION

[0052] The application will be further described below in conjunction with the drawings and examples in the specification.

[0053] Example 1

[0054] As shown in Figure 1 , a TBM electric drive system fault diagnosis method of an adaptive sparse fusion network includes the following steps:

[0055] The original data is denoised and effective signals are obtained, time domain features and frequency domain features are extracted based on signal characteristics, and are fused into a multi-modal fusion data set;

[0056] A multi-task diagnosis model including a temperature anomaly detection auxiliary module and multiple task feature recognition modules is built, and each module shares a core diagnosis module while retaining an independent feature recognition branch;

[0057] The multi-task diagnostic model is trained, and a learnable modality selection weight vector is introduced for each module. The optimal modality data combination is dynamically assigned to the corresponding module using the modality selection weight vector.

[0058] The temperature anomaly detection auxiliary module uses a dual support vector quantile regression algorithm to fit the temperature change threshold and establishes a dynamic threshold update mechanism.

[0059] The task feature recognition module calculates the conflict degree between tasks based on task gradient similarity and constructs a task conflict graph; the task conflict graph guides the generation of tasks. sparse mask parameters Preliminary decoding results were obtained by regulating the activation range of neurons;

[0060] A temperature monitoring and fault correlation mechanism is established, and the initial decoding results are weighted by confidence and compensated by a fusion modulator to generate the final fusion decision.

[0061] It should be noted that the method of the present invention consists of the following tasks:

[0062] (1) Short circuit fault diagnosis: mainly manifested as a sudden change or asymmetrical abnormality in the current of a certain phase, with current data as the main mode and transient detection used.

[0063] (2) Open circuit fault diagnosis: It is generally manifested as the disappearance of single-phase current and the increase of torque. The torque is the main mode, and its steady-state characteristic changes are explored.

[0064] (3) Temperature anomaly diagnosis: This task is a separate branch.

[0065] In a further embodiment, the raw data is collected from the TBM electric drive system switch cabinet, including current data, torque data, and temperature data; furthermore, each type of data may have 3600 samples, and each fault task type may have 600 samples.

[0066] To address the severe noise interference during TBM operation, this embodiment employs wavelet filtering technology to denoise the original data. Denoising experiments were conducted on the original data of the electric drive system by experimenting with different combinations of wavelet bases and decomposition levels. Simultaneously, the signal-to-noise ratio (SNR) of the denoised signal under each combination was calculated. Finally, the optimal db8 wavelet base and 3-level decomposition parameters were selected to complete the denoising operation on the original data. A comparison between the effective signal and the original signal is shown in Figure 2(a) and... Figure 2 As shown in (b) of the table, the commonly used wavelet functions and wavelet bases are listed in Table 1.

[0067] Table 1 Common wavelet functions and wavelet bases

[0068]

[0069] According to the signal characteristics of the TBM electric drive system, the time domain and frequency domain characteristics of the signal are calculated respectively. The time domain characteristics include: peak value, skewness and kurtosis; the frequency domain characteristics include: spectral energy, spectral bandwidth, peak frequency.

[0070] Based on the above effective signal, time domain feature, frequency domain feature fusion obtains a multi-modal fusion data set. In order to facilitate subsequent processing, each sample in the multi-modal fusion data set is added with a corresponding fault label, and is divided according to the proportion of 70% training set, 20% validation set and 10% test set. Ensure that the distribution of each task category in each data set is balanced to meet the model training requirements.

[0071] In further embodiments, based on the aforementioned multi-modal fusion data set, the multi-task diagnosis model is trained. During the training process, a learnable modal selection weight vector is introduced for each task in the model; the vector is optimized by gradient descent algorithm, which can automatically learn and determine the optimal modal combination for a specific task to improve the accuracy of task feature extraction, and the modal with greater weight indicates that it is more important for the task.

[0072] It is further understood that the allocation process of the optimal modal data combination is as follows:

[0073] The selection weight vector is defined by the task The selection weight of the modal is composed of, and the following formula is used to realize the weighted fusion of multi-modal features:

[0074] ; in the formula, is the weighted fusion mechanism of the multi-modal features of the task , is the number of modalities, , is the output of the feature extractor of the modal , is the input of the feature extractor of the modal ;

[0075] The selection weight vector is optimized by the gradient descent algorithm, combined with the loss function of each task, and the total loss is minimized by using the back propagation algorithm and the Adam optimizer, and the selection weight is updated; wherein the loss function of each task can be a predefined loss function, such as cross-entropy loss for fault task and binary cross-entropy loss for temperature anomaly detection auxiliary module. Therefore, when updating, the parameters of the core diagnosis module and the parameters of each feature recognition branch are updated at the same time. Finally, the updated selection weight greater than the weight threshold is determined as the task ​optimal modal data combination.

[0076] In a further embodiment, the fitting process of the temperature change threshold is as follows:

[0077] A TSVQR algorithm (double support vector quantile regression algorithm) supporting double quantile parallel calculation is constructed, and a double-objective boundary function of the TSVQR algorithm is established, which is a nonlinear fitting of a kernel function mapping; in this embodiment, the double-objective boundary function is:

[0078] ;

[0079] ; wherein, indicates the input data of the temperature anomaly detection auxiliary module, such as the optimal modal data combination allocated above, and are a temperature upper boundary function and a temperature lower boundary function, respectively, and are weight coefficients, and are bias tops.

[0080] According to the time sequence characteristics of the temperature data, the lower quantile and the upper quantile are constrained; in combination with the application of this embodiment, the temperature change of the TBM electric drive system is complex and nonlinear, and the existing conventional algorithm is difficult to fit the change rule of the specified position. According to the temperature characteristics of the TBM electric drive system, the quantiles are set to 0.1 and 0.99, which correspond to the upper and lower parts of the data, respectively, so as to make the algorithm pay attention to the boundary of the data change.

[0081] The boundary constraint condition of the TSVQR algorithm is defined, the parameters of the double-objective boundary function are optimized simultaneously by using the boundary constraint condition, and the change data of the double-objective boundary function are recorded; the upper limit value and the lower limit value of the change data are saved as the temperature change threshold.

[0082] In this embodiment, for the boundary constraint condition of the temperature upper boundary function , the expression form is as follows:

[0083] ;

[0084] For the boundary constraint condition of the temperature lower boundary function , the expression form is as follows:

[0085] .

[0086] In the boundary constraint condition formula of the upper and lower boundary functions, and weight vector representing the upper bound function and the lower bound function of temperature respectively, and bias representing the upper bound function and the lower bound function of temperature respectively, τ is a quantile parameter, ranging from 0 τ <1; and are regularization parameters, controlling the complexity of the model; and represent insensitive parameters, controlling the regression bandwidth; are slack variables representing the upper and lower bound errors of the temperature change, respectively; A is an input matrix , each row being a sample; Y is a response variable , each element being the output value of the corresponding sample; is an all-one vector , is the number of training samples, n is the feature dimension, is the transpose of a vector, is constrained to.

[0087] Therefore, the obtained upper and lower bounds of data variation are saved as temperature variation thresholds. Correspondingly, the update process of the threshold dynamic update mechanism is as follows: if the number of temperature data exceeds a specified number, the input temperature data is updated and the TSVQR algorithm is re-run to update the temperature variation thresholds.

[0088] Using the TSVQR algorithm, by setting the quantile, the algorithm focuses on different parts of the data, thereby obtaining the upper and lower bounds of temperature variation. If the temperature exceeds the upper and lower bounds, an alarm is triggered. At the same time, temperature is also used as an auxiliary feature to judge motor failure. If the temperature is abnormal, a diagnostic program is run. In this way, the normal signal can be better weakened to cover the fault signal.

[0089] Based on the training of the temperature anomaly detection auxiliary module, the model parameters are fixed when the training reaches a preset number of iterations. For the task feature recognition module, the construction process of the task conflict graph is as follows:

[0090] For any task and task , calculate the cosine similarity between the backbone network gradient vectors and , and calculate the conflict degree between task and task , ​and the gradient vector of the backbone network The angle between them;

[0091] Furthermore, cosine similarity The calculation formula is as follows:

[0092] .

[0093] Then, the degree of conflict The following formula is used to obtain: .

[0094] Based on the conflict level Construct a task conflict matrix , R is the real number field, and N is the total number of tasks;

[0095] Set conflict threshold p Traverse the task conflict matrix Each element in, if Then, in representing the task and tasks Add a conflict edge between nodes and the degree of conflict Assigning weights to the edges creates a task conflict graph. : ; A collection of task nodes. It is a set of weighted conflicting edges.

[0096] Using the constructed task conflict graph as a guide, a task-related sparse mask parameter is introduced into the shared backbone network. This parameter controls whether neurons are active in gradient updates and predictions for specific tasks, so that neurons for tasks with high conflict overlap less, and tasks with low conflict share more neurons.

[0097] Therefore, the sparse mask parameters in this embodiment Express it in the following form:

[0098] ,in, The elements representing the mask parameters are either 0 or 1, where 0 represents masking irrelevant features and 1 represents preserving relevant features. The dimension of the mask.

[0099] Correspondingly, the process of obtaining the preliminary decoding results is as follows:

[0100] ;

[0101] In the formula, Represents element-wise product. The basic feature extraction function, for task optimal modal data combination of input, model parameters of multi-task diagnosis model.

[0102] Based on this, the temperature monitoring and fault association mechanism is: if the current temperature exceeds the temperature change threshold, the temperature-fault association judgment process is triggered, the preliminary decoding results of each task feature recognition module are combined, and the following association verification is performed:

[0103] Check the diagnosis confidence of each task. If the diagnosis confidence of a task is higher than the corresponding confidence threshold, it is determined that the temperature anomaly is caused by the task, and the association result containing the fault type and the temperature anomaly is output.

[0104] If the diagnosis confidence of all tasks is lower than the corresponding confidence threshold, an early warning is issued and manual intervention is prompted.

[0105] Finally, for the case where the single task result is unreliable, the reliability is increased by weighting the confidence of the preliminary decoding result and decision compensation, and the process is as follows: in the preliminary decoding result, the task result of task , the diagnosis confidence of task result is obtained, and the following formula is used to generate the final fusion decision :

[0106] , wherein, is the normalized weight of task .

[0107] In summary, the embodiment proposes a TBM electric drive system multi-task fault diagnosis method based on an adaptive sparse fusion network. The optimal modal combination is dynamically allocated to short circuit fault, open circuit fault and temperature anomaly task through a learnable modal weight vector. The upper and lower threshold values of temperature change are fitted by proposing a double support vector quantile regression algorithm to realize real-time warning of temperature anomaly. The task conflict graph is constructed based on the task gradient similarity, and the neuron activation range of the conflict task is controlled by using dynamic sparse path. Finally, the confidence weighting and decision compensation of each task result are performed by the fusion modulator to output the fault type and diagnosis basis. The invention solves the problem that weak fault signals are masked under strong coupling of multiple motors, and the problem of high misdiagnosis and missed diagnosis rate. The robustness and accuracy of fault diagnosis are improved, and the continuity and safety of TBM tunneling operation are ensured.

[0108] Embodiment 2

[0109] ​​This embodiment discloses a TBM electric drive system fault diagnosis system, used to implement the TBM electric drive system fault diagnosis method based on the adaptive sparse fusion network described in Embodiment 1, including:

[0110] The first module is configured to denoise the original data and obtain effective signals, extract time-domain features and frequency-domain features based on the signal features, and fuse them into a multimodal fusion dataset.

[0111] The second module is set up to build a multi-task diagnostic model that includes a temperature anomaly detection auxiliary module and multiple task feature recognition modules. Each module shares the core diagnostic module while retaining an independent feature recognition branch.

[0112] The third module is designed to train the multi-task diagnostic model, introducing a learnable modality selection weight vector for each module and dynamically assigning the optimal modality data combination to the corresponding module using the modality selection weight vector. The temperature anomaly detection auxiliary module uses a dual support vector quantile regression algorithm to fit the temperature change threshold and establishes a dynamic threshold update mechanism. The task feature recognition module calculates the conflict degree between tasks based on task gradient similarity and constructs a task conflict graph. The task conflict graph guides the generation of tasks. sparse mask parameters Preliminary decoding results were obtained by regulating the activation range of neurons;

[0113] The fourth module is set up to establish a temperature monitoring and fault correlation mechanism. It uses a fusion modulator to perform confidence weighting and decision compensation on the preliminary decoding results to generate the final fusion decision.

[0114] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A TBM electric drive system fault diagnosis method of an adaptive sparse fusion network, characterized in that, The method comprises the following steps: The original data is denoised and effective signals are obtained, time domain features and frequency domain features are extracted based on signal characteristics, and multi-modal fusion data sets are fused; A multi-task diagnosis model including a temperature anomaly detection auxiliary module and multiple task feature recognition modules is built, and each module shares a core diagnosis module while retaining an independent feature recognition branch; The multi-task diagnosis model is trained, a learnable modal selection weight vector is introduced for each module, and the modal selection weight vector is used to dynamically allocate the optimal modal data combination to the corresponding module; The temperature anomaly detection auxiliary module uses a double support vector quantile regression algorithm to fit the temperature change threshold and establish a threshold dynamic updating mechanism; The task feature recognition module calculates the conflict degree between tasks based on task gradient similarity, and constructs a task conflict graph; Task conflict graph for guided generation of tasks Sparse mask parameters for guided generation of tasks Modulating the range of neuron activations yields preliminary decoding results A temperature monitoring and fault correlation mechanism is established, the preliminary decoding result is weighted and decision compensated by a fusion modulator, and a final fusion decision is generated.

2. The TBM electric drive system fault diagnosis method of the self-adaptive sparse fusion network according to claim 1, characterized in that, The original data includes current data, torque data, and temperature data; The time domain features include peak value, skewness, and kurtosis; The frequency domain features include spectral energy, spectral bandwidth, and peak frequency.

3. The TBM electric drive system fault diagnosis method of the self-adaptive sparse fusion network according to claim 1, characterized in that, For task , the allocation procedure of the optimal modality data combination is as follows: Define the selection weight vector by the task For modes Selection weight The composition is achieved by weighted fusion of multimodal features using the following formula: ； wherein is a task weighted fusion mechanism of multi-modal features, is the number of modalities, , is the output of a feature extractor of a modality is the input of a feature extractor of a modality is the input of a feature extractor of a modality is the input of a feature extractor of a modality The selection weight vector is optimized by a gradient descent algorithm, combined with the loss function of each task, the total loss is minimized by using a back propagation algorithm and an Adam optimizer, and the selection weight is updated ; updating the selection weights determining the modality corresponding to the task as the optimal modality data combination.

4. The TBM electric drive system fault diagnosis method of the self-adaptive sparse fusion network according to claim 1, characterized in that, The fitting process of the temperature change threshold is as follows: A TSVQR algorithm with double quantile parallel calculation is constructed, a double objective boundary function of the TSVQR algorithm is established, and the double objective boundary function is a nonlinear fitting of a kernel function mapping; the lower bound quantile and the upper bound quantile are constrained according to the time sequence characteristics of the temperature data; The boundary constraint condition of the TSVQR algorithm is defined, the parameters of the double objective boundary function are optimized simultaneously using the boundary constraint condition, and the change data of the double objective boundary function is recorded; the upper limit value and the lower limit value of the change data are saved as the temperature change threshold; Correspondingly, the updating process of the threshold dynamic updating mechanism is as follows: If the number of temperature data exceeds a specified number, the input temperature data is updated to re-run the TSVQR algorithm to update the temperature change threshold.

5. The TBM electric drive system fault diagnosis method of the self-adaptive sparse fusion network according to claim 1, characterized in that, The construction process of the task conflict graph is as follows: For any task and task , compute cosine similarity between backbone network gradient vector and backbone network gradient vector , and compute conflict degree between task and task , is the angle between backbone network gradient vector and backbone network gradient vector ;​​ based on the conflict degree constructing a task conflict matrix , R is a real number field, and N is the total number of tasks. Set conflict threshold Traverse the task conflict matrix Each element in, if Then, in representing the task and tasks Add a conflict edge between nodes and the degree of conflict Assigning weights to the edges creates a task conflict graph. : , A collection of task nodes. It is a set of weighted conflicting edges.

6. The TBM electric drive system fault diagnosis method of the self-adaptive sparse fusion network according to claim 1, characterized in that, The sparse mask parameter Is expressed in the following form: wherein, represents an element of the mask parameter being 0 or 1, 0 representing a masking of irrelevant features, 1 representing a preservation of relevant features, is the dimension of the mask.

7. The TBM electric drive system fault diagnosis method of claim 1, wherein, The acquisition process of the preliminary decoding result is as follows: ； wherein denotes an element-wise multiplication, is a base feature extraction function, is a task optimal modality data combination for the input, is a model parameter of the multi-task diagnostic model, is a task a preliminary decoding result.

8. The TBM electric drive system fault diagnosis method of claim 1, wherein, The temperature monitoring and fault correlation mechanism is as follows: If the current temperature If the temperature change exceeds the threshold, a temperature-fault correlation determination process is triggered, and the following correlation verification is performed in combination with the preliminary decoding results of each task feature recognition module: The diagnostic confidence of each task is checked, if the diagnostic confidence of a certain task is higher than the corresponding confidence threshold, it is determined that the temperature anomaly is caused by the task, and the correlation result containing the fault type and the temperature anomaly is output; If the diagnostic confidence of all tasks is lower than the corresponding confidence threshold, an early warning is issued and manual intervention is prompted.

9. The TBM electric drive system fault diagnosis method of claim 1, wherein, The process of confidence weighting and decision compensation of the preliminary decoding result is as follows: corresponding to the acquisition task the task result the task result the diagnostic confidence The final fusion decision is generated using the following formula : wherein, is the normalized weight for task , N is the total number of tasks.

Images