Non-intrusive intelligent diagnosis method for early turn-to-turn short circuit of servo motor
By combining self-supervised comparative pre-training and supervised fine-tuning frameworks with timing consistency comparison and control-sensor consistency fusion, and utilizing the angular position pulses and three-phase current signals of the servo motor controller, the problem of efficient diagnosis of early inter-turn short circuit faults in servo motors is solved, achieving non-invasive diagnosis with high accuracy and generalization ability.
Patent Information
- Application Number
- CN202511065218.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-31
- Publication Date
- 2025-11-14
AI Technical Summary
Existing methods for diagnosing early inter-turn short circuit faults in servo motors mostly involve multi-source information fusion, requiring additional sensing equipment. Furthermore, they are difficult to effectively identify early inter-turn short circuit faults under non-stationary operating conditions. Existing non-invasive diagnostic methods have limited effectiveness on current signals.
A self-supervised comparative pre-training and supervised fine-tuning framework is adopted, which combines timing consistency comparison and control-sensor consistency fusion. By utilizing the angular position pulse of the servo motor controller and the three-phase current signal of the external sensor, a lightweight residual network and a self-supervised comparative loss function are used to achieve efficient fusion and diagnosis of fault features.
Under non-invasive monitoring, the model's generalization ability under different fault levels is improved, the model complexity is reduced, and efficient early inter-turn short circuit fault diagnosis of servo motors is achieved, with a diagnostic accuracy of over 95% and a Matthews correlation coefficient of over 90%.
Smart Images
Figure CN120949034A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of servo motor fault diagnosis technology, and specifically to a non-invasive intelligent diagnostic method for early inter-turn short circuits in servo motors. Background Technology
[0002] Servo motors, as control actuators that achieve precise motion through reference and feedback signals, play a crucial role in industrial robots, electric vehicles, and other high-end equipment. In practical industrial applications, servo motors mainly operate under continuous non-stationary conditions, requiring frequent acceleration, deceleration, and reversal. These harsh operating conditions often accelerate equipment failures; therefore, improving the reliability and safety of servo motors has become a major research focus.
[0003] Among the many types of faults in servo motors, early inter-turn short circuit faults have become one of the key challenges in servo motor fault diagnosis due to their high probability of occurrence, weak fault characteristics, and rapid propagation into phase-to-phase faults during continuous operation. Existing servo motor fault diagnosis methods are mostly based on deep learning-based multi-source information fusion methods, which identify faults such as inter-turn short circuits by collecting signals such as vibration, current, motor torque, and speed. However, acquiring these multi-source signals often requires additional sensing equipment, which greatly limits the practical application of these multi-source information fusion methods.
[0004] Non-invasive monitoring features non-destructive testing of existing equipment structure and uninterrupted real-time operation, avoiding secondary damage caused by disassembly, reducing operating costs, and meeting the safety and economic requirements of practical industrial applications. Current research on non-invasive diagnosis of inter-turn short-circuit faults in servo motors mainly focuses on current signals; however, it is difficult to identify early-stage inter-turn short-circuit faults using only current signals under non-stationary operating conditions. In contrast, control signals generated by motor controllers have significant advantages under non-stationary operating conditions. Therefore, developing a series of non-invasive diagnostic methods based on control signals for early-stage inter-turn short-circuit faults in servo motors has become an urgent problem to be solved. Summary of the Invention
[0005] To overcome the shortcomings of the prior art, the present invention aims to provide a non-intrusive intelligent diagnostic method for early inter-turn short circuits in servo motors, which enhances the generalization ability of the model under different fault levels, achieves efficient feature fusion, and does not increase the model complexity in practical applications.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A non-intrusive intelligent diagnostic method for early inter-turn short circuits in servo motors is proposed. The angular position pulse from the servo motor controller is directly converted into an instantaneous angular velocity signal and integrated with the current signal to achieve multi-view diagnosis. Then, a diagnostic model is constructed using a self-supervised comparative pre-training and supervised fine-tuning framework to diagnose early inter-turn short circuit faults in servo motors.
[0008] A non-intrusive intelligent diagnostic method for early inter-turn short circuits in servo motors includes the following steps:
[0009] 1) Acquisition of instantaneous angular velocity signal: In the servo motor system, the angular position pulses collected by the built-in encoder are converted into instantaneous angular velocity signals to capture the weak speed fluctuations caused by inter-turn short circuit faults;
[0010] 2) Timing consistency comparison: Timing consistency refers to the fact that when the equipment is in a healthy state, the monitoring signal will show similar states or trends in adjacent time periods. Therefore, abrupt changes in signal amplitude or phase are used as the basis for fault diagnosis. The timing consistency comparison module is used to extract fault feature components from the monitoring signal.
[0011] After converting the discrete pulse signal into a continuous-time signal, two lightweight residual networks with the same architecture are used as dual-branch feature encoders. Then, the high-dimensional features of the encoder are input into a projection layer containing a nonlinear fully connected network and L2 regularization.
[0012] Let i∈I≡{1…N} denote the index of the input sequence, z i =Proj(Encoder(u i )) represents the features extracted by the encoder, then the self-supervised contrastive loss L CL,i (·) represents the following:
[0013]
[0014] In the formula, z p For z i The positive sample features obtained after data augmentation, z a Not containing z i Features in the feature set The transpose of the features extracted by the encoder, where a∈A(i) is z a The index, τ m ∈R + This is a scalar temperature parameter;
[0015] L CL_Con and L CL_Sen The self-monitored comparison losses of the encoder diagonal position pulse and phase current are respectively calculated from the above formula, and therefore the total loss L of the timing consistency comparison module is... TCC (·) is defined as:
[0016]
[0017] 3) Control-sensor consistency fusion: The angular position pulse provided by the motor controller and the three-phase current obtained by the external sensor are both used to identify non-intrusive inter-turn short circuit faults. Within the same time period, the fault characteristics of these monitoring signals will characterize the equipment operating status from different perspectives. Control-sensor consistency indicates that within the same time interval, the fault characteristics extracted from the control signal and the sensor signal will be closer to each other in the potential feature space. Therefore, various fault characteristics are matched with control-sensor consistency.
[0018] To match the N pairs of sensing features extracted by the dual-branch encoder, the features C that satisfy the sensing consistency will be... i and S i Considered as positive sample pairs, while C i and S j To form negative sample pairs, where i ≠ j; introduce the self-supervised contrastive loss in formula (5) to make the positive sample pairs C i With S i Maximize the similarity between them, while also making the negative sample pair C i With S j The similarity between samples is minimized, thereby aligning control and sensing features. Furthermore, a symmetric loss structure is introduced to facilitate similarity calculation between positive and negative sample pairs.
[0019]
[0020] In the formula, L CSCF,i (·) represents the loss of the control-sensor consistency fusion module, L CCF,i (·) represents the self-supervised comparison loss of the control signal, L SCF,i (·) represents the self-supervised contrast loss of the sensor information;
[0021] The timing consistency comparison module and the control-sensing consistency fusion module complement each other's losses, using L TCC and L CSCF The pre-training loss function L of the linear combination composition model pretra The specific description is as follows:
[0022] L pretra =βL TCC (·)+(1-β)L CSCF (·) (9) In the formula, β is the penalty coefficient;
[0023] 4) Supervised fine-tuning: After self-supervised pre-training of the temporal consistency comparison module and the control-sensory consistency fusion module using unlabeled control-sensory sample pairs, the dual-branch feature encoder will initially master the ability to extract control-sensory feature pairs that meet the consistency metric. Then, based on the training subset of labeled samples, a supervised fine-tuning module is introduced to enable the diagnostic model to be fine-tuned and quickly deployed under specific tasks.
[0024] In step 1), the instantaneous angular velocity signal is acquired by performing forward differential analysis on each rising edge of the angular position pulse. The specific process is as follows:
[0025]
[0026] In the formula, ω m Let t be the mechanical angular velocity. i and θ i N represents the time point and the current rotor mechanical angle, respectively. R For built-in encoder resolution, 2π / N R t(θ) represents the angle of the encoder grating associated with pulses at adjacent positions. i ) indicates that the rotor mechanical angle is θ i The time point of the rising edge of the position pulse.
[0027] Step 2) extracts fault features for temporal consistency by combining data augmentation operations. Based on the principle of not changing the semantics of the original data, Gaussian white noise, random scaling, and masked noise are used to augment the temporal data. The specific operations are as follows:
[0028]
[0029] In the formula, For the enhanced signal, u i Given the original input signal, the noise G follows a Gaussian distribution of N(0,0.01), the scaling factor λ follows a Gaussian distribution of N(1,0.01), and the elements of the mask matrix M follow a Bernoulli distribution. The symbols are... This indicates element-wise multiplication; furthermore, the augmentation probability of all time-based data augmentation methods is 0.5.
[0030] In step 3), to improve the fusion efficiency of the sensing samples, a hierarchical strategy is incorporated into the loss calculation process. This involves applying max pooling to the extracted features along the time axis and recursively calculating the loss proposed in the above formula, thereby achieving feature alignment across granular levels. The hierarchical loss L of the sensing consistency fusion module... CSCF,i (·) is represented as:
[0031]
[0032] In the formula, m is the feature order after max pooling, and M represents the cumulative number of max pooling operations.
[0033] Step 4) Set the fault diagnosis of inter-turn short circuit of servo motor as a binary classification problem, freeze the model parameters of the pre-trained dual-branch feature encoder and transfer them to the downstream task; in addition, use a linear layer to map the control sensing features to a unified embedding space; then, fuse the aligned control sensing features through element-level addition and input them into the classifier layer; finally, based on limited labeled data, use cross-entropy loss to fine-tune the linear layer and classifier layer in the supervised fine-tuning module.
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] 1) Under the constraint of non-intrusive monitoring, the present invention utilizes the angular position pulse of the motor controller and the three-phase current signal of the external sensor to realize the diagnosis of inter-turn short circuit fault. The angular position pulse is converted into an instantaneous angular velocity signal to reduce the influence of non-steady operating conditions, while the current signal is used to compensate for the information loss caused by the short recording time of the motor controller.
[0036] 2) Based on the self-supervised comparative pre-training and supervised fine-tuning framework, this invention designs an intelligent fusion strategy for control and sensing signals. This strategy extracts common features of multiple monitoring signals, reduces human intervention, and enhances the model's generalization ability under different fault levels.
[0037] 3) This invention combines temporal consistency comparison with control-sensory consistency fusion to establish a self-supervised pre-training framework for inter-turn short-circuit faults, optimizes the learning strategy and loss function to ensure efficient feature fusion, and does not increase model complexity in practical applications. Attached Figure Description
[0038] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.
[0039] Figure 2 This is a schematic diagram showing the distribution of health status features and fault status features in two-dimensional space according to an embodiment of the present invention.
[0040] Figure 3 The diagram shows the diagnostic results of the model in this embodiment of the invention for operating conditions 4, 5, and 6. Detailed Implementation
[0041] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings.
[0042] like Figure 1 As shown, a non-intrusive intelligent diagnostic method for early inter-turn short circuits in servo motors includes the following steps:
[0043] 1) Acquisition of instantaneous angular velocity signal: The data used in this embodiment are the angular position pulse signal and phase current signal collected by the servo motor drive platform. This platform simulates the servo motor driving the industrial robotic arm to reciprocate within the range of -90° to 90° to realize the acquisition of non-steady-state operation data. The specific working conditions for acquisition are shown in Table 1.
[0044] Table 1. Operating Conditions of the Servo Motor Drive Platform
[0045]
[0046] In a servo motor system, the angular position pulses acquired by the built-in encoder can be converted into instantaneous angular velocity signals to capture the subtle speed fluctuations caused by inter-turn short-circuit faults. The instantaneous angular velocity signal is obtained by performing forward differential analysis on each rising edge of the angular position pulse, as detailed below:
[0047]
[0048] In the formula, ω m Let t be the mechanical angular velocity. i and θ i N represents the time point and the current rotor mechanical angle, respectively. R For built-in encoder resolution, 2π / N R t(θ) represents the angle of the encoder grating associated with pulses at adjacent positions. i ) indicates that the rotor mechanical angle is θ i The time point of the rising edge of the position pulse;
[0049] 2) Timing consistency comparison: Timing consistency refers to the fact that when the equipment is in a healthy state, the monitoring signal will show similar state or trend in adjacent time periods. Therefore, abrupt changes in signal amplitude or phase can be used as a basis for fault judgment. Contrastive learning can maximize the intra-class similarity and inter-class differences between features. Therefore, this method uses the timing consistency comparison module to extract fault feature components in the monitoring signal.
[0050] The extraction of fault features with temporal consistency benefits from a reasonable combination of data augmentation operations. Based on the principle of not changing the semantics of the original data, Gaussian white noise, random scaling, and masked noise are used to complete temporal data augmentation. The specific operations are as follows:
[0051]
[0052] In the formula, For the enhanced signal, u i Given the original input signal, the noise G follows a Gaussian distribution of N(0,0.01), the scaling factor λ follows a Gaussian distribution of N(1,0.01), and the elements of the mask matrix M follow a Bernoulli distribution. The symbols are... This indicates element-wise multiplication; furthermore, the augmentation probability of all time-based data augmentation methods is 0.5.
[0053] After converting the discrete pulse signal into a continuous-time signal, two lightweight residual networks with the same architecture are used as bi-branch feature encoders to reduce model complexity. Then, the high-dimensional features of the encoder are input into a projection layer containing a nonlinear fully connected network and L2 regularization.
[0054] Let i∈I≡{1…N} denote the index of the input sequence, z i =Proj(Encoder(u i )) represents the features extracted by the encoder, then the self-supervised contrastive loss L CL,i (·) can be represented as follows:
[0055]
[0056] In the formula, z p For z i The positive sample features obtained after data augmentation, z a Not containing z i Features in the feature set The transpose of the features extracted by the encoder, where a∈A(i) is z a The index, τ m ∈R + This is a scalar temperature parameter;
[0057] L CL_Con and L CL_Sen The self-monitored comparison losses of the encoder diagonal position pulse and phase current are respectively calculated from the above formula, and therefore the total loss L of the timing consistency comparison module is... TCC (·) is defined as:
[0058]
[0059] 3) Control-Sensing Consistency Fusion: Both the angular position pulses provided by the motor controller and the three-phase currents acquired by external sensors are used to identify non-intrusive inter-turn short-circuit faults. Within the same time frame, the fault characteristics of these monitoring signals will characterize the equipment's operating status from different perspectives. Control-sensing consistency indicates that, within the same time interval, fault features extracted from control signals and sensor signals will be closer to each other in the potential feature space. Therefore, matching various fault features with control-sensing consistency can maximize the utilization of fault information and avoid misjudgments caused by feature fusion.
[0060] To match the N pairs of sensing features extracted by the dual-branch encoder, the features C that satisfy the sensing consistency will be... i and S iConsidered as positive sample pairs, while C i and S j (where i≠j) constitute negative sample pairs; further, the self-supervised contrast loss in formula (5) is introduced, making the positive sample pairs C i With S i Maximize the similarity between them, while also making the negative sample pair C i With S j The similarity between samples is minimized, thereby aligning control and sensing features. Furthermore, a symmetric loss structure is introduced to facilitate similarity calculation between positive and negative sample pairs.
[0061]
[0062] In the formula, L CSCF,i (·) represents the loss of the control-sensor consistency fusion module, L CCF,i (·) represents the self-supervised comparison loss of the control signal, L SCF,i (·) represents the self-supervised contrast loss of the sensor information;
[0063] To improve the fusion efficiency of sensing samples, a hierarchical strategy is incorporated into the loss calculation process. This involves applying max pooling to the extracted features along the time axis and recursively calculating the loss expressed in the above formula, thereby achieving feature alignment across granular levels. The hierarchical loss L of the sensing consistency fusion module... CSCF,i (·) can be represented as:
[0064]
[0065] In the formula, m is the feature order after max pooling, and M represents the cumulative number of max pooling operations;
[0066] The losses of the timing consistency comparison module and the control-sensing consistency fusion module mentioned above are complementary, therefore L is used. TCC and L CSCF The pre-training loss function L of the linear combination composition model pretra The specific description is as follows:
[0067] L pretra =βL TCC (·)+(1-β)L CSCF (·) (9) In the formula, β is the penalty coefficient;
[0068] 4) Supervised fine-tuning: After self-supervised pre-training of the temporal consistency comparison module and the control-sensory consistency fusion module using unlabeled control-sensory sample pairs, the dual-branch feature encoder will initially master the ability to extract control-sensory feature pairs that meet the consistency metric; then, based on a small training subset of labeled samples, a supervised fine-tuning module is introduced to enable the diagnostic model to be fine-tuned and quickly deployed under specific tasks.
[0069] During the pre-training phase, the initial learning rate and weight decay of the Adam optimizer are set to 1e-4 and 5e-4, respectively, and the mini-batch size is 128. After 50 pre-training cycles, the optimal pre-trained diagnostic model is transferred to the fine-tuning phase, where the initial learning rate is set to 1e-4, the mini-batch size is set to 64, and the fine-tuning cycle is set to 10.
[0070] The fault diagnosis of inter-turn short circuit in servo motors is set as a binary classification problem. The model parameters of the pre-trained dual-branch feature encoder are frozen and transferred to the downstream task. Furthermore, a simple linear layer is used to map the sensing features to a unified embedding space, thereby avoiding misalignment of feature pairs extracted during the fine-tuning stage. A t-distributed random neighborhood embedding technique is used to map the high-dimensional features extracted from operating conditions 1, 2, and 3 to a two-dimensional space, specifically as follows... Figure 2 As shown, there are clear boundaries between fault features and normal features under the three different operating conditions, and each clusters well. This indicates that the method can learn to obtain appropriate decision boundaries under different fault levels, which can be used for the diagnosis of early inter-turn short circuit faults.
[0071] Subsequently, the aligned control features are fused through element-level addition and input into the classifier layer; finally, based on limited labeled data, cross-entropy loss is used to fine-tune the linear layer and classifier layer in the supervised fine-tuning module.
[0072] To verify the generalization ability of the method of the present invention under different short-circuit scenarios, the model was trained using data from scenario 4, and tested using data from scenarios 5 and 6. Specific diagnostic results are as follows: Figure 3 As shown, the diagnostic accuracy of the model under both operating conditions 5 and 6 is higher than 95%, and the Matthews correlation coefficient is higher than 90%. The results indicate that by combining the pre-training framework with self-supervised contrastive learning, common features can be extracted from multiple monitoring signals without human intervention, thus ensuring the model's generalization ability under different fault levels.
[0073] The present invention has been disclosed above with reference to preferred embodiments, but it is not intended to limit the present invention. Any simple modifications, equivalent changes and alterations made by those skilled in the art to the above embodiments without departing from the scope of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.
Claims
1. A non-invasive intelligent diagnostic method for early inter-turn short circuits in servo motors, characterized in that: The angular position pulse from the servo motor controller is directly converted into an instantaneous angular velocity signal and integrated with the current signal to achieve multi-view diagnosis. Then, a diagnostic model is built using a self-supervised comparative pre-training and supervised fine-tuning framework to achieve the diagnosis of early inter-turn short circuit faults in the servo motor.
2. The non-intrusive intelligent diagnostic method for early inter-turn short circuits in servo motors according to claim 1, characterized in that, Includes the following steps: 1) Acquisition of instantaneous angular velocity signal: In the servo motor system, the angular position pulses collected by the built-in encoder are converted into instantaneous angular velocity signals to capture the weak speed fluctuations caused by inter-turn short circuit faults; 2) Timing consistency comparison: Timing consistency refers to the fact that when the equipment is in a healthy state, the monitoring signal will show similar states or trends in adjacent time periods. Therefore, abrupt changes in signal amplitude or phase are used as the basis for fault diagnosis. The timing consistency comparison module is used to extract fault feature components from the monitoring signal. After converting the discrete pulse signal into a continuous-time signal, two lightweight residual networks with the same architecture are used as dual-branch feature encoders. Then, the high-dimensional features of the encoder are input into a projection layer containing a nonlinear fully connected network and L2 regularization. Let i∈I≡{1…N} denote the index of the input sequence, z i =Proj(Encoder(u i )) represents the features extracted by the encoder, then the self-supervised contrastive loss L CL,i (·) represents the following: In the formula, z p For z i The positive sample features obtained after data augmentation, z a Not containing z i Features in the feature set The transpose of the features extracted by the encoder, where a∈A(i) is z a The index, τ m ∈R + This is a scalar temperature parameter; L CL_Con and L CL_Sen The self-monitored comparison losses of the encoder diagonal position pulse and phase current are respectively calculated from the above formula, and therefore the total loss L of the timing consistency comparison module is... TCC (·) is defined as: 3) Control-sensor consistency fusion: The angular position pulse provided by the motor controller and the three-phase current obtained by the external sensor are both used to identify non-intrusive inter-turn short circuit faults. Within the same time period, the fault characteristics of these monitoring signals will characterize the equipment operating status from different perspectives. Control-sensor consistency indicates that within the same time interval, the fault characteristics extracted from the control signal and the sensor signal will be closer to each other in the potential feature space. Therefore, various fault characteristics are matched with control-sensor consistency. To match the N pairs of sensing features extracted by the dual-branch encoder, the features C that satisfy the sensing consistency will be... i and S i Considered as positive sample pairs, while C i and S j To form negative sample pairs, where i ≠ j; introduce the self-supervised contrastive loss in formula (5) to make the positive sample pairs C i With S i Maximize the similarity between them, while also making the negative sample pair C i With S j Minimize the similarity between them, thereby achieving alignment between control and sensing features; Furthermore, a symmetric loss structure is introduced to facilitate the calculation of similarity between positive and negative sample pairs: In the formula, L CSCF,i (·) represents the loss of the control-sensor consistency fusion module, L CCF,i (·) represents the self-supervised comparison loss of the control signal, L SCF,i (·) represents the self-supervised contrast loss of the sensor information; The timing consistency comparison module and the control-sensing consistency fusion module complement each other's losses, using L TCC and L CSCF The pre-training loss function L of the linear combination composition model pretra The specific description is as follows: L pretra =βL TCC (·)+(1-β)L CSCF (·) (9) In the formula, β is the penalty coefficient; 4) Supervised fine-tuning: After self-supervised pre-training of the temporal consistency comparison module and the control-sensory consistency fusion module using unlabeled control-sensory sample pairs, the dual-branch feature encoder will initially master the ability to extract control-sensory feature pairs that meet the consistency metric. Then, based on the training subset of labeled samples, a supervised fine-tuning module is introduced to enable the diagnostic model to be fine-tuned and quickly deployed under specific tasks.
3. The method according to claim 2, characterized in that: In step 1), the instantaneous angular velocity signal is acquired by performing forward differential analysis on each rising edge of the angular position pulse. The specific process is as follows: In the formula, ω m Let t be the mechanical angular velocity. i and θ i N represents the time point and the current rotor mechanical angle, respectively. R For built-in encoder resolution, 2π / N R t(θ) represents the angle of the encoder grating associated with pulses at adjacent positions. i ) indicates that the rotor mechanical angle is θ i The time point of the rising edge of the position pulse.
4. The method according to claim 2, characterized in that: Step 2) extracts fault features for temporal consistency by combining data augmentation operations. Based on the principle of not changing the semantics of the original data, Gaussian white noise, random scaling, and masked noise are used to augment the temporal data. The specific operations are as follows: In the formula, For the enhanced signal, u i Given the original input signal, the noise G follows a Gaussian distribution of N(0,0.01), the scaling factor λ follows a Gaussian distribution of N(1,0.01), and the elements of the mask matrix M follow a Bernoulli distribution. The symbols are... This indicates element-wise multiplication; furthermore, the augmentation probability of all time-based data augmentation methods is 0.
5.
5. The method according to claim 2, characterized in that: In step 3), to improve the fusion efficiency of the sensing samples, a hierarchical strategy is incorporated into the loss calculation process. This involves applying max pooling to the extracted features along the time axis and recursively calculating the loss proposed in the above formula, thereby achieving feature alignment across granular levels. The hierarchical loss L of the sensing consistency fusion module... CSCF,i (·) is represented as: In the formula, m is the feature order after max pooling, and M represents the cumulative number of max pooling operations.
6. The method according to claim 2, characterized in that: Step 4) Set the fault diagnosis of inter-turn short circuit of servo motor as a binary classification problem, freeze the model parameters of the pre-trained dual-branch feature encoder and transfer them to the downstream task; in addition, use a linear layer to map the control sensing features to a unified embedding space. Subsequently, the aligned control features are fused through element-level addition and input into the classifier layer; finally, based on limited labeled data, cross-entropy loss is used to fine-tune the linear layer and classifier layer in the supervised fine-tuning module.